diff --git "a/ir_dump/debug-final.py" "b/ir_dump/debug-final.py" new file mode 100644--- /dev/null +++ "b/ir_dump/debug-final.py" @@ -0,0 +1,11932 @@ +# from tvm.script import ir as I +# from tvm.script import tir as T +# from tvm.script import relax as R + +@I.ir_module +class Module: + I.module_attrs({"system_lib_prefix": "gpt_neox_q0f16_"}) + @T.prim_func + def gpt_neox_q0f16_apply_bitmask_inplace(var_logits: T.handle, var_seq_ids: T.handle, var_bitmask: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": T.bool(True), "tir.noalias": T.bool(True)}) + batch_size, vocab_size = T.int32(is_size_var=True), T.int32(is_size_var=True) + logits = T.match_buffer(var_logits, (batch_size, vocab_size)) + num_seq = T.int32(is_size_var=True) + seq_ids = T.match_buffer(var_seq_ids, (num_seq,), "int32") + bitmask = T.match_buffer(var_bitmask, (batch_size, (vocab_size + 31) // 32), "int32") + # with T.block("root"): + for fused_s_v_0 in T.thread_binding((num_seq * vocab_size + 255) // 256, thread="blockIdx.x"): + for fused_s_v_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("block"): + vs = T.axis.spatial(num_seq, (fused_s_v_0 * 256 + fused_s_v_1) // vocab_size) + vv = T.axis.spatial(vocab_size, (fused_s_v_0 * 256 + fused_s_v_1) % vocab_size) + T.where(fused_s_v_0 * 256 + fused_s_v_1 < num_seq * vocab_size) + T.reads(bitmask[seq_ids[vs], vv // 32], seq_ids[vs], logits[seq_ids[vs], vv]) + T.writes(logits[seq_ids[vs], vv]) + logits[seq_ids[vs], vv] = T.if_then_else(T.bitwise_and(T.shift_right(bitmask[seq_ids[vs], vv // 32], vv % 32), 1) == 1, logits[seq_ids[vs], vv], T.float32(-340282346638528859811704183484516925440.0)) + + @T.prim_func + def gpt_neox_q0f16_apply_logit_bias_inplace(var_logits: T.handle, var_pos2seq_id: T.handle, var_token_ids: T.handle, var_logit_bias: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": T.bool(True), "tir.noalias": T.bool(True)}) + batch_size, vocab_size = T.int32(is_size_var=True), T.int32(is_size_var=True) + logits = T.match_buffer(var_logits, (batch_size, vocab_size)) + num_token = T.int32(is_size_var=True) + pos2seq_id = T.match_buffer(var_pos2seq_id, (num_token,), "int32") + token_ids = T.match_buffer(var_token_ids, (num_token,), "int32") + logit_bias = T.match_buffer(var_logit_bias, (num_token,)) + # with T.block("root"): + for p0 in T.thread_binding((num_token + 255) // 256, thread="blockIdx.x"): + for p1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("block"): + vp = T.axis.spatial(num_token, p0 * 256 + p1) + T.where(p0 * 256 + p1 < num_token) + T.reads(logits[pos2seq_id[vp], token_ids[vp]], pos2seq_id[vp], token_ids[vp], logit_bias[vp]) + T.writes(logits[pos2seq_id[vp], token_ids[vp]]) + logits[pos2seq_id[vp], token_ids[vp]] = logits[pos2seq_id[vp], token_ids[vp]] + logit_bias[vp] + + @T.prim_func + def gpt_neox_q0f16_apply_penalty_inplace(var_logits: T.handle, var_seq_ids: T.handle, var_pos2seq_id: T.handle, var_token_ids: T.handle, var_token_cnt: T.handle, var_penalties: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": T.bool(True), "tir.noalias": T.bool(True)}) + batch_size, vocab_size = T.int32(is_size_var=True), T.int32(is_size_var=True) + logits = T.match_buffer(var_logits, (batch_size, vocab_size)) + num_seq = T.int32(is_size_var=True) + seq_ids = T.match_buffer(var_seq_ids, (num_seq,), "int32") + num_token = T.int32(is_size_var=True) + pos2seq_id = T.match_buffer(var_pos2seq_id, (num_token,), "int32") + token_ids = T.match_buffer(var_token_ids, (num_token,), "int32") + token_cnt = T.match_buffer(var_token_cnt, (num_token,), "int32") + penalties = T.match_buffer(var_penalties, (num_seq, 3)) + # with T.block("root"): + for p0 in T.thread_binding((num_token + 255) // 256, thread="blockIdx.x"): + for p1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("block"): + vp = T.axis.spatial(num_token, p0 * 256 + p1) + T.where(p0 * 256 + p1 < num_token) + T.reads(logits[seq_ids[pos2seq_id[vp]], token_ids[vp]], seq_ids[pos2seq_id[vp]], pos2seq_id[vp], token_ids[vp], penalties[pos2seq_id[vp], 0:3], token_cnt[vp]) + T.writes(logits[seq_ids[pos2seq_id[vp]], token_ids[vp]]) + logits[seq_ids[pos2seq_id[vp]], token_ids[vp]] = logits[seq_ids[pos2seq_id[vp]], token_ids[vp]] - (penalties[pos2seq_id[vp], 0] + T.Cast("float32", token_cnt[vp]) * penalties[pos2seq_id[vp], 1]) + logits[seq_ids[pos2seq_id[vp]], token_ids[vp]] = T.if_then_else(logits[seq_ids[pos2seq_id[vp]], token_ids[vp]] < T.float32(0.0), logits[seq_ids[pos2seq_id[vp]], token_ids[vp]] * penalties[pos2seq_id[vp], 2], logits[seq_ids[pos2seq_id[vp]], token_ids[vp]] / penalties[pos2seq_id[vp], 2]) + + @T.prim_func + def gpt_neox_q0f16_batch_decode_paged_kv(_0: T.int32, Q_handle: T.handle, pages_handle: T.handle, page_table_indptr_handle: T.handle, page_table_values_handle: T.handle, var_length_info: T.handle, k_rope_pos_offset_handle: T.handle, q_rope_position_handle: T.handle, output_handle: T.handle, lse_handle: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + B = T.int32(is_size_var=True) + Q = T.match_buffer(Q_handle, (B, 16, 64), "float16") + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(pages_handle, (max_num_pages, 2, 16, 16, 64), "float16", offset_factor=1) + page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (B,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) + output = T.match_buffer(output_handle, (B, 16, 64), "float16") + lse = T.match_buffer(lse_handle, (B, 16)) + # with T.block("root"): + sm_scale: T.float32 = T.float32(0.18033688011112042) + for bx in T.thread_binding(B, thread="blockIdx.x"): + for fused_by_bz in T.thread_binding(16, thread="blockIdx.y"): + for ty in T.thread_binding(1, thread="threadIdx.y"): + for tx in T.thread_binding(16, thread="threadIdx.x"): + for tz in T.thread_binding(16, thread="threadIdx.z"): + with T.block("attn"): + T.reads(page_table_indptr[bx:bx + 2], length_info[bx], q_rope_position[bx], Q[bx, fused_by_bz // 16 + ty + fused_by_bz % 16, tx * 4 - 32:tx * 4 - 32 + 68]) + T.writes(output[bx, fused_by_bz % 16 + fused_by_bz // 16 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 16 + fused_by_bz // 16 + ty]) + Q_local = T.alloc_buffer((4,), "float16", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + O_allreduce = T.alloc_buffer((16, 1, 64), scope="shared") + md_allreduce = T.alloc_buffer((16, 1, 2), scope="shared") + S_reduce_local = T.alloc_buffer((1,), scope="local") + t0 = T.alloc_buffer((1,), scope="local") + S_local = T.alloc_buffer((1,), scope="local") + QK_local = T.alloc_buffer((4,), scope="local") + V_local = T.alloc_buffer((4,), "float16", scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_prev = T.alloc_buffer((1,), scope="local") + other_m = T.alloc_buffer((1,), scope="local") + other_d = T.alloc_buffer((1,), scope="local") + exp_mprev = T.alloc_buffer((1,), scope="local") + exp_otherm = T.alloc_buffer((1,), scope="local") + other_o = T.alloc_buffer((4,), scope="local") + st_m = T.alloc_buffer((1,), scope="local") + st_d = T.alloc_buffer((1,), scope="local") + O_local = T.alloc_buffer((4,), scope="local") + by: T.int32 = fused_by_bz % 16 + bz: T.int32 = fused_by_bz // 16 + batch_idx: T.int32 = bx + cur_page_indptr_begin: T.int32 = page_table_indptr[batch_idx] + cur_page_indptr_end: T.int32 = page_table_indptr[batch_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[batch_idx], 0) + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for vec in T.vectorized(4): + freq = T.float32() + Q_local[vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", Q[bx, by + bz + ty, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, Q[bx, by + bz + ty, tx * 4 + vec + 32] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) + for iterator in range((kv_chunk_len[0] + 15) // 16): + tile_start_s: T.int32 = tz + ty + tile_start_g: T.int32 = iterator * 16 + tz + ty + for j in range(1): + with T.block("KV_load"): + T.reads() + T.writes() + row_g: T.int32 = tile_start_g + j + if row_g < kv_chunk_len[0]: + seq_offset: T.int32 = row_g + page_no: T.int32 = page_table_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + for vec in T.vectorized(4): + freq = T.float32() + K_smem[tile_start_s + j, tx * 4 + vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, pages[page_no, 0, by, page_offset, tx * 4 + vec + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[batch_idx] + row_g) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, tx * 4 + vec]) + V_smem[tile_start_s + j, tx * 4 + vec] = pages[page_no, 1, by, page_offset, tx * 4 + vec] + else: + for vec in T.vectorized(4): + K_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + V_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + T.tvm_storage_sync("shared") + m_prev[0] = st_m[0] + for j in range(1): + for vec in T.vectorized(4): + QK_local[vec] = T.Cast("float32", Q_local[vec]) * T.Cast("float32", K_smem[tz + j, tx * 4 + vec]) * attn_score_scaling_factor * sm_scale + S_reduce_local[0] = T.float32(0.0) + for vec in T.unroll(4): + S_reduce_local[0] = S_reduce_local[0] + QK_local[vec] + with T.block("block_cross_thread"): + T.reads(S_reduce_local[0]) + T.writes(t0[0]) + T.attr(T.comm_reducer(lambda x0, y0: x0 + y0, [T.float32(0.0)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))) + T.tvm_thread_allreduce(T.uint32(1), S_reduce_local[0], T.bool(True), t0[0], tx) + S_local[j] = T.float32(-50000.0) + if iterator * 16 + tz + j < kv_chunk_len[0]: + S_local[j] = t0[0] + st_m[0] = T.max(st_m[0], S_local[j]) + o_scale: T.float32 = T.exp2(m_prev[0] - st_m[0]) + st_d[0] = st_d[0] * o_scale + for j in range(1): + S_local[j] = T.exp2(S_local[j] - st_m[0]) + st_d[0] = st_d[0] + S_local[j] + for j in T.vectorized(4): + O_local[j] = O_local[j] * o_scale + for j in range(1): + for vec in T.vectorized(4): + V_local[vec] = V_smem[tz + j, tx * 4 + vec] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] + T.Cast("float32", V_local[vec]) * S_local[j] + for vec in T.vectorized(4): + O_allreduce[tz, ty, tx * 4 + vec] = O_local[vec] + md_allreduce[tz, ty, 0] = st_m[0] + md_allreduce[tz, ty, 1] = st_d[0] + T.tvm_storage_sync("shared") + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for j in range(16): + m_prev[0] = st_m[0] + d_prev[0] = st_d[0] + other_m[0] = md_allreduce[j, ty, 0] + other_d[0] = md_allreduce[j, ty, 1] + for vec in T.vectorized(4): + other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] + st_m[0] = T.max(st_m[0], other_m[0]) + st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) + exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) + exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] / st_d[0] + for vec in T.vectorized(4): + output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) + lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) + + @T.prim_func + def gpt_neox_q0f16_batch_decode_paged_kv_sliding_window(_0: T.int32, Q_handle: T.handle, pages_handle: T.handle, page_table_indptr_handle: T.handle, page_table_values_handle: T.handle, var_length_info: T.handle, k_rope_pos_offset_handle: T.handle, q_rope_position_handle: T.handle, output_handle: T.handle, lse_handle: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + B = T.int32(is_size_var=True) + Q = T.match_buffer(Q_handle, (B, 16, 64), "float16") + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(pages_handle, (max_num_pages, 2, 16, 16, 64), "float16", offset_factor=1) + page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (3, B), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) + output = T.match_buffer(output_handle, (B, 16, 64), "float16") + lse = T.match_buffer(lse_handle, (B, 16)) + # with T.block("root"): + sm_scale: T.float32 = T.float32(0.18033688011112042) + for bx in T.thread_binding(B, thread="blockIdx.x"): + for fused_by_bz in T.thread_binding(16, thread="blockIdx.y"): + for ty in T.thread_binding(1, thread="threadIdx.y"): + for tx in T.thread_binding(16, thread="threadIdx.x"): + for tz in T.thread_binding(16, thread="threadIdx.z"): + with T.block("attn"): + T.reads(page_table_indptr[bx:bx + 2], length_info[0:3, bx], q_rope_position[bx], Q[bx, fused_by_bz // 16 + ty + fused_by_bz % 16, tx * 4 - 32:tx * 4 - 32 + 68]) + T.writes(output[bx, fused_by_bz % 16 + fused_by_bz // 16 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 16 + fused_by_bz // 16 + ty]) + Q_local = T.alloc_buffer((4,), "float16", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + O_allreduce = T.alloc_buffer((16, 1, 64), scope="shared") + md_allreduce = T.alloc_buffer((16, 1, 2), scope="shared") + S_reduce_local = T.alloc_buffer((1,), scope="local") + t0 = T.alloc_buffer((1,), scope="local") + S_local = T.alloc_buffer((1,), scope="local") + QK_local = T.alloc_buffer((4,), scope="local") + V_local = T.alloc_buffer((4,), "float16", scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_prev = T.alloc_buffer((1,), scope="local") + other_m = T.alloc_buffer((1,), scope="local") + other_d = T.alloc_buffer((1,), scope="local") + exp_mprev = T.alloc_buffer((1,), scope="local") + exp_otherm = T.alloc_buffer((1,), scope="local") + other_o = T.alloc_buffer((4,), scope="local") + st_m = T.alloc_buffer((1,), scope="local") + st_d = T.alloc_buffer((1,), scope="local") + O_local = T.alloc_buffer((4,), scope="local") + by: T.int32 = fused_by_bz % 16 + bz: T.int32 = fused_by_bz // 16 + batch_idx: T.int32 = bx + cur_page_indptr_begin: T.int32 = page_table_indptr[batch_idx] + cur_page_indptr_end: T.int32 = page_table_indptr[batch_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[0, batch_idx] - length_info[1, batch_idx] + length_info[2, batch_idx], 0) + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for vec in T.vectorized(4): + freq = T.float32() + Q_local[vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", Q[bx, by + bz + ty, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, Q[bx, by + bz + ty, tx * 4 + vec + 32] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) + for iterator in range((kv_chunk_len[0] + 15) // 16): + tile_start_s: T.int32 = tz + ty + tile_start_g: T.int32 = iterator * 16 + tz + ty + for j in range(1): + with T.block("KV_load"): + T.reads() + T.writes() + row_g: T.int32 = tile_start_g + j + if row_g < kv_chunk_len[0]: + seq_offset: T.int32 = T.if_then_else(row_g < length_info[2, batch_idx], row_g, row_g - length_info[2, batch_idx] + length_info[1, batch_idx]) + page_no: T.int32 = page_table_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + for vec in T.vectorized(4): + freq = T.float32() + K_smem[tile_start_s + j, tx * 4 + vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, pages[page_no, 0, by, page_offset, tx * 4 + vec + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[batch_idx] + row_g) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, tx * 4 + vec]) + V_smem[tile_start_s + j, tx * 4 + vec] = pages[page_no, 1, by, page_offset, tx * 4 + vec] + else: + for vec in T.vectorized(4): + K_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + V_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + T.tvm_storage_sync("shared") + m_prev[0] = st_m[0] + for j in range(1): + for vec in T.vectorized(4): + QK_local[vec] = T.Cast("float32", Q_local[vec]) * T.Cast("float32", K_smem[tz + j, tx * 4 + vec]) * attn_score_scaling_factor * sm_scale + S_reduce_local[0] = T.float32(0.0) + for vec in T.unroll(4): + S_reduce_local[0] = S_reduce_local[0] + QK_local[vec] + with T.block("block_cross_thread"): + T.reads(S_reduce_local[0]) + T.writes(t0[0]) + T.attr(T.comm_reducer(lambda x0, y0: x0 + y0, [T.float32(0.0)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))) + T.tvm_thread_allreduce(T.uint32(1), S_reduce_local[0], T.bool(True), t0[0], tx) + S_local[j] = T.float32(-50000.0) + if iterator * 16 + tz + j < kv_chunk_len[0]: + S_local[j] = t0[0] + st_m[0] = T.max(st_m[0], S_local[j]) + o_scale: T.float32 = T.exp2(m_prev[0] - st_m[0]) + st_d[0] = st_d[0] * o_scale + for j in range(1): + S_local[j] = T.exp2(S_local[j] - st_m[0]) + st_d[0] = st_d[0] + S_local[j] + for j in T.vectorized(4): + O_local[j] = O_local[j] * o_scale + for j in range(1): + for vec in T.vectorized(4): + V_local[vec] = V_smem[tz + j, tx * 4 + vec] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] + T.Cast("float32", V_local[vec]) * S_local[j] + for vec in T.vectorized(4): + O_allreduce[tz, ty, tx * 4 + vec] = O_local[vec] + md_allreduce[tz, ty, 0] = st_m[0] + md_allreduce[tz, ty, 1] = st_d[0] + T.tvm_storage_sync("shared") + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for j in range(16): + m_prev[0] = st_m[0] + d_prev[0] = st_d[0] + other_m[0] = md_allreduce[j, ty, 0] + other_d[0] = md_allreduce[j, ty, 1] + for vec in T.vectorized(4): + other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] + st_m[0] = T.max(st_m[0], other_m[0]) + st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) + exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) + exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] / st_d[0] + for vec in T.vectorized(4): + output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) + lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) + + @T.prim_func + def gpt_neox_q0f16_batch_prefill_paged_kv(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + total_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (total_len, 16, 64), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 16, 16, 64), "float16", offset_factor=1) + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 16, 64), "float16") + lse = T.match_buffer(var_lse, (total_len, 16)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(16, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 16), scope="shared") + S_local = T.alloc_buffer((32, 16), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(4): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) + j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 15) // 16): + L_kv_start: T.int32 = iterator_1 * 16 + for lz_ly_fused_0 in range(2): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + freq = T.float32() + K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, pages[page_no, 0, by, page_offset, j + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, j]) + else: + K_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_ly_fused_0 in range(2): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:16, 0:64]) + T.writes(S_local[0:32, 0:16]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + k = T.axis.reduce(64, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 2): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(16): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) + T.writes(S_smem[row, 0:16]) + for j in range(16): + if row < 32: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(16): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:16], V_smem[0:16, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(2, 8, 4, 4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + k = T.axis.reduce(16, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def gpt_neox_q0f16_batch_prefill_paged_kv_sliding_window(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + total_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (total_len, 16, 64), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 16, 16, 64), "float16", offset_factor=1) + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (3, batch_size), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 16, 64), "float16") + lse = T.match_buffer(var_lse, (total_len, 16)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(16, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 16), scope="shared") + S_local = T.alloc_buffer((32, 16), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[0, b_idx] - length_info[1, b_idx] + length_info[2, b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(4): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) + j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 15) // 16): + L_kv_start: T.int32 = iterator_1 * 16 + for lz_ly_fused_0 in range(2): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = T.if_then_else(cur_L < length_info[2, b_idx], cur_L, cur_L - length_info[2, b_idx] + length_info[1, b_idx]) + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + freq = T.float32() + K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, pages[page_no, 0, by, page_offset, j + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, j]) + else: + K_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_ly_fused_0 in range(2): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = T.if_then_else(cur_L < length_info[2, b_idx], cur_L, cur_L - length_info[2, b_idx] + length_info[1, b_idx]) + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:16, 0:64]) + T.writes(S_local[0:32, 0:16]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + k = T.axis.reduce(64, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 2): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(16): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) + T.writes(S_smem[row, 0:16]) + for j in range(16): + if row < 32: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(16): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:16], V_smem[0:16, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(2, 8, 4, 4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + k = T.axis.reduce(16, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def gpt_neox_q0f16_batch_prefill_ragged_kv(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_k_rope_pos_offset: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + qo_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (qo_len, 16, 64), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + kv_len = T.int32(is_size_var=True) + k = T.match_buffer(var_k, (kv_len, 16, 64), "float16") + v = T.match_buffer(var_v, (kv_len, 16, 64), "float16") + kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (qo_len, 16, 64), "float16") + lse = T.match_buffer(var_lse, (qo_len, 16)) + # with T.block("root"): + for lbx in T.thread_binding(8, thread="blockIdx.x"): + for lby in T.thread_binding(16, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((64, 32), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 32), scope="shared") + S_local = T.alloc_buffer((32, 32), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + q_indptr_val: T.int32 = q_indptr[b_idx] + LH_start: T.int32 = tile_id[0] * 32 + kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1_0 * 4 + lj_1_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_1_lj_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_1_lj_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_0, lj_0_0 in T.grid(1, 2): + for lj_1 in T.vectorized(8): + with T.block("Q_load"): + i = T.axis.spatial(32, li_0 * 32 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) // 4) + j = T.axis.spatial(64, lj_0_0 * 32 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) % 4 * 8 + lj_1) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 31) // 32): + L_kv_start: T.int32 = iterator_1 * 32 + L_kv_base: T.int32 = kv_indptr[b_idx] + for lz_1_ly_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for lz_1_ly_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lz_0, ly_0_0 in T.grid(1, 2): + for ly_1 in T.vectorized(8): + with T.block("K_load"): + i = T.axis.spatial(32, lz_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 4) + j = T.axis.spatial(64, ly_0_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 4 * 8 + ly_1) + T.reads(kv_chunk_len[0], k_rope_pos_offset[b_idx], k[L_kv_base + L_kv_start + i, by, j - 32:j - 32 + 65]) + T.writes(K_smem[j, i]) + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + freq = T.float32() + K_smem[j, i] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", k[L_kv_base + cur_L, by, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, k[L_kv_base + cur_L, by, j + 32] * T.float16(-1.0), k[L_kv_base + cur_L, by, j - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), k[L_kv_base + cur_L, by, j]) + else: + K_smem[j, i] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_1_ly_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for lz_1_ly_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lz_0, ly_0_0 in T.grid(1, 2): + for ly_1 in T.vectorized(8): + with T.block("V_load"): + i = T.axis.spatial(32, lz_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 4) + j = T.axis.spatial(64, ly_0_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 4 * 8 + ly_1) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + V_smem[i, j] = v[L_kv_base + cur_L, by, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:64, 0:32]) + T.writes(S_local[0:32, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init in T.unroll(2): + for lj_1_0_init in T.unroll(1): + for lj_1_1_init in T.vectorized(4): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_0_init * 4 + lj_1_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0 in range(4): + for li_1 in T.unroll(2): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + for lk_1 in range(16): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1_0 * 4 + lj_1_1) + k_1 = T.axis.reduce(64, lk_0 * 16 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k_1], K_smem[k_1, j]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k_1]) * T.Cast("float32", K_smem[k_1, j]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(2): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1_0 * 4 + lj_1_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 32: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init in T.unroll(4): + for lj_1_0_init in T.unroll(1): + for lj_1_1_init in T.vectorized(4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_0_init * 4 + lj_1_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1 in T.grid(2, 16): + for li_1 in T.unroll(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1_0 * 4 + lj_1_1) + k_1 = T.axis.reduce(32, lk_0 * 16 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k_1], V_smem[k_1, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k_1] * T.Cast("float32", V_smem[k_1, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1_0 * 4 + lj_1_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 8 + + @T.prim_func + def gpt_neox_q0f16_batch_tree_attn(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_mn_indptr: T.handle, var_mask: T.handle, var_output: T.handle, var_lse: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32, batch_size: T.int32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + qo_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (qo_len, 16, 64), "float16") + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + kv_len = T.int32(is_size_var=True) + k = T.match_buffer(var_k, (kv_len, 16, 64), "float16") + v = T.match_buffer(var_v, (kv_len, 16, 64), "float16") + kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) + mn_indptr = T.match_buffer(var_mn_indptr, (batch_size + 1,), "int32", offset_factor=1) + tree_size = T.int32(is_size_var=True) + mask = T.match_buffer(var_mask, (tree_size, 2), "int32", offset_factor=1) + output = T.match_buffer(var_output, (qo_len, 16, 64), "float16") + lse = T.match_buffer(var_lse, (qo_len, 16)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(16, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 32), scope="shared") + S_local = T.alloc_buffer((32, 32), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(4): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) + j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 31) // 32): + L_kv_start: T.int32 = iterator_1 * 32 + L_kv_base: T.int32 = kv_indptr[b_idx] + for lz_ly_fused_0 in range(4): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("KV_load"): + i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_base + L_kv_start + i + if L_kv_start + i < kv_chunk_len[0]: + freq = T.float32() + K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", k[cur_L, by, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, k[cur_L, by, j + 32] * T.float16(-1.0), k[cur_L, by, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), k[cur_L, by, j]) + V_smem[i, j] = v[cur_L, by, j] + else: + K_smem[i, j] = T.float16(0.0) + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:32, 0:64]) + T.writes(S_local[0:32, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 4): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 4, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + k_1 = T.axis.reduce(64, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k_1], K_smem[j, k_1]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k_1]) * T.Cast("float32", K_smem[j, k_1]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 4): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 32: + row_: T.int32 = LH_start + row + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(4, 8, 4, 4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + k_1 = T.axis.reduce(32, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k_1], V_smem[k_1, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k_1] * T.Cast("float32", V_smem[k_1, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def gpt_neox_q0f16_chunk_lse(var_A: T.handle, var_temperature: T.handle, var_chunked_sum: T.handle, var_chunked_max: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size, vocab_size = T.int32(is_size_var=True), T.int32(is_size_var=True) + A = T.match_buffer(var_A, (batch_size, vocab_size)) + temperature = T.match_buffer(var_temperature, (batch_size,)) + num_chunks = T.int32(is_size_var=True) + chunked_sum = T.match_buffer(var_chunked_sum, (batch_size, num_chunks)) + chunked_max = T.match_buffer(var_chunked_max, (batch_size, num_chunks)) + # with T.block("root"): + temp_max_shared = T.alloc_buffer((batch_size, num_chunks), scope="shared") + temp_sum_shared = T.alloc_buffer((batch_size, num_chunks), scope="shared") + for ax0_ax1_fused in T.thread_binding(batch_size * num_chunks, thread="blockIdx.x"): + for ax0, ax1 in T.grid(1, 1): + for ax2_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + for ax2_fused_0 in T.serial(16, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("max"): + v0 = T.axis.spatial(batch_size, ax0_ax1_fused % (num_chunks * batch_size) // num_chunks + ax0) + v1 = T.axis.spatial(num_chunks, ax0_ax1_fused % num_chunks + ax1) + v2 = T.axis.reduce(4096, ax2_fused_0 * 256 + ax2_fused_1) + T.reads(temperature[v0], A[v0, v1 * 4096 + v2]) + T.writes(temp_max_shared[v0, v1]) + with T.init(): + temp_max_shared[v0, v1] = T.float32(-340282346638528859811704183484516925440.0) + temp_max_shared[v0, v1] = T.max(temp_max_shared[v0, v1], T.if_then_else(v1 * 4096 + v2 < vocab_size, T.if_then_else(temperature[v0] > T.float32(1.0000000000000001e-05), A[v0, v1 * 4096 + v2] / temperature[v0], A[v0, v1 * 4096 + v2]), T.float32(-340282346638528859811704183484516925440.0))) + for ax0, ax1 in T.grid(1, 1): + for ax2_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + for ax2_fused_0 in T.serial(16, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("sum_exp"): + v0 = T.axis.spatial(batch_size, ax0_ax1_fused % (num_chunks * batch_size) // num_chunks + ax0) + v1 = T.axis.spatial(num_chunks, ax0_ax1_fused % num_chunks + ax1) + v2 = T.axis.reduce(4096, ax2_fused_0 * 256 + ax2_fused_1) + T.reads(temperature[v0], A[v0, v1 * 4096 + v2], temp_max_shared[v0, v1]) + T.writes(temp_sum_shared[v0, v1]) + with T.init(): + temp_sum_shared[v0, v1] = T.float32(0.0) + temp_sum_shared[v0, v1] = temp_sum_shared[v0, v1] + T.if_then_else(v1 * 4096 + v2 < vocab_size, T.Select(temperature[v0] > T.float32(1.0000000000000001e-05), T.exp(T.if_then_else(v1 * 4096 + v2 < vocab_size, T.if_then_else(temperature[v0] > T.float32(1.0000000000000001e-05), A[v0, v1 * 4096 + v2] / temperature[v0], A[v0, v1 * 4096 + v2]), T.float32(-340282346638528859811704183484516925440.0)) - temp_max_shared[v0, v1]), T.Cast("float32", T.if_then_else(v1 * 4096 + v2 < vocab_size, T.if_then_else(temperature[v0] > T.float32(1.0000000000000001e-05), A[v0, v1 * 4096 + v2] / temperature[v0], A[v0, v1 * 4096 + v2]), T.float32(-340282346638528859811704183484516925440.0)) == temp_max_shared[v0, v1])), T.float32(0.0)) + for ax2_1 in T.thread_binding(256, thread="threadIdx.x"): + for ax2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("log"): + v0 = T.axis.spatial(batch_size, ax0_ax1_fused % (num_chunks * batch_size) // num_chunks) + v1 = T.axis.spatial(num_chunks, ax0_ax1_fused % num_chunks) + v2 = T.axis.spatial(1, ax2_0 * 256 + ax2_1) + T.where(ax2_0 * 256 + ax2_1 < 1) + T.reads(temperature[v0], temp_sum_shared[v0, v1], temp_max_shared[v0, v1]) + T.writes(chunked_sum[v0, v1], chunked_max[v0, v1]) + chunked_sum[v0, v1] = T.Select(temperature[v0] > T.float32(1.0000000000000001e-05), T.log(temp_sum_shared[v0, v1]), temp_sum_shared[v0, v1]) + chunked_max[v0, v1] = temp_max_shared[v0, v1] + + @T.prim_func + def gpt_neox_q0f16_compact_kv_copy(var_pages: T.handle, var_copy_length_indptr: T.handle, var_copy_src_dst_pos: T.handle, batch_size: T.int32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + num_pages = T.int32() + pages = T.match_buffer(var_pages, (num_pages, 2, 16, 16, 64), "float16", offset_factor=1) + copy_length_indptr = T.match_buffer(var_copy_length_indptr, (batch_size + 1,), "int32", offset_factor=1) + total_copy_length = T.int32() + copy_src_dst_pos = T.match_buffer(var_copy_src_dst_pos, (2, total_copy_length), "int32", offset_factor=1) + with T.block("root"): + T.reads() + T.writes() + for bhd_o in T.thread_binding(batch_size * 4, thread="blockIdx.x"): + for bhd_i in T.thread_binding(256, thread="threadIdx.x"): + b: T.int32 = (bhd_o * 256 + bhd_i) // 1024 + h: T.int32 = (bhd_o * 256 + bhd_i) // 64 % 16 + d: T.int32 = (bhd_o * 256 + bhd_i) % 64 + if bhd_o * 256 + bhd_i < batch_size * 16 * 64: + for i in range(copy_length_indptr[b + 1] - copy_length_indptr[b]): + src_pos: T.int32 = copy_src_dst_pos[0, copy_length_indptr[b] + i] + dst_pos: T.int32 = copy_src_dst_pos[1, copy_length_indptr[b] + i] + pages[dst_pos // 16, 0, h, dst_pos % 16, d] = pages[src_pos // 16, 0, h, src_pos % 16, d] + pages[dst_pos // 16, 1, h, dst_pos % 16, d] = pages[src_pos // 16, 1, h, src_pos % 16, d] + + @T.prim_func + def gpt_neox_q0f16_copy_single_page(var_pages: T.handle, src_page_id: T.int32, tgt_page_id: T.int32, copy_length: T.int32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + num_pages, page_size = T.int32(), T.int32() + pages = T.match_buffer(var_pages, (num_pages, 2, 16, page_size, 64), "float16", offset_factor=1) + # with T.block("root"): + for b in T.thread_binding(copy_length * 4, thread="blockIdx.x"): + for t in T.thread_binding(256, thread="threadIdx.x"): + with T.block("copy"): + vh = T.axis.spatial(16, (b * 256 + t) // (copy_length * 64)) + vp = T.axis.spatial(copy_length, (b * 256 + t) % (copy_length * 64) // 64) + vd = T.axis.spatial(64, (b * 256 + t) % 64) + T.where(b * 256 + t < copy_length * 16 * 64) + T.reads(pages[src_page_id, 0:2, vh, vp, vd]) + T.writes(pages[tgt_page_id, 0:2, vh, vp, vd]) + pages[tgt_page_id, 0, vh, vp, vd] = pages[src_page_id, 0, vh, vp, vd] + pages[tgt_page_id, 1, vh, vp, vd] = pages[src_page_id, 1, vh, vp, vd] + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul10_add10(layer_norm49: T.Buffer((1, 1, 1024), "float16"), gpt_neox_layers_0_attention_query_key_value_weight2: T.Buffer((3072, 1024), "float16"), gpt_neox_layers_0_attention_query_key_value_bias2: T.Buffer((3072,), "float16"), T_add_intermediate: T.Buffer((1, 1, 3072), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + NT_matmul_intermediate_local = T.alloc_buffer((1, 1, 3072), "float16", scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((256, 1, 1, 3072), "float16", scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((32, 1, 1, 3072), "float16", scope="local") + gpt_neox_layers_0_attention_query_key_value_weight2_local = T.alloc_buffer((3072, 1024), "float16", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(1536, thread="blockIdx.x"): + for u_fused_ax0_fused_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for u_fused_ax0_fused_fused_2_init in range(1): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused = T.axis.spatial(256, ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1_init) + v0 = T.axis.spatial(3072, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0]) + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] = T.float16(0.0) + for ax1_fused_u_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_ax1_fused_0 in range(8): + for ax0_ax1_fused_1 in T.vectorized(1): + with T.block("gpt_neox_layers_0_attention_query_key_value_weight2_local"): + v0 = T.axis.spatial(3072, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(1024, ax1_fused_u_fused_0 * 256 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax0_ax1_fused_0 + ax0_ax1_fused_1) + T.reads(gpt_neox_layers_0_attention_query_key_value_weight2[v0, v1]) + T.writes(gpt_neox_layers_0_attention_query_key_value_weight2_local[v0, v1]) + gpt_neox_layers_0_attention_query_key_value_weight2_local[v0, v1] = gpt_neox_layers_0_attention_query_key_value_weight2[v0, v1] + for u_fused_ax0_fused_fused_2, ax1_fused_u_fused_2 in T.grid(1, 1): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused = T.axis.spatial(256, ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1) + v0 = T.axis.spatial(3072, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2) + vax1_fused_u_fused_2, vax1_fused_u_fused_0 = T.axis.remap("RR", [ax1_fused_u_fused_2, ax1_fused_u_fused_0]) + T.reads(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0], layer_norm49[0, 0, vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused], gpt_neox_layers_0_attention_query_key_value_weight2_local[v0, vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused]) + T.writes(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0]) + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] = NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] + layer_norm49[0, 0, vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused] * gpt_neox_layers_0_attention_query_key_value_weight2_local[v0, vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused] + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax2_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2_fused_2_1 in T.vectorized(1): + with T.block("NT_matmul_rf_init"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial(3072, u_fused_ax0_fused_fused_0 * 2 + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] = T.float16(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial(3072, u_fused_ax0_fused_fused_0 * 2 + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) + T.reads(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0], NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1, 0, 0, v0]) + T.writes(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] = NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1, 0, 0, v0] + for ax1_fused_2 in range(1): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial(3072, u_fused_ax0_fused_fused_0 * 2 + ax1_fused_0_ax1_fused_1_fused + ax1_fused_2) + T.reads(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + T.writes(NT_matmul_intermediate_local[0, 0, v0]) + with T.init(): + NT_matmul_intermediate_local[0, 0, v0] = T.float16(0.0) + NT_matmul_intermediate_local[0, 0, v0] = NT_matmul_intermediate_local[0, 0, v0] + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] + for ax0_fused_0_ax0_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0_fused_2 in range(1): + with T.block("T_add"): + v0 = T.axis.spatial(3072, u_fused_ax0_fused_fused_0 * 2 + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) + T.reads(NT_matmul_intermediate_local[0, 0, v0], gpt_neox_layers_0_attention_query_key_value_bias2[v0]) + T.writes(T_add_intermediate[0, 0, v0]) + T_add_intermediate[0, 0, v0] = NT_matmul_intermediate_local[0, 0, v0] + gpt_neox_layers_0_attention_query_key_value_bias2[v0] + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(reshape99: T.Buffer((1, 1, 1024), "float16"), gpt_neox_layers_0_attention_dense_weight2: T.Buffer((1024, 1024), "float16"), gpt_neox_layers_0_attention_dense_bias2: T.Buffer((1024,), "float16"), astype50: T.Buffer((1, 1, 1024), "float16"), input_embed: T.Buffer((1, 1, 1024), "float16"), T_add_intermediate_1_2: T.Buffer((1, 1, 1024), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + NT_matmul_intermediate_local = T.alloc_buffer((1, 1, 1024), "float16", scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((256, 1, 1, 1024), "float16", scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((32, 1, 1, 1024), "float16", scope="local") + gpt_neox_layers_0_attention_dense_weight2_local = T.alloc_buffer((1024, 1024), "float16", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(512, thread="blockIdx.x"): + for u_fused_ax0_fused_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for u_fused_ax0_fused_fused_2_init in range(1): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused = T.axis.spatial(256, ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1_init) + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0]) + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] = T.float16(0.0) + for ax1_fused_u_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_ax1_fused_0 in range(8): + for ax0_ax1_fused_1 in T.vectorized(1): + with T.block("gpt_neox_layers_0_attention_dense_weight2_local"): + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(1024, ax1_fused_u_fused_0 * 256 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax0_ax1_fused_0 + ax0_ax1_fused_1) + T.reads(gpt_neox_layers_0_attention_dense_weight2[v0, v1]) + T.writes(gpt_neox_layers_0_attention_dense_weight2_local[v0, v1]) + gpt_neox_layers_0_attention_dense_weight2_local[v0, v1] = gpt_neox_layers_0_attention_dense_weight2[v0, v1] + for u_fused_ax0_fused_fused_2, ax1_fused_u_fused_2 in T.grid(1, 1): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused = T.axis.spatial(256, ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1) + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2) + vax1_fused_u_fused_2, vax1_fused_u_fused_0 = T.axis.remap("RR", [ax1_fused_u_fused_2, ax1_fused_u_fused_0]) + T.reads(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0], reshape99[0, 0, vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused], gpt_neox_layers_0_attention_dense_weight2_local[v0, vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused]) + T.writes(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0]) + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] = NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] + reshape99[0, 0, vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused] * gpt_neox_layers_0_attention_dense_weight2_local[v0, vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused] + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax2_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2_fused_2_1 in T.vectorized(1): + with T.block("NT_matmul_rf_init"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] = T.float16(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) + T.reads(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0], NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1, 0, 0, v0]) + T.writes(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] = NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1, 0, 0, v0] + for ax1_fused_2 in range(1): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + ax1_fused_0_ax1_fused_1_fused + ax1_fused_2) + T.reads(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + T.writes(NT_matmul_intermediate_local[0, 0, v0]) + with T.init(): + NT_matmul_intermediate_local[0, 0, v0] = T.float16(0.0) + NT_matmul_intermediate_local[0, 0, v0] = NT_matmul_intermediate_local[0, 0, v0] + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] + for ax0_fused_0_ax0_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0_fused_2 in range(1): + with T.block("T_add_2"): + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) + T.reads(astype50[0, 0, v0], NT_matmul_intermediate_local[0, 0, v0], gpt_neox_layers_0_attention_dense_bias2[v0], input_embed[0, 0, v0]) + T.writes(T_add_intermediate_1_2[0, 0, v0]) + T_add_intermediate_1_2[0, 0, v0] = astype50[0, 0, v0] + (NT_matmul_intermediate_local[0, 0, v0] + gpt_neox_layers_0_attention_dense_bias2[v0]) + input_embed[0, 0, v0] + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(layer_norm50: T.Buffer((1, 1, 1024), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_weight2: T.Buffer((4096, 1024), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: T.Buffer((4096,), "float32"), compute_intermediate: T.Buffer((1, 1, 4096), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + NT_matmul_intermediate_local = T.alloc_buffer((1, 1, 4096), scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((256, 1, 1, 4096), scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((32, 1, 1, 4096), scope="local") + gpt_neox_layers_0_mlp_dense_h_to_4h_weight2_local = T.alloc_buffer((4096, 1024), "float16", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(2048, thread="blockIdx.x"): + for u_fused_ax0_fused_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for u_fused_ax0_fused_fused_2_init in range(1): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused = T.axis.spatial(256, ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1_init) + v0 = T.axis.spatial(4096, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0]) + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] = T.float32(0.0) + for ax1_fused_u_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_ax1_fused_0 in range(8): + for ax0_ax1_fused_1 in T.vectorized(1): + with T.block("gpt_neox_layers_0_mlp_dense_h_to_4h_weight2_local"): + v0 = T.axis.spatial(4096, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(1024, ax1_fused_u_fused_0 * 256 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax0_ax1_fused_0 + ax0_ax1_fused_1) + T.reads(gpt_neox_layers_0_mlp_dense_h_to_4h_weight2[v0, v1]) + T.writes(gpt_neox_layers_0_mlp_dense_h_to_4h_weight2_local[v0, v1]) + gpt_neox_layers_0_mlp_dense_h_to_4h_weight2_local[v0, v1] = gpt_neox_layers_0_mlp_dense_h_to_4h_weight2[v0, v1] + for u_fused_ax0_fused_fused_2, ax1_fused_u_fused_2 in T.grid(1, 1): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused = T.axis.spatial(256, ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1) + v0 = T.axis.spatial(4096, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2) + vax1_fused_u_fused_2, vax1_fused_u_fused_0 = T.axis.remap("RR", [ax1_fused_u_fused_2, ax1_fused_u_fused_0]) + T.reads(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0], layer_norm50[0, 0, vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused], gpt_neox_layers_0_mlp_dense_h_to_4h_weight2_local[v0, vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused]) + T.writes(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0]) + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] = NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] + T.Cast("float32", layer_norm50[0, 0, vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_h_to_4h_weight2_local[v0, vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused]) + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax2_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2_fused_2_1 in T.vectorized(1): + with T.block("NT_matmul_rf_init"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial(4096, u_fused_ax0_fused_fused_0 * 2 + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] = T.float32(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial(4096, u_fused_ax0_fused_fused_0 * 2 + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) + T.reads(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0], NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1, 0, 0, v0]) + T.writes(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] = NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1, 0, 0, v0] + for ax1_fused_2 in range(1): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial(4096, u_fused_ax0_fused_fused_0 * 2 + ax1_fused_0_ax1_fused_1_fused + ax1_fused_2) + T.reads(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + T.writes(NT_matmul_intermediate_local[0, 0, v0]) + with T.init(): + NT_matmul_intermediate_local[0, 0, v0] = T.float32(0.0) + NT_matmul_intermediate_local[0, 0, v0] = NT_matmul_intermediate_local[0, 0, v0] + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] + for ax0_fused_0_ax0_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0_fused_2 in range(1): + with T.block("compute_1"): + v0 = T.axis.spatial(4096, u_fused_ax0_fused_fused_0 * 2 + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) + T.reads(NT_matmul_intermediate_local[0, 0, v0], gpt_neox_layers_0_mlp_dense_h_to_4h_bias2[v0]) + T.writes(compute_intermediate[0, 0, v0]) + compute_intermediate[0, 0, v0] = T.Cast("float16", (NT_matmul_intermediate_local[0, 0, v0] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias2[v0]) * (T.float32(0.5) + T.erf((NT_matmul_intermediate_local[0, 0, v0] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias2[v0]) * T.float32(0.70710678118654757)) * T.float32(0.5))) + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(astype49: T.Buffer((1, 1, 4096), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_weight2: T.Buffer((1024, 4096), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: T.Buffer((1024,), "float32"), compute_intermediate: T.Buffer((1, 1, 1024), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + NT_matmul_intermediate_local = T.alloc_buffer((1, 1, 1024), scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((256, 1, 1, 1024), scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((32, 1, 1, 1024), scope="local") + gpt_neox_layers_0_mlp_dense_4h_to_h_weight2_local = T.alloc_buffer((1024, 4096), "float16", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(512, thread="blockIdx.x"): + for u_fused_ax0_fused_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for u_fused_ax0_fused_fused_2_init in range(1): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused = T.axis.spatial(256, ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1_init) + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0]) + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] = T.float32(0.0) + for ax1_fused_u_fused_0 in T.serial(16, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_ax1_fused_0 in range(8): + for ax0_ax1_fused_1 in T.vectorized(1): + with T.block("gpt_neox_layers_0_mlp_dense_4h_to_h_weight2_local"): + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(4096, ax1_fused_u_fused_0 * 256 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax0_ax1_fused_0 + ax0_ax1_fused_1) + T.reads(gpt_neox_layers_0_mlp_dense_4h_to_h_weight2[v0, v1]) + T.writes(gpt_neox_layers_0_mlp_dense_4h_to_h_weight2_local[v0, v1]) + gpt_neox_layers_0_mlp_dense_4h_to_h_weight2_local[v0, v1] = gpt_neox_layers_0_mlp_dense_4h_to_h_weight2[v0, v1] + for u_fused_ax0_fused_fused_2, ax1_fused_u_fused_2 in T.grid(1, 1): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused = T.axis.spatial(256, ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1) + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2) + vax1_fused_u_fused_2, vax1_fused_u_fused_0 = T.axis.remap("RR", [ax1_fused_u_fused_2, ax1_fused_u_fused_0]) + T.reads(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0], astype49[0, 0, vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused], gpt_neox_layers_0_mlp_dense_4h_to_h_weight2_local[v0, vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused]) + T.writes(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0]) + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] = NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] + T.Cast("float32", astype49[0, 0, vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_4h_to_h_weight2_local[v0, vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused]) + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax2_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2_fused_2_1 in T.vectorized(1): + with T.block("NT_matmul_rf_init"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] = T.float32(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) + T.reads(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0], NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1, 0, 0, v0]) + T.writes(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] = NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1, 0, 0, v0] + for ax1_fused_2 in range(1): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + ax1_fused_0_ax1_fused_1_fused + ax1_fused_2) + T.reads(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + T.writes(NT_matmul_intermediate_local[0, 0, v0]) + with T.init(): + NT_matmul_intermediate_local[0, 0, v0] = T.float32(0.0) + NT_matmul_intermediate_local[0, 0, v0] = NT_matmul_intermediate_local[0, 0, v0] + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] + for ax0_fused_0_ax0_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0_fused_2 in range(1): + with T.block("compute"): + v0 = T.axis.spatial(1024, u_fused_ax0_fused_fused_0 * 2 + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) + T.reads(NT_matmul_intermediate_local[0, 0, v0], gpt_neox_layers_0_mlp_dense_4h_to_h_bias2[v0]) + T.writes(compute_intermediate[0, 0, v0]) + compute_intermediate[0, 0, v0] = T.Cast("float16", NT_matmul_intermediate_local[0, 0, v0] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias2[v0]) + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul14_cast8(layer_norm97: T.Buffer((1, 1, 1024), "float16"), p_embed_out_weight2: T.handle, p_output0: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + vocab_size = T.int32() + embed_out_weight2 = T.match_buffer(p_embed_out_weight2, (vocab_size, 1024), "float16") + compute_intermediate = T.match_buffer(p_output0, (1, 1, vocab_size)) + # with T.block("root"): + NT_matmul_intermediate_local = T.alloc_buffer((1, 1, vocab_size), "float16", scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((256, 1, 1, vocab_size), "float16", scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((32, 1, 1, vocab_size), "float16", scope="local") + embed_out_weight2_local = T.alloc_buffer((vocab_size, 1024), "float16", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding((vocab_size + 1) // 2, thread="blockIdx.x"): + for u_fused_ax0_fused_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for u_fused_ax0_fused_fused_2_init in range(1): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused = T.axis.spatial(256, ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1_init) + v0 = T.axis.spatial(vocab_size, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2_init) + T.where(u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2_init < vocab_size) + T.reads() + T.writes(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0]) + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] = T.float16(0.0) + for ax1_fused_u_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_ax1_fused_0 in range(8): + for ax0_ax1_fused_1 in T.vectorized(1): + with T.block("embed_out_weight2_local"): + v0 = T.axis.spatial(vocab_size, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(1024, ax1_fused_u_fused_0 * 256 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax0_ax1_fused_0 + ax0_ax1_fused_1) + T.where(u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 < vocab_size) + T.reads(embed_out_weight2[v0, v1]) + T.writes(embed_out_weight2_local[v0, v1]) + embed_out_weight2_local[v0, v1] = embed_out_weight2[v0, v1] + for u_fused_ax0_fused_fused_2, ax1_fused_u_fused_2 in T.grid(1, 1): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused = T.axis.spatial(256, ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1) + v0 = T.axis.spatial(vocab_size, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2) + vax1_fused_u_fused_2, vax1_fused_u_fused_0 = T.axis.remap("RR", [ax1_fused_u_fused_2, ax1_fused_u_fused_0]) + T.where(u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2 < vocab_size) + T.reads(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0], layer_norm97[0, 0, vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused], embed_out_weight2_local[v0, vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused]) + T.writes(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0]) + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] = NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] + layer_norm97[0, 0, vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused] * embed_out_weight2_local[v0, vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused] + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax2_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2_fused_2_1 in T.vectorized(1): + with T.block("NT_matmul_rf_init"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial(vocab_size, u_fused_ax0_fused_fused_0 * 2 + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) + T.where(u_fused_ax0_fused_fused_0 * 2 + (ax2_fused_0_ax2_fused_1_fused % 2 + (ax2_fused_2_0 + ax2_fused_2_1)) < vocab_size) + T.reads() + T.writes(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] = T.float16(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial(vocab_size, u_fused_ax0_fused_fused_0 * 2 + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) + T.where(u_fused_ax0_fused_fused_0 * 2 + (ax2_fused_0_ax2_fused_1_fused % 2 + (ax2_fused_2_0 + ax2_fused_2_1)) < vocab_size) + T.reads(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0], NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1, 0, 0, v0]) + T.writes(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] = NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1, 0, 0, v0] + for ax1_fused_2 in range(1): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial(vocab_size, u_fused_ax0_fused_fused_0 * 2 + ax1_fused_0_ax1_fused_1_fused + ax1_fused_2) + T.where(u_fused_ax0_fused_fused_0 * 2 + (ax1_fused_0_ax1_fused_1_fused % 2 + ax1_fused_2) < vocab_size) + T.reads(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + T.writes(NT_matmul_intermediate_local[0, 0, v0]) + with T.init(): + NT_matmul_intermediate_local[0, 0, v0] = T.float16(0.0) + NT_matmul_intermediate_local[0, 0, v0] = NT_matmul_intermediate_local[0, 0, v0] + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] + for ax0_fused_0_ax0_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0_fused_2 in range(1): + with T.block("compute"): + v0 = T.axis.spatial(vocab_size, u_fused_ax0_fused_fused_0 * 2 + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) + T.where(u_fused_ax0_fused_fused_0 * 2 + (ax0_fused_0_ax0_fused_1_fused % 2 + ax0_fused_2) < vocab_size) + T.reads(NT_matmul_intermediate_local[0, 0, v0]) + T.writes(compute_intermediate[0, 0, v0]) + compute_intermediate[0, 0, v0] = T.Cast("float32", NT_matmul_intermediate_local[0, 0, v0]) + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(p_reshape291: T.handle, gpt_neox_layers_0_attention_dense_weight4: T.Buffer((1024, 1024), "float16"), gpt_neox_layers_0_attention_dense_bias4: T.Buffer((1024,), "float16"), p_astype148: T.handle, p_input_embeds: T.handle, p_output0: T.handle): + T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + reshape291 = T.match_buffer(p_reshape291, (batch_size, 1, 1024), "float16") + astype148 = T.match_buffer(p_astype148, (batch_size, 1, 1024), "float16") + input_embeds = T.match_buffer(p_input_embeds, (batch_size, 1, 1024), "float16") + T_add_intermediate_1_2 = T.match_buffer(p_output0, (batch_size, 1, 1024), "float16") + # with T.block("root"): + if T.tvm_thread_invariant(batch_size <= 2): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + 1) // 2 * 2, 1, 1024), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 1) // 2 * 2, 1, 1024), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 1) // 2 * 2, 1, 1024), "float16", scope="local") + for ax0_0 in T.thread_binding((batch_size + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(256, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(2, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax0_1_init) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = T.float16(0.0) + for ax2_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(2, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax0_1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1], reshape291[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_attention_dense_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] + T.if_then_else(v0 < batch_size, reshape291[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * gpt_neox_layers_0_attention_dense_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused] + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(2): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = T.float16(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1] + for ax2_fused_2, ax1 in T.grid(2, 2): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_local[v0, 0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[v0, 0, v1] = T.float16(0.0) + NT_matmul_intermediate_pad_local[v0, 0, v1] = NT_matmul_intermediate_pad_local[v0, 0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + for ax0 in range(2): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(batch_size, ax0_0 * 2 + ax0) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (batch_size + 1) // 2 < 0 or ax0_0 == 0) and ax0_0 * 2 + ax0 < batch_size) + T.reads(astype148[v0, 0, v1], NT_matmul_intermediate_pad_local[v0, 0, v1], gpt_neox_layers_0_attention_dense_bias4[v1], input_embeds[v0, 0, v1]) + T.writes(T_add_intermediate_1_2[v0, 0, v1]) + T_add_intermediate_1_2[v0, 0, v1] = astype148[v0, 0, v1] + (NT_matmul_intermediate_pad_local[v0, 0, v1] + gpt_neox_layers_0_attention_dense_bias4[v1]) + input_embeds[v0, 0, v1] + else: + if T.tvm_thread_invariant(batch_size <= 8): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + 3) // 4 * 4, 1, 1024), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 3) // 4 * 4, 1, 1024), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 3) // 4 * 4, 1, 1024), "float16", scope="local") + for ax0_0 in T.thread_binding((batch_size + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(256, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(4, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax0_1_init) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = T.float16(0.0) + for ax2_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(4, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax0_1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1], reshape291[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_attention_dense_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] + T.if_then_else(v0 < batch_size, reshape291[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * gpt_neox_layers_0_attention_dense_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused] + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(4): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = T.float16(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1] + for ax2_fused_2, ax1 in T.grid(2, 4): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_local[v0, 0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[v0, 0, v1] = T.float16(0.0) + NT_matmul_intermediate_pad_local[v0, 0, v1] = NT_matmul_intermediate_pad_local[v0, 0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + for ax0 in range(4): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(batch_size, ax0_0 * 4 + ax0) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (batch_size + 3) // 4 < 0 or ax0_0 == 0) and ax0_0 * 4 + ax0 < batch_size) + T.reads(astype148[v0, 0, v1], NT_matmul_intermediate_pad_local[v0, 0, v1], gpt_neox_layers_0_attention_dense_bias4[v1], input_embeds[v0, 0, v1]) + T.writes(T_add_intermediate_1_2[v0, 0, v1]) + T_add_intermediate_1_2[v0, 0, v1] = astype148[v0, 0, v1] + (NT_matmul_intermediate_pad_local[v0, 0, v1] + gpt_neox_layers_0_attention_dense_bias4[v1]) + input_embeds[v0, 0, v1] + else: + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((1, (batch_size + 7) // 8 * 8, 1024), "float16", scope="local") + for ax0_ax1_0_fused in T.thread_binding((batch_size + 7) // 8, thread="blockIdx.y"): + for ax2_0 in T.thread_binding(4, thread="blockIdx.x"): + for ax1_1 in T.thread_binding(1, thread="vthread.y"): + for ax2_1 in T.thread_binding(1, thread="vthread.x"): + for ax1_2 in T.thread_binding(4, thread="threadIdx.y"): + for ax2_2 in T.thread_binding(32, thread="threadIdx.x", annotations={"pragma_auto_unroll_max_step": 16, "pragma_unroll_explicit": 1}): + for ax1_3_init, ax2_3_0_init in T.grid(2, 1): + for ax2_3_1_init in T.vectorized(8): + with T.block("NT_matmul_init"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3_init) + v2 = T.axis.spatial(1024, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0_init * 8 + ax2_3_1_init) + T.reads() + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = T.float16(0.0) + for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(64, 16, 2, 1): + for ax2_3_1 in T.vectorized(8): + with T.block("NT_matmul_update"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3) + v2 = T.axis.spatial(1024, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(1024, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], reshape291[v1, 0, v3], gpt_neox_layers_0_attention_dense_weight4[v2, v3]) + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = NT_matmul_intermediate_reindex_pad_local[0, v1, v2] + T.if_then_else(v1 < batch_size, reshape291[v1, 0, v3], T.float16(0.0)) * gpt_neox_layers_0_attention_dense_weight4[v2, v3] + for ax0, ax1, ax2_0_1 in T.grid(1, 2, 1): + for ax2_1_1 in T.vectorized(8): + with T.block("NT_matmul_intermediate_reindex_pad_local"): + v0 = T.axis.spatial(1, ax0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1) + v2 = T.axis.spatial(1024, ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1) + T.where(ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1 < batch_size) + T.reads(astype148[v1, 0, v2], NT_matmul_intermediate_reindex_pad_local[v0, v1, v2], gpt_neox_layers_0_attention_dense_bias4[v2], input_embeds[v1, 0, v2]) + T.writes(T_add_intermediate_1_2[v1, 0, v2]) + T_add_intermediate_1_2[v1, 0, v2] = astype148[v1, 0, v2] + (NT_matmul_intermediate_reindex_pad_local[v0, v1, v2] + gpt_neox_layers_0_attention_dense_bias4[v2]) + input_embeds[v1, 0, v2] + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(p_layer_norm148: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_weight4: T.Buffer((4096, 1024), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: T.Buffer((4096,), "float32"), p_output0: T.handle): + T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + layer_norm148 = T.match_buffer(p_layer_norm148, (batch_size, 1, 1024), "float16") + compute_intermediate = T.match_buffer(p_output0, (batch_size, 1, 4096), "float16") + # with T.block("root"): + if T.tvm_thread_invariant(batch_size <= 2): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + 1) // 2 * 2, 1, 4096), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 1) // 2 * 2, 1, 4096), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 1) // 2 * 2, 1, 4096), scope="local") + for ax0_0 in T.thread_binding((batch_size + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(1024, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(2, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax0_1_init) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = T.float32(0.0) + for ax2_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(2, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax0_1) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1], layer_norm148[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_mlp_dense_h_to_4h_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] + T.Cast("float32", T.if_then_else(v0 < batch_size, layer_norm148[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_h_to_4h_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(2): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = T.float32(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1] + for ax2_fused_2, ax1 in T.grid(2, 2): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax1) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_local[v0, 0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[v0, 0, v1] = T.float32(0.0) + NT_matmul_intermediate_pad_local[v0, 0, v1] = NT_matmul_intermediate_pad_local[v0, 0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + for ax0 in range(2): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(batch_size, ax0_0 * 2 + ax0) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (batch_size + 1) // 2 < 0 or ax0_0 == 0) and ax0_0 * 2 + ax0 < batch_size) + T.reads(NT_matmul_intermediate_pad_local[v0, 0, v1], gpt_neox_layers_0_mlp_dense_h_to_4h_bias4[v1]) + T.writes(compute_intermediate[v0, 0, v1]) + compute_intermediate[v0, 0, v1] = T.Cast("float16", (NT_matmul_intermediate_pad_local[v0, 0, v1] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias4[v1]) * (T.float32(0.5) + T.erf((NT_matmul_intermediate_pad_local[v0, 0, v1] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias4[v1]) * T.float32(0.70710678118654757)) * T.float32(0.5))) + else: + if T.tvm_thread_invariant(batch_size <= 8): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + 3) // 4 * 4, 1, 4096), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 3) // 4 * 4, 1, 4096), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 3) // 4 * 4, 1, 4096), scope="local") + for ax0_0 in T.thread_binding((batch_size + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(1024, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(4, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax0_1_init) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = T.float32(0.0) + for ax2_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(4, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax0_1) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1], layer_norm148[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_mlp_dense_h_to_4h_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] + T.Cast("float32", T.if_then_else(v0 < batch_size, layer_norm148[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_h_to_4h_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(4): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = T.float32(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1] + for ax2_fused_2, ax1 in T.grid(2, 4): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax1) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_local[v0, 0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[v0, 0, v1] = T.float32(0.0) + NT_matmul_intermediate_pad_local[v0, 0, v1] = NT_matmul_intermediate_pad_local[v0, 0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + for ax0 in range(4): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(batch_size, ax0_0 * 4 + ax0) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (batch_size + 3) // 4 < 0 or ax0_0 == 0) and ax0_0 * 4 + ax0 < batch_size) + T.reads(NT_matmul_intermediate_pad_local[v0, 0, v1], gpt_neox_layers_0_mlp_dense_h_to_4h_bias4[v1]) + T.writes(compute_intermediate[v0, 0, v1]) + compute_intermediate[v0, 0, v1] = T.Cast("float16", (NT_matmul_intermediate_pad_local[v0, 0, v1] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias4[v1]) * (T.float32(0.5) + T.erf((NT_matmul_intermediate_pad_local[v0, 0, v1] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias4[v1]) * T.float32(0.70710678118654757)) * T.float32(0.5))) + else: + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((1, (batch_size + 7) // 8 * 8, 4096), scope="local") + for ax0_ax1_0_fused in T.thread_binding((batch_size + 7) // 8, thread="blockIdx.y"): + for ax2_0 in T.thread_binding(16, thread="blockIdx.x"): + for ax1_1 in T.thread_binding(1, thread="vthread.y"): + for ax2_1 in T.thread_binding(1, thread="vthread.x"): + for ax1_2 in T.thread_binding(4, thread="threadIdx.y"): + for ax2_2 in T.thread_binding(32, thread="threadIdx.x", annotations={"pragma_auto_unroll_max_step": 16, "pragma_unroll_explicit": 1}): + for ax1_3_init, ax2_3_0_init in T.grid(2, 1): + for ax2_3_1_init in T.vectorized(8): + with T.block("NT_matmul_init"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3_init) + v2 = T.axis.spatial(4096, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0_init * 8 + ax2_3_1_init) + T.reads() + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = T.float32(0.0) + for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(64, 16, 2, 1): + for ax2_3_1 in T.vectorized(8): + with T.block("NT_matmul_update"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3) + v2 = T.axis.spatial(4096, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(1024, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm148[v1, 0, v3], gpt_neox_layers_0_mlp_dense_h_to_4h_weight4[v2, v3]) + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = NT_matmul_intermediate_reindex_pad_local[0, v1, v2] + T.Cast("float32", T.if_then_else(v1 < batch_size, layer_norm148[v1, 0, v3], T.float16(0.0))) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_h_to_4h_weight4[v2, v3]) + for ax0, ax1, ax2_0_1 in T.grid(1, 2, 1): + for ax2_1_1 in T.vectorized(8): + with T.block("NT_matmul_intermediate_reindex_pad_local"): + v0 = T.axis.spatial(1, ax0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1) + v2 = T.axis.spatial(4096, ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1) + T.where(ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1 < batch_size) + T.reads(NT_matmul_intermediate_reindex_pad_local[v0, v1, v2], gpt_neox_layers_0_mlp_dense_h_to_4h_bias4[v2]) + T.writes(compute_intermediate[v1, 0, v2]) + compute_intermediate[v1, 0, v2] = T.Cast("float16", (NT_matmul_intermediate_reindex_pad_local[v0, v1, v2] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias4[v2]) * (T.float32(0.5) + T.erf((NT_matmul_intermediate_reindex_pad_local[v0, v1, v2] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias4[v2]) * T.float32(0.70710678118654757)) * T.float32(0.5))) + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(p_astype147: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_weight4: T.Buffer((1024, 4096), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: T.Buffer((1024,), "float32"), p_output0: T.handle): + T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + astype147 = T.match_buffer(p_astype147, (batch_size, 1, 4096), "float16") + compute_intermediate = T.match_buffer(p_output0, (batch_size, 1, 1024), "float16") + # with T.block("root"): + if T.tvm_thread_invariant(batch_size <= 2): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + 1) // 2 * 2, 1, 1024), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 1) // 2 * 2, 1, 1024), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 1) // 2 * 2, 1, 1024), scope="local") + for ax0_0 in T.thread_binding((batch_size + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(256, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(2, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax0_1_init) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = T.float32(0.0) + for ax2_fused_0 in T.serial(16, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(2, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax0_1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1], astype147[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_mlp_dense_4h_to_h_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] + T.Cast("float32", T.if_then_else(v0 < batch_size, astype147[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_4h_to_h_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(2): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = T.float32(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1] + for ax2_fused_2, ax1 in T.grid(2, 2): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_local[v0, 0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[v0, 0, v1] = T.float32(0.0) + NT_matmul_intermediate_pad_local[v0, 0, v1] = NT_matmul_intermediate_pad_local[v0, 0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + for ax0 in range(2): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(batch_size, ax0_0 * 2 + ax0) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (batch_size + 1) // 2 < 0 or ax0_0 == 0) and ax0_0 * 2 + ax0 < batch_size) + T.reads(NT_matmul_intermediate_pad_local[v0, 0, v1], gpt_neox_layers_0_mlp_dense_4h_to_h_bias4[v1]) + T.writes(compute_intermediate[v0, 0, v1]) + compute_intermediate[v0, 0, v1] = T.Cast("float16", NT_matmul_intermediate_pad_local[v0, 0, v1] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias4[v1]) + else: + if T.tvm_thread_invariant(batch_size <= 8): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + 3) // 4 * 4, 1, 1024), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 3) // 4 * 4, 1, 1024), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 3) // 4 * 4, 1, 1024), scope="local") + for ax0_0 in T.thread_binding((batch_size + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(256, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(4, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax0_1_init) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = T.float32(0.0) + for ax2_fused_0 in T.serial(16, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(4, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax0_1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1], astype147[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_mlp_dense_4h_to_h_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] + T.Cast("float32", T.if_then_else(v0 < batch_size, astype147[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_4h_to_h_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(4): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = T.float32(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1] + for ax2_fused_2, ax1 in T.grid(2, 4): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_local[v0, 0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[v0, 0, v1] = T.float32(0.0) + NT_matmul_intermediate_pad_local[v0, 0, v1] = NT_matmul_intermediate_pad_local[v0, 0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + for ax0 in range(4): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(batch_size, ax0_0 * 4 + ax0) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (batch_size + 3) // 4 < 0 or ax0_0 == 0) and ax0_0 * 4 + ax0 < batch_size) + T.reads(NT_matmul_intermediate_pad_local[v0, 0, v1], gpt_neox_layers_0_mlp_dense_4h_to_h_bias4[v1]) + T.writes(compute_intermediate[v0, 0, v1]) + compute_intermediate[v0, 0, v1] = T.Cast("float16", NT_matmul_intermediate_pad_local[v0, 0, v1] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias4[v1]) + else: + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((1, (batch_size + 7) // 8 * 8, 1024), scope="local") + for ax0_ax1_0_fused in T.thread_binding((batch_size + 7) // 8, thread="blockIdx.y"): + for ax2_0 in T.thread_binding(4, thread="blockIdx.x"): + for ax1_1 in T.thread_binding(1, thread="vthread.y"): + for ax2_1 in T.thread_binding(1, thread="vthread.x"): + for ax1_2 in T.thread_binding(4, thread="threadIdx.y"): + for ax2_2 in T.thread_binding(32, thread="threadIdx.x", annotations={"pragma_auto_unroll_max_step": 16, "pragma_unroll_explicit": 1}): + for ax1_3_init, ax2_3_0_init in T.grid(2, 1): + for ax2_3_1_init in T.vectorized(8): + with T.block("NT_matmul_init"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3_init) + v2 = T.axis.spatial(1024, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0_init * 8 + ax2_3_1_init) + T.reads() + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = T.float32(0.0) + for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(256, 16, 2, 1): + for ax2_3_1 in T.vectorized(8): + with T.block("NT_matmul_update"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3) + v2 = T.axis.spatial(1024, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(4096, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], astype147[v1, 0, v3], gpt_neox_layers_0_mlp_dense_4h_to_h_weight4[v2, v3]) + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = NT_matmul_intermediate_reindex_pad_local[0, v1, v2] + T.Cast("float32", T.if_then_else(v1 < batch_size, astype147[v1, 0, v3], T.float16(0.0))) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_4h_to_h_weight4[v2, v3]) + for ax0, ax1, ax2_0_1 in T.grid(1, 2, 1): + for ax2_1_1 in T.vectorized(8): + with T.block("NT_matmul_intermediate_reindex_pad_local"): + v0 = T.axis.spatial(1, ax0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1) + v2 = T.axis.spatial(1024, ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1) + T.where(ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1 < batch_size) + T.reads(NT_matmul_intermediate_reindex_pad_local[v0, v1, v2], gpt_neox_layers_0_mlp_dense_4h_to_h_bias4[v2]) + T.writes(compute_intermediate[v1, 0, v2]) + compute_intermediate[v1, 0, v2] = T.Cast("float16", NT_matmul_intermediate_reindex_pad_local[v0, v1, v2] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias4[v2]) + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul4_cast2(p_layer_norm195: T.handle, p_embed_out_weight4: T.handle, p_output0: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + layer_norm195 = T.match_buffer(p_layer_norm195, (batch_size, 1, 1024), "float16") + vocab_size = T.int32() + embed_out_weight4 = T.match_buffer(p_embed_out_weight4, (vocab_size, 1024), "float16") + compute_intermediate = T.match_buffer(p_output0, (batch_size, 1, vocab_size)) + # with T.block("root"): + NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((1, (batch_size + 7) // 8 * 8, (vocab_size + 255) // 256 * 256), "float16", scope="local") + for ax0_ax1_0_fused in T.thread_binding((batch_size + 7) // 8, thread="blockIdx.y"): + for ax2_0 in T.thread_binding((vocab_size + 255) // 256, thread="blockIdx.x"): + for ax1_1 in T.thread_binding(1, thread="vthread.y"): + for ax2_1 in T.thread_binding(1, thread="vthread.x"): + for ax1_2 in T.thread_binding(4, thread="threadIdx.y"): + for ax2_2 in T.thread_binding(32, thread="threadIdx.x", annotations={"pragma_auto_unroll_max_step": 16, "pragma_unroll_explicit": 1}): + for ax1_3_init, ax2_3_0_init in T.grid(2, 1): + for ax2_3_1_init in T.vectorized(8): + with T.block("NT_matmul_init"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3_init) + v2 = T.axis.spatial((vocab_size + 255) // 256 * 256, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0_init * 8 + ax2_3_1_init) + T.reads() + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = T.float16(0.0) + for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(64, 16, 2, 1): + for ax2_3_1 in T.vectorized(8): + with T.block("NT_matmul_update"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3) + v2 = T.axis.spatial((vocab_size + 255) // 256 * 256, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(1024, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm195[v1, 0, v3], embed_out_weight4[v2, v3]) + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = NT_matmul_intermediate_reindex_pad_local[0, v1, v2] + T.if_then_else(v1 < batch_size, layer_norm195[v1, 0, v3], T.float16(0.0)) * T.if_then_else(v2 < vocab_size, embed_out_weight4[v2, v3], T.float16(0.0)) + for ax0, ax1, ax2_0_1 in T.grid(1, 2, 1): + for ax2_1_1 in T.vectorized(8): + with T.block("NT_matmul_intermediate_reindex_pad_local"): + v0 = T.axis.spatial(1, ax0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1) + v2 = T.axis.spatial((vocab_size + 255) // 256 * 256, ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1) + T.where(ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1 < batch_size and ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1 < vocab_size) + T.reads(NT_matmul_intermediate_reindex_pad_local[v0, v1, v2]) + T.writes(compute_intermediate[v1, 0, v2]) + compute_intermediate[v1, 0, v2] = T.Cast("float32", NT_matmul_intermediate_reindex_pad_local[v0, v1, v2]) + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul5_add5(p_layer_norm98: T.handle, gpt_neox_layers_0_attention_query_key_value_weight3: T.Buffer((3072, 1024), "float16"), gpt_neox_layers_0_attention_query_key_value_bias3: T.Buffer((3072,), "float16"), p_output0: T.handle): + T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + layer_norm98 = T.match_buffer(p_layer_norm98, (1, seq_len, 1024), "float16") + T_add_intermediate = T.match_buffer(p_output0, (1, seq_len, 3072), "float16") + # with T.block("root"): + if T.tvm_thread_invariant(seq_len <= 2): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer((1, (seq_len + 1) // 2 * 2, 3072), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 1) // 2 * 2, 3072), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 1) // 2 * 2, 3072), "float16", scope="local") + for ax0_0 in T.thread_binding((seq_len + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(768, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(2, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax0_1_init) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = T.float16(0.0) + for ax2_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(2, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax0_1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1], layer_norm98[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_attention_query_key_value_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] + T.if_then_else(v0 < seq_len, layer_norm98[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * gpt_neox_layers_0_attention_query_key_value_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused] + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(2): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = T.float16(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1] + for ax2_fused_2, ax1 in T.grid(2, 2): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_local[0, v0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[0, v0, v1] = T.float16(0.0) + NT_matmul_intermediate_pad_local[0, v0, v1] = NT_matmul_intermediate_pad_local[0, v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + for ax0 in range(2): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(seq_len, ax0_0 * 2 + ax0) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (seq_len + 1) // 2 < 0 or ax0_0 == 0) and ax0_0 * 2 + ax0 < seq_len) + T.reads(NT_matmul_intermediate_pad_local[0, v0, v1], gpt_neox_layers_0_attention_query_key_value_bias3[v1]) + T.writes(T_add_intermediate[0, v0, v1]) + T_add_intermediate[0, v0, v1] = NT_matmul_intermediate_pad_local[0, v0, v1] + gpt_neox_layers_0_attention_query_key_value_bias3[v1] + else: + if T.tvm_thread_invariant(seq_len <= 8): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer((1, (seq_len + 3) // 4 * 4, 3072), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 3) // 4 * 4, 3072), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 3) // 4 * 4, 3072), "float16", scope="local") + for ax0_0 in T.thread_binding((seq_len + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(768, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(4, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax0_1_init) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = T.float16(0.0) + for ax2_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(4, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax0_1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1], layer_norm98[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_attention_query_key_value_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] + T.if_then_else(v0 < seq_len, layer_norm98[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * gpt_neox_layers_0_attention_query_key_value_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused] + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(4): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = T.float16(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1] + for ax2_fused_2, ax1 in T.grid(2, 4): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_local[0, v0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[0, v0, v1] = T.float16(0.0) + NT_matmul_intermediate_pad_local[0, v0, v1] = NT_matmul_intermediate_pad_local[0, v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + for ax0 in range(4): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(seq_len, ax0_0 * 4 + ax0) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (seq_len + 3) // 4 < 0 or ax0_0 == 0) and ax0_0 * 4 + ax0 < seq_len) + T.reads(NT_matmul_intermediate_pad_local[0, v0, v1], gpt_neox_layers_0_attention_query_key_value_bias3[v1]) + T.writes(T_add_intermediate[0, v0, v1]) + T_add_intermediate[0, v0, v1] = NT_matmul_intermediate_pad_local[0, v0, v1] + gpt_neox_layers_0_attention_query_key_value_bias3[v1] + else: + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((1, (seq_len + 7) // 8 * 8, 3072), "float16", scope="local") + for ax0_ax1_0_fused in T.thread_binding((seq_len + 7) // 8, thread="blockIdx.y"): + for ax2_0 in T.thread_binding(12, thread="blockIdx.x"): + for ax1_1 in T.thread_binding(1, thread="vthread.y"): + for ax2_1 in T.thread_binding(1, thread="vthread.x"): + for ax1_2 in T.thread_binding(4, thread="threadIdx.y"): + for ax2_2 in T.thread_binding(32, thread="threadIdx.x", annotations={"pragma_auto_unroll_max_step": 16, "pragma_unroll_explicit": 1}): + for ax1_3_init, ax2_3_0_init in T.grid(2, 1): + for ax2_3_1_init in T.vectorized(8): + with T.block("NT_matmul_init"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3_init) + v2 = T.axis.spatial(3072, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0_init * 8 + ax2_3_1_init) + T.reads() + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = T.float16(0.0) + for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(64, 16, 2, 1): + for ax2_3_1 in T.vectorized(8): + with T.block("NT_matmul_update"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3) + v2 = T.axis.spatial(3072, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(1024, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm98[0, v1, v3], gpt_neox_layers_0_attention_query_key_value_weight3[v2, v3]) + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = NT_matmul_intermediate_reindex_pad_local[0, v1, v2] + T.if_then_else(v1 < seq_len, layer_norm98[0, v1, v3], T.float16(0.0)) * gpt_neox_layers_0_attention_query_key_value_weight3[v2, v3] + for ax0, ax1, ax2_0_1 in T.grid(1, 2, 1): + for ax2_1_1 in T.vectorized(8): + with T.block("NT_matmul_intermediate_reindex_pad_local"): + v0 = T.axis.spatial(1, ax0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1) + v2 = T.axis.spatial(3072, ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1) + T.where(ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1 < seq_len) + T.reads(NT_matmul_intermediate_reindex_pad_local[v0, v1, v2], gpt_neox_layers_0_attention_query_key_value_bias3[v2]) + T.writes(T_add_intermediate[0, v1, v2]) + T_add_intermediate[0, v1, v2] = NT_matmul_intermediate_reindex_pad_local[v0, v1, v2] + gpt_neox_layers_0_attention_query_key_value_bias3[v2] + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(p_reshape195: T.handle, gpt_neox_layers_0_attention_dense_weight3: T.Buffer((1024, 1024), "float16"), gpt_neox_layers_0_attention_dense_bias3: T.Buffer((1024,), "float16"), p_astype99: T.handle, p_input_embeds: T.handle, p_output0: T.handle): + T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + reshape195 = T.match_buffer(p_reshape195, (1, seq_len, 1024), "float16") + astype99 = T.match_buffer(p_astype99, (1, seq_len, 1024), "float16") + input_embeds = T.match_buffer(p_input_embeds, (1, seq_len, 1024), "float16") + T_add_intermediate_1_2 = T.match_buffer(p_output0, (1, seq_len, 1024), "float16") + # with T.block("root"): + if T.tvm_thread_invariant(seq_len <= 2): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer((1, (seq_len + 1) // 2 * 2, 1024), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 1) // 2 * 2, 1024), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 1) // 2 * 2, 1024), "float16", scope="local") + for ax0_0 in T.thread_binding((seq_len + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(256, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(2, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax0_1_init) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = T.float16(0.0) + for ax2_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(2, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax0_1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1], reshape195[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_attention_dense_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] + T.if_then_else(v0 < seq_len, reshape195[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * gpt_neox_layers_0_attention_dense_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused] + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(2): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = T.float16(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1] + for ax2_fused_2, ax1 in T.grid(2, 2): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_local[0, v0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[0, v0, v1] = T.float16(0.0) + NT_matmul_intermediate_pad_local[0, v0, v1] = NT_matmul_intermediate_pad_local[0, v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + for ax0 in range(2): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(seq_len, ax0_0 * 2 + ax0) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (seq_len + 1) // 2 < 0 or ax0_0 == 0) and ax0_0 * 2 + ax0 < seq_len) + T.reads(astype99[0, v0, v1], NT_matmul_intermediate_pad_local[0, v0, v1], gpt_neox_layers_0_attention_dense_bias3[v1], input_embeds[0, v0, v1]) + T.writes(T_add_intermediate_1_2[0, v0, v1]) + T_add_intermediate_1_2[0, v0, v1] = astype99[0, v0, v1] + (NT_matmul_intermediate_pad_local[0, v0, v1] + gpt_neox_layers_0_attention_dense_bias3[v1]) + input_embeds[0, v0, v1] + else: + if T.tvm_thread_invariant(seq_len <= 8): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer((1, (seq_len + 3) // 4 * 4, 1024), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 3) // 4 * 4, 1024), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 3) // 4 * 4, 1024), "float16", scope="local") + for ax0_0 in T.thread_binding((seq_len + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(256, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(4, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax0_1_init) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = T.float16(0.0) + for ax2_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(4, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax0_1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1], reshape195[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_attention_dense_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] + T.if_then_else(v0 < seq_len, reshape195[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * gpt_neox_layers_0_attention_dense_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused] + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(4): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = T.float16(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1] + for ax2_fused_2, ax1 in T.grid(2, 4): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_local[0, v0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[0, v0, v1] = T.float16(0.0) + NT_matmul_intermediate_pad_local[0, v0, v1] = NT_matmul_intermediate_pad_local[0, v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + for ax0 in range(4): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(seq_len, ax0_0 * 4 + ax0) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (seq_len + 3) // 4 < 0 or ax0_0 == 0) and ax0_0 * 4 + ax0 < seq_len) + T.reads(astype99[0, v0, v1], NT_matmul_intermediate_pad_local[0, v0, v1], gpt_neox_layers_0_attention_dense_bias3[v1], input_embeds[0, v0, v1]) + T.writes(T_add_intermediate_1_2[0, v0, v1]) + T_add_intermediate_1_2[0, v0, v1] = astype99[0, v0, v1] + (NT_matmul_intermediate_pad_local[0, v0, v1] + gpt_neox_layers_0_attention_dense_bias3[v1]) + input_embeds[0, v0, v1] + else: + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((1, (seq_len + 7) // 8 * 8, 1024), "float16", scope="local") + for ax0_ax1_0_fused in T.thread_binding((seq_len + 7) // 8, thread="blockIdx.y"): + for ax2_0 in T.thread_binding(4, thread="blockIdx.x"): + for ax1_1 in T.thread_binding(1, thread="vthread.y"): + for ax2_1 in T.thread_binding(1, thread="vthread.x"): + for ax1_2 in T.thread_binding(4, thread="threadIdx.y"): + for ax2_2 in T.thread_binding(32, thread="threadIdx.x", annotations={"pragma_auto_unroll_max_step": 16, "pragma_unroll_explicit": 1}): + for ax1_3_init, ax2_3_0_init in T.grid(2, 1): + for ax2_3_1_init in T.vectorized(8): + with T.block("NT_matmul_init"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3_init) + v2 = T.axis.spatial(1024, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0_init * 8 + ax2_3_1_init) + T.reads() + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = T.float16(0.0) + for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(64, 16, 2, 1): + for ax2_3_1 in T.vectorized(8): + with T.block("NT_matmul_update"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3) + v2 = T.axis.spatial(1024, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(1024, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], reshape195[0, v1, v3], gpt_neox_layers_0_attention_dense_weight3[v2, v3]) + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = NT_matmul_intermediate_reindex_pad_local[0, v1, v2] + T.if_then_else(v1 < seq_len, reshape195[0, v1, v3], T.float16(0.0)) * gpt_neox_layers_0_attention_dense_weight3[v2, v3] + for ax0, ax1, ax2_0_1 in T.grid(1, 2, 1): + for ax2_1_1 in T.vectorized(8): + with T.block("NT_matmul_intermediate_reindex_pad_local"): + v0 = T.axis.spatial(1, ax0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1) + v2 = T.axis.spatial(1024, ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1) + T.where(ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1 < seq_len) + T.reads(astype99[0, v1, v2], NT_matmul_intermediate_reindex_pad_local[v0, v1, v2], gpt_neox_layers_0_attention_dense_bias3[v2], input_embeds[0, v1, v2]) + T.writes(T_add_intermediate_1_2[0, v1, v2]) + T_add_intermediate_1_2[0, v1, v2] = astype99[0, v1, v2] + (NT_matmul_intermediate_reindex_pad_local[v0, v1, v2] + gpt_neox_layers_0_attention_dense_bias3[v2]) + input_embeds[0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(p_layer_norm99: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_weight3: T.Buffer((4096, 1024), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: T.Buffer((4096,), "float32"), p_output0: T.handle): + T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + layer_norm99 = T.match_buffer(p_layer_norm99, (1, seq_len, 1024), "float16") + compute_intermediate = T.match_buffer(p_output0, (1, seq_len, 4096), "float16") + # with T.block("root"): + if T.tvm_thread_invariant(seq_len <= 2): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer((1, (seq_len + 1) // 2 * 2, 4096), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 1) // 2 * 2, 4096), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 1) // 2 * 2, 4096), scope="local") + for ax0_0 in T.thread_binding((seq_len + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(1024, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(2, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax0_1_init) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = T.float32(0.0) + for ax2_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(2, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax0_1) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1], layer_norm99[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_mlp_dense_h_to_4h_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] + T.Cast("float32", T.if_then_else(v0 < seq_len, layer_norm99[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_h_to_4h_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(2): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = T.float32(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1] + for ax2_fused_2, ax1 in T.grid(2, 2): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax1) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_local[0, v0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[0, v0, v1] = T.float32(0.0) + NT_matmul_intermediate_pad_local[0, v0, v1] = NT_matmul_intermediate_pad_local[0, v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + for ax0 in range(2): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(seq_len, ax0_0 * 2 + ax0) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (seq_len + 1) // 2 < 0 or ax0_0 == 0) and ax0_0 * 2 + ax0 < seq_len) + T.reads(NT_matmul_intermediate_pad_local[0, v0, v1], gpt_neox_layers_0_mlp_dense_h_to_4h_bias3[v1]) + T.writes(compute_intermediate[0, v0, v1]) + compute_intermediate[0, v0, v1] = T.Cast("float16", (NT_matmul_intermediate_pad_local[0, v0, v1] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias3[v1]) * (T.float32(0.5) + T.erf((NT_matmul_intermediate_pad_local[0, v0, v1] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias3[v1]) * T.float32(0.70710678118654757)) * T.float32(0.5))) + else: + if T.tvm_thread_invariant(seq_len <= 8): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer((1, (seq_len + 3) // 4 * 4, 4096), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 3) // 4 * 4, 4096), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 3) // 4 * 4, 4096), scope="local") + for ax0_0 in T.thread_binding((seq_len + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(1024, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(4, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax0_1_init) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = T.float32(0.0) + for ax2_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(4, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax0_1) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1], layer_norm99[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_mlp_dense_h_to_4h_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] + T.Cast("float32", T.if_then_else(v0 < seq_len, layer_norm99[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_h_to_4h_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(4): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = T.float32(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1] + for ax2_fused_2, ax1 in T.grid(2, 4): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax1) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_local[0, v0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[0, v0, v1] = T.float32(0.0) + NT_matmul_intermediate_pad_local[0, v0, v1] = NT_matmul_intermediate_pad_local[0, v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + for ax0 in range(4): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(seq_len, ax0_0 * 4 + ax0) + v1 = T.axis.spatial(4096, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (seq_len + 3) // 4 < 0 or ax0_0 == 0) and ax0_0 * 4 + ax0 < seq_len) + T.reads(NT_matmul_intermediate_pad_local[0, v0, v1], gpt_neox_layers_0_mlp_dense_h_to_4h_bias3[v1]) + T.writes(compute_intermediate[0, v0, v1]) + compute_intermediate[0, v0, v1] = T.Cast("float16", (NT_matmul_intermediate_pad_local[0, v0, v1] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias3[v1]) * (T.float32(0.5) + T.erf((NT_matmul_intermediate_pad_local[0, v0, v1] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias3[v1]) * T.float32(0.70710678118654757)) * T.float32(0.5))) + else: + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((1, (seq_len + 7) // 8 * 8, 4096), scope="local") + for ax0_ax1_0_fused in T.thread_binding((seq_len + 7) // 8, thread="blockIdx.y"): + for ax2_0 in T.thread_binding(16, thread="blockIdx.x"): + for ax1_1 in T.thread_binding(1, thread="vthread.y"): + for ax2_1 in T.thread_binding(1, thread="vthread.x"): + for ax1_2 in T.thread_binding(4, thread="threadIdx.y"): + for ax2_2 in T.thread_binding(32, thread="threadIdx.x", annotations={"pragma_auto_unroll_max_step": 16, "pragma_unroll_explicit": 1}): + for ax1_3_init, ax2_3_0_init in T.grid(2, 1): + for ax2_3_1_init in T.vectorized(8): + with T.block("NT_matmul_init"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3_init) + v2 = T.axis.spatial(4096, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0_init * 8 + ax2_3_1_init) + T.reads() + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = T.float32(0.0) + for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(64, 16, 2, 1): + for ax2_3_1 in T.vectorized(8): + with T.block("NT_matmul_update"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3) + v2 = T.axis.spatial(4096, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(1024, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm99[0, v1, v3], gpt_neox_layers_0_mlp_dense_h_to_4h_weight3[v2, v3]) + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = NT_matmul_intermediate_reindex_pad_local[0, v1, v2] + T.Cast("float32", T.if_then_else(v1 < seq_len, layer_norm99[0, v1, v3], T.float16(0.0))) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_h_to_4h_weight3[v2, v3]) + for ax0, ax1, ax2_0_1 in T.grid(1, 2, 1): + for ax2_1_1 in T.vectorized(8): + with T.block("NT_matmul_intermediate_reindex_pad_local"): + v0 = T.axis.spatial(1, ax0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1) + v2 = T.axis.spatial(4096, ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1) + T.where(ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1 < seq_len) + T.reads(NT_matmul_intermediate_reindex_pad_local[v0, v1, v2], gpt_neox_layers_0_mlp_dense_h_to_4h_bias3[v2]) + T.writes(compute_intermediate[0, v1, v2]) + compute_intermediate[0, v1, v2] = T.Cast("float16", (NT_matmul_intermediate_reindex_pad_local[v0, v1, v2] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias3[v2]) * (T.float32(0.5) + T.erf((NT_matmul_intermediate_reindex_pad_local[v0, v1, v2] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias3[v2]) * T.float32(0.70710678118654757)) * T.float32(0.5))) + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(p_astype98: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_weight3: T.Buffer((1024, 4096), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: T.Buffer((1024,), "float32"), p_output0: T.handle): + T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + astype98 = T.match_buffer(p_astype98, (1, seq_len, 4096), "float16") + compute_intermediate = T.match_buffer(p_output0, (1, seq_len, 1024), "float16") + # with T.block("root"): + if T.tvm_thread_invariant(seq_len <= 2): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer((1, (seq_len + 1) // 2 * 2, 1024), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 1) // 2 * 2, 1024), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 1) // 2 * 2, 1024), scope="local") + for ax0_0 in T.thread_binding((seq_len + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(256, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(2, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax0_1_init) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = T.float32(0.0) + for ax2_fused_0 in T.serial(16, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(2, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax0_1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1], astype98[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_mlp_dense_4h_to_h_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] + T.Cast("float32", T.if_then_else(v0 < seq_len, astype98[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_4h_to_h_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(2): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = T.float32(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1] + for ax2_fused_2, ax1 in T.grid(2, 2): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_local[0, v0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[0, v0, v1] = T.float32(0.0) + NT_matmul_intermediate_pad_local[0, v0, v1] = NT_matmul_intermediate_pad_local[0, v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + for ax0 in range(2): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(seq_len, ax0_0 * 2 + ax0) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (seq_len + 1) // 2 < 0 or ax0_0 == 0) and ax0_0 * 2 + ax0 < seq_len) + T.reads(NT_matmul_intermediate_pad_local[0, v0, v1], gpt_neox_layers_0_mlp_dense_4h_to_h_bias3[v1]) + T.writes(compute_intermediate[0, v0, v1]) + compute_intermediate[0, v0, v1] = T.Cast("float16", NT_matmul_intermediate_pad_local[0, v0, v1] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias3[v1]) + else: + if T.tvm_thread_invariant(seq_len <= 8): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer((1, (seq_len + 3) // 4 * 4, 1024), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 3) // 4 * 4, 1024), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 3) // 4 * 4, 1024), scope="local") + for ax0_0 in T.thread_binding((seq_len + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(256, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(4, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax0_1_init) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = T.float32(0.0) + for ax2_fused_0 in T.serial(16, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(4, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax0_1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1], astype98[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_mlp_dense_4h_to_h_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] + T.Cast("float32", T.if_then_else(v0 < seq_len, astype98[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_4h_to_h_weight3[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(4): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = T.float32(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1] + for ax2_fused_2, ax1 in T.grid(2, 4): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax1) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_local[0, v0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[0, v0, v1] = T.float32(0.0) + NT_matmul_intermediate_pad_local[0, v0, v1] = NT_matmul_intermediate_pad_local[0, v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + for ax0 in range(4): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(seq_len, ax0_0 * 4 + ax0) + v1 = T.axis.spatial(1024, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (seq_len + 3) // 4 < 0 or ax0_0 == 0) and ax0_0 * 4 + ax0 < seq_len) + T.reads(NT_matmul_intermediate_pad_local[0, v0, v1], gpt_neox_layers_0_mlp_dense_4h_to_h_bias3[v1]) + T.writes(compute_intermediate[0, v0, v1]) + compute_intermediate[0, v0, v1] = T.Cast("float16", NT_matmul_intermediate_pad_local[0, v0, v1] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias3[v1]) + else: + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((1, (seq_len + 7) // 8 * 8, 1024), scope="local") + for ax0_ax1_0_fused in T.thread_binding((seq_len + 7) // 8, thread="blockIdx.y"): + for ax2_0 in T.thread_binding(4, thread="blockIdx.x"): + for ax1_1 in T.thread_binding(1, thread="vthread.y"): + for ax2_1 in T.thread_binding(1, thread="vthread.x"): + for ax1_2 in T.thread_binding(4, thread="threadIdx.y"): + for ax2_2 in T.thread_binding(32, thread="threadIdx.x", annotations={"pragma_auto_unroll_max_step": 16, "pragma_unroll_explicit": 1}): + for ax1_3_init, ax2_3_0_init in T.grid(2, 1): + for ax2_3_1_init in T.vectorized(8): + with T.block("NT_matmul_init"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3_init) + v2 = T.axis.spatial(1024, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0_init * 8 + ax2_3_1_init) + T.reads() + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = T.float32(0.0) + for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(256, 16, 2, 1): + for ax2_3_1 in T.vectorized(8): + with T.block("NT_matmul_update"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3) + v2 = T.axis.spatial(1024, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(4096, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], astype98[0, v1, v3], gpt_neox_layers_0_mlp_dense_4h_to_h_weight3[v2, v3]) + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = NT_matmul_intermediate_reindex_pad_local[0, v1, v2] + T.Cast("float32", T.if_then_else(v1 < seq_len, astype98[0, v1, v3], T.float16(0.0))) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_4h_to_h_weight3[v2, v3]) + for ax0, ax1, ax2_0_1 in T.grid(1, 2, 1): + for ax2_1_1 in T.vectorized(8): + with T.block("NT_matmul_intermediate_reindex_pad_local"): + v0 = T.axis.spatial(1, ax0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1) + v2 = T.axis.spatial(1024, ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1) + T.where(ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1 < seq_len) + T.reads(NT_matmul_intermediate_reindex_pad_local[v0, v1, v2], gpt_neox_layers_0_mlp_dense_4h_to_h_bias3[v2]) + T.writes(compute_intermediate[0, v1, v2]) + compute_intermediate[0, v1, v2] = T.Cast("float16", NT_matmul_intermediate_reindex_pad_local[v0, v1, v2] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias3[v2]) + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul9_cast5(p_take1: T.handle, p_embed_out_weight3: T.handle, p_output0: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + take1 = T.match_buffer(p_take1, (1, batch_size, 1024), "float16") + vocab_size = T.int32() + embed_out_weight3 = T.match_buffer(p_embed_out_weight3, (vocab_size, 1024), "float16") + compute_intermediate = T.match_buffer(p_output0, (1, batch_size, vocab_size)) + # with T.block("root"): + NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((1, (batch_size + 7) // 8 * 8, (vocab_size + 255) // 256 * 256), "float16", scope="local") + for ax0_ax1_0_fused in T.thread_binding((batch_size + 7) // 8, thread="blockIdx.y"): + for ax2_0 in T.thread_binding((vocab_size + 255) // 256, thread="blockIdx.x"): + for ax1_1 in T.thread_binding(1, thread="vthread.y"): + for ax2_1 in T.thread_binding(1, thread="vthread.x"): + for ax1_2 in T.thread_binding(4, thread="threadIdx.y"): + for ax2_2 in T.thread_binding(32, thread="threadIdx.x", annotations={"pragma_auto_unroll_max_step": 16, "pragma_unroll_explicit": 1}): + for ax1_3_init, ax2_3_0_init in T.grid(2, 1): + for ax2_3_1_init in T.vectorized(8): + with T.block("NT_matmul_init"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3_init) + v2 = T.axis.spatial((vocab_size + 255) // 256 * 256, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0_init * 8 + ax2_3_1_init) + T.reads() + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = T.float16(0.0) + for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(64, 16, 2, 1): + for ax2_3_1 in T.vectorized(8): + with T.block("NT_matmul_update"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3) + v2 = T.axis.spatial((vocab_size + 255) // 256 * 256, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(1024, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], take1[0, v1, v3], embed_out_weight3[v2, v3]) + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = NT_matmul_intermediate_reindex_pad_local[0, v1, v2] + T.if_then_else(v1 < batch_size, take1[0, v1, v3], T.float16(0.0)) * T.if_then_else(v2 < vocab_size, embed_out_weight3[v2, v3], T.float16(0.0)) + for ax0, ax1, ax2_0_1 in T.grid(1, 2, 1): + for ax2_1_1 in T.vectorized(8): + with T.block("NT_matmul_intermediate_reindex_pad_local"): + v0 = T.axis.spatial(1, ax0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1) + v2 = T.axis.spatial((vocab_size + 255) // 256 * 256, ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1) + T.where(ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1 < batch_size and ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1 < vocab_size) + T.reads(NT_matmul_intermediate_reindex_pad_local[v0, v1, v2]) + T.writes(compute_intermediate[0, v1, v2]) + compute_intermediate[0, v1, v2] = T.Cast("float32", NT_matmul_intermediate_reindex_pad_local[v0, v1, v2]) + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul_add(p_layer_norm147: T.handle, gpt_neox_layers_0_attention_query_key_value_weight4: T.Buffer((3072, 1024), "float16"), gpt_neox_layers_0_attention_query_key_value_bias4: T.Buffer((3072,), "float16"), p_output0: T.handle): + T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + layer_norm147 = T.match_buffer(p_layer_norm147, (batch_size, 1, 1024), "float16") + T_add_intermediate = T.match_buffer(p_output0, (batch_size, 1, 3072), "float16") + # with T.block("root"): + if T.tvm_thread_invariant(batch_size <= 2): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + 1) // 2 * 2, 1, 3072), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 1) // 2 * 2, 1, 3072), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 1) // 2 * 2, 1, 3072), "float16", scope="local") + for ax0_0 in T.thread_binding((batch_size + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(768, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(2, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax0_1_init) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = T.float16(0.0) + for ax2_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(2, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax0_1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1], layer_norm147[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_attention_query_key_value_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] + T.if_then_else(v0 < batch_size, layer_norm147[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * gpt_neox_layers_0_attention_query_key_value_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused] + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(2): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = T.float16(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax2) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1] + for ax2_fused_2, ax1 in T.grid(2, 2): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_local[v0, 0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[v0, 0, v1] = T.float16(0.0) + NT_matmul_intermediate_pad_local[v0, 0, v1] = NT_matmul_intermediate_pad_local[v0, 0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + for ax0 in range(2): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(batch_size, ax0_0 * 2 + ax0) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (batch_size + 1) // 2 < 0 or ax0_0 == 0) and ax0_0 * 2 + ax0 < batch_size) + T.reads(NT_matmul_intermediate_pad_local[v0, 0, v1], gpt_neox_layers_0_attention_query_key_value_bias4[v1]) + T.writes(T_add_intermediate[v0, 0, v1]) + T_add_intermediate[v0, 0, v1] = NT_matmul_intermediate_pad_local[v0, 0, v1] + gpt_neox_layers_0_attention_query_key_value_bias4[v1] + else: + if T.tvm_thread_invariant(batch_size <= 8): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + 3) // 4 * 4, 1, 3072), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 3) // 4 * 4, 1, 3072), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 3) // 4 * 4, 1, 3072), "float16", scope="local") + for ax0_0 in T.thread_binding((batch_size + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(768, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(4, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax0_1_init) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = T.float16(0.0) + for ax2_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(4, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax0_1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1], layer_norm147[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], gpt_neox_layers_0_attention_query_key_value_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused]) + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] + T.if_then_else(v0 < batch_size, layer_norm147[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * gpt_neox_layers_0_attention_query_key_value_weight4[v1, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused] + for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + for ax3_fused_2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax2 in range(4): + for ax3_fused_2_1 in T.vectorized(2): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(32, ax0) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = T.float16(0.0) + for ax1 in range(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused_0, vax2_fused_1_ax2_fused_3_fused_1 = T.axis.remap("SR", [ax0, ax1]) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax2) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1] + for ax2_fused_2, ax1 in T.grid(2, 4): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_local[v0, 0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[v0, 0, v1] = T.float16(0.0) + NT_matmul_intermediate_pad_local[v0, 0, v1] = NT_matmul_intermediate_pad_local[v0, 0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + for ax0 in range(4): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_2 in range(2): + with T.block("NT_matmul_intermediate_pad"): + v0 = T.axis.spatial(batch_size, ax0_0 * 4 + ax0) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (batch_size + 3) // 4 < 0 or ax0_0 == 0) and ax0_0 * 4 + ax0 < batch_size) + T.reads(NT_matmul_intermediate_pad_local[v0, 0, v1], gpt_neox_layers_0_attention_query_key_value_bias4[v1]) + T.writes(T_add_intermediate[v0, 0, v1]) + T_add_intermediate[v0, 0, v1] = NT_matmul_intermediate_pad_local[v0, 0, v1] + gpt_neox_layers_0_attention_query_key_value_bias4[v1] + else: + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((1, (batch_size + 7) // 8 * 8, 3072), "float16", scope="local") + for ax0_ax1_0_fused in T.thread_binding((batch_size + 7) // 8, thread="blockIdx.y"): + for ax2_0 in T.thread_binding(12, thread="blockIdx.x"): + for ax1_1 in T.thread_binding(1, thread="vthread.y"): + for ax2_1 in T.thread_binding(1, thread="vthread.x"): + for ax1_2 in T.thread_binding(4, thread="threadIdx.y"): + for ax2_2 in T.thread_binding(32, thread="threadIdx.x", annotations={"pragma_auto_unroll_max_step": 16, "pragma_unroll_explicit": 1}): + for ax1_3_init, ax2_3_0_init in T.grid(2, 1): + for ax2_3_1_init in T.vectorized(8): + with T.block("NT_matmul_init"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3_init) + v2 = T.axis.spatial(3072, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0_init * 8 + ax2_3_1_init) + T.reads() + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = T.float16(0.0) + for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(64, 16, 2, 1): + for ax2_3_1 in T.vectorized(8): + with T.block("NT_matmul_update"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3) + v2 = T.axis.spatial(3072, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(1024, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm147[v1, 0, v3], gpt_neox_layers_0_attention_query_key_value_weight4[v2, v3]) + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = NT_matmul_intermediate_reindex_pad_local[0, v1, v2] + T.if_then_else(v1 < batch_size, layer_norm147[v1, 0, v3], T.float16(0.0)) * gpt_neox_layers_0_attention_query_key_value_weight4[v2, v3] + for ax0, ax1, ax2_0_1 in T.grid(1, 2, 1): + for ax2_1_1 in T.vectorized(8): + with T.block("NT_matmul_intermediate_reindex_pad_local"): + v0 = T.axis.spatial(1, ax0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1) + v2 = T.axis.spatial(3072, ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1) + T.where(ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1 < batch_size) + T.reads(NT_matmul_intermediate_reindex_pad_local[v0, v1, v2], gpt_neox_layers_0_attention_query_key_value_bias4[v2]) + T.writes(T_add_intermediate[v1, 0, v2]) + T_add_intermediate[v1, 0, v2] = NT_matmul_intermediate_reindex_pad_local[v0, v1, v2] + gpt_neox_layers_0_attention_query_key_value_bias4[v2] + + @T.prim_func + def gpt_neox_q0f16_fused_reshape10_reshape11(lv25: T.Buffer((1, 16, 64), "float16"), T_reshape_intermediate_1: T.Buffer((1, 1, 1024), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + for ax0_fused_0 in T.thread_binding(4, thread="blockIdx.x"): + for ax0_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("T_reshape_1"): + v0 = T.axis.spatial(1024, ax0_fused_0 * 256 + ax0_fused_1) + T.reads(lv25[0, v0 // 64, v0 % 64]) + T.writes(T_reshape_intermediate_1[0, 0, v0]) + T_reshape_intermediate_1[0, 0, v0] = lv25[0, v0 // 64, v0 % 64] + + @T.prim_func + def gpt_neox_q0f16_fused_reshape8_reshape9(add144: T.Buffer((1, 1, 3072), "float16"), T_reshape_intermediate_1: T.Buffer((1, 48, 64), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(12, thread="blockIdx.x"): + for ax0_ax1_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("T_reshape_1"): + v0 = T.axis.spatial(48, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) // 64) + v1 = T.axis.spatial(64, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 64) + T.reads(add144[0, 0, v0 * 64 + v1]) + T.writes(T_reshape_intermediate_1[0, v0, v1]) + T_reshape_intermediate_1[0, v0, v1] = add144[0, 0, v0 * 64 + v1] + + @T.prim_func + def gpt_neox_q0f16_fused_rope(var_qkv: T.handle, var_position_map: T.handle, var_q: T.handle, var_k: T.handle, var_v: T.handle, apply_rope: T.int32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + qkv = T.match_buffer(var_qkv, (seq_len, 48, 64), "float16") + position_map = T.match_buffer(var_position_map, (seq_len,), "int32", offset_factor=1) + q = T.match_buffer(var_q, (seq_len, 16, 64), "float16") + k = T.match_buffer(var_k, (seq_len, 16, 64), "float16") + v = T.match_buffer(var_v, (seq_len, 16, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 12, thread="blockIdx.x"): + for ax0_ax1_ax2_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("llama_fused_rope"): + v0 = T.axis.spatial(seq_len, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) // 3072) + v1 = T.axis.spatial(48, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 3072 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(position_map[v0], qkv[v0, v1, v2 + -8:v2 + -8 + 17]) + T.writes(q[v0, v1, v2], k[v0, v1 + -16, v2], v[v0, v1 + -32, v2]) + if v1 < 16: + freq = T.float32() + q[v0, v1, v2] = T.if_then_else(apply_rope > 0 and v2 < 16, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", qkv[v0, v1, v2]) + T.sin(freq) * T.Cast("float32", T.if_then_else(v2 < 8, qkv[v0, v1, v2 + 8] * T.float16(-1.0), qkv[v0, v1, v2 + -8]))), where={freq: T.Cast("float32", position_map[v0]) / T.pow(T.float32(10000.0), T.Cast("float32", v2 * 2 % 16) / T.float32(16.0))}), qkv[v0, v1, v2]) + else: + if v1 < 32: + freq = T.float32() + k[v0, v1 + -16, v2] = T.if_then_else(apply_rope > 0 and v2 < 16, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", qkv[v0, v1, v2]) + T.sin(freq) * T.Cast("float32", T.if_then_else(v2 < 8, qkv[v0, v1, v2 + 8] * T.float16(-1.0), qkv[v0, v1, v2 + -8]))), where={freq: T.Cast("float32", position_map[v0]) / T.pow(T.float32(10000.0), T.Cast("float32", v2 * 2 % 16) / T.float32(16.0))}), qkv[v0, v1, v2]) + else: + v[v0, v1 + -32, v2] = qkv[v0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_gather_probs(var_src: T.handle, var_indices: T.handle, var_dst: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + m, n = T.int32(is_size_var=True), T.int32(is_size_var=True) + src = T.match_buffer(var_src, (m, n)) + batch_size = T.int32(is_size_var=True) + indices = T.match_buffer(var_indices, (batch_size,), "int32") + dst = T.match_buffer(var_dst, (batch_size, n)) + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding((batch_size * n + 255) // 256, thread="blockIdx.x"): + for ax0_ax1_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("gather_2d"): + v0 = T.axis.spatial(batch_size, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % (n * batch_size) // n) + v1 = T.axis.spatial(n, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % n) + T.where(ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1 < batch_size * n) + T.reads(src[indices[v0], v1], indices[v0]) + T.writes(dst[v0, v1]) + dst[v0, v1] = src[indices[v0], v1] + + @T.prim_func + def gpt_neox_q0f16_index(var_layer_norm48: T.handle, index: T.Buffer((1, 1, 1024), "float16")): + T.func_attr({"target": T.target({"keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + layer_norm48 = T.match_buffer(var_layer_norm48, (1, seq_len, 1024), "float16") + # with T.block("root"): + for ax0_fused_0 in T.thread_binding(4, thread="blockIdx.x"): + for ax0_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("index"): + v0 = T.axis.spatial(1024, ax0_fused_0 * 256 + ax0_fused_1) + T.reads(layer_norm48[0, seq_len - 1, v0]) + T.writes(index[0, 0, v0]) + index[0, 0, v0] = layer_norm48[0, seq_len - 1, v0] + + @T.prim_func + def gpt_neox_q0f16_layer_norm(var_input_embeds: T.handle, gpt_neox_layers_0_input_layernorm_weight4: T.Buffer((1024,), "float16"), gpt_neox_layers_0_input_layernorm_bias4: T.Buffer((1024,), "float16"), var_T_layer_norm: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + input_embeds = T.match_buffer(var_input_embeds, (batch_size, 1, 1024), "float16") + T_layer_norm = T.match_buffer(var_T_layer_norm, (batch_size, 1, 1024), "float16") + # with T.block("root"): + input_embeds_red_temp_v0_shared = T.alloc_buffer((batch_size, 1), scope="shared") + input_embeds_red_temp_v1_shared = T.alloc_buffer((batch_size, 1), scope="shared") + for ax0_fused in T.thread_binding(batch_size, thread="blockIdx.x"): + for ax0 in range(1): + for ax1_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + for ax1_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("input_embeds_red_temp"): + v0 = T.axis.spatial(batch_size, ax0_fused + ax0) + v1 = T.axis.reduce(1024, ax1_fused_0 * 256 + ax1_fused_1) + T.reads(input_embeds[v0, 0, v1]) + T.writes(input_embeds_red_temp_v0_shared[v0, 0], input_embeds_red_temp_v1_shared[v0, 0]) + with T.init(): + input_embeds_red_temp_v0_shared[v0, 0] = T.float32(0.0) + input_embeds_red_temp_v1_shared[v0, 0] = T.float32(0.0) + v_input_embeds_red_temp_v0: T.float32 = input_embeds_red_temp_v0_shared[v0, 0] + T.Cast("float32", input_embeds[v0, 0, v1]) + v_input_embeds_red_temp_v1: T.float32 = input_embeds_red_temp_v1_shared[v0, 0] + T.Cast("float32", input_embeds[v0, 0, v1]) * T.Cast("float32", input_embeds[v0, 0, v1]) + input_embeds_red_temp_v0_shared[v0, 0] = v_input_embeds_red_temp_v0 + input_embeds_red_temp_v1_shared[v0, 0] = v_input_embeds_red_temp_v1 + for ax1_1 in T.thread_binding(256, thread="threadIdx.x"): + for ax1_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("T_layer_norm"): + v0 = T.axis.spatial(batch_size, ax0_fused) + v1 = T.axis.spatial(1024, ax1_0 * 256 + ax1_1) + T.reads(input_embeds[v0, 0, v1], input_embeds_red_temp_v0_shared[v0, 0], input_embeds_red_temp_v1_shared[v0, 0], gpt_neox_layers_0_input_layernorm_weight4[v1], gpt_neox_layers_0_input_layernorm_bias4[v1]) + T.writes(T_layer_norm[v0, 0, v1]) + T_layer_norm[v0, 0, v1] = T.Cast("float16", (T.Cast("float32", input_embeds[v0, 0, v1]) - input_embeds_red_temp_v0_shared[v0, 0] * T.float32(0.0009765625)) * T.rsqrt(input_embeds_red_temp_v1_shared[v0, 0] * T.float32(0.0009765625) - input_embeds_red_temp_v0_shared[v0, 0] * T.float32(0.0009765625) * (input_embeds_red_temp_v0_shared[v0, 0] * T.float32(0.0009765625)) + T.float32(1.0000000000000001e-05))) * gpt_neox_layers_0_input_layernorm_weight4[v1] + gpt_neox_layers_0_input_layernorm_bias4[v1] + + @T.prim_func + def gpt_neox_q0f16_layer_norm1(var_input_embeds: T.handle, gpt_neox_layers_0_input_layernorm_weight3: T.Buffer((1024,), "float16"), gpt_neox_layers_0_input_layernorm_bias3: T.Buffer((1024,), "float16"), var_T_layer_norm: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + input_embeds = T.match_buffer(var_input_embeds, (1, seq_len, 1024), "float16") + T_layer_norm = T.match_buffer(var_T_layer_norm, (1, seq_len, 1024), "float16") + # with T.block("root"): + input_embeds_red_temp_v0_shared = T.alloc_buffer((1, seq_len), scope="shared") + input_embeds_red_temp_v1_shared = T.alloc_buffer((1, seq_len), scope="shared") + for ax0_fused in T.thread_binding(seq_len, thread="blockIdx.x"): + for ax0 in range(1): + for ax1_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + for ax1_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("input_embeds_red_temp"): + v0 = T.axis.spatial(seq_len, ax0_fused + ax0) + v1 = T.axis.reduce(1024, ax1_fused_0 * 256 + ax1_fused_1) + T.reads(input_embeds[0, v0, v1]) + T.writes(input_embeds_red_temp_v0_shared[0, v0], input_embeds_red_temp_v1_shared[0, v0]) + with T.init(): + input_embeds_red_temp_v0_shared[0, v0] = T.float32(0.0) + input_embeds_red_temp_v1_shared[0, v0] = T.float32(0.0) + v_input_embeds_red_temp_v0: T.float32 = input_embeds_red_temp_v0_shared[0, v0] + T.Cast("float32", input_embeds[0, v0, v1]) + v_input_embeds_red_temp_v1: T.float32 = input_embeds_red_temp_v1_shared[0, v0] + T.Cast("float32", input_embeds[0, v0, v1]) * T.Cast("float32", input_embeds[0, v0, v1]) + input_embeds_red_temp_v0_shared[0, v0] = v_input_embeds_red_temp_v0 + input_embeds_red_temp_v1_shared[0, v0] = v_input_embeds_red_temp_v1 + for ax1_1 in T.thread_binding(256, thread="threadIdx.x"): + for ax1_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("T_layer_norm"): + v0 = T.axis.spatial(seq_len, ax0_fused) + v1 = T.axis.spatial(1024, ax1_0 * 256 + ax1_1) + T.reads(input_embeds[0, v0, v1], input_embeds_red_temp_v0_shared[0, v0], input_embeds_red_temp_v1_shared[0, v0], gpt_neox_layers_0_input_layernorm_weight3[v1], gpt_neox_layers_0_input_layernorm_bias3[v1]) + T.writes(T_layer_norm[0, v0, v1]) + T_layer_norm[0, v0, v1] = T.Cast("float16", (T.Cast("float32", input_embeds[0, v0, v1]) - input_embeds_red_temp_v0_shared[0, v0] * T.float32(0.0009765625)) * T.rsqrt(input_embeds_red_temp_v1_shared[0, v0] * T.float32(0.0009765625) - input_embeds_red_temp_v0_shared[0, v0] * T.float32(0.0009765625) * (input_embeds_red_temp_v0_shared[0, v0] * T.float32(0.0009765625)) + T.float32(1.0000000000000001e-05))) * gpt_neox_layers_0_input_layernorm_weight3[v1] + gpt_neox_layers_0_input_layernorm_bias3[v1] + + @T.prim_func + def gpt_neox_q0f16_layer_norm2(input_embed: T.Buffer((1, 1, 1024), "float16"), gpt_neox_layers_0_input_layernorm_weight2: T.Buffer((1024,), "float16"), gpt_neox_layers_0_input_layernorm_bias2: T.Buffer((1024,), "float16"), T_layer_norm: T.Buffer((1, 1, 1024), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + input_embed_red_temp_v0_shared = T.alloc_buffer((1, 1), scope="shared") + input_embed_red_temp_v1_shared = T.alloc_buffer((1, 1), scope="shared") + for ax0_fused in T.thread_binding(1, thread="blockIdx.x"): + for ax0 in range(1): + for ax1_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + for ax1_fused_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("input_embed_red_temp"): + v0 = T.axis.spatial(1, ax0) + v1 = T.axis.reduce(1024, ax1_fused_0 * 256 + ax1_fused_1) + T.reads(input_embed[0, 0, v1]) + T.writes(input_embed_red_temp_v0_shared[0, 0], input_embed_red_temp_v1_shared[0, 0]) + with T.init(): + input_embed_red_temp_v0_shared[0, 0] = T.float32(0.0) + input_embed_red_temp_v1_shared[0, 0] = T.float32(0.0) + v_input_embed_red_temp_v0: T.float32 = input_embed_red_temp_v0_shared[0, 0] + T.Cast("float32", input_embed[0, 0, v1]) + v_input_embed_red_temp_v1: T.float32 = input_embed_red_temp_v1_shared[0, 0] + T.Cast("float32", input_embed[0, 0, v1]) * T.Cast("float32", input_embed[0, 0, v1]) + input_embed_red_temp_v0_shared[0, 0] = v_input_embed_red_temp_v0 + input_embed_red_temp_v1_shared[0, 0] = v_input_embed_red_temp_v1 + for ax1_1 in T.thread_binding(256, thread="threadIdx.x"): + for ax1_0 in T.serial(4, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("T_layer_norm"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial(1024, ax1_0 * 256 + ax1_1) + T.reads(input_embed[0, 0, v1], input_embed_red_temp_v0_shared[0, 0], input_embed_red_temp_v1_shared[0, 0], gpt_neox_layers_0_input_layernorm_weight2[v1], gpt_neox_layers_0_input_layernorm_bias2[v1]) + T.writes(T_layer_norm[0, 0, v1]) + T_layer_norm[0, 0, v1] = T.Cast("float16", (T.Cast("float32", input_embed[0, 0, v1]) - input_embed_red_temp_v0_shared[0, 0] * T.float32(0.0009765625)) * T.rsqrt(input_embed_red_temp_v1_shared[0, 0] * T.float32(0.0009765625) - input_embed_red_temp_v0_shared[0, 0] * T.float32(0.0009765625) * (input_embed_red_temp_v0_shared[0, 0] * T.float32(0.0009765625)) + T.float32(1.0000000000000001e-05))) * gpt_neox_layers_0_input_layernorm_weight2[v1] + gpt_neox_layers_0_input_layernorm_bias2[v1] + + @T.prim_func + def gpt_neox_q0f16_merge_state_inplace(v: T.handle, s: T.handle, v_other: T.handle, s_other: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + N, H, D = T.int32(is_size_var=True), T.int32(is_size_var=True), T.int32(is_size_var=True) + V = T.match_buffer(v, (N, H, D), "float16") + S = T.match_buffer(s, (N, H)) + V_other = T.match_buffer(v_other, (N, H, D), "float16") + S_other = T.match_buffer(s_other, (N, H)) + # with T.block("root"): + for bx in T.thread_binding(N, thread="blockIdx.x"): + for by in T.thread_binding(1, thread="blockIdx.y"): + for ty in T.thread_binding(16, thread="threadIdx.y"): + for tx in T.thread_binding(16, thread="threadIdx.x"): + with T.block("merge"): + T.reads(S[bx, ty + by * 16], S_other[bx, ty + by * 16], V[bx, ty + by * 16, tx * 4:tx * 4 + 4], V_other[bx, ty + by * 16, tx * 4:tx * 4 + 4]) + T.writes(V[bx, ty + by * 16, tx * 4:tx * 4 + 4], S[bx, ty + by * 16]) + s_val = T.alloc_buffer((1,), scope="local") + s_other_val = T.alloc_buffer((1,), scope="local") + s_max = T.alloc_buffer((1,), scope="local") + scale = T.alloc_buffer((1,), scope="local") + other_scale = T.alloc_buffer((1,), scope="local") + v_vec = T.alloc_buffer((4,), "float16", scope="local") + v_other_vec = T.alloc_buffer((4,), "float16", scope="local") + s_val[0] = S[bx, ty + by * 16] + s_other_val[0] = S_other[bx, ty + by * 16] + s_max[0] = T.max(s_val[0], s_other_val[0]) + s_val[0] = T.exp2(s_val[0] - s_max[0]) + s_other_val[0] = T.exp2(s_other_val[0] - s_max[0]) + scale[0] = s_val[0] / (s_val[0] + s_other_val[0]) + other_scale[0] = s_other_val[0] / (s_val[0] + s_other_val[0]) + for vec in T.vectorized(4): + v_vec[vec] = V[bx, ty + by * 16, tx * 4 + vec] + for vec in T.vectorized(4): + v_other_vec[vec] = V_other[bx, ty + by * 16, tx * 4 + vec] + for vec in range(4): + v_vec[vec] = T.Cast("float16", T.Cast("float32", v_vec[vec]) * scale[0] + T.Cast("float32", v_other_vec[vec]) * other_scale[0]) + for vec in T.vectorized(4): + V[bx, ty + by * 16, tx * 4 + vec] = v_vec[vec] + S[bx, ty + by * 16] = T.log2(s_val[0] + s_other_val[0]) + s_max[0] + + @T.prim_func + def gpt_neox_q0f16_reshape(var_add432: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + add432 = T.match_buffer(var_add432, (batch_size, 1, 3072), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, 1, 48, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * 12, thread="blockIdx.x"): + for ax0_ax1_ax2_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("T_reshape"): + v0 = T.axis.spatial(batch_size, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) // 3072) + v1 = T.axis.spatial(48, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 3072 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(add432[v0, 0, v1 * 64 + v2]) + T.writes(T_reshape[v0, 0, v1, v2]) + T_reshape[v0, 0, v1, v2] = add432[v0, 0, v1 * 64 + v2] + + @T.prim_func + def gpt_neox_q0f16_reshape1(var_reshape288: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + reshape288 = T.match_buffer(var_reshape288, (batch_size, 1, 48, 64), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, 48, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * 12, thread="blockIdx.x"): + for ax0_ax1_ax2_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("T_reshape"): + v0 = T.axis.spatial(batch_size, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) // 3072) + v1 = T.axis.spatial(48, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 3072 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(reshape288[v0, 0, v1, v2]) + T.writes(T_reshape[v0, v1, v2]) + T_reshape[v0, v1, v2] = reshape288[v0, 0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_reshape2(var_lv73: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + lv73 = T.match_buffer(var_lv73, (batch_size, 16, 64), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, 1, 16, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * 4, thread="blockIdx.x"): + for ax0_ax1_ax2_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("T_reshape"): + v0 = T.axis.spatial(batch_size, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) // 1024) + v1 = T.axis.spatial(16, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 1024 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(lv73[v0, v1, v2]) + T.writes(T_reshape[v0, 0, v1, v2]) + T_reshape[v0, 0, v1, v2] = lv73[v0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_reshape3(var_reshape290: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + reshape290 = T.match_buffer(var_reshape290, (batch_size, 1, 16, 64), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, 1, 1024), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(batch_size * 4, thread="blockIdx.x"): + for ax0_ax1_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("T_reshape"): + v0 = T.axis.spatial(batch_size, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) // 1024) + v1 = T.axis.spatial(1024, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 1024) + T.reads(reshape290[v0, 0, v1 // 64, v1 % 64]) + T.writes(T_reshape[v0, 0, v1]) + T_reshape[v0, 0, v1] = reshape290[v0, 0, v1 // 64, v1 % 64] + + @T.prim_func + def gpt_neox_q0f16_reshape4(var_add288: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + add288 = T.match_buffer(var_add288, (1, seq_len, 3072), "float16") + T_reshape = T.match_buffer(var_T_reshape, (1, seq_len, 48, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 12, thread="blockIdx.x"): + for ax0_ax1_ax2_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("T_reshape"): + v0 = T.axis.spatial(seq_len, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) // 3072) + v1 = T.axis.spatial(48, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 3072 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(add288[0, v0, v1 * 64 + v2]) + T.writes(T_reshape[0, v0, v1, v2]) + T_reshape[0, v0, v1, v2] = add288[0, v0, v1 * 64 + v2] + + @T.prim_func + def gpt_neox_q0f16_reshape5(var_reshape192: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + reshape192 = T.match_buffer(var_reshape192, (1, seq_len, 48, 64), "float16") + T_reshape = T.match_buffer(var_T_reshape, (seq_len, 48, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 12, thread="blockIdx.x"): + for ax0_ax1_ax2_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("T_reshape"): + v0 = T.axis.spatial(seq_len, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) // 3072) + v1 = T.axis.spatial(48, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 3072 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(reshape192[0, v0, v1, v2]) + T.writes(T_reshape[v0, v1, v2]) + T_reshape[v0, v1, v2] = reshape192[0, v0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_reshape6(var_lv49: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + lv49 = T.match_buffer(var_lv49, (seq_len, 16, 64), "float16") + T_reshape = T.match_buffer(var_T_reshape, (1, seq_len, 16, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 4, thread="blockIdx.x"): + for ax0_ax1_ax2_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("T_reshape"): + v0 = T.axis.spatial(seq_len, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) // 1024) + v1 = T.axis.spatial(16, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 1024 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(lv49[v0, v1, v2]) + T.writes(T_reshape[0, v0, v1, v2]) + T_reshape[0, v0, v1, v2] = lv49[v0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_reshape7(var_reshape194: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + reshape194 = T.match_buffer(var_reshape194, (1, seq_len, 16, 64), "float16") + T_reshape = T.match_buffer(var_T_reshape, (1, seq_len, 1024), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(seq_len * 4, thread="blockIdx.x"): + for ax0_ax1_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("T_reshape"): + v0 = T.axis.spatial(seq_len, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) // 1024) + v1 = T.axis.spatial(1024, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 1024) + T.reads(reshape194[0, v0, v1 // 64, v1 % 64]) + T.writes(T_reshape[0, v0, v1]) + T_reshape[0, v0, v1] = reshape194[0, v0, v1 // 64, v1 % 64] + + @T.prim_func + def gpt_neox_q0f16_scatter_probs(var_src: T.handle, var_indices: T.handle, var_dst: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size, n = T.int32(is_size_var=True), T.int32(is_size_var=True) + src = T.match_buffer(var_src, (batch_size, n)) + indices = T.match_buffer(var_indices, (batch_size,), "int32") + m = T.int32(is_size_var=True) + dst = T.match_buffer(var_dst, (m, n)) + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding((batch_size * n + 255) // 256, thread="blockIdx.x"): + for ax0_ax1_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("scatter_2d"): + v0 = T.axis.spatial(batch_size, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % (n * batch_size) // n) + v1 = T.axis.spatial(n, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % n) + T.where(ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1 < batch_size * n) + T.reads(src[v0, v1], indices[v0]) + T.writes(dst[indices[v0], v1]) + dst[indices[v0], v1] = src[v0, v1] + + @T.prim_func + def gpt_neox_q0f16_shape_func(H: T.Buffer((T.int64(5),), "int64")): + T.func_attr({"tir.is_host_func": 1}) + H[T.int64(3)] = (H[T.int64(1)] + T.int64(4096) - T.int64(1)) // T.int64(4096) + H[T.int64(2)] = T.int64(512) * ((H[T.int64(1)] + T.int64(4096) - T.int64(1)) // T.int64(4096)) + H[T.int64(4)] = T.int64(512) * H[T.int64(1)] + + @T.prim_func + def gpt_neox_q0f16_shape_func1(H: T.Buffer((T.int64(4),), "int64")): + T.func_attr({"tir.is_host_func": 1}) + H[T.int64(3)] = T.int64(128) * H[T.int64(1)] * T.int64(4) + + @T.prim_func + def gpt_neox_q0f16_shape_func2(H: T.Buffer((T.int64(4),), "int64")): + T.func_attr({"tir.is_host_func": 1}) + H[T.int64(3)] = H[T.int64(1)] * T.int64(4) + + @T.prim_func + def gpt_neox_q0f16_shape_func3(H: T.Buffer((T.int64(5),), "int64")): + T.func_attr({"tir.is_host_func": 1}) + H[T.int64(4)] = T.int64(128) * H[T.int64(2)] * T.int64(4) + + @T.prim_func + def gpt_neox_q0f16_shape_func4(H: T.Buffer((T.int64(4),), "int64")): + T.func_attr({"tir.is_host_func": 1}) + H[T.int64(3)] = T.int64(2048) * H[T.int64(1)] * T.int64(4) + + @T.prim_func + def gpt_neox_q0f16_shape_func5(H: T.Buffer((T.int64(3),), "int64")): + T.func_attr({"tir.is_host_func": 1}) + H[T.int64(2)] = H[T.int64(0)] * T.int64(4) + + @T.prim_func + def gpt_neox_q0f16_softmax_with_chunked_sum(var_A: T.handle, var_temperature: T.handle, var_chunked_sum: T.handle, var_chunked_max: T.handle, var_softmax: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size, vocab_size = T.int32(is_size_var=True), T.int32(is_size_var=True) + A = T.match_buffer(var_A, (batch_size, vocab_size)) + temperature = T.match_buffer(var_temperature, (batch_size,)) + num_chunks = T.int32(is_size_var=True) + chunked_sum = T.match_buffer(var_chunked_sum, (batch_size, num_chunks)) + chunked_max = T.match_buffer(var_chunked_max, (batch_size, num_chunks)) + softmax = T.match_buffer(var_softmax, (batch_size, vocab_size)) + # with T.block("root"): + temp_max_shared = T.alloc_buffer((batch_size,), scope="shared") + temp_sum_shared = T.alloc_buffer((batch_size,), scope="shared") + for l0_l1_fused in T.thread_binding(batch_size * num_chunks, thread="blockIdx.x"): + for ax0_1 in T.thread_binding(32, thread="threadIdx.x"): + for ax0_0 in T.serial((num_chunks + 31) // 32, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("max"): + v0 = T.axis.spatial(batch_size, l0_l1_fused % (num_chunks * batch_size) // num_chunks) + v1 = T.axis.reduce(num_chunks, ax0_0 * 32 + ax0_1) + T.where(ax0_0 * 32 + ax0_1 < num_chunks) + T.reads(chunked_max[v0, v1]) + T.writes(temp_max_shared[v0]) + with T.init(): + temp_max_shared[v0] = T.float32(-340282346638528859811704183484516925440.0) + temp_max_shared[v0] = T.max(temp_max_shared[v0], chunked_max[v0, v1]) + for ax0_1 in T.thread_binding(32, thread="threadIdx.x"): + for ax0_0 in T.serial((num_chunks + 31) // 32, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("sum_exp"): + v0 = T.axis.spatial(batch_size, l0_l1_fused % (num_chunks * batch_size) // num_chunks) + v1 = T.axis.reduce(num_chunks, ax0_0 * 32 + ax0_1) + T.where(ax0_0 * 32 + ax0_1 < num_chunks) + T.reads(temperature[v0], chunked_sum[v0, v1], chunked_max[v0, v1], temp_max_shared[v0]) + T.writes(temp_sum_shared[v0]) + with T.init(): + temp_sum_shared[v0] = T.float32(0.0) + temp_sum_shared[v0] = temp_sum_shared[v0] + T.Select(temperature[v0] > T.float32(1.0000000000000001e-05), T.exp(chunked_sum[v0, v1] + chunked_max[v0, v1] - temp_max_shared[v0]), T.Cast("float32", chunked_max[v0, v1] == temp_max_shared[v0]) * chunked_sum[v0, v1]) + for l2_0 in T.serial(16, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + for l2_1 in T.thread_binding(8, thread="threadIdx.y"): + for l2_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("log_pad"): + v0 = T.axis.spatial(batch_size, l0_l1_fused % (num_chunks * batch_size) // num_chunks) + v1 = T.axis.spatial(num_chunks, l0_l1_fused % num_chunks) + v2 = T.axis.spatial(4096, l2_0 * 256 + l2_1 * 32 + l2_2) + T.reads(temperature[v0], A[v0, v1 * 4096 + v2], temp_sum_shared[v0], temp_max_shared[v0]) + T.writes(softmax[v0, v1 * 4096 + v2]) + if v1 * 4096 + v2 < vocab_size: + softmax[v0, v1 * 4096 + v2] = T.if_then_else(temperature[v0] > T.float32(1.0000000000000001e-05), T.exp(A[v0, v1 * 4096 + v2] / temperature[v0] - (T.log(temp_sum_shared[v0]) + temp_max_shared[v0])), T.Cast("float32", A[v0, v1 * 4096 + v2] == temp_max_shared[v0]) / temp_sum_shared[v0]) + + @T.prim_func + def gpt_neox_q0f16_take(var_layer_norm146: T.handle, var_logit_positions: T.handle, var_T_take: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + layer_norm146 = T.match_buffer(var_layer_norm146, (1, seq_len, 1024), "float16") + batch_size = T.int32() + logit_positions = T.match_buffer(var_logit_positions, (batch_size,), "int32") + T_take = T.match_buffer(var_T_take, (1, batch_size, 1024), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(batch_size * 4, thread="blockIdx.x"): + for ax0_ax1_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("T_take"): + v0 = T.axis.spatial(batch_size, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) // 1024) + v1 = T.axis.spatial(1024, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 1024) + T.reads(layer_norm146[0, logit_positions[v0], v1], logit_positions[v0]) + T.writes(T_take[0, v0, v1]) + T_take[0, v0, v1] = layer_norm146[0, logit_positions[v0], v1] + + @T.prim_func + def gpt_neox_q0f16_take1(var_gpt_neox_embed_in_weight: T.handle, var_input_ids: T.handle, var_T_take: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + vocab_size = T.int32() + gpt_neox_embed_in_weight = T.match_buffer(var_gpt_neox_embed_in_weight, (vocab_size, 1024), "float16") + seq_len = T.int32() + input_ids = T.match_buffer(var_input_ids, (seq_len,), "int32") + T_take = T.match_buffer(var_T_take, (seq_len, 1024), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(seq_len * 4, thread="blockIdx.x"): + for ax0_ax1_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("T_take"): + v0 = T.axis.spatial(seq_len, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) // 1024) + v1 = T.axis.spatial(1024, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 1024) + T.reads(gpt_neox_embed_in_weight[input_ids[v0], v1], input_ids[v0]) + T.writes(T_take[v0, v1]) + T_take[v0, v1] = gpt_neox_embed_in_weight[input_ids[v0], v1] + + @T.prim_func + def gpt_neox_q0f16_tir_kv_cache_debug_get_kv(var_pages: T.handle, var_position_map: T.handle, var_k_data: T.handle, var_v_data: T.handle, layer_id: T.int32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + num_pages, page_size = T.int32(), T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (num_pages, 2, 16, page_size, 64), "float16", offset_factor=1) + seqlen = T.int32(is_size_var=True) + position_map = T.match_buffer(var_position_map, (seqlen,), "int32", offset_factor=1) + k_data = T.match_buffer(var_k_data, (24, seqlen, 16, 64), "float16") + v_data = T.match_buffer(var_v_data, (24, seqlen, 16, 64), "float16") + # with T.block("root"): + for p_h_d_fused_0 in T.thread_binding(seqlen * 4, thread="blockIdx.x"): + for p_h_d_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("copy0"): + vp = T.axis.spatial(seqlen, (p_h_d_fused_0 * 256 + p_h_d_fused_1) // 1024) + vh = T.axis.spatial(16, (p_h_d_fused_0 * 256 + p_h_d_fused_1) % 1024 // 64) + vd = T.axis.spatial(64, (p_h_d_fused_0 * 256 + p_h_d_fused_1) % 64) + T.reads(position_map[vp], pages[position_map[vp] // page_size, 0:2, vh, position_map[vp] % page_size, vd]) + T.writes(k_data[layer_id, vp, vh, vd], v_data[layer_id, vp, vh, vd]) + position: T.int32 = position_map[vp] + k_data[layer_id, vp, vh, vd] = pages[position // page_size, 0, vh, position % page_size, vd] + v_data[layer_id, vp, vh, vd] = pages[position // page_size, 1, vh, position % page_size, vd] + + @T.prim_func + def gpt_neox_q0f16_tir_kv_cache_transpose_append(var_pages: T.handle, var_k_data: T.handle, var_v_data: T.handle, var_position_map: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + num_pages = T.int32() + pages = T.match_buffer(var_pages, (num_pages, 2, 16, 16, 64), "float16", offset_factor=1) + ntoken = T.int32(is_size_var=True) + k_data = T.match_buffer(var_k_data, (ntoken, 16, 64), "float16") + v_data = T.match_buffer(var_v_data, (ntoken, 16, 64), "float16") + position_map = T.match_buffer(var_position_map, (ntoken,), "int32", offset_factor=1) + # with T.block("root"): + for global_pos_h_f_fused_0 in T.thread_binding(ntoken * 4, thread="blockIdx.x"): + for global_pos_h_f_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + if position_map[(global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) // 1024] != -1: + with T.block("k_transpose_append"): + vgpos = T.axis.spatial(ntoken, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) // 1024) + vh = T.axis.spatial(16, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) % 1024 // 64) + vf = T.axis.spatial(64, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) % 64) + T.reads(position_map[vgpos], k_data[vgpos, vh, vf]) + T.writes(pages[position_map[vgpos] // 16, 0, vh, position_map[vgpos] % 16, vf]) + position: T.int32 = position_map[vgpos] + pages[position // 16, 0, vh, position % 16, vf] = k_data[vgpos, vh, vf] + with T.block("v_transpose_append"): + vgpos = T.axis.spatial(ntoken, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) // 1024) + vh = T.axis.spatial(16, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) % 1024 // 64) + vf = T.axis.spatial(64, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) % 64) + T.reads(position_map[vgpos], v_data[vgpos, vh, vf]) + T.writes(pages[position_map[vgpos] // 16, 1, vh, position_map[vgpos] % 16, vf]) + position: T.int32 = position_map[vgpos] + pages[position // 16, 1, vh, position % 16, vf] = v_data[vgpos, vh, vf] + + @T.prim_func + def gpt_neox_q0f16_tree_attn_paged_kv(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32, tree_order_indptr_handle: T.handle, tree_order_handle: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + total_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (total_len, 16, 64), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 16, 32, 64), "float16") + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 16, 64), "float16") + lse = T.match_buffer(var_lse, (total_len, 16)) + tree_order_indptr = T.match_buffer(tree_order_indptr_handle, (batch_size + 1,), "int32", offset_factor=1) + total_tree_order_len = T.int32(is_size_var=True) + tree_order = T.match_buffer(tree_order_handle, (total_tree_order_len, 2), "int32", offset_factor=1) + # with T.block("root"): + assert rotary_mode == 0, "Inline rotary mode is not supported in tree attention." + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(16, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 32), scope="shared") + S_local = T.alloc_buffer((32, 32), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 32 + length_info[b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(4): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) + j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 31) // 32): + L_kv_start: T.int32 = iterator_1 * 32 + for lz_ly_fused_0 in range(4): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 32] + page_offset: T.int32 = seq_offset % 32 + K_smem[i, j] = pages[page_no, 0, by, page_offset, j] + else: + K_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_ly_fused_0 in range(4): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 32] + page_offset: T.int32 = seq_offset % 32 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:32, 0:64]) + T.writes(S_local[0:32, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 4): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 4, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + k = T.axis.reduce(64, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 4): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 32: + row_: T.int32 = LH_start + row + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(4, 8, 4, 4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + k = T.axis.reduce(32, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @R.function + def _metadata() -> R.Object: + shape_heap: R.Object = R.null_value() + return R.str("{\"model_type\": \"gpt_neox\", \"quantization\": \"q0f16\", \"context_window_size\": 2048, \"sliding_window_size\": -1, \"attention_sink_size\": -1, \"prefill_chunk_size\": 2048, \"tensor_parallel_shards\": 1, \"pipeline_parallel_stages\": 1, \"disaggregation\": false, \"kv_state_kind\": \"kv_cache\", \"max_batch_size\": 128, \"params\": [{\"name\": \"gpt_neox.embed_in.weight\", \"shape\": [\"vocab_size\", 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.16.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.16.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.16.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.16.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.16.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.16.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.16.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.16.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.16.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.16.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.16.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.16.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.17.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.17.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.17.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.17.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.17.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.17.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.17.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.17.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.17.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.17.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.17.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.17.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.18.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.18.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.18.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.18.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.18.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.18.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.18.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.18.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.18.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.18.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.18.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.18.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.19.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.19.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.19.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.19.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.19.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.19.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.19.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.19.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.19.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.19.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.19.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.19.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.20.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.20.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.20.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.20.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.20.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.20.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.20.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.20.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.20.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.20.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.20.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.20.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.21.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.21.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.21.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.21.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.21.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.21.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.21.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.21.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.21.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.21.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.21.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.21.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.22.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.22.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.22.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.22.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.22.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.22.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.22.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.22.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.22.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.22.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.22.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.22.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.23.input_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.23.input_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.23.post_attention_layernorm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.23.post_attention_layernorm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.23.attention.query_key_value.weight\", \"shape\": [3072, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.23.attention.query_key_value.bias\", \"shape\": [3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.23.attention.dense.weight\", \"shape\": [1024, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.23.attention.dense.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.23.mlp.dense_h_to_4h.weight\", \"shape\": [4096, 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.23.mlp.dense_h_to_4h.bias\", \"shape\": [4096], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.23.mlp.dense_4h_to_h.weight\", \"shape\": [1024, 4096], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.23.mlp.dense_4h_to_h.bias\", \"shape\": [1024], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.final_layer_norm.weight\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.final_layer_norm.bias\", \"shape\": [1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"embed_out.weight\", \"shape\": [\"vocab_size\", 1024], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}], \"kv_cache\": {\"num_hidden_layers\": 24, \"num_attention_heads\": 16, \"num_key_value_heads\": 16, \"head_dim\": 64}, \"memory_usage\": {\"alloc_embedding_tensor\": 4194304, \"batch_decode\": 3145728, \"batch_prefill\": 50593792, \"batch_verify\": 50331648, \"create_tir_paged_kv_cache\": 0, \"decode\": 24576, \"embed\": 4194304, \"prefill\": 50333696, \"softmax_with_temperature\": 0}}") + + @R.function + def alloc_embedding_tensor() -> R.Tensor((2048, 1024), dtype="float16"): + R.func_attr({"relax.force_pure": True}) + shape_heap: R.Object = R.null_value() + storage: R.Object = R.vm.alloc_storage(R.shape([4194304]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv: R.Tensor((2048, 1024), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([2048, 1024]), R.dtype("float16")) + R.vm.kill_object(storage) + return gv + + @R.function + def batch_decode(input_embeds: R.Tensor(("batch_size", 1, 1024), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor(("vocab_size", 1024), dtype="float16"))) -> R.Tuple(R.Tensor(("batch_size", 1, "vocab_size"), dtype="float32"), R.Object): + batch_size = T.int64() + vocab_size = T.int64() + R.func_attr({"num_input": 2, "pipeline_parallel_stages": 1, "relax.force_pure": True, "relax.rewrite_cuda_graph.capture_symbolic_vars": ["batch_size"], "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) + cls = Module + shape_heap: R.Tensor(dtype="int64", ndim=1) = R.call_builtin_with_ctx("vm.builtin.alloc_shape_heap", (R.prim_value(4),), sinfo_args=(R.Tensor(dtype="int64", ndim=1),)) + R.call_packed("vm.builtin.check_tensor_info", input_embeds, R.prim_value(3), R.dtype("float16"), R.str("ErrorContext(fn=batch_decode, loc=param[0], param=input_embeds, annotation=R.Tensor((batch_size, 1, 1024), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(292), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv9: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv9, R.prim_value(2), R.dtype("float16"), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv10: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[291] + R.call_packed("vm.builtin.check_tensor_info", gv10, R.prim_value(2), R.dtype("float16"), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", input_embeds, shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=batch_decode, loc=param[0], param=input_embeds, annotation=R.Tensor((batch_size, 1, 1024), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv9, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv10, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func1(shape_heap) + gpt_neox_layers_0_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[2] + storage3: R.Object = R.vm.alloc_storage(R.shape([262144]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv11: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc3: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv11, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(input_embeds, gpt_neox_layers_0_input_layernorm_weight4, gpt_neox_layers_0_input_layernorm_bias4, alloc3) + R.vm.kill_object(gpt_neox_layers_0_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_0_input_layernorm_bias4) + gpt_neox_layers_0_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[6] + storage4: R.Object = R.vm.alloc_storage(R.shape([786432]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv12: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc4: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv12, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc3, gpt_neox_layers_0_attention_query_key_value_weight4, gpt_neox_layers_0_attention_query_key_value_bias4, alloc4) + R.vm.kill_object(alloc3) + R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_bias4) + gv13: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape288: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc4, gv13, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc4) + gv14: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape289: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape288, gv14, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape288) + gv15: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc5: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv15, R.dtype("float16")) + _4: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), reshape289, alloc5) + R.vm.kill_object(reshape289) + gv16: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape290: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc5, gv16, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc5) + gv17: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape291: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape290, gv17, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape290) + gpt_neox_layers_0_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[4] + gv18: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc6: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv18, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(input_embeds, gpt_neox_layers_0_post_attention_layernorm_weight4, gpt_neox_layers_0_post_attention_layernorm_bias4, alloc6) + R.vm.kill_object(gpt_neox_layers_0_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_0_post_attention_layernorm_bias4) + gpt_neox_layers_0_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[10] + storage5: R.Object = R.vm.alloc_storage(R.shape([1048576]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv19: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc7: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv19, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc6, gpt_neox_layers_0_mlp_dense_h_to_4h_weight4, gpt_neox_layers_0_mlp_dense_h_to_4h_bias4, alloc7) + R.vm.kill_object(alloc6) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_0_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[12] + gv20: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc8: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv20, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc7, gpt_neox_layers_0_mlp_dense_4h_to_h_weight4, gpt_neox_layers_0_mlp_dense_4h_to_h_bias4, alloc8) + R.vm.kill_object(alloc7) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_0_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[8] + gv21: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc9: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv21, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape291, gpt_neox_layers_0_attention_dense_weight4, gpt_neox_layers_0_attention_dense_bias4, alloc8, input_embeds, alloc9) + R.vm.kill_object(reshape291) + R.vm.kill_object(alloc8) + R.vm.kill_object(gpt_neox_layers_0_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_0_attention_dense_bias4) + gpt_neox_layers_1_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[14] + gv22: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc10: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv22, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc9, gpt_neox_layers_1_input_layernorm_weight4, gpt_neox_layers_1_input_layernorm_bias4, alloc10) + R.vm.kill_object(gpt_neox_layers_1_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_1_input_layernorm_bias4) + gpt_neox_layers_1_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[18] + gv23: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc11: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv23, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc10, gpt_neox_layers_1_attention_query_key_value_weight4, gpt_neox_layers_1_attention_query_key_value_bias4, alloc11) + R.vm.kill_object(alloc10) + R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_bias4) + gv24: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape292: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc11, gv24, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc11) + gv25: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape293: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape292, gv25, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape292) + gv26: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc12: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv26, R.dtype("float16")) + _11: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), reshape293, alloc12) + R.vm.kill_object(reshape293) + gv27: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape294: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc12, gv27, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc12) + gv28: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape295: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape294, gv28, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape294) + gpt_neox_layers_1_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[16] + gv29: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc13: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv29, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc9, gpt_neox_layers_1_post_attention_layernorm_weight4, gpt_neox_layers_1_post_attention_layernorm_bias4, alloc13) + R.vm.kill_object(gpt_neox_layers_1_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_1_post_attention_layernorm_bias4) + gpt_neox_layers_1_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[22] + storage6: R.Object = R.vm.alloc_storage(R.shape([1048576]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv30: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc14: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv30, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc13, gpt_neox_layers_1_mlp_dense_h_to_4h_weight4, gpt_neox_layers_1_mlp_dense_h_to_4h_bias4, alloc14) + R.vm.kill_object(alloc13) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_1_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[24] + gv31: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc15: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv31, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc14, gpt_neox_layers_1_mlp_dense_4h_to_h_weight4, gpt_neox_layers_1_mlp_dense_4h_to_h_bias4, alloc15) + R.vm.kill_object(alloc14) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_1_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[20] + gv32: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc16: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv32, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape295, gpt_neox_layers_1_attention_dense_weight4, gpt_neox_layers_1_attention_dense_bias4, alloc15, alloc9, alloc16) + R.vm.kill_object(alloc9) + R.vm.kill_object(reshape295) + R.vm.kill_object(alloc15) + R.vm.kill_object(gpt_neox_layers_1_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_1_attention_dense_bias4) + gpt_neox_layers_2_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[26] + gv33: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc17: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv33, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc16, gpt_neox_layers_2_input_layernorm_weight4, gpt_neox_layers_2_input_layernorm_bias4, alloc17) + R.vm.kill_object(gpt_neox_layers_2_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_2_input_layernorm_bias4) + gpt_neox_layers_2_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[30] + gv34: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc18: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv34, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc17, gpt_neox_layers_2_attention_query_key_value_weight4, gpt_neox_layers_2_attention_query_key_value_bias4, alloc18) + R.vm.kill_object(alloc17) + R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_bias4) + gv35: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape296: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc18, gv35, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc18) + gv36: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape297: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape296, gv36, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape296) + gv37: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc19: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv37, R.dtype("float16")) + _18: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), reshape297, alloc19) + R.vm.kill_object(reshape297) + gv38: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape298: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc19, gv38, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc19) + gv39: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape299: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape298, gv39, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape298) + gpt_neox_layers_2_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[28] + gv40: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc20: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv40, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc16, gpt_neox_layers_2_post_attention_layernorm_weight4, gpt_neox_layers_2_post_attention_layernorm_bias4, alloc20) + R.vm.kill_object(gpt_neox_layers_2_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_2_post_attention_layernorm_bias4) + gpt_neox_layers_2_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[34] + gv41: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc21: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv41, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc20, gpt_neox_layers_2_mlp_dense_h_to_4h_weight4, gpt_neox_layers_2_mlp_dense_h_to_4h_bias4, alloc21) + R.vm.kill_object(alloc20) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_2_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[36] + gv42: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc22: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv42, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc21, gpt_neox_layers_2_mlp_dense_4h_to_h_weight4, gpt_neox_layers_2_mlp_dense_4h_to_h_bias4, alloc22) + R.vm.kill_object(alloc21) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_2_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[32] + gv43: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc23: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv43, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape299, gpt_neox_layers_2_attention_dense_weight4, gpt_neox_layers_2_attention_dense_bias4, alloc22, alloc16, alloc23) + R.vm.kill_object(alloc16) + R.vm.kill_object(reshape299) + R.vm.kill_object(alloc22) + R.vm.kill_object(gpt_neox_layers_2_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_2_attention_dense_bias4) + gpt_neox_layers_3_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[38] + gv44: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc24: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv44, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc23, gpt_neox_layers_3_input_layernorm_weight4, gpt_neox_layers_3_input_layernorm_bias4, alloc24) + R.vm.kill_object(gpt_neox_layers_3_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_3_input_layernorm_bias4) + gpt_neox_layers_3_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[42] + gv45: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc25: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv45, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc24, gpt_neox_layers_3_attention_query_key_value_weight4, gpt_neox_layers_3_attention_query_key_value_bias4, alloc25) + R.vm.kill_object(alloc24) + R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_bias4) + gv46: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape300: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc25, gv46, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc25) + gv47: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape301: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape300, gv47, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape300) + gv48: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc26: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv48, R.dtype("float16")) + _25: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), reshape301, alloc26) + R.vm.kill_object(reshape301) + gv49: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape302: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc26, gv49, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc26) + gv50: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape303: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape302, gv50, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape302) + gpt_neox_layers_3_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[40] + gv51: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc27: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv51, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc23, gpt_neox_layers_3_post_attention_layernorm_weight4, gpt_neox_layers_3_post_attention_layernorm_bias4, alloc27) + R.vm.kill_object(gpt_neox_layers_3_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_3_post_attention_layernorm_bias4) + gpt_neox_layers_3_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[46] + gv52: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc28: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv52, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc27, gpt_neox_layers_3_mlp_dense_h_to_4h_weight4, gpt_neox_layers_3_mlp_dense_h_to_4h_bias4, alloc28) + R.vm.kill_object(alloc27) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_3_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[48] + gv53: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc29: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv53, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc28, gpt_neox_layers_3_mlp_dense_4h_to_h_weight4, gpt_neox_layers_3_mlp_dense_4h_to_h_bias4, alloc29) + R.vm.kill_object(alloc28) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_3_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[44] + gv54: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc30: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv54, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape303, gpt_neox_layers_3_attention_dense_weight4, gpt_neox_layers_3_attention_dense_bias4, alloc29, alloc23, alloc30) + R.vm.kill_object(alloc23) + R.vm.kill_object(reshape303) + R.vm.kill_object(alloc29) + R.vm.kill_object(gpt_neox_layers_3_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_3_attention_dense_bias4) + gpt_neox_layers_4_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[50] + gv55: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc31: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv55, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc30, gpt_neox_layers_4_input_layernorm_weight4, gpt_neox_layers_4_input_layernorm_bias4, alloc31) + R.vm.kill_object(gpt_neox_layers_4_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_4_input_layernorm_bias4) + gpt_neox_layers_4_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[54] + gv56: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc32: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv56, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc31, gpt_neox_layers_4_attention_query_key_value_weight4, gpt_neox_layers_4_attention_query_key_value_bias4, alloc32) + R.vm.kill_object(alloc31) + R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_bias4) + gv57: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape304: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc32, gv57, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc32) + gv58: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape305: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape304, gv58, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape304) + gv59: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc33: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv59, R.dtype("float16")) + _32: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), reshape305, alloc33) + R.vm.kill_object(reshape305) + gv60: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape306: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc33, gv60, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc33) + gv61: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape307: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape306, gv61, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape306) + gpt_neox_layers_4_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[52] + gv62: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc34: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv62, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc30, gpt_neox_layers_4_post_attention_layernorm_weight4, gpt_neox_layers_4_post_attention_layernorm_bias4, alloc34) + R.vm.kill_object(gpt_neox_layers_4_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_4_post_attention_layernorm_bias4) + gpt_neox_layers_4_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[58] + gv63: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc35: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv63, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc34, gpt_neox_layers_4_mlp_dense_h_to_4h_weight4, gpt_neox_layers_4_mlp_dense_h_to_4h_bias4, alloc35) + R.vm.kill_object(alloc34) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_4_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[60] + gv64: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc36: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv64, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc35, gpt_neox_layers_4_mlp_dense_4h_to_h_weight4, gpt_neox_layers_4_mlp_dense_4h_to_h_bias4, alloc36) + R.vm.kill_object(alloc35) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_4_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[56] + gv65: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc37: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv65, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape307, gpt_neox_layers_4_attention_dense_weight4, gpt_neox_layers_4_attention_dense_bias4, alloc36, alloc30, alloc37) + R.vm.kill_object(alloc30) + R.vm.kill_object(reshape307) + R.vm.kill_object(alloc36) + R.vm.kill_object(gpt_neox_layers_4_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_4_attention_dense_bias4) + gpt_neox_layers_5_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[62] + gv66: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc38: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv66, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc37, gpt_neox_layers_5_input_layernorm_weight4, gpt_neox_layers_5_input_layernorm_bias4, alloc38) + R.vm.kill_object(gpt_neox_layers_5_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_5_input_layernorm_bias4) + gpt_neox_layers_5_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[66] + gv67: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc39: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv67, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc38, gpt_neox_layers_5_attention_query_key_value_weight4, gpt_neox_layers_5_attention_query_key_value_bias4, alloc39) + R.vm.kill_object(alloc38) + R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_bias4) + gv68: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape308: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc39, gv68, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc39) + gv69: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape309: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape308, gv69, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape308) + gv70: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc40: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv70, R.dtype("float16")) + _39: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), reshape309, alloc40) + R.vm.kill_object(reshape309) + gv71: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape310: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc40, gv71, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc40) + gv72: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape311: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape310, gv72, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape310) + gpt_neox_layers_5_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[64] + gv73: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc41: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv73, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc37, gpt_neox_layers_5_post_attention_layernorm_weight4, gpt_neox_layers_5_post_attention_layernorm_bias4, alloc41) + R.vm.kill_object(gpt_neox_layers_5_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_5_post_attention_layernorm_bias4) + gpt_neox_layers_5_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[70] + gv74: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc42: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv74, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc41, gpt_neox_layers_5_mlp_dense_h_to_4h_weight4, gpt_neox_layers_5_mlp_dense_h_to_4h_bias4, alloc42) + R.vm.kill_object(alloc41) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_5_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[72] + gv75: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc43: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv75, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc42, gpt_neox_layers_5_mlp_dense_4h_to_h_weight4, gpt_neox_layers_5_mlp_dense_4h_to_h_bias4, alloc43) + R.vm.kill_object(alloc42) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_5_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[68] + gv76: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc44: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv76, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape311, gpt_neox_layers_5_attention_dense_weight4, gpt_neox_layers_5_attention_dense_bias4, alloc43, alloc37, alloc44) + R.vm.kill_object(alloc37) + R.vm.kill_object(reshape311) + R.vm.kill_object(alloc43) + R.vm.kill_object(gpt_neox_layers_5_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_5_attention_dense_bias4) + gpt_neox_layers_6_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[74] + gv77: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc45: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv77, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc44, gpt_neox_layers_6_input_layernorm_weight4, gpt_neox_layers_6_input_layernorm_bias4, alloc45) + R.vm.kill_object(gpt_neox_layers_6_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_6_input_layernorm_bias4) + gpt_neox_layers_6_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[78] + gv78: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc46: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv78, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc45, gpt_neox_layers_6_attention_query_key_value_weight4, gpt_neox_layers_6_attention_query_key_value_bias4, alloc46) + R.vm.kill_object(alloc45) + R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_bias4) + gv79: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape312: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc46, gv79, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc46) + gv80: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape313: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape312, gv80, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape312) + gv81: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc47: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv81, R.dtype("float16")) + _46: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), reshape313, alloc47) + R.vm.kill_object(reshape313) + gv82: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape314: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc47, gv82, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc47) + gv83: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape315: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape314, gv83, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape314) + gpt_neox_layers_6_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[76] + gv84: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc48: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv84, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc44, gpt_neox_layers_6_post_attention_layernorm_weight4, gpt_neox_layers_6_post_attention_layernorm_bias4, alloc48) + R.vm.kill_object(gpt_neox_layers_6_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_6_post_attention_layernorm_bias4) + gpt_neox_layers_6_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[82] + gv85: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc49: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv85, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc48, gpt_neox_layers_6_mlp_dense_h_to_4h_weight4, gpt_neox_layers_6_mlp_dense_h_to_4h_bias4, alloc49) + R.vm.kill_object(alloc48) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_6_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[84] + gv86: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc50: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv86, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc49, gpt_neox_layers_6_mlp_dense_4h_to_h_weight4, gpt_neox_layers_6_mlp_dense_4h_to_h_bias4, alloc50) + R.vm.kill_object(alloc49) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_6_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[80] + gv87: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc51: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv87, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape315, gpt_neox_layers_6_attention_dense_weight4, gpt_neox_layers_6_attention_dense_bias4, alloc50, alloc44, alloc51) + R.vm.kill_object(alloc44) + R.vm.kill_object(reshape315) + R.vm.kill_object(alloc50) + R.vm.kill_object(gpt_neox_layers_6_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_6_attention_dense_bias4) + gpt_neox_layers_7_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[86] + gv88: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc52: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv88, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc51, gpt_neox_layers_7_input_layernorm_weight4, gpt_neox_layers_7_input_layernorm_bias4, alloc52) + R.vm.kill_object(gpt_neox_layers_7_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_7_input_layernorm_bias4) + gpt_neox_layers_7_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[90] + gv89: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc53: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv89, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc52, gpt_neox_layers_7_attention_query_key_value_weight4, gpt_neox_layers_7_attention_query_key_value_bias4, alloc53) + R.vm.kill_object(alloc52) + R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_bias4) + gv90: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape316: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc53, gv90, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc53) + gv91: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape317: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape316, gv91, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape316) + gv92: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc54: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv92, R.dtype("float16")) + _53: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), reshape317, alloc54) + R.vm.kill_object(reshape317) + gv93: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape318: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc54, gv93, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc54) + gv94: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape319: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape318, gv94, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape318) + gpt_neox_layers_7_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[88] + gv95: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc55: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv95, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc51, gpt_neox_layers_7_post_attention_layernorm_weight4, gpt_neox_layers_7_post_attention_layernorm_bias4, alloc55) + R.vm.kill_object(gpt_neox_layers_7_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_7_post_attention_layernorm_bias4) + gpt_neox_layers_7_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[94] + gv96: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc56: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv96, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc55, gpt_neox_layers_7_mlp_dense_h_to_4h_weight4, gpt_neox_layers_7_mlp_dense_h_to_4h_bias4, alloc56) + R.vm.kill_object(alloc55) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_7_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[96] + gv97: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc57: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv97, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc56, gpt_neox_layers_7_mlp_dense_4h_to_h_weight4, gpt_neox_layers_7_mlp_dense_4h_to_h_bias4, alloc57) + R.vm.kill_object(alloc56) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_7_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[92] + gv98: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc58: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv98, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape319, gpt_neox_layers_7_attention_dense_weight4, gpt_neox_layers_7_attention_dense_bias4, alloc57, alloc51, alloc58) + R.vm.kill_object(alloc51) + R.vm.kill_object(reshape319) + R.vm.kill_object(alloc57) + R.vm.kill_object(gpt_neox_layers_7_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_7_attention_dense_bias4) + gpt_neox_layers_8_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[98] + gv99: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc59: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv99, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc58, gpt_neox_layers_8_input_layernorm_weight4, gpt_neox_layers_8_input_layernorm_bias4, alloc59) + R.vm.kill_object(gpt_neox_layers_8_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_8_input_layernorm_bias4) + gpt_neox_layers_8_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[102] + gv100: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc60: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv100, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc59, gpt_neox_layers_8_attention_query_key_value_weight4, gpt_neox_layers_8_attention_query_key_value_bias4, alloc60) + R.vm.kill_object(alloc59) + R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_bias4) + gv101: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape320: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc60, gv101, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc60) + gv102: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape321: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape320, gv102, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape320) + gv103: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc61: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv103, R.dtype("float16")) + _60: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), reshape321, alloc61) + R.vm.kill_object(reshape321) + gv104: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape322: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc61, gv104, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc61) + gv105: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape323: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape322, gv105, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape322) + gpt_neox_layers_8_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[100] + gv106: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc62: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv106, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc58, gpt_neox_layers_8_post_attention_layernorm_weight4, gpt_neox_layers_8_post_attention_layernorm_bias4, alloc62) + R.vm.kill_object(gpt_neox_layers_8_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_8_post_attention_layernorm_bias4) + gpt_neox_layers_8_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[106] + gv107: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc63: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv107, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc62, gpt_neox_layers_8_mlp_dense_h_to_4h_weight4, gpt_neox_layers_8_mlp_dense_h_to_4h_bias4, alloc63) + R.vm.kill_object(alloc62) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_8_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[108] + gv108: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc64: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv108, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc63, gpt_neox_layers_8_mlp_dense_4h_to_h_weight4, gpt_neox_layers_8_mlp_dense_4h_to_h_bias4, alloc64) + R.vm.kill_object(alloc63) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_8_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[104] + gv109: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc65: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv109, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape323, gpt_neox_layers_8_attention_dense_weight4, gpt_neox_layers_8_attention_dense_bias4, alloc64, alloc58, alloc65) + R.vm.kill_object(alloc58) + R.vm.kill_object(reshape323) + R.vm.kill_object(alloc64) + R.vm.kill_object(gpt_neox_layers_8_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_8_attention_dense_bias4) + gpt_neox_layers_9_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[110] + gv110: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc66: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv110, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc65, gpt_neox_layers_9_input_layernorm_weight4, gpt_neox_layers_9_input_layernorm_bias4, alloc66) + R.vm.kill_object(gpt_neox_layers_9_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_9_input_layernorm_bias4) + gpt_neox_layers_9_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[114] + gv111: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc67: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv111, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc66, gpt_neox_layers_9_attention_query_key_value_weight4, gpt_neox_layers_9_attention_query_key_value_bias4, alloc67) + R.vm.kill_object(alloc66) + R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_bias4) + gv112: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape324: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc67, gv112, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc67) + gv113: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape325: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape324, gv113, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape324) + gv114: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc68: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv114, R.dtype("float16")) + _67: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), reshape325, alloc68) + R.vm.kill_object(reshape325) + gv115: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape326: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc68, gv115, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc68) + gv116: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape327: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape326, gv116, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape326) + gpt_neox_layers_9_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[112] + gv117: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc69: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv117, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc65, gpt_neox_layers_9_post_attention_layernorm_weight4, gpt_neox_layers_9_post_attention_layernorm_bias4, alloc69) + R.vm.kill_object(gpt_neox_layers_9_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_9_post_attention_layernorm_bias4) + gpt_neox_layers_9_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[118] + gv118: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc70: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv118, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc69, gpt_neox_layers_9_mlp_dense_h_to_4h_weight4, gpt_neox_layers_9_mlp_dense_h_to_4h_bias4, alloc70) + R.vm.kill_object(alloc69) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_9_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[120] + gv119: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc71: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv119, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc70, gpt_neox_layers_9_mlp_dense_4h_to_h_weight4, gpt_neox_layers_9_mlp_dense_4h_to_h_bias4, alloc71) + R.vm.kill_object(alloc70) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_9_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[116] + gv120: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc72: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv120, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape327, gpt_neox_layers_9_attention_dense_weight4, gpt_neox_layers_9_attention_dense_bias4, alloc71, alloc65, alloc72) + R.vm.kill_object(alloc65) + R.vm.kill_object(reshape327) + R.vm.kill_object(alloc71) + R.vm.kill_object(gpt_neox_layers_9_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_9_attention_dense_bias4) + gpt_neox_layers_10_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[122] + gv121: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc73: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv121, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc72, gpt_neox_layers_10_input_layernorm_weight4, gpt_neox_layers_10_input_layernorm_bias4, alloc73) + R.vm.kill_object(gpt_neox_layers_10_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_10_input_layernorm_bias4) + gpt_neox_layers_10_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[126] + gv122: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc74: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv122, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc73, gpt_neox_layers_10_attention_query_key_value_weight4, gpt_neox_layers_10_attention_query_key_value_bias4, alloc74) + R.vm.kill_object(alloc73) + R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_bias4) + gv123: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape328: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc74, gv123, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc74) + gv124: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape329: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape328, gv124, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape328) + gv125: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc75: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv125, R.dtype("float16")) + _74: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), reshape329, alloc75) + R.vm.kill_object(reshape329) + gv126: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape330: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc75, gv126, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc75) + gv127: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape331: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape330, gv127, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape330) + gpt_neox_layers_10_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[124] + gv128: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc76: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv128, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc72, gpt_neox_layers_10_post_attention_layernorm_weight4, gpt_neox_layers_10_post_attention_layernorm_bias4, alloc76) + R.vm.kill_object(gpt_neox_layers_10_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_10_post_attention_layernorm_bias4) + gpt_neox_layers_10_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[130] + gv129: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc77: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv129, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc76, gpt_neox_layers_10_mlp_dense_h_to_4h_weight4, gpt_neox_layers_10_mlp_dense_h_to_4h_bias4, alloc77) + R.vm.kill_object(alloc76) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_10_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[132] + gv130: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc78: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv130, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc77, gpt_neox_layers_10_mlp_dense_4h_to_h_weight4, gpt_neox_layers_10_mlp_dense_4h_to_h_bias4, alloc78) + R.vm.kill_object(alloc77) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_10_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[128] + gv131: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc79: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv131, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape331, gpt_neox_layers_10_attention_dense_weight4, gpt_neox_layers_10_attention_dense_bias4, alloc78, alloc72, alloc79) + R.vm.kill_object(alloc72) + R.vm.kill_object(reshape331) + R.vm.kill_object(alloc78) + R.vm.kill_object(gpt_neox_layers_10_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_10_attention_dense_bias4) + gpt_neox_layers_11_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[134] + gv132: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc80: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv132, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc79, gpt_neox_layers_11_input_layernorm_weight4, gpt_neox_layers_11_input_layernorm_bias4, alloc80) + R.vm.kill_object(gpt_neox_layers_11_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_11_input_layernorm_bias4) + gpt_neox_layers_11_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[138] + gv133: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc81: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv133, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc80, gpt_neox_layers_11_attention_query_key_value_weight4, gpt_neox_layers_11_attention_query_key_value_bias4, alloc81) + R.vm.kill_object(alloc80) + R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_bias4) + gv134: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape332: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc81, gv134, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc81) + gv135: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape333: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape332, gv135, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape332) + gv136: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc82: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv136, R.dtype("float16")) + _81: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), reshape333, alloc82) + R.vm.kill_object(reshape333) + gv137: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape334: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc82, gv137, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc82) + gv138: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape335: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape334, gv138, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape334) + gpt_neox_layers_11_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[136] + gv139: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc83: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv139, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc79, gpt_neox_layers_11_post_attention_layernorm_weight4, gpt_neox_layers_11_post_attention_layernorm_bias4, alloc83) + R.vm.kill_object(gpt_neox_layers_11_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_11_post_attention_layernorm_bias4) + gpt_neox_layers_11_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[142] + gv140: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc84: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv140, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc83, gpt_neox_layers_11_mlp_dense_h_to_4h_weight4, gpt_neox_layers_11_mlp_dense_h_to_4h_bias4, alloc84) + R.vm.kill_object(alloc83) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_11_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[144] + gv141: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc85: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv141, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc84, gpt_neox_layers_11_mlp_dense_4h_to_h_weight4, gpt_neox_layers_11_mlp_dense_4h_to_h_bias4, alloc85) + R.vm.kill_object(alloc84) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_11_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[140] + gv142: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc86: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv142, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape335, gpt_neox_layers_11_attention_dense_weight4, gpt_neox_layers_11_attention_dense_bias4, alloc85, alloc79, alloc86) + R.vm.kill_object(alloc79) + R.vm.kill_object(reshape335) + R.vm.kill_object(alloc85) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias4) + gpt_neox_layers_12_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[145] + gpt_neox_layers_12_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[146] + gv143: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc87: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv143, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc86, gpt_neox_layers_12_input_layernorm_weight4, gpt_neox_layers_12_input_layernorm_bias4, alloc87) + R.vm.kill_object(gpt_neox_layers_12_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_12_input_layernorm_bias4) + gpt_neox_layers_12_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[149] + gpt_neox_layers_12_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[150] + gv144: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc88: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv144, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc87, gpt_neox_layers_12_attention_query_key_value_weight4, gpt_neox_layers_12_attention_query_key_value_bias4, alloc88) + R.vm.kill_object(alloc87) + R.vm.kill_object(gpt_neox_layers_12_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_12_attention_query_key_value_bias4) + gv145: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape336: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc88, gv145, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc88) + gv146: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape337: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape336, gv146, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape336) + gv147: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc89: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv147, R.dtype("float16")) + _88: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape337, alloc89) + R.vm.kill_object(reshape337) + gv148: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape338: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc89, gv148, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc89) + gv149: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape339: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape338, gv149, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape338) + gpt_neox_layers_12_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[147] + gpt_neox_layers_12_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[148] + gv150: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc90: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv150, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc86, gpt_neox_layers_12_post_attention_layernorm_weight4, gpt_neox_layers_12_post_attention_layernorm_bias4, alloc90) + R.vm.kill_object(gpt_neox_layers_12_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_12_post_attention_layernorm_bias4) + gpt_neox_layers_12_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[153] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[154] + gv151: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc91: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv151, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc90, gpt_neox_layers_12_mlp_dense_h_to_4h_weight4, gpt_neox_layers_12_mlp_dense_h_to_4h_bias4, alloc91) + R.vm.kill_object(alloc90) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_12_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[155] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[156] + gv152: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc92: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv152, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc91, gpt_neox_layers_12_mlp_dense_4h_to_h_weight4, gpt_neox_layers_12_mlp_dense_4h_to_h_bias4, alloc92) + R.vm.kill_object(alloc91) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_12_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[151] + gpt_neox_layers_12_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[152] + gv153: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc93: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv153, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape339, gpt_neox_layers_12_attention_dense_weight4, gpt_neox_layers_12_attention_dense_bias4, alloc92, alloc86, alloc93) + R.vm.kill_object(alloc86) + R.vm.kill_object(reshape339) + R.vm.kill_object(alloc92) + R.vm.kill_object(gpt_neox_layers_12_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_12_attention_dense_bias4) + gpt_neox_layers_13_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[157] + gpt_neox_layers_13_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[158] + gv154: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc94: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv154, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc93, gpt_neox_layers_13_input_layernorm_weight4, gpt_neox_layers_13_input_layernorm_bias4, alloc94) + R.vm.kill_object(gpt_neox_layers_13_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_13_input_layernorm_bias4) + gpt_neox_layers_13_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[161] + gpt_neox_layers_13_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[162] + gv155: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc95: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv155, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc94, gpt_neox_layers_13_attention_query_key_value_weight4, gpt_neox_layers_13_attention_query_key_value_bias4, alloc95) + R.vm.kill_object(alloc94) + R.vm.kill_object(gpt_neox_layers_13_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_13_attention_query_key_value_bias4) + gv156: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape340: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc95, gv156, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc95) + gv157: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape341: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape340, gv157, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape340) + gv158: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc96: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv158, R.dtype("float16")) + _95: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape341, alloc96) + R.vm.kill_object(reshape341) + gv159: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape342: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc96, gv159, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc96) + gv160: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape343: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape342, gv160, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape342) + gpt_neox_layers_13_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[159] + gpt_neox_layers_13_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[160] + gv161: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc97: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv161, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc93, gpt_neox_layers_13_post_attention_layernorm_weight4, gpt_neox_layers_13_post_attention_layernorm_bias4, alloc97) + R.vm.kill_object(gpt_neox_layers_13_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_13_post_attention_layernorm_bias4) + gpt_neox_layers_13_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[165] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[166] + gv162: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc98: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv162, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc97, gpt_neox_layers_13_mlp_dense_h_to_4h_weight4, gpt_neox_layers_13_mlp_dense_h_to_4h_bias4, alloc98) + R.vm.kill_object(alloc97) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_13_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[167] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[168] + gv163: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc99: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv163, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc98, gpt_neox_layers_13_mlp_dense_4h_to_h_weight4, gpt_neox_layers_13_mlp_dense_4h_to_h_bias4, alloc99) + R.vm.kill_object(alloc98) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_13_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[163] + gpt_neox_layers_13_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[164] + gv164: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc100: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv164, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape343, gpt_neox_layers_13_attention_dense_weight4, gpt_neox_layers_13_attention_dense_bias4, alloc99, alloc93, alloc100) + R.vm.kill_object(alloc93) + R.vm.kill_object(reshape343) + R.vm.kill_object(alloc99) + R.vm.kill_object(gpt_neox_layers_13_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_13_attention_dense_bias4) + gpt_neox_layers_14_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[169] + gpt_neox_layers_14_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[170] + gv165: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc101: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv165, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc100, gpt_neox_layers_14_input_layernorm_weight4, gpt_neox_layers_14_input_layernorm_bias4, alloc101) + R.vm.kill_object(gpt_neox_layers_14_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_14_input_layernorm_bias4) + gpt_neox_layers_14_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[173] + gpt_neox_layers_14_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[174] + gv166: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc102: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv166, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc101, gpt_neox_layers_14_attention_query_key_value_weight4, gpt_neox_layers_14_attention_query_key_value_bias4, alloc102) + R.vm.kill_object(alloc101) + R.vm.kill_object(gpt_neox_layers_14_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_14_attention_query_key_value_bias4) + gv167: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape344: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc102, gv167, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc102) + gv168: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape345: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape344, gv168, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape344) + gv169: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc103: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv169, R.dtype("float16")) + _102: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape345, alloc103) + R.vm.kill_object(reshape345) + gv170: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape346: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc103, gv170, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc103) + gv171: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape347: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape346, gv171, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape346) + gpt_neox_layers_14_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[171] + gpt_neox_layers_14_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[172] + gv172: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc104: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv172, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc100, gpt_neox_layers_14_post_attention_layernorm_weight4, gpt_neox_layers_14_post_attention_layernorm_bias4, alloc104) + R.vm.kill_object(gpt_neox_layers_14_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_14_post_attention_layernorm_bias4) + gpt_neox_layers_14_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[177] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[178] + gv173: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc105: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv173, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc104, gpt_neox_layers_14_mlp_dense_h_to_4h_weight4, gpt_neox_layers_14_mlp_dense_h_to_4h_bias4, alloc105) + R.vm.kill_object(alloc104) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_14_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[179] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[180] + gv174: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc106: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv174, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc105, gpt_neox_layers_14_mlp_dense_4h_to_h_weight4, gpt_neox_layers_14_mlp_dense_4h_to_h_bias4, alloc106) + R.vm.kill_object(alloc105) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_14_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[175] + gpt_neox_layers_14_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[176] + gv175: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc107: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv175, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape347, gpt_neox_layers_14_attention_dense_weight4, gpt_neox_layers_14_attention_dense_bias4, alloc106, alloc100, alloc107) + R.vm.kill_object(alloc100) + R.vm.kill_object(reshape347) + R.vm.kill_object(alloc106) + R.vm.kill_object(gpt_neox_layers_14_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_14_attention_dense_bias4) + gpt_neox_layers_15_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[181] + gpt_neox_layers_15_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[182] + gv176: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc108: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv176, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc107, gpt_neox_layers_15_input_layernorm_weight4, gpt_neox_layers_15_input_layernorm_bias4, alloc108) + R.vm.kill_object(gpt_neox_layers_15_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_15_input_layernorm_bias4) + gpt_neox_layers_15_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[185] + gpt_neox_layers_15_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[186] + gv177: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc109: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv177, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc108, gpt_neox_layers_15_attention_query_key_value_weight4, gpt_neox_layers_15_attention_query_key_value_bias4, alloc109) + R.vm.kill_object(alloc108) + R.vm.kill_object(gpt_neox_layers_15_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_15_attention_query_key_value_bias4) + gv178: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape348: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc109, gv178, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc109) + gv179: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape349: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape348, gv179, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape348) + gv180: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc110: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv180, R.dtype("float16")) + _109: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape349, alloc110) + R.vm.kill_object(reshape349) + gv181: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape350: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc110, gv181, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc110) + gv182: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape351: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape350, gv182, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape350) + gpt_neox_layers_15_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[183] + gpt_neox_layers_15_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[184] + gv183: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc111: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv183, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc107, gpt_neox_layers_15_post_attention_layernorm_weight4, gpt_neox_layers_15_post_attention_layernorm_bias4, alloc111) + R.vm.kill_object(gpt_neox_layers_15_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_15_post_attention_layernorm_bias4) + gpt_neox_layers_15_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[189] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[190] + gv184: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc112: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv184, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc111, gpt_neox_layers_15_mlp_dense_h_to_4h_weight4, gpt_neox_layers_15_mlp_dense_h_to_4h_bias4, alloc112) + R.vm.kill_object(alloc111) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_15_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[191] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[192] + gv185: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc113: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv185, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc112, gpt_neox_layers_15_mlp_dense_4h_to_h_weight4, gpt_neox_layers_15_mlp_dense_4h_to_h_bias4, alloc113) + R.vm.kill_object(alloc112) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_15_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[187] + gpt_neox_layers_15_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[188] + gv186: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc114: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv186, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape351, gpt_neox_layers_15_attention_dense_weight4, gpt_neox_layers_15_attention_dense_bias4, alloc113, alloc107, alloc114) + R.vm.kill_object(alloc107) + R.vm.kill_object(reshape351) + R.vm.kill_object(alloc113) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_bias4) + gpt_neox_layers_16_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[193] + gpt_neox_layers_16_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[194] + gv187: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc115: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv187, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc114, gpt_neox_layers_16_input_layernorm_weight4, gpt_neox_layers_16_input_layernorm_bias4, alloc115) + R.vm.kill_object(gpt_neox_layers_16_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_16_input_layernorm_bias4) + gpt_neox_layers_16_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[197] + gpt_neox_layers_16_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[198] + gv188: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc116: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv188, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc115, gpt_neox_layers_16_attention_query_key_value_weight4, gpt_neox_layers_16_attention_query_key_value_bias4, alloc116) + R.vm.kill_object(alloc115) + R.vm.kill_object(gpt_neox_layers_16_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_16_attention_query_key_value_bias4) + gv189: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape352: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc116, gv189, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc116) + gv190: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape353: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape352, gv190, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape352) + gv191: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc117: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv191, R.dtype("float16")) + _116: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), reshape353, alloc117) + R.vm.kill_object(reshape353) + gv192: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape354: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc117, gv192, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc117) + gv193: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape355: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape354, gv193, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape354) + gpt_neox_layers_16_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[195] + gpt_neox_layers_16_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[196] + gv194: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc118: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv194, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc114, gpt_neox_layers_16_post_attention_layernorm_weight4, gpt_neox_layers_16_post_attention_layernorm_bias4, alloc118) + R.vm.kill_object(gpt_neox_layers_16_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_16_post_attention_layernorm_bias4) + gpt_neox_layers_16_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[201] + gpt_neox_layers_16_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[202] + gv195: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc119: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv195, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc118, gpt_neox_layers_16_mlp_dense_h_to_4h_weight4, gpt_neox_layers_16_mlp_dense_h_to_4h_bias4, alloc119) + R.vm.kill_object(alloc118) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_16_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[203] + gpt_neox_layers_16_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[204] + gv196: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc120: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv196, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc119, gpt_neox_layers_16_mlp_dense_4h_to_h_weight4, gpt_neox_layers_16_mlp_dense_4h_to_h_bias4, alloc120) + R.vm.kill_object(alloc119) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_16_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[199] + gpt_neox_layers_16_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[200] + gv197: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc121: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv197, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape355, gpt_neox_layers_16_attention_dense_weight4, gpt_neox_layers_16_attention_dense_bias4, alloc120, alloc114, alloc121) + R.vm.kill_object(alloc114) + R.vm.kill_object(reshape355) + R.vm.kill_object(alloc120) + R.vm.kill_object(gpt_neox_layers_16_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_16_attention_dense_bias4) + gpt_neox_layers_17_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[205] + gpt_neox_layers_17_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[206] + gv198: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc122: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv198, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc121, gpt_neox_layers_17_input_layernorm_weight4, gpt_neox_layers_17_input_layernorm_bias4, alloc122) + R.vm.kill_object(gpt_neox_layers_17_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_17_input_layernorm_bias4) + gpt_neox_layers_17_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[209] + gpt_neox_layers_17_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[210] + gv199: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc123: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv199, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc122, gpt_neox_layers_17_attention_query_key_value_weight4, gpt_neox_layers_17_attention_query_key_value_bias4, alloc123) + R.vm.kill_object(alloc122) + R.vm.kill_object(gpt_neox_layers_17_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_17_attention_query_key_value_bias4) + gv200: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape356: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc123, gv200, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc123) + gv201: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape357: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape356, gv201, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape356) + gv202: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc124: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv202, R.dtype("float16")) + _123: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), reshape357, alloc124) + R.vm.kill_object(reshape357) + gv203: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape358: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc124, gv203, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc124) + gv204: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape359: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape358, gv204, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape358) + gpt_neox_layers_17_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[207] + gpt_neox_layers_17_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[208] + gv205: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc125: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv205, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc121, gpt_neox_layers_17_post_attention_layernorm_weight4, gpt_neox_layers_17_post_attention_layernorm_bias4, alloc125) + R.vm.kill_object(gpt_neox_layers_17_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_17_post_attention_layernorm_bias4) + gpt_neox_layers_17_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[213] + gpt_neox_layers_17_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[214] + gv206: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc126: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv206, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc125, gpt_neox_layers_17_mlp_dense_h_to_4h_weight4, gpt_neox_layers_17_mlp_dense_h_to_4h_bias4, alloc126) + R.vm.kill_object(alloc125) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_17_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[215] + gpt_neox_layers_17_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[216] + gv207: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc127: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv207, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc126, gpt_neox_layers_17_mlp_dense_4h_to_h_weight4, gpt_neox_layers_17_mlp_dense_4h_to_h_bias4, alloc127) + R.vm.kill_object(alloc126) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_17_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[211] + gpt_neox_layers_17_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[212] + gv208: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc128: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv208, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape359, gpt_neox_layers_17_attention_dense_weight4, gpt_neox_layers_17_attention_dense_bias4, alloc127, alloc121, alloc128) + R.vm.kill_object(alloc121) + R.vm.kill_object(reshape359) + R.vm.kill_object(alloc127) + R.vm.kill_object(gpt_neox_layers_17_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_17_attention_dense_bias4) + gpt_neox_layers_18_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[217] + gpt_neox_layers_18_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[218] + gv209: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc129: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv209, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc128, gpt_neox_layers_18_input_layernorm_weight4, gpt_neox_layers_18_input_layernorm_bias4, alloc129) + R.vm.kill_object(gpt_neox_layers_18_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_18_input_layernorm_bias4) + gpt_neox_layers_18_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[221] + gpt_neox_layers_18_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[222] + gv210: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc130: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv210, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc129, gpt_neox_layers_18_attention_query_key_value_weight4, gpt_neox_layers_18_attention_query_key_value_bias4, alloc130) + R.vm.kill_object(alloc129) + R.vm.kill_object(gpt_neox_layers_18_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_18_attention_query_key_value_bias4) + gv211: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape360: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc130, gv211, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc130) + gv212: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape361: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape360, gv212, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape360) + gv213: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc131: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv213, R.dtype("float16")) + _130: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), reshape361, alloc131) + R.vm.kill_object(reshape361) + gv214: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape362: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc131, gv214, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc131) + gv215: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape363: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape362, gv215, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape362) + gpt_neox_layers_18_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[219] + gpt_neox_layers_18_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[220] + gv216: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc132: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv216, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc128, gpt_neox_layers_18_post_attention_layernorm_weight4, gpt_neox_layers_18_post_attention_layernorm_bias4, alloc132) + R.vm.kill_object(gpt_neox_layers_18_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_18_post_attention_layernorm_bias4) + gpt_neox_layers_18_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[225] + gpt_neox_layers_18_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[226] + gv217: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc133: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv217, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc132, gpt_neox_layers_18_mlp_dense_h_to_4h_weight4, gpt_neox_layers_18_mlp_dense_h_to_4h_bias4, alloc133) + R.vm.kill_object(alloc132) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_18_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[227] + gpt_neox_layers_18_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[228] + gv218: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc134: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv218, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc133, gpt_neox_layers_18_mlp_dense_4h_to_h_weight4, gpt_neox_layers_18_mlp_dense_4h_to_h_bias4, alloc134) + R.vm.kill_object(alloc133) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_18_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[223] + gpt_neox_layers_18_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[224] + gv219: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc135: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv219, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape363, gpt_neox_layers_18_attention_dense_weight4, gpt_neox_layers_18_attention_dense_bias4, alloc134, alloc128, alloc135) + R.vm.kill_object(alloc128) + R.vm.kill_object(reshape363) + R.vm.kill_object(alloc134) + R.vm.kill_object(gpt_neox_layers_18_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_18_attention_dense_bias4) + gpt_neox_layers_19_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[229] + gpt_neox_layers_19_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[230] + gv220: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc136: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv220, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc135, gpt_neox_layers_19_input_layernorm_weight4, gpt_neox_layers_19_input_layernorm_bias4, alloc136) + R.vm.kill_object(gpt_neox_layers_19_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_19_input_layernorm_bias4) + gpt_neox_layers_19_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[233] + gpt_neox_layers_19_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[234] + gv221: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc137: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv221, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc136, gpt_neox_layers_19_attention_query_key_value_weight4, gpt_neox_layers_19_attention_query_key_value_bias4, alloc137) + R.vm.kill_object(alloc136) + R.vm.kill_object(gpt_neox_layers_19_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_19_attention_query_key_value_bias4) + gv222: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape364: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc137, gv222, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc137) + gv223: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape365: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape364, gv223, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape364) + gv224: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc138: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv224, R.dtype("float16")) + _137: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), reshape365, alloc138) + R.vm.kill_object(reshape365) + gv225: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape366: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc138, gv225, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc138) + gv226: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape367: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape366, gv226, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape366) + gpt_neox_layers_19_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[231] + gpt_neox_layers_19_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[232] + gv227: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc139: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv227, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc135, gpt_neox_layers_19_post_attention_layernorm_weight4, gpt_neox_layers_19_post_attention_layernorm_bias4, alloc139) + R.vm.kill_object(gpt_neox_layers_19_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_19_post_attention_layernorm_bias4) + gpt_neox_layers_19_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[237] + gpt_neox_layers_19_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[238] + gv228: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc140: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv228, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc139, gpt_neox_layers_19_mlp_dense_h_to_4h_weight4, gpt_neox_layers_19_mlp_dense_h_to_4h_bias4, alloc140) + R.vm.kill_object(alloc139) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_19_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[239] + gpt_neox_layers_19_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[240] + gv229: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc141: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv229, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc140, gpt_neox_layers_19_mlp_dense_4h_to_h_weight4, gpt_neox_layers_19_mlp_dense_4h_to_h_bias4, alloc141) + R.vm.kill_object(alloc140) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_19_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[235] + gpt_neox_layers_19_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[236] + gv230: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc142: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv230, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape367, gpt_neox_layers_19_attention_dense_weight4, gpt_neox_layers_19_attention_dense_bias4, alloc141, alloc135, alloc142) + R.vm.kill_object(alloc135) + R.vm.kill_object(reshape367) + R.vm.kill_object(alloc141) + R.vm.kill_object(gpt_neox_layers_19_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_19_attention_dense_bias4) + gpt_neox_layers_20_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[241] + gpt_neox_layers_20_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[242] + gv231: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc143: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv231, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc142, gpt_neox_layers_20_input_layernorm_weight4, gpt_neox_layers_20_input_layernorm_bias4, alloc143) + R.vm.kill_object(gpt_neox_layers_20_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_20_input_layernorm_bias4) + gpt_neox_layers_20_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[245] + gpt_neox_layers_20_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[246] + gv232: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc144: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv232, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc143, gpt_neox_layers_20_attention_query_key_value_weight4, gpt_neox_layers_20_attention_query_key_value_bias4, alloc144) + R.vm.kill_object(alloc143) + R.vm.kill_object(gpt_neox_layers_20_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_20_attention_query_key_value_bias4) + gv233: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape368: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc144, gv233, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc144) + gv234: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape369: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape368, gv234, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape368) + gv235: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc145: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv235, R.dtype("float16")) + _144: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), reshape369, alloc145) + R.vm.kill_object(reshape369) + gv236: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape370: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc145, gv236, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc145) + gv237: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape371: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape370, gv237, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape370) + gpt_neox_layers_20_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[243] + gpt_neox_layers_20_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[244] + gv238: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc146: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv238, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc142, gpt_neox_layers_20_post_attention_layernorm_weight4, gpt_neox_layers_20_post_attention_layernorm_bias4, alloc146) + R.vm.kill_object(gpt_neox_layers_20_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_20_post_attention_layernorm_bias4) + gpt_neox_layers_20_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[249] + gpt_neox_layers_20_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[250] + gv239: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc147: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv239, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc146, gpt_neox_layers_20_mlp_dense_h_to_4h_weight4, gpt_neox_layers_20_mlp_dense_h_to_4h_bias4, alloc147) + R.vm.kill_object(alloc146) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_20_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[251] + gpt_neox_layers_20_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[252] + gv240: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc148: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv240, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc147, gpt_neox_layers_20_mlp_dense_4h_to_h_weight4, gpt_neox_layers_20_mlp_dense_4h_to_h_bias4, alloc148) + R.vm.kill_object(alloc147) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_20_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[247] + gpt_neox_layers_20_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[248] + gv241: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc149: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv241, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape371, gpt_neox_layers_20_attention_dense_weight4, gpt_neox_layers_20_attention_dense_bias4, alloc148, alloc142, alloc149) + R.vm.kill_object(alloc142) + R.vm.kill_object(reshape371) + R.vm.kill_object(alloc148) + R.vm.kill_object(gpt_neox_layers_20_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_20_attention_dense_bias4) + gpt_neox_layers_21_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[253] + gpt_neox_layers_21_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[254] + gv242: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc150: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv242, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc149, gpt_neox_layers_21_input_layernorm_weight4, gpt_neox_layers_21_input_layernorm_bias4, alloc150) + R.vm.kill_object(gpt_neox_layers_21_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_21_input_layernorm_bias4) + gpt_neox_layers_21_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[257] + gpt_neox_layers_21_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[258] + gv243: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc151: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv243, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc150, gpt_neox_layers_21_attention_query_key_value_weight4, gpt_neox_layers_21_attention_query_key_value_bias4, alloc151) + R.vm.kill_object(alloc150) + R.vm.kill_object(gpt_neox_layers_21_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_21_attention_query_key_value_bias4) + gv244: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape372: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc151, gv244, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc151) + gv245: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape373: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape372, gv245, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape372) + gv246: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc152: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv246, R.dtype("float16")) + _151: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), reshape373, alloc152) + R.vm.kill_object(reshape373) + gv247: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape374: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc152, gv247, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc152) + gv248: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape375: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape374, gv248, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape374) + gpt_neox_layers_21_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[255] + gpt_neox_layers_21_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[256] + gv249: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc153: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv249, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc149, gpt_neox_layers_21_post_attention_layernorm_weight4, gpt_neox_layers_21_post_attention_layernorm_bias4, alloc153) + R.vm.kill_object(gpt_neox_layers_21_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_21_post_attention_layernorm_bias4) + gpt_neox_layers_21_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[261] + gpt_neox_layers_21_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[262] + gv250: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc154: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv250, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc153, gpt_neox_layers_21_mlp_dense_h_to_4h_weight4, gpt_neox_layers_21_mlp_dense_h_to_4h_bias4, alloc154) + R.vm.kill_object(alloc153) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_21_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[263] + gpt_neox_layers_21_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[264] + gv251: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc155: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv251, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc154, gpt_neox_layers_21_mlp_dense_4h_to_h_weight4, gpt_neox_layers_21_mlp_dense_4h_to_h_bias4, alloc155) + R.vm.kill_object(alloc154) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_21_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[259] + gpt_neox_layers_21_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[260] + gv252: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc156: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv252, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape375, gpt_neox_layers_21_attention_dense_weight4, gpt_neox_layers_21_attention_dense_bias4, alloc155, alloc149, alloc156) + R.vm.kill_object(alloc149) + R.vm.kill_object(reshape375) + R.vm.kill_object(alloc155) + R.vm.kill_object(gpt_neox_layers_21_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_21_attention_dense_bias4) + gpt_neox_layers_22_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[265] + gpt_neox_layers_22_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[266] + gv253: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc157: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv253, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc156, gpt_neox_layers_22_input_layernorm_weight4, gpt_neox_layers_22_input_layernorm_bias4, alloc157) + R.vm.kill_object(gpt_neox_layers_22_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_22_input_layernorm_bias4) + gpt_neox_layers_22_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[269] + gpt_neox_layers_22_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[270] + gv254: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc158: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv254, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc157, gpt_neox_layers_22_attention_query_key_value_weight4, gpt_neox_layers_22_attention_query_key_value_bias4, alloc158) + R.vm.kill_object(alloc157) + R.vm.kill_object(gpt_neox_layers_22_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_22_attention_query_key_value_bias4) + gv255: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape376: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc158, gv255, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc158) + gv256: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape377: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape376, gv256, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape376) + gv257: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc159: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv257, R.dtype("float16")) + _158: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), reshape377, alloc159) + R.vm.kill_object(reshape377) + gv258: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape378: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc159, gv258, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc159) + gv259: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape379: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape378, gv259, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape378) + gpt_neox_layers_22_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[267] + gpt_neox_layers_22_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[268] + gv260: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc160: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv260, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc156, gpt_neox_layers_22_post_attention_layernorm_weight4, gpt_neox_layers_22_post_attention_layernorm_bias4, alloc160) + R.vm.kill_object(gpt_neox_layers_22_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_22_post_attention_layernorm_bias4) + gpt_neox_layers_22_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[273] + gpt_neox_layers_22_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[274] + gv261: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc161: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv261, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc160, gpt_neox_layers_22_mlp_dense_h_to_4h_weight4, gpt_neox_layers_22_mlp_dense_h_to_4h_bias4, alloc161) + R.vm.kill_object(alloc160) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_22_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[275] + gpt_neox_layers_22_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[276] + gv262: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc162: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv262, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc161, gpt_neox_layers_22_mlp_dense_4h_to_h_weight4, gpt_neox_layers_22_mlp_dense_4h_to_h_bias4, alloc162) + R.vm.kill_object(alloc161) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_22_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[271] + gpt_neox_layers_22_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[272] + gv263: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc163: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv263, R.dtype("float16")) + R.vm.kill_object(storage5) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape379, gpt_neox_layers_22_attention_dense_weight4, gpt_neox_layers_22_attention_dense_bias4, alloc162, alloc156, alloc163) + R.vm.kill_object(alloc156) + R.vm.kill_object(reshape379) + R.vm.kill_object(alloc162) + R.vm.kill_object(gpt_neox_layers_22_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_22_attention_dense_bias4) + gpt_neox_layers_23_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[277] + gpt_neox_layers_23_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[278] + gv264: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc164: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv264, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc163, gpt_neox_layers_23_input_layernorm_weight4, gpt_neox_layers_23_input_layernorm_bias4, alloc164) + R.vm.kill_object(gpt_neox_layers_23_input_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_23_input_layernorm_bias4) + gpt_neox_layers_23_attention_query_key_value_weight4: R.Tensor((3072, 1024), dtype="float16") = packed_params[281] + gpt_neox_layers_23_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[282] + gv265: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc165: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv265, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc164, gpt_neox_layers_23_attention_query_key_value_weight4, gpt_neox_layers_23_attention_query_key_value_bias4, alloc165) + R.vm.kill_object(alloc164) + R.vm.kill_object(gpt_neox_layers_23_attention_query_key_value_weight4) + R.vm.kill_object(gpt_neox_layers_23_attention_query_key_value_bias4) + gv266: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape380: R.Tensor((batch_size, 1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc165, gv266, sinfo_args=(R.Tensor((batch_size, 1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc165) + gv267: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape381: R.Tensor((batch_size, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape380, gv267, sinfo_args=(R.Tensor((batch_size, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape380) + gv268: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc166: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv268, R.dtype("float16")) + _165: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), reshape381, alloc166) + R.vm.kill_object(reshape381) + gv269: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape382: R.Tensor((batch_size, 1, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc166, gv269, sinfo_args=(R.Tensor((batch_size, 1, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc166) + gv270: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape383: R.Tensor((batch_size, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape382, gv270, sinfo_args=(R.Tensor((batch_size, 1, 1024), dtype="float16"),)) + R.vm.kill_object(reshape382) + gpt_neox_layers_23_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[279] + gpt_neox_layers_23_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[280] + gv271: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc167: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv271, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc163, gpt_neox_layers_23_post_attention_layernorm_weight4, gpt_neox_layers_23_post_attention_layernorm_bias4, alloc167) + R.vm.kill_object(gpt_neox_layers_23_post_attention_layernorm_weight4) + R.vm.kill_object(gpt_neox_layers_23_post_attention_layernorm_bias4) + gpt_neox_layers_23_mlp_dense_h_to_4h_weight4: R.Tensor((4096, 1024), dtype="float16") = packed_params[285] + gpt_neox_layers_23_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[286] + gv272: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc168: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv272, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc167, gpt_neox_layers_23_mlp_dense_h_to_4h_weight4, gpt_neox_layers_23_mlp_dense_h_to_4h_bias4, alloc168) + R.vm.kill_object(alloc167) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_h_to_4h_weight4) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_h_to_4h_bias4) + gpt_neox_layers_23_mlp_dense_4h_to_h_weight4: R.Tensor((1024, 4096), dtype="float16") = packed_params[287] + gpt_neox_layers_23_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[288] + gv273: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc169: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv273, R.dtype("float16")) + R.vm.kill_object(storage4) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc168, gpt_neox_layers_23_mlp_dense_4h_to_h_weight4, gpt_neox_layers_23_mlp_dense_4h_to_h_bias4, alloc169) + R.vm.kill_object(alloc168) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_4h_to_h_weight4) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_4h_to_h_bias4) + gpt_neox_layers_23_attention_dense_weight4: R.Tensor((1024, 1024), dtype="float16") = packed_params[283] + gpt_neox_layers_23_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[284] + gv274: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc170: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv274, R.dtype("float16")) + R.vm.kill_object(storage6) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape383, gpt_neox_layers_23_attention_dense_weight4, gpt_neox_layers_23_attention_dense_bias4, alloc169, alloc163, alloc170) + R.vm.kill_object(alloc163) + R.vm.kill_object(reshape383) + R.vm.kill_object(alloc169) + R.vm.kill_object(gpt_neox_layers_23_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_23_attention_dense_bias4) + gpt_neox_final_layer_norm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[289] + gpt_neox_final_layer_norm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[290] + gv275: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc171: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv275, R.dtype("float16")) + R.vm.kill_object(storage3) + cls.gpt_neox_q0f16_layer_norm(alloc170, gpt_neox_final_layer_norm_weight4, gpt_neox_final_layer_norm_bias4, alloc171) + R.vm.kill_object(alloc170) + R.vm.kill_object(gpt_neox_final_layer_norm_weight4) + R.vm.kill_object(gpt_neox_final_layer_norm_bias4) + embed_out_weight4: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[291] + gv276: R.Shape(ndim=1) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(3), sinfo_args=(R.Shape(ndim=1),)) + storage7: R.Object = R.vm.alloc_storage(gv276, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv277: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(1), sinfo_args=(R.Shape(ndim=3),)) + alloc172: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv277, R.dtype("float32")) + R.vm.kill_object(storage7) + cls.gpt_neox_q0f16_fused_NT_matmul4_cast2(alloc171, embed_out_weight4, alloc172) + R.vm.kill_object(alloc171) + R.vm.kill_object(embed_out_weight4) + gv4: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc172, paged_kv_cache + R.vm.kill_object(alloc172) + gv278: R.Tensor(dtype="float32", ndim=3) = gv4[0] + R.call_packed("vm.builtin.match_shape", gv278, shape_heap, R.prim_value(3), R.prim_value(3), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(3), R.prim_value(1), R.str("ErrorContext(fn=batch_decode, loc=return, annotation=R.Tuple(R.Tensor((batch_size, 1, vocab_size), dtype=\"float32\"), R.Object)) "), sinfo_args=(R.Tuple,)) + return gv4 + + @R.function + def batch_prefill(input_embeds: R.Tensor((1, "seq_len", 1024), dtype="float16"), logit_positions: R.Tensor(("batch_size",), dtype="int32"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor(("vocab_size", 1024), dtype="float16"))) -> R.Tuple(R.Tensor((1, "batch_size", "vocab_size"), dtype="float32"), R.Object): + batch_size = T.int64() + vocab_size = T.int64() + seq_len = T.int64() + R.func_attr({"num_input": 3, "pipeline_parallel_stages": 1, "relax.force_pure": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) + cls = Module + shape_heap: R.Tensor(dtype="int64", ndim=1) = R.call_builtin_with_ctx("vm.builtin.alloc_shape_heap", (R.prim_value(5),), sinfo_args=(R.Tensor(dtype="int64", ndim=1),)) + R.call_packed("vm.builtin.check_tensor_info", input_embeds, R.prim_value(3), R.dtype("float16"), R.str("ErrorContext(fn=batch_prefill, loc=param[0], param=input_embeds, annotation=R.Tensor((1, seq_len, 1024), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tensor_info", logit_positions, R.prim_value(1), R.dtype("int32"), R.str("ErrorContext(fn=batch_prefill, loc=param[1], param=logit_positions, annotation=R.Tensor((batch_size,), dtype=\"int32\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(292), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv549: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv549, R.prim_value(2), R.dtype("float16"), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv550: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[291] + R.call_packed("vm.builtin.check_tensor_info", gv550, R.prim_value(2), R.dtype("float16"), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", input_embeds, shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=batch_prefill, loc=param[0], param=input_embeds, annotation=R.Tensor((1, seq_len, 1024), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", logit_positions, shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(1), R.str("ErrorContext(fn=batch_prefill, loc=param[1], param=logit_positions, annotation=R.Tensor((batch_size,), dtype=\"int32\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv549, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(2), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv550, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(2), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func3(shape_heap) + gpt_neox_layers_0_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[2] + storage14: R.Object = R.vm.alloc_storage(R.shape([4194304]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv551: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc344: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv551, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(input_embeds, gpt_neox_layers_0_input_layernorm_weight3, gpt_neox_layers_0_input_layernorm_bias3, alloc344) + R.vm.kill_object(gpt_neox_layers_0_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_0_input_layernorm_bias3) + gpt_neox_layers_0_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[6] + storage15: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv552: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc345: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv552, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc344, gpt_neox_layers_0_attention_query_key_value_weight3, gpt_neox_layers_0_attention_query_key_value_bias3, alloc345) + R.vm.kill_object(alloc344) + R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_bias3) + gv553: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape192: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc345, gv553, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc345) + gv554: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape193: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape192, gv554, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape192) + gv555: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc346: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv555, R.dtype("float16")) + _345: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), reshape193, alloc346) + R.vm.kill_object(reshape193) + gv556: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape194: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc346, gv556, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc346) + gv557: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape195: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape194, gv557, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape194) + gpt_neox_layers_0_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[4] + gv558: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc347: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv558, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(input_embeds, gpt_neox_layers_0_post_attention_layernorm_weight3, gpt_neox_layers_0_post_attention_layernorm_bias3, alloc347) + R.vm.kill_object(gpt_neox_layers_0_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_0_post_attention_layernorm_bias3) + gpt_neox_layers_0_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[10] + storage16: R.Object = R.vm.alloc_storage(R.shape([16777216]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv559: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc348: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv559, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc347, gpt_neox_layers_0_mlp_dense_h_to_4h_weight3, gpt_neox_layers_0_mlp_dense_h_to_4h_bias3, alloc348) + R.vm.kill_object(alloc347) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_0_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[12] + gv560: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc349: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv560, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc348, gpt_neox_layers_0_mlp_dense_4h_to_h_weight3, gpt_neox_layers_0_mlp_dense_4h_to_h_bias3, alloc349) + R.vm.kill_object(alloc348) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_0_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[8] + gv561: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc350: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv561, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape195, gpt_neox_layers_0_attention_dense_weight3, gpt_neox_layers_0_attention_dense_bias3, alloc349, input_embeds, alloc350) + R.vm.kill_object(reshape195) + R.vm.kill_object(alloc349) + R.vm.kill_object(gpt_neox_layers_0_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_0_attention_dense_bias3) + gpt_neox_layers_1_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[14] + gv562: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc351: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv562, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc350, gpt_neox_layers_1_input_layernorm_weight3, gpt_neox_layers_1_input_layernorm_bias3, alloc351) + R.vm.kill_object(gpt_neox_layers_1_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_1_input_layernorm_bias3) + gpt_neox_layers_1_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[18] + gv563: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc352: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv563, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc351, gpt_neox_layers_1_attention_query_key_value_weight3, gpt_neox_layers_1_attention_query_key_value_bias3, alloc352) + R.vm.kill_object(alloc351) + R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_bias3) + gv564: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape196: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc352, gv564, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc352) + gv565: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape197: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape196, gv565, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape196) + gv566: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc353: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv566, R.dtype("float16")) + _352: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), reshape197, alloc353) + R.vm.kill_object(reshape197) + gv567: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape198: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc353, gv567, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc353) + gv568: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape199: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape198, gv568, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape198) + gpt_neox_layers_1_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[16] + gv569: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc354: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv569, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc350, gpt_neox_layers_1_post_attention_layernorm_weight3, gpt_neox_layers_1_post_attention_layernorm_bias3, alloc354) + R.vm.kill_object(gpt_neox_layers_1_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_1_post_attention_layernorm_bias3) + gpt_neox_layers_1_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[22] + storage17: R.Object = R.vm.alloc_storage(R.shape([16777216]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv570: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc355: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv570, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc354, gpt_neox_layers_1_mlp_dense_h_to_4h_weight3, gpt_neox_layers_1_mlp_dense_h_to_4h_bias3, alloc355) + R.vm.kill_object(alloc354) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_1_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[24] + gv571: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc356: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv571, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc355, gpt_neox_layers_1_mlp_dense_4h_to_h_weight3, gpt_neox_layers_1_mlp_dense_4h_to_h_bias3, alloc356) + R.vm.kill_object(alloc355) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_1_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[20] + gv572: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc357: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv572, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape199, gpt_neox_layers_1_attention_dense_weight3, gpt_neox_layers_1_attention_dense_bias3, alloc356, alloc350, alloc357) + R.vm.kill_object(alloc350) + R.vm.kill_object(reshape199) + R.vm.kill_object(alloc356) + R.vm.kill_object(gpt_neox_layers_1_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_1_attention_dense_bias3) + gpt_neox_layers_2_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[26] + gv573: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc358: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv573, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc357, gpt_neox_layers_2_input_layernorm_weight3, gpt_neox_layers_2_input_layernorm_bias3, alloc358) + R.vm.kill_object(gpt_neox_layers_2_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_2_input_layernorm_bias3) + gpt_neox_layers_2_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[30] + gv574: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc359: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv574, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc358, gpt_neox_layers_2_attention_query_key_value_weight3, gpt_neox_layers_2_attention_query_key_value_bias3, alloc359) + R.vm.kill_object(alloc358) + R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_bias3) + gv575: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape200: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc359, gv575, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc359) + gv576: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape201: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape200, gv576, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape200) + gv577: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc360: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv577, R.dtype("float16")) + _359: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), reshape201, alloc360) + R.vm.kill_object(reshape201) + gv578: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape202: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc360, gv578, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc360) + gv579: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape203: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape202, gv579, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape202) + gpt_neox_layers_2_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[28] + gv580: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc361: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv580, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc357, gpt_neox_layers_2_post_attention_layernorm_weight3, gpt_neox_layers_2_post_attention_layernorm_bias3, alloc361) + R.vm.kill_object(gpt_neox_layers_2_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_2_post_attention_layernorm_bias3) + gpt_neox_layers_2_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[34] + gv581: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc362: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv581, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc361, gpt_neox_layers_2_mlp_dense_h_to_4h_weight3, gpt_neox_layers_2_mlp_dense_h_to_4h_bias3, alloc362) + R.vm.kill_object(alloc361) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_2_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[36] + gv582: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc363: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv582, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc362, gpt_neox_layers_2_mlp_dense_4h_to_h_weight3, gpt_neox_layers_2_mlp_dense_4h_to_h_bias3, alloc363) + R.vm.kill_object(alloc362) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_2_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[32] + gv583: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc364: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv583, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape203, gpt_neox_layers_2_attention_dense_weight3, gpt_neox_layers_2_attention_dense_bias3, alloc363, alloc357, alloc364) + R.vm.kill_object(alloc357) + R.vm.kill_object(reshape203) + R.vm.kill_object(alloc363) + R.vm.kill_object(gpt_neox_layers_2_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_2_attention_dense_bias3) + gpt_neox_layers_3_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[38] + gv584: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc365: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv584, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc364, gpt_neox_layers_3_input_layernorm_weight3, gpt_neox_layers_3_input_layernorm_bias3, alloc365) + R.vm.kill_object(gpt_neox_layers_3_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_3_input_layernorm_bias3) + gpt_neox_layers_3_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[42] + gv585: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc366: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv585, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc365, gpt_neox_layers_3_attention_query_key_value_weight3, gpt_neox_layers_3_attention_query_key_value_bias3, alloc366) + R.vm.kill_object(alloc365) + R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_bias3) + gv586: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape204: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc366, gv586, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc366) + gv587: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape205: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape204, gv587, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape204) + gv588: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc367: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv588, R.dtype("float16")) + _366: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), reshape205, alloc367) + R.vm.kill_object(reshape205) + gv589: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape206: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc367, gv589, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc367) + gv590: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape207: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape206, gv590, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape206) + gpt_neox_layers_3_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[40] + gv591: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc368: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv591, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc364, gpt_neox_layers_3_post_attention_layernorm_weight3, gpt_neox_layers_3_post_attention_layernorm_bias3, alloc368) + R.vm.kill_object(gpt_neox_layers_3_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_3_post_attention_layernorm_bias3) + gpt_neox_layers_3_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[46] + gv592: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc369: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv592, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc368, gpt_neox_layers_3_mlp_dense_h_to_4h_weight3, gpt_neox_layers_3_mlp_dense_h_to_4h_bias3, alloc369) + R.vm.kill_object(alloc368) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_3_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[48] + gv593: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc370: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv593, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc369, gpt_neox_layers_3_mlp_dense_4h_to_h_weight3, gpt_neox_layers_3_mlp_dense_4h_to_h_bias3, alloc370) + R.vm.kill_object(alloc369) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_3_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[44] + gv594: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc371: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv594, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape207, gpt_neox_layers_3_attention_dense_weight3, gpt_neox_layers_3_attention_dense_bias3, alloc370, alloc364, alloc371) + R.vm.kill_object(alloc364) + R.vm.kill_object(reshape207) + R.vm.kill_object(alloc370) + R.vm.kill_object(gpt_neox_layers_3_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_3_attention_dense_bias3) + gpt_neox_layers_4_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[50] + gv595: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc372: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv595, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc371, gpt_neox_layers_4_input_layernorm_weight3, gpt_neox_layers_4_input_layernorm_bias3, alloc372) + R.vm.kill_object(gpt_neox_layers_4_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_4_input_layernorm_bias3) + gpt_neox_layers_4_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[54] + gv596: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc373: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv596, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc372, gpt_neox_layers_4_attention_query_key_value_weight3, gpt_neox_layers_4_attention_query_key_value_bias3, alloc373) + R.vm.kill_object(alloc372) + R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_bias3) + gv597: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape208: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc373, gv597, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc373) + gv598: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape209: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape208, gv598, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape208) + gv599: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc374: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv599, R.dtype("float16")) + _373: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), reshape209, alloc374) + R.vm.kill_object(reshape209) + gv600: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape210: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc374, gv600, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc374) + gv601: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape211: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape210, gv601, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape210) + gpt_neox_layers_4_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[52] + gv602: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc375: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv602, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc371, gpt_neox_layers_4_post_attention_layernorm_weight3, gpt_neox_layers_4_post_attention_layernorm_bias3, alloc375) + R.vm.kill_object(gpt_neox_layers_4_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_4_post_attention_layernorm_bias3) + gpt_neox_layers_4_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[58] + gv603: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc376: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv603, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc375, gpt_neox_layers_4_mlp_dense_h_to_4h_weight3, gpt_neox_layers_4_mlp_dense_h_to_4h_bias3, alloc376) + R.vm.kill_object(alloc375) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_4_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[60] + gv604: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc377: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv604, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc376, gpt_neox_layers_4_mlp_dense_4h_to_h_weight3, gpt_neox_layers_4_mlp_dense_4h_to_h_bias3, alloc377) + R.vm.kill_object(alloc376) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_4_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[56] + gv605: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc378: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv605, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape211, gpt_neox_layers_4_attention_dense_weight3, gpt_neox_layers_4_attention_dense_bias3, alloc377, alloc371, alloc378) + R.vm.kill_object(alloc371) + R.vm.kill_object(reshape211) + R.vm.kill_object(alloc377) + R.vm.kill_object(gpt_neox_layers_4_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_4_attention_dense_bias3) + gpt_neox_layers_5_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[62] + gv606: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc379: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv606, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc378, gpt_neox_layers_5_input_layernorm_weight3, gpt_neox_layers_5_input_layernorm_bias3, alloc379) + R.vm.kill_object(gpt_neox_layers_5_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_5_input_layernorm_bias3) + gpt_neox_layers_5_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[66] + gv607: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc380: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv607, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc379, gpt_neox_layers_5_attention_query_key_value_weight3, gpt_neox_layers_5_attention_query_key_value_bias3, alloc380) + R.vm.kill_object(alloc379) + R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_bias3) + gv608: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape212: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc380, gv608, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc380) + gv609: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape213: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape212, gv609, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape212) + gv610: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc381: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv610, R.dtype("float16")) + _380: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), reshape213, alloc381) + R.vm.kill_object(reshape213) + gv611: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape214: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc381, gv611, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc381) + gv612: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape215: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape214, gv612, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape214) + gpt_neox_layers_5_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[64] + gv613: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc382: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv613, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc378, gpt_neox_layers_5_post_attention_layernorm_weight3, gpt_neox_layers_5_post_attention_layernorm_bias3, alloc382) + R.vm.kill_object(gpt_neox_layers_5_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_5_post_attention_layernorm_bias3) + gpt_neox_layers_5_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[70] + gv614: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc383: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv614, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc382, gpt_neox_layers_5_mlp_dense_h_to_4h_weight3, gpt_neox_layers_5_mlp_dense_h_to_4h_bias3, alloc383) + R.vm.kill_object(alloc382) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_5_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[72] + gv615: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc384: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv615, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc383, gpt_neox_layers_5_mlp_dense_4h_to_h_weight3, gpt_neox_layers_5_mlp_dense_4h_to_h_bias3, alloc384) + R.vm.kill_object(alloc383) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_5_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[68] + gv616: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc385: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv616, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape215, gpt_neox_layers_5_attention_dense_weight3, gpt_neox_layers_5_attention_dense_bias3, alloc384, alloc378, alloc385) + R.vm.kill_object(alloc378) + R.vm.kill_object(reshape215) + R.vm.kill_object(alloc384) + R.vm.kill_object(gpt_neox_layers_5_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_5_attention_dense_bias3) + gpt_neox_layers_6_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[74] + gv617: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc386: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv617, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc385, gpt_neox_layers_6_input_layernorm_weight3, gpt_neox_layers_6_input_layernorm_bias3, alloc386) + R.vm.kill_object(gpt_neox_layers_6_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_6_input_layernorm_bias3) + gpt_neox_layers_6_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[78] + gv618: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc387: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv618, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc386, gpt_neox_layers_6_attention_query_key_value_weight3, gpt_neox_layers_6_attention_query_key_value_bias3, alloc387) + R.vm.kill_object(alloc386) + R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_bias3) + gv619: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape216: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc387, gv619, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc387) + gv620: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape217: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape216, gv620, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape216) + gv621: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc388: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv621, R.dtype("float16")) + _387: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), reshape217, alloc388) + R.vm.kill_object(reshape217) + gv622: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape218: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc388, gv622, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc388) + gv623: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape219: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape218, gv623, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape218) + gpt_neox_layers_6_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[76] + gv624: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc389: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv624, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc385, gpt_neox_layers_6_post_attention_layernorm_weight3, gpt_neox_layers_6_post_attention_layernorm_bias3, alloc389) + R.vm.kill_object(gpt_neox_layers_6_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_6_post_attention_layernorm_bias3) + gpt_neox_layers_6_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[82] + gv625: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc390: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv625, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc389, gpt_neox_layers_6_mlp_dense_h_to_4h_weight3, gpt_neox_layers_6_mlp_dense_h_to_4h_bias3, alloc390) + R.vm.kill_object(alloc389) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_6_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[84] + gv626: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc391: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv626, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc390, gpt_neox_layers_6_mlp_dense_4h_to_h_weight3, gpt_neox_layers_6_mlp_dense_4h_to_h_bias3, alloc391) + R.vm.kill_object(alloc390) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_6_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[80] + gv627: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc392: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv627, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape219, gpt_neox_layers_6_attention_dense_weight3, gpt_neox_layers_6_attention_dense_bias3, alloc391, alloc385, alloc392) + R.vm.kill_object(alloc385) + R.vm.kill_object(reshape219) + R.vm.kill_object(alloc391) + R.vm.kill_object(gpt_neox_layers_6_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_6_attention_dense_bias3) + gpt_neox_layers_7_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[86] + gv628: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc393: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv628, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc392, gpt_neox_layers_7_input_layernorm_weight3, gpt_neox_layers_7_input_layernorm_bias3, alloc393) + R.vm.kill_object(gpt_neox_layers_7_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_7_input_layernorm_bias3) + gpt_neox_layers_7_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[90] + gv629: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc394: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv629, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc393, gpt_neox_layers_7_attention_query_key_value_weight3, gpt_neox_layers_7_attention_query_key_value_bias3, alloc394) + R.vm.kill_object(alloc393) + R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_bias3) + gv630: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape220: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc394, gv630, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc394) + gv631: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape221: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape220, gv631, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape220) + gv632: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc395: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv632, R.dtype("float16")) + _394: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), reshape221, alloc395) + R.vm.kill_object(reshape221) + gv633: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape222: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc395, gv633, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc395) + gv634: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape223: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape222, gv634, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape222) + gpt_neox_layers_7_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[88] + gv635: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc396: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv635, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc392, gpt_neox_layers_7_post_attention_layernorm_weight3, gpt_neox_layers_7_post_attention_layernorm_bias3, alloc396) + R.vm.kill_object(gpt_neox_layers_7_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_7_post_attention_layernorm_bias3) + gpt_neox_layers_7_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[94] + gv636: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc397: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv636, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc396, gpt_neox_layers_7_mlp_dense_h_to_4h_weight3, gpt_neox_layers_7_mlp_dense_h_to_4h_bias3, alloc397) + R.vm.kill_object(alloc396) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_7_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[96] + gv637: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc398: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv637, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc397, gpt_neox_layers_7_mlp_dense_4h_to_h_weight3, gpt_neox_layers_7_mlp_dense_4h_to_h_bias3, alloc398) + R.vm.kill_object(alloc397) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_7_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[92] + gv638: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc399: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv638, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape223, gpt_neox_layers_7_attention_dense_weight3, gpt_neox_layers_7_attention_dense_bias3, alloc398, alloc392, alloc399) + R.vm.kill_object(alloc392) + R.vm.kill_object(reshape223) + R.vm.kill_object(alloc398) + R.vm.kill_object(gpt_neox_layers_7_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_7_attention_dense_bias3) + gpt_neox_layers_8_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[98] + gv639: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc400: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv639, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc399, gpt_neox_layers_8_input_layernorm_weight3, gpt_neox_layers_8_input_layernorm_bias3, alloc400) + R.vm.kill_object(gpt_neox_layers_8_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_8_input_layernorm_bias3) + gpt_neox_layers_8_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[102] + gv640: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc401: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv640, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc400, gpt_neox_layers_8_attention_query_key_value_weight3, gpt_neox_layers_8_attention_query_key_value_bias3, alloc401) + R.vm.kill_object(alloc400) + R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_bias3) + gv641: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape224: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc401, gv641, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc401) + gv642: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape225: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape224, gv642, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape224) + gv643: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc402: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv643, R.dtype("float16")) + _401: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), reshape225, alloc402) + R.vm.kill_object(reshape225) + gv644: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape226: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc402, gv644, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc402) + gv645: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape227: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape226, gv645, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape226) + gpt_neox_layers_8_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[100] + gv646: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc403: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv646, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc399, gpt_neox_layers_8_post_attention_layernorm_weight3, gpt_neox_layers_8_post_attention_layernorm_bias3, alloc403) + R.vm.kill_object(gpt_neox_layers_8_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_8_post_attention_layernorm_bias3) + gpt_neox_layers_8_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[106] + gv647: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc404: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv647, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc403, gpt_neox_layers_8_mlp_dense_h_to_4h_weight3, gpt_neox_layers_8_mlp_dense_h_to_4h_bias3, alloc404) + R.vm.kill_object(alloc403) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_8_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[108] + gv648: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc405: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv648, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc404, gpt_neox_layers_8_mlp_dense_4h_to_h_weight3, gpt_neox_layers_8_mlp_dense_4h_to_h_bias3, alloc405) + R.vm.kill_object(alloc404) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_8_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[104] + gv649: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc406: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv649, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape227, gpt_neox_layers_8_attention_dense_weight3, gpt_neox_layers_8_attention_dense_bias3, alloc405, alloc399, alloc406) + R.vm.kill_object(alloc399) + R.vm.kill_object(reshape227) + R.vm.kill_object(alloc405) + R.vm.kill_object(gpt_neox_layers_8_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_8_attention_dense_bias3) + gpt_neox_layers_9_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[110] + gv650: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc407: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv650, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc406, gpt_neox_layers_9_input_layernorm_weight3, gpt_neox_layers_9_input_layernorm_bias3, alloc407) + R.vm.kill_object(gpt_neox_layers_9_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_9_input_layernorm_bias3) + gpt_neox_layers_9_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[114] + gv651: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc408: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv651, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc407, gpt_neox_layers_9_attention_query_key_value_weight3, gpt_neox_layers_9_attention_query_key_value_bias3, alloc408) + R.vm.kill_object(alloc407) + R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_bias3) + gv652: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape228: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc408, gv652, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc408) + gv653: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape229: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape228, gv653, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape228) + gv654: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc409: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv654, R.dtype("float16")) + _408: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), reshape229, alloc409) + R.vm.kill_object(reshape229) + gv655: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape230: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc409, gv655, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc409) + gv656: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape231: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape230, gv656, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape230) + gpt_neox_layers_9_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[112] + gv657: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc410: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv657, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc406, gpt_neox_layers_9_post_attention_layernorm_weight3, gpt_neox_layers_9_post_attention_layernorm_bias3, alloc410) + R.vm.kill_object(gpt_neox_layers_9_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_9_post_attention_layernorm_bias3) + gpt_neox_layers_9_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[118] + gv658: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc411: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv658, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc410, gpt_neox_layers_9_mlp_dense_h_to_4h_weight3, gpt_neox_layers_9_mlp_dense_h_to_4h_bias3, alloc411) + R.vm.kill_object(alloc410) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_9_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[120] + gv659: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc412: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv659, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc411, gpt_neox_layers_9_mlp_dense_4h_to_h_weight3, gpt_neox_layers_9_mlp_dense_4h_to_h_bias3, alloc412) + R.vm.kill_object(alloc411) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_9_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[116] + gv660: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc413: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv660, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape231, gpt_neox_layers_9_attention_dense_weight3, gpt_neox_layers_9_attention_dense_bias3, alloc412, alloc406, alloc413) + R.vm.kill_object(alloc406) + R.vm.kill_object(reshape231) + R.vm.kill_object(alloc412) + R.vm.kill_object(gpt_neox_layers_9_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_9_attention_dense_bias3) + gpt_neox_layers_10_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[122] + gv661: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc414: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv661, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc413, gpt_neox_layers_10_input_layernorm_weight3, gpt_neox_layers_10_input_layernorm_bias3, alloc414) + R.vm.kill_object(gpt_neox_layers_10_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_10_input_layernorm_bias3) + gpt_neox_layers_10_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[126] + gv662: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc415: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv662, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc414, gpt_neox_layers_10_attention_query_key_value_weight3, gpt_neox_layers_10_attention_query_key_value_bias3, alloc415) + R.vm.kill_object(alloc414) + R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_bias3) + gv663: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape232: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc415, gv663, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc415) + gv664: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape233: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape232, gv664, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape232) + gv665: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc416: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv665, R.dtype("float16")) + _415: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), reshape233, alloc416) + R.vm.kill_object(reshape233) + gv666: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape234: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc416, gv666, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc416) + gv667: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape235: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape234, gv667, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape234) + gpt_neox_layers_10_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[124] + gv668: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc417: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv668, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc413, gpt_neox_layers_10_post_attention_layernorm_weight3, gpt_neox_layers_10_post_attention_layernorm_bias3, alloc417) + R.vm.kill_object(gpt_neox_layers_10_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_10_post_attention_layernorm_bias3) + gpt_neox_layers_10_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[130] + gv669: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc418: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv669, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc417, gpt_neox_layers_10_mlp_dense_h_to_4h_weight3, gpt_neox_layers_10_mlp_dense_h_to_4h_bias3, alloc418) + R.vm.kill_object(alloc417) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_10_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[132] + gv670: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc419: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv670, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc418, gpt_neox_layers_10_mlp_dense_4h_to_h_weight3, gpt_neox_layers_10_mlp_dense_4h_to_h_bias3, alloc419) + R.vm.kill_object(alloc418) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_10_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[128] + gv671: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc420: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv671, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape235, gpt_neox_layers_10_attention_dense_weight3, gpt_neox_layers_10_attention_dense_bias3, alloc419, alloc413, alloc420) + R.vm.kill_object(alloc413) + R.vm.kill_object(reshape235) + R.vm.kill_object(alloc419) + R.vm.kill_object(gpt_neox_layers_10_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_10_attention_dense_bias3) + gpt_neox_layers_11_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[134] + gv672: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc421: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv672, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc420, gpt_neox_layers_11_input_layernorm_weight3, gpt_neox_layers_11_input_layernorm_bias3, alloc421) + R.vm.kill_object(gpt_neox_layers_11_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_11_input_layernorm_bias3) + gpt_neox_layers_11_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[138] + gv673: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc422: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv673, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc421, gpt_neox_layers_11_attention_query_key_value_weight3, gpt_neox_layers_11_attention_query_key_value_bias3, alloc422) + R.vm.kill_object(alloc421) + R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_bias3) + gv674: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape236: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc422, gv674, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc422) + gv675: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape237: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape236, gv675, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape236) + gv676: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc423: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv676, R.dtype("float16")) + _422: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), reshape237, alloc423) + R.vm.kill_object(reshape237) + gv677: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape238: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc423, gv677, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc423) + gv678: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape239: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape238, gv678, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape238) + gpt_neox_layers_11_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[136] + gv679: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc424: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv679, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc420, gpt_neox_layers_11_post_attention_layernorm_weight3, gpt_neox_layers_11_post_attention_layernorm_bias3, alloc424) + R.vm.kill_object(gpt_neox_layers_11_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_11_post_attention_layernorm_bias3) + gpt_neox_layers_11_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[142] + gv680: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc425: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv680, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc424, gpt_neox_layers_11_mlp_dense_h_to_4h_weight3, gpt_neox_layers_11_mlp_dense_h_to_4h_bias3, alloc425) + R.vm.kill_object(alloc424) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_11_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[144] + gv681: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc426: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv681, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc425, gpt_neox_layers_11_mlp_dense_4h_to_h_weight3, gpt_neox_layers_11_mlp_dense_4h_to_h_bias3, alloc426) + R.vm.kill_object(alloc425) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_11_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[140] + gv682: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc427: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv682, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape239, gpt_neox_layers_11_attention_dense_weight3, gpt_neox_layers_11_attention_dense_bias3, alloc426, alloc420, alloc427) + R.vm.kill_object(alloc420) + R.vm.kill_object(reshape239) + R.vm.kill_object(alloc426) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias3) + gpt_neox_layers_12_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[145] + gpt_neox_layers_12_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[146] + gv683: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc428: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv683, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc427, gpt_neox_layers_12_input_layernorm_weight3, gpt_neox_layers_12_input_layernorm_bias3, alloc428) + R.vm.kill_object(gpt_neox_layers_12_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_12_input_layernorm_bias3) + gpt_neox_layers_12_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[149] + gpt_neox_layers_12_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[150] + gv684: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc429: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv684, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc428, gpt_neox_layers_12_attention_query_key_value_weight3, gpt_neox_layers_12_attention_query_key_value_bias3, alloc429) + R.vm.kill_object(alloc428) + R.vm.kill_object(gpt_neox_layers_12_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_12_attention_query_key_value_bias3) + gv685: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape240: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc429, gv685, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc429) + gv686: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape241: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape240, gv686, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape240) + gv687: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc430: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv687, R.dtype("float16")) + _429: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape241, alloc430) + R.vm.kill_object(reshape241) + gv688: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape242: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc430, gv688, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc430) + gv689: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape243: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape242, gv689, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape242) + gpt_neox_layers_12_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[147] + gpt_neox_layers_12_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[148] + gv690: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc431: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv690, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc427, gpt_neox_layers_12_post_attention_layernorm_weight3, gpt_neox_layers_12_post_attention_layernorm_bias3, alloc431) + R.vm.kill_object(gpt_neox_layers_12_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_12_post_attention_layernorm_bias3) + gpt_neox_layers_12_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[153] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[154] + gv691: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc432: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv691, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc431, gpt_neox_layers_12_mlp_dense_h_to_4h_weight3, gpt_neox_layers_12_mlp_dense_h_to_4h_bias3, alloc432) + R.vm.kill_object(alloc431) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_12_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[155] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[156] + gv692: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc433: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv692, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc432, gpt_neox_layers_12_mlp_dense_4h_to_h_weight3, gpt_neox_layers_12_mlp_dense_4h_to_h_bias3, alloc433) + R.vm.kill_object(alloc432) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_12_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[151] + gpt_neox_layers_12_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[152] + gv693: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc434: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv693, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape243, gpt_neox_layers_12_attention_dense_weight3, gpt_neox_layers_12_attention_dense_bias3, alloc433, alloc427, alloc434) + R.vm.kill_object(alloc427) + R.vm.kill_object(reshape243) + R.vm.kill_object(alloc433) + R.vm.kill_object(gpt_neox_layers_12_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_12_attention_dense_bias3) + gpt_neox_layers_13_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[157] + gpt_neox_layers_13_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[158] + gv694: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc435: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv694, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc434, gpt_neox_layers_13_input_layernorm_weight3, gpt_neox_layers_13_input_layernorm_bias3, alloc435) + R.vm.kill_object(gpt_neox_layers_13_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_13_input_layernorm_bias3) + gpt_neox_layers_13_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[161] + gpt_neox_layers_13_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[162] + gv695: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc436: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv695, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc435, gpt_neox_layers_13_attention_query_key_value_weight3, gpt_neox_layers_13_attention_query_key_value_bias3, alloc436) + R.vm.kill_object(alloc435) + R.vm.kill_object(gpt_neox_layers_13_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_13_attention_query_key_value_bias3) + gv696: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape244: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc436, gv696, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc436) + gv697: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape245: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape244, gv697, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape244) + gv698: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc437: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv698, R.dtype("float16")) + _436: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape245, alloc437) + R.vm.kill_object(reshape245) + gv699: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape246: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc437, gv699, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc437) + gv700: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape247: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape246, gv700, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape246) + gpt_neox_layers_13_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[159] + gpt_neox_layers_13_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[160] + gv701: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc438: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv701, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc434, gpt_neox_layers_13_post_attention_layernorm_weight3, gpt_neox_layers_13_post_attention_layernorm_bias3, alloc438) + R.vm.kill_object(gpt_neox_layers_13_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_13_post_attention_layernorm_bias3) + gpt_neox_layers_13_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[165] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[166] + gv702: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc439: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv702, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc438, gpt_neox_layers_13_mlp_dense_h_to_4h_weight3, gpt_neox_layers_13_mlp_dense_h_to_4h_bias3, alloc439) + R.vm.kill_object(alloc438) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_13_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[167] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[168] + gv703: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc440: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv703, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc439, gpt_neox_layers_13_mlp_dense_4h_to_h_weight3, gpt_neox_layers_13_mlp_dense_4h_to_h_bias3, alloc440) + R.vm.kill_object(alloc439) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_13_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[163] + gpt_neox_layers_13_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[164] + gv704: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc441: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv704, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape247, gpt_neox_layers_13_attention_dense_weight3, gpt_neox_layers_13_attention_dense_bias3, alloc440, alloc434, alloc441) + R.vm.kill_object(alloc434) + R.vm.kill_object(reshape247) + R.vm.kill_object(alloc440) + R.vm.kill_object(gpt_neox_layers_13_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_13_attention_dense_bias3) + gpt_neox_layers_14_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[169] + gpt_neox_layers_14_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[170] + gv705: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc442: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv705, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc441, gpt_neox_layers_14_input_layernorm_weight3, gpt_neox_layers_14_input_layernorm_bias3, alloc442) + R.vm.kill_object(gpt_neox_layers_14_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_14_input_layernorm_bias3) + gpt_neox_layers_14_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[173] + gpt_neox_layers_14_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[174] + gv706: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc443: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv706, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc442, gpt_neox_layers_14_attention_query_key_value_weight3, gpt_neox_layers_14_attention_query_key_value_bias3, alloc443) + R.vm.kill_object(alloc442) + R.vm.kill_object(gpt_neox_layers_14_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_14_attention_query_key_value_bias3) + gv707: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape248: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc443, gv707, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc443) + gv708: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape249: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape248, gv708, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape248) + gv709: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc444: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv709, R.dtype("float16")) + _443: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape249, alloc444) + R.vm.kill_object(reshape249) + gv710: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape250: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc444, gv710, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc444) + gv711: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape251: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape250, gv711, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape250) + gpt_neox_layers_14_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[171] + gpt_neox_layers_14_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[172] + gv712: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc445: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv712, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc441, gpt_neox_layers_14_post_attention_layernorm_weight3, gpt_neox_layers_14_post_attention_layernorm_bias3, alloc445) + R.vm.kill_object(gpt_neox_layers_14_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_14_post_attention_layernorm_bias3) + gpt_neox_layers_14_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[177] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[178] + gv713: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc446: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv713, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc445, gpt_neox_layers_14_mlp_dense_h_to_4h_weight3, gpt_neox_layers_14_mlp_dense_h_to_4h_bias3, alloc446) + R.vm.kill_object(alloc445) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_14_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[179] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[180] + gv714: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc447: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv714, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc446, gpt_neox_layers_14_mlp_dense_4h_to_h_weight3, gpt_neox_layers_14_mlp_dense_4h_to_h_bias3, alloc447) + R.vm.kill_object(alloc446) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_14_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[175] + gpt_neox_layers_14_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[176] + gv715: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc448: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv715, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape251, gpt_neox_layers_14_attention_dense_weight3, gpt_neox_layers_14_attention_dense_bias3, alloc447, alloc441, alloc448) + R.vm.kill_object(alloc441) + R.vm.kill_object(reshape251) + R.vm.kill_object(alloc447) + R.vm.kill_object(gpt_neox_layers_14_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_14_attention_dense_bias3) + gpt_neox_layers_15_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[181] + gpt_neox_layers_15_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[182] + gv716: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc449: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv716, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc448, gpt_neox_layers_15_input_layernorm_weight3, gpt_neox_layers_15_input_layernorm_bias3, alloc449) + R.vm.kill_object(gpt_neox_layers_15_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_15_input_layernorm_bias3) + gpt_neox_layers_15_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[185] + gpt_neox_layers_15_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[186] + gv717: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc450: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv717, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc449, gpt_neox_layers_15_attention_query_key_value_weight3, gpt_neox_layers_15_attention_query_key_value_bias3, alloc450) + R.vm.kill_object(alloc449) + R.vm.kill_object(gpt_neox_layers_15_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_15_attention_query_key_value_bias3) + gv718: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape252: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc450, gv718, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc450) + gv719: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape253: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape252, gv719, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape252) + gv720: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc451: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv720, R.dtype("float16")) + _450: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape253, alloc451) + R.vm.kill_object(reshape253) + gv721: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape254: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc451, gv721, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc451) + gv722: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape255: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape254, gv722, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape254) + gpt_neox_layers_15_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[183] + gpt_neox_layers_15_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[184] + gv723: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc452: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv723, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc448, gpt_neox_layers_15_post_attention_layernorm_weight3, gpt_neox_layers_15_post_attention_layernorm_bias3, alloc452) + R.vm.kill_object(gpt_neox_layers_15_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_15_post_attention_layernorm_bias3) + gpt_neox_layers_15_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[189] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[190] + gv724: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc453: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv724, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc452, gpt_neox_layers_15_mlp_dense_h_to_4h_weight3, gpt_neox_layers_15_mlp_dense_h_to_4h_bias3, alloc453) + R.vm.kill_object(alloc452) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_15_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[191] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[192] + gv725: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc454: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv725, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc453, gpt_neox_layers_15_mlp_dense_4h_to_h_weight3, gpt_neox_layers_15_mlp_dense_4h_to_h_bias3, alloc454) + R.vm.kill_object(alloc453) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_15_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[187] + gpt_neox_layers_15_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[188] + gv726: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc455: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv726, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape255, gpt_neox_layers_15_attention_dense_weight3, gpt_neox_layers_15_attention_dense_bias3, alloc454, alloc448, alloc455) + R.vm.kill_object(alloc448) + R.vm.kill_object(reshape255) + R.vm.kill_object(alloc454) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_bias3) + gpt_neox_layers_16_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[193] + gpt_neox_layers_16_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[194] + gv727: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc456: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv727, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc455, gpt_neox_layers_16_input_layernorm_weight3, gpt_neox_layers_16_input_layernorm_bias3, alloc456) + R.vm.kill_object(gpt_neox_layers_16_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_16_input_layernorm_bias3) + gpt_neox_layers_16_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[197] + gpt_neox_layers_16_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[198] + gv728: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc457: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv728, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc456, gpt_neox_layers_16_attention_query_key_value_weight3, gpt_neox_layers_16_attention_query_key_value_bias3, alloc457) + R.vm.kill_object(alloc456) + R.vm.kill_object(gpt_neox_layers_16_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_16_attention_query_key_value_bias3) + gv729: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape256: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc457, gv729, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc457) + gv730: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape257: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape256, gv730, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape256) + gv731: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc458: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv731, R.dtype("float16")) + _457: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), reshape257, alloc458) + R.vm.kill_object(reshape257) + gv732: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape258: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc458, gv732, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc458) + gv733: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape259: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape258, gv733, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape258) + gpt_neox_layers_16_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[195] + gpt_neox_layers_16_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[196] + gv734: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc459: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv734, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc455, gpt_neox_layers_16_post_attention_layernorm_weight3, gpt_neox_layers_16_post_attention_layernorm_bias3, alloc459) + R.vm.kill_object(gpt_neox_layers_16_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_16_post_attention_layernorm_bias3) + gpt_neox_layers_16_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[201] + gpt_neox_layers_16_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[202] + gv735: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc460: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv735, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc459, gpt_neox_layers_16_mlp_dense_h_to_4h_weight3, gpt_neox_layers_16_mlp_dense_h_to_4h_bias3, alloc460) + R.vm.kill_object(alloc459) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_16_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[203] + gpt_neox_layers_16_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[204] + gv736: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc461: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv736, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc460, gpt_neox_layers_16_mlp_dense_4h_to_h_weight3, gpt_neox_layers_16_mlp_dense_4h_to_h_bias3, alloc461) + R.vm.kill_object(alloc460) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_16_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[199] + gpt_neox_layers_16_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[200] + gv737: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc462: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv737, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape259, gpt_neox_layers_16_attention_dense_weight3, gpt_neox_layers_16_attention_dense_bias3, alloc461, alloc455, alloc462) + R.vm.kill_object(alloc455) + R.vm.kill_object(reshape259) + R.vm.kill_object(alloc461) + R.vm.kill_object(gpt_neox_layers_16_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_16_attention_dense_bias3) + gpt_neox_layers_17_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[205] + gpt_neox_layers_17_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[206] + gv738: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc463: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv738, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc462, gpt_neox_layers_17_input_layernorm_weight3, gpt_neox_layers_17_input_layernorm_bias3, alloc463) + R.vm.kill_object(gpt_neox_layers_17_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_17_input_layernorm_bias3) + gpt_neox_layers_17_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[209] + gpt_neox_layers_17_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[210] + gv739: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc464: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv739, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc463, gpt_neox_layers_17_attention_query_key_value_weight3, gpt_neox_layers_17_attention_query_key_value_bias3, alloc464) + R.vm.kill_object(alloc463) + R.vm.kill_object(gpt_neox_layers_17_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_17_attention_query_key_value_bias3) + gv740: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape260: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc464, gv740, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc464) + gv741: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape261: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape260, gv741, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape260) + gv742: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc465: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv742, R.dtype("float16")) + _464: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), reshape261, alloc465) + R.vm.kill_object(reshape261) + gv743: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape262: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc465, gv743, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc465) + gv744: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape263: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape262, gv744, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape262) + gpt_neox_layers_17_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[207] + gpt_neox_layers_17_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[208] + gv745: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc466: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv745, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc462, gpt_neox_layers_17_post_attention_layernorm_weight3, gpt_neox_layers_17_post_attention_layernorm_bias3, alloc466) + R.vm.kill_object(gpt_neox_layers_17_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_17_post_attention_layernorm_bias3) + gpt_neox_layers_17_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[213] + gpt_neox_layers_17_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[214] + gv746: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc467: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv746, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc466, gpt_neox_layers_17_mlp_dense_h_to_4h_weight3, gpt_neox_layers_17_mlp_dense_h_to_4h_bias3, alloc467) + R.vm.kill_object(alloc466) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_17_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[215] + gpt_neox_layers_17_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[216] + gv747: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc468: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv747, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc467, gpt_neox_layers_17_mlp_dense_4h_to_h_weight3, gpt_neox_layers_17_mlp_dense_4h_to_h_bias3, alloc468) + R.vm.kill_object(alloc467) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_17_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[211] + gpt_neox_layers_17_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[212] + gv748: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc469: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv748, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape263, gpt_neox_layers_17_attention_dense_weight3, gpt_neox_layers_17_attention_dense_bias3, alloc468, alloc462, alloc469) + R.vm.kill_object(alloc462) + R.vm.kill_object(reshape263) + R.vm.kill_object(alloc468) + R.vm.kill_object(gpt_neox_layers_17_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_17_attention_dense_bias3) + gpt_neox_layers_18_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[217] + gpt_neox_layers_18_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[218] + gv749: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc470: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv749, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc469, gpt_neox_layers_18_input_layernorm_weight3, gpt_neox_layers_18_input_layernorm_bias3, alloc470) + R.vm.kill_object(gpt_neox_layers_18_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_18_input_layernorm_bias3) + gpt_neox_layers_18_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[221] + gpt_neox_layers_18_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[222] + gv750: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc471: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv750, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc470, gpt_neox_layers_18_attention_query_key_value_weight3, gpt_neox_layers_18_attention_query_key_value_bias3, alloc471) + R.vm.kill_object(alloc470) + R.vm.kill_object(gpt_neox_layers_18_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_18_attention_query_key_value_bias3) + gv751: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape264: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc471, gv751, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc471) + gv752: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape265: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape264, gv752, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape264) + gv753: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc472: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv753, R.dtype("float16")) + _471: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), reshape265, alloc472) + R.vm.kill_object(reshape265) + gv754: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape266: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc472, gv754, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc472) + gv755: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape267: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape266, gv755, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape266) + gpt_neox_layers_18_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[219] + gpt_neox_layers_18_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[220] + gv756: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc473: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv756, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc469, gpt_neox_layers_18_post_attention_layernorm_weight3, gpt_neox_layers_18_post_attention_layernorm_bias3, alloc473) + R.vm.kill_object(gpt_neox_layers_18_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_18_post_attention_layernorm_bias3) + gpt_neox_layers_18_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[225] + gpt_neox_layers_18_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[226] + gv757: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc474: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv757, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc473, gpt_neox_layers_18_mlp_dense_h_to_4h_weight3, gpt_neox_layers_18_mlp_dense_h_to_4h_bias3, alloc474) + R.vm.kill_object(alloc473) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_18_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[227] + gpt_neox_layers_18_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[228] + gv758: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc475: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv758, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc474, gpt_neox_layers_18_mlp_dense_4h_to_h_weight3, gpt_neox_layers_18_mlp_dense_4h_to_h_bias3, alloc475) + R.vm.kill_object(alloc474) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_18_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[223] + gpt_neox_layers_18_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[224] + gv759: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc476: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv759, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape267, gpt_neox_layers_18_attention_dense_weight3, gpt_neox_layers_18_attention_dense_bias3, alloc475, alloc469, alloc476) + R.vm.kill_object(alloc469) + R.vm.kill_object(reshape267) + R.vm.kill_object(alloc475) + R.vm.kill_object(gpt_neox_layers_18_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_18_attention_dense_bias3) + gpt_neox_layers_19_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[229] + gpt_neox_layers_19_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[230] + gv760: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc477: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv760, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc476, gpt_neox_layers_19_input_layernorm_weight3, gpt_neox_layers_19_input_layernorm_bias3, alloc477) + R.vm.kill_object(gpt_neox_layers_19_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_19_input_layernorm_bias3) + gpt_neox_layers_19_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[233] + gpt_neox_layers_19_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[234] + gv761: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc478: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv761, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc477, gpt_neox_layers_19_attention_query_key_value_weight3, gpt_neox_layers_19_attention_query_key_value_bias3, alloc478) + R.vm.kill_object(alloc477) + R.vm.kill_object(gpt_neox_layers_19_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_19_attention_query_key_value_bias3) + gv762: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape268: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc478, gv762, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc478) + gv763: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape269: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape268, gv763, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape268) + gv764: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc479: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv764, R.dtype("float16")) + _478: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), reshape269, alloc479) + R.vm.kill_object(reshape269) + gv765: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape270: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc479, gv765, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc479) + gv766: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape271: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape270, gv766, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape270) + gpt_neox_layers_19_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[231] + gpt_neox_layers_19_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[232] + gv767: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc480: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv767, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc476, gpt_neox_layers_19_post_attention_layernorm_weight3, gpt_neox_layers_19_post_attention_layernorm_bias3, alloc480) + R.vm.kill_object(gpt_neox_layers_19_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_19_post_attention_layernorm_bias3) + gpt_neox_layers_19_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[237] + gpt_neox_layers_19_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[238] + gv768: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc481: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv768, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc480, gpt_neox_layers_19_mlp_dense_h_to_4h_weight3, gpt_neox_layers_19_mlp_dense_h_to_4h_bias3, alloc481) + R.vm.kill_object(alloc480) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_19_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[239] + gpt_neox_layers_19_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[240] + gv769: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc482: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv769, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc481, gpt_neox_layers_19_mlp_dense_4h_to_h_weight3, gpt_neox_layers_19_mlp_dense_4h_to_h_bias3, alloc482) + R.vm.kill_object(alloc481) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_19_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[235] + gpt_neox_layers_19_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[236] + gv770: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc483: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv770, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape271, gpt_neox_layers_19_attention_dense_weight3, gpt_neox_layers_19_attention_dense_bias3, alloc482, alloc476, alloc483) + R.vm.kill_object(alloc476) + R.vm.kill_object(reshape271) + R.vm.kill_object(alloc482) + R.vm.kill_object(gpt_neox_layers_19_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_19_attention_dense_bias3) + gpt_neox_layers_20_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[241] + gpt_neox_layers_20_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[242] + gv771: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc484: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv771, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc483, gpt_neox_layers_20_input_layernorm_weight3, gpt_neox_layers_20_input_layernorm_bias3, alloc484) + R.vm.kill_object(gpt_neox_layers_20_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_20_input_layernorm_bias3) + gpt_neox_layers_20_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[245] + gpt_neox_layers_20_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[246] + gv772: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc485: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv772, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc484, gpt_neox_layers_20_attention_query_key_value_weight3, gpt_neox_layers_20_attention_query_key_value_bias3, alloc485) + R.vm.kill_object(alloc484) + R.vm.kill_object(gpt_neox_layers_20_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_20_attention_query_key_value_bias3) + gv773: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape272: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc485, gv773, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc485) + gv774: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape273: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape272, gv774, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape272) + gv775: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc486: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv775, R.dtype("float16")) + _485: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), reshape273, alloc486) + R.vm.kill_object(reshape273) + gv776: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape274: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc486, gv776, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc486) + gv777: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape275: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape274, gv777, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape274) + gpt_neox_layers_20_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[243] + gpt_neox_layers_20_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[244] + gv778: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc487: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv778, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc483, gpt_neox_layers_20_post_attention_layernorm_weight3, gpt_neox_layers_20_post_attention_layernorm_bias3, alloc487) + R.vm.kill_object(gpt_neox_layers_20_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_20_post_attention_layernorm_bias3) + gpt_neox_layers_20_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[249] + gpt_neox_layers_20_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[250] + gv779: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc488: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv779, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc487, gpt_neox_layers_20_mlp_dense_h_to_4h_weight3, gpt_neox_layers_20_mlp_dense_h_to_4h_bias3, alloc488) + R.vm.kill_object(alloc487) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_20_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[251] + gpt_neox_layers_20_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[252] + gv780: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc489: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv780, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc488, gpt_neox_layers_20_mlp_dense_4h_to_h_weight3, gpt_neox_layers_20_mlp_dense_4h_to_h_bias3, alloc489) + R.vm.kill_object(alloc488) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_20_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[247] + gpt_neox_layers_20_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[248] + gv781: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc490: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv781, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape275, gpt_neox_layers_20_attention_dense_weight3, gpt_neox_layers_20_attention_dense_bias3, alloc489, alloc483, alloc490) + R.vm.kill_object(alloc483) + R.vm.kill_object(reshape275) + R.vm.kill_object(alloc489) + R.vm.kill_object(gpt_neox_layers_20_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_20_attention_dense_bias3) + gpt_neox_layers_21_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[253] + gpt_neox_layers_21_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[254] + gv782: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc491: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv782, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc490, gpt_neox_layers_21_input_layernorm_weight3, gpt_neox_layers_21_input_layernorm_bias3, alloc491) + R.vm.kill_object(gpt_neox_layers_21_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_21_input_layernorm_bias3) + gpt_neox_layers_21_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[257] + gpt_neox_layers_21_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[258] + gv783: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc492: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv783, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc491, gpt_neox_layers_21_attention_query_key_value_weight3, gpt_neox_layers_21_attention_query_key_value_bias3, alloc492) + R.vm.kill_object(alloc491) + R.vm.kill_object(gpt_neox_layers_21_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_21_attention_query_key_value_bias3) + gv784: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape276: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc492, gv784, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc492) + gv785: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape277: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape276, gv785, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape276) + gv786: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc493: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv786, R.dtype("float16")) + _492: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), reshape277, alloc493) + R.vm.kill_object(reshape277) + gv787: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape278: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc493, gv787, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc493) + gv788: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape279: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape278, gv788, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape278) + gpt_neox_layers_21_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[255] + gpt_neox_layers_21_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[256] + gv789: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc494: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv789, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc490, gpt_neox_layers_21_post_attention_layernorm_weight3, gpt_neox_layers_21_post_attention_layernorm_bias3, alloc494) + R.vm.kill_object(gpt_neox_layers_21_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_21_post_attention_layernorm_bias3) + gpt_neox_layers_21_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[261] + gpt_neox_layers_21_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[262] + gv790: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc495: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv790, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc494, gpt_neox_layers_21_mlp_dense_h_to_4h_weight3, gpt_neox_layers_21_mlp_dense_h_to_4h_bias3, alloc495) + R.vm.kill_object(alloc494) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_21_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[263] + gpt_neox_layers_21_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[264] + gv791: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc496: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv791, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc495, gpt_neox_layers_21_mlp_dense_4h_to_h_weight3, gpt_neox_layers_21_mlp_dense_4h_to_h_bias3, alloc496) + R.vm.kill_object(alloc495) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_21_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[259] + gpt_neox_layers_21_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[260] + gv792: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc497: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv792, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape279, gpt_neox_layers_21_attention_dense_weight3, gpt_neox_layers_21_attention_dense_bias3, alloc496, alloc490, alloc497) + R.vm.kill_object(alloc490) + R.vm.kill_object(reshape279) + R.vm.kill_object(alloc496) + R.vm.kill_object(gpt_neox_layers_21_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_21_attention_dense_bias3) + gpt_neox_layers_22_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[265] + gpt_neox_layers_22_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[266] + gv793: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc498: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv793, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc497, gpt_neox_layers_22_input_layernorm_weight3, gpt_neox_layers_22_input_layernorm_bias3, alloc498) + R.vm.kill_object(gpt_neox_layers_22_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_22_input_layernorm_bias3) + gpt_neox_layers_22_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[269] + gpt_neox_layers_22_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[270] + gv794: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc499: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv794, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc498, gpt_neox_layers_22_attention_query_key_value_weight3, gpt_neox_layers_22_attention_query_key_value_bias3, alloc499) + R.vm.kill_object(alloc498) + R.vm.kill_object(gpt_neox_layers_22_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_22_attention_query_key_value_bias3) + gv795: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape280: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc499, gv795, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc499) + gv796: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape281: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape280, gv796, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape280) + gv797: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc500: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv797, R.dtype("float16")) + _499: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), reshape281, alloc500) + R.vm.kill_object(reshape281) + gv798: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape282: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc500, gv798, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc500) + gv799: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape283: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape282, gv799, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape282) + gpt_neox_layers_22_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[267] + gpt_neox_layers_22_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[268] + gv800: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc501: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv800, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc497, gpt_neox_layers_22_post_attention_layernorm_weight3, gpt_neox_layers_22_post_attention_layernorm_bias3, alloc501) + R.vm.kill_object(gpt_neox_layers_22_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_22_post_attention_layernorm_bias3) + gpt_neox_layers_22_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[273] + gpt_neox_layers_22_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[274] + gv801: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc502: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv801, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc501, gpt_neox_layers_22_mlp_dense_h_to_4h_weight3, gpt_neox_layers_22_mlp_dense_h_to_4h_bias3, alloc502) + R.vm.kill_object(alloc501) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_22_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[275] + gpt_neox_layers_22_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[276] + gv802: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc503: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv802, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc502, gpt_neox_layers_22_mlp_dense_4h_to_h_weight3, gpt_neox_layers_22_mlp_dense_4h_to_h_bias3, alloc503) + R.vm.kill_object(alloc502) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_22_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[271] + gpt_neox_layers_22_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[272] + gv803: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc504: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv803, R.dtype("float16")) + R.vm.kill_object(storage16) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape283, gpt_neox_layers_22_attention_dense_weight3, gpt_neox_layers_22_attention_dense_bias3, alloc503, alloc497, alloc504) + R.vm.kill_object(alloc497) + R.vm.kill_object(reshape283) + R.vm.kill_object(alloc503) + R.vm.kill_object(gpt_neox_layers_22_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_22_attention_dense_bias3) + gpt_neox_layers_23_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[277] + gpt_neox_layers_23_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[278] + gv804: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc505: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv804, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc504, gpt_neox_layers_23_input_layernorm_weight3, gpt_neox_layers_23_input_layernorm_bias3, alloc505) + R.vm.kill_object(gpt_neox_layers_23_input_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_23_input_layernorm_bias3) + gpt_neox_layers_23_attention_query_key_value_weight3: R.Tensor((3072, 1024), dtype="float16") = packed_params[281] + gpt_neox_layers_23_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[282] + gv805: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc506: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv805, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc505, gpt_neox_layers_23_attention_query_key_value_weight3, gpt_neox_layers_23_attention_query_key_value_bias3, alloc506) + R.vm.kill_object(alloc505) + R.vm.kill_object(gpt_neox_layers_23_attention_query_key_value_weight3) + R.vm.kill_object(gpt_neox_layers_23_attention_query_key_value_bias3) + gv806: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape284: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc506, gv806, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc506) + gv807: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape285: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape284, gv807, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape284) + gv808: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc507: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv808, R.dtype("float16")) + _506: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), reshape285, alloc507) + R.vm.kill_object(reshape285) + gv809: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape286: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc507, gv809, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc507) + gv810: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape287: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape286, gv810, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape286) + gpt_neox_layers_23_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[279] + gpt_neox_layers_23_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[280] + gv811: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc508: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv811, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc504, gpt_neox_layers_23_post_attention_layernorm_weight3, gpt_neox_layers_23_post_attention_layernorm_bias3, alloc508) + R.vm.kill_object(gpt_neox_layers_23_post_attention_layernorm_weight3) + R.vm.kill_object(gpt_neox_layers_23_post_attention_layernorm_bias3) + gpt_neox_layers_23_mlp_dense_h_to_4h_weight3: R.Tensor((4096, 1024), dtype="float16") = packed_params[285] + gpt_neox_layers_23_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[286] + gv812: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc509: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv812, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc508, gpt_neox_layers_23_mlp_dense_h_to_4h_weight3, gpt_neox_layers_23_mlp_dense_h_to_4h_bias3, alloc509) + R.vm.kill_object(alloc508) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_h_to_4h_weight3) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_h_to_4h_bias3) + gpt_neox_layers_23_mlp_dense_4h_to_h_weight3: R.Tensor((1024, 4096), dtype="float16") = packed_params[287] + gpt_neox_layers_23_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[288] + gv813: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc510: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv813, R.dtype("float16")) + R.vm.kill_object(storage15) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc509, gpt_neox_layers_23_mlp_dense_4h_to_h_weight3, gpt_neox_layers_23_mlp_dense_4h_to_h_bias3, alloc510) + R.vm.kill_object(alloc509) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_4h_to_h_weight3) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_4h_to_h_bias3) + gpt_neox_layers_23_attention_dense_weight3: R.Tensor((1024, 1024), dtype="float16") = packed_params[283] + gpt_neox_layers_23_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[284] + gv814: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc511: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv814, R.dtype("float16")) + R.vm.kill_object(storage17) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape287, gpt_neox_layers_23_attention_dense_weight3, gpt_neox_layers_23_attention_dense_bias3, alloc510, alloc504, alloc511) + R.vm.kill_object(alloc504) + R.vm.kill_object(reshape287) + R.vm.kill_object(alloc510) + R.vm.kill_object(gpt_neox_layers_23_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_23_attention_dense_bias3) + gpt_neox_final_layer_norm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[289] + gpt_neox_final_layer_norm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[290] + gv815: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc512: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv815, R.dtype("float16")) + R.vm.kill_object(storage14) + cls.gpt_neox_q0f16_layer_norm1(alloc511, gpt_neox_final_layer_norm_weight3, gpt_neox_final_layer_norm_bias3, alloc512) + R.vm.kill_object(alloc511) + R.vm.kill_object(gpt_neox_final_layer_norm_weight3) + R.vm.kill_object(gpt_neox_final_layer_norm_bias3) + storage18: R.Object = R.vm.alloc_storage(R.shape([262144]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv816: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc513: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage18, R.prim_value(0), gv816, R.dtype("float16")) + R.vm.kill_object(storage18) + cls.gpt_neox_q0f16_take(alloc512, logit_positions, alloc513) + R.vm.kill_object(alloc512) + embed_out_weight3: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[291] + gv817: R.Shape(ndim=1) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(4), sinfo_args=(R.Shape(ndim=1),)) + storage19: R.Object = R.vm.alloc_storage(gv817, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv818: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(1), R.prim_value(1), R.prim_value(2), sinfo_args=(R.Shape(ndim=3),)) + alloc514: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage19, R.prim_value(0), gv818, R.dtype("float32")) + R.vm.kill_object(storage19) + cls.gpt_neox_q0f16_fused_NT_matmul9_cast5(alloc513, embed_out_weight3, alloc514) + R.vm.kill_object(alloc513) + R.vm.kill_object(embed_out_weight3) + gv3: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc514, paged_kv_cache + R.vm.kill_object(alloc514) + gv819: R.Tensor(dtype="float32", ndim=3) = gv3[0] + R.call_packed("vm.builtin.match_shape", gv819, shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(3), R.prim_value(1), R.prim_value(3), R.prim_value(2), R.str("ErrorContext(fn=batch_prefill, loc=return, annotation=R.Tuple(R.Tensor((1, batch_size, vocab_size), dtype=\"float32\"), R.Object)) "), sinfo_args=(R.Tuple,)) + return gv3 + + @R.function + def batch_verify(input_embeds: R.Tensor((1, "seq_len", 1024), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor(("vocab_size", 1024), dtype="float16"))) -> R.Tuple(R.Tensor((1, "seq_len", "vocab_size"), dtype="float32"), R.Object): + seq_len = T.int64() + vocab_size = T.int64() + R.func_attr({"num_input": 2, "pipeline_parallel_stages": 1, "relax.force_pure": True, "relax.rewrite_cuda_graph.capture_symbolic_vars": ["batch_size", "seq_len"], "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) + cls = Module + shape_heap: R.Tensor(dtype="int64", ndim=1) = R.call_builtin_with_ctx("vm.builtin.alloc_shape_heap", (R.prim_value(4),), sinfo_args=(R.Tensor(dtype="int64", ndim=1),)) + R.call_packed("vm.builtin.check_tensor_info", input_embeds, R.prim_value(3), R.dtype("float16"), R.str("ErrorContext(fn=batch_verify, loc=param[0], param=input_embeds, annotation=R.Tensor((1, seq_len, 1024), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(292), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv823: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv823, R.prim_value(2), R.dtype("float16"), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv824: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[291] + R.call_packed("vm.builtin.check_tensor_info", gv824, R.prim_value(2), R.dtype("float16"), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", input_embeds, shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=batch_verify, loc=param[0], param=input_embeds, annotation=R.Tensor((1, seq_len, 1024), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv823, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv824, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func4(shape_heap) + gpt_neox_layers_0_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[2] + storage21: R.Object = R.vm.alloc_storage(R.shape([4194304]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv825: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc516: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv825, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(input_embeds, gpt_neox_layers_0_input_layernorm_weight5, gpt_neox_layers_0_input_layernorm_bias5, alloc516) + R.vm.kill_object(gpt_neox_layers_0_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_0_input_layernorm_bias5) + gpt_neox_layers_0_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[6] + storage22: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv826: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc517: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv826, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc516, gpt_neox_layers_0_attention_query_key_value_weight5, gpt_neox_layers_0_attention_query_key_value_bias5, alloc517) + R.vm.kill_object(alloc516) + R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_bias5) + gv827: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape384: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc517, gv827, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc517) + gv828: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape385: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape384, gv828, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape384) + gv829: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc518: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv829, R.dtype("float16")) + _517: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), reshape385, alloc518) + R.vm.kill_object(reshape385) + gv830: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape386: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc518, gv830, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc518) + gv831: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape387: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape386, gv831, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape386) + gpt_neox_layers_0_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[4] + gv832: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc519: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv832, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(input_embeds, gpt_neox_layers_0_post_attention_layernorm_weight5, gpt_neox_layers_0_post_attention_layernorm_bias5, alloc519) + R.vm.kill_object(gpt_neox_layers_0_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_0_post_attention_layernorm_bias5) + gpt_neox_layers_0_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[10] + storage23: R.Object = R.vm.alloc_storage(R.shape([16777216]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv833: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc520: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv833, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc519, gpt_neox_layers_0_mlp_dense_h_to_4h_weight5, gpt_neox_layers_0_mlp_dense_h_to_4h_bias5, alloc520) + R.vm.kill_object(alloc519) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_0_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[12] + gv834: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc521: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv834, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc520, gpt_neox_layers_0_mlp_dense_4h_to_h_weight5, gpt_neox_layers_0_mlp_dense_4h_to_h_bias5, alloc521) + R.vm.kill_object(alloc520) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_0_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[8] + gv835: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc522: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv835, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape387, gpt_neox_layers_0_attention_dense_weight5, gpt_neox_layers_0_attention_dense_bias5, alloc521, input_embeds, alloc522) + R.vm.kill_object(reshape387) + R.vm.kill_object(alloc521) + R.vm.kill_object(gpt_neox_layers_0_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_0_attention_dense_bias5) + gpt_neox_layers_1_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[14] + gv836: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc523: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv836, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc522, gpt_neox_layers_1_input_layernorm_weight5, gpt_neox_layers_1_input_layernorm_bias5, alloc523) + R.vm.kill_object(gpt_neox_layers_1_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_1_input_layernorm_bias5) + gpt_neox_layers_1_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[18] + gv837: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc524: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv837, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc523, gpt_neox_layers_1_attention_query_key_value_weight5, gpt_neox_layers_1_attention_query_key_value_bias5, alloc524) + R.vm.kill_object(alloc523) + R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_bias5) + gv838: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape388: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc524, gv838, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc524) + gv839: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape389: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape388, gv839, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape388) + gv840: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc525: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv840, R.dtype("float16")) + _524: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), reshape389, alloc525) + R.vm.kill_object(reshape389) + gv841: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape390: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc525, gv841, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc525) + gv842: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape391: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape390, gv842, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape390) + gpt_neox_layers_1_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[16] + gv843: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc526: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv843, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc522, gpt_neox_layers_1_post_attention_layernorm_weight5, gpt_neox_layers_1_post_attention_layernorm_bias5, alloc526) + R.vm.kill_object(gpt_neox_layers_1_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_1_post_attention_layernorm_bias5) + gpt_neox_layers_1_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[22] + storage24: R.Object = R.vm.alloc_storage(R.shape([16777216]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv844: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc527: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv844, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc526, gpt_neox_layers_1_mlp_dense_h_to_4h_weight5, gpt_neox_layers_1_mlp_dense_h_to_4h_bias5, alloc527) + R.vm.kill_object(alloc526) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_1_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[24] + gv845: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc528: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv845, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc527, gpt_neox_layers_1_mlp_dense_4h_to_h_weight5, gpt_neox_layers_1_mlp_dense_4h_to_h_bias5, alloc528) + R.vm.kill_object(alloc527) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_1_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[20] + gv846: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc529: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv846, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape391, gpt_neox_layers_1_attention_dense_weight5, gpt_neox_layers_1_attention_dense_bias5, alloc528, alloc522, alloc529) + R.vm.kill_object(alloc522) + R.vm.kill_object(reshape391) + R.vm.kill_object(alloc528) + R.vm.kill_object(gpt_neox_layers_1_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_1_attention_dense_bias5) + gpt_neox_layers_2_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[26] + gv847: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc530: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv847, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc529, gpt_neox_layers_2_input_layernorm_weight5, gpt_neox_layers_2_input_layernorm_bias5, alloc530) + R.vm.kill_object(gpt_neox_layers_2_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_2_input_layernorm_bias5) + gpt_neox_layers_2_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[30] + gv848: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc531: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv848, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc530, gpt_neox_layers_2_attention_query_key_value_weight5, gpt_neox_layers_2_attention_query_key_value_bias5, alloc531) + R.vm.kill_object(alloc530) + R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_bias5) + gv849: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape392: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc531, gv849, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc531) + gv850: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape393: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape392, gv850, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape392) + gv851: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc532: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv851, R.dtype("float16")) + _531: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), reshape393, alloc532) + R.vm.kill_object(reshape393) + gv852: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape394: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc532, gv852, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc532) + gv853: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape395: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape394, gv853, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape394) + gpt_neox_layers_2_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[28] + gv854: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc533: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv854, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc529, gpt_neox_layers_2_post_attention_layernorm_weight5, gpt_neox_layers_2_post_attention_layernorm_bias5, alloc533) + R.vm.kill_object(gpt_neox_layers_2_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_2_post_attention_layernorm_bias5) + gpt_neox_layers_2_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[34] + gv855: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc534: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv855, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc533, gpt_neox_layers_2_mlp_dense_h_to_4h_weight5, gpt_neox_layers_2_mlp_dense_h_to_4h_bias5, alloc534) + R.vm.kill_object(alloc533) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_2_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[36] + gv856: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc535: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv856, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc534, gpt_neox_layers_2_mlp_dense_4h_to_h_weight5, gpt_neox_layers_2_mlp_dense_4h_to_h_bias5, alloc535) + R.vm.kill_object(alloc534) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_2_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[32] + gv857: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc536: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv857, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape395, gpt_neox_layers_2_attention_dense_weight5, gpt_neox_layers_2_attention_dense_bias5, alloc535, alloc529, alloc536) + R.vm.kill_object(alloc529) + R.vm.kill_object(reshape395) + R.vm.kill_object(alloc535) + R.vm.kill_object(gpt_neox_layers_2_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_2_attention_dense_bias5) + gpt_neox_layers_3_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[38] + gv858: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc537: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv858, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc536, gpt_neox_layers_3_input_layernorm_weight5, gpt_neox_layers_3_input_layernorm_bias5, alloc537) + R.vm.kill_object(gpt_neox_layers_3_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_3_input_layernorm_bias5) + gpt_neox_layers_3_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[42] + gv859: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc538: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv859, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc537, gpt_neox_layers_3_attention_query_key_value_weight5, gpt_neox_layers_3_attention_query_key_value_bias5, alloc538) + R.vm.kill_object(alloc537) + R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_bias5) + gv860: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape396: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc538, gv860, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc538) + gv861: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape397: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape396, gv861, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape396) + gv862: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc539: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv862, R.dtype("float16")) + _538: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), reshape397, alloc539) + R.vm.kill_object(reshape397) + gv863: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape398: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc539, gv863, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc539) + gv864: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape399: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape398, gv864, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape398) + gpt_neox_layers_3_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[40] + gv865: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc540: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv865, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc536, gpt_neox_layers_3_post_attention_layernorm_weight5, gpt_neox_layers_3_post_attention_layernorm_bias5, alloc540) + R.vm.kill_object(gpt_neox_layers_3_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_3_post_attention_layernorm_bias5) + gpt_neox_layers_3_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[46] + gv866: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc541: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv866, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc540, gpt_neox_layers_3_mlp_dense_h_to_4h_weight5, gpt_neox_layers_3_mlp_dense_h_to_4h_bias5, alloc541) + R.vm.kill_object(alloc540) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_3_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[48] + gv867: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc542: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv867, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc541, gpt_neox_layers_3_mlp_dense_4h_to_h_weight5, gpt_neox_layers_3_mlp_dense_4h_to_h_bias5, alloc542) + R.vm.kill_object(alloc541) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_3_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[44] + gv868: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc543: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv868, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape399, gpt_neox_layers_3_attention_dense_weight5, gpt_neox_layers_3_attention_dense_bias5, alloc542, alloc536, alloc543) + R.vm.kill_object(alloc536) + R.vm.kill_object(reshape399) + R.vm.kill_object(alloc542) + R.vm.kill_object(gpt_neox_layers_3_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_3_attention_dense_bias5) + gpt_neox_layers_4_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[50] + gv869: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc544: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv869, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc543, gpt_neox_layers_4_input_layernorm_weight5, gpt_neox_layers_4_input_layernorm_bias5, alloc544) + R.vm.kill_object(gpt_neox_layers_4_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_4_input_layernorm_bias5) + gpt_neox_layers_4_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[54] + gv870: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc545: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv870, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc544, gpt_neox_layers_4_attention_query_key_value_weight5, gpt_neox_layers_4_attention_query_key_value_bias5, alloc545) + R.vm.kill_object(alloc544) + R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_bias5) + gv871: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape400: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc545, gv871, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc545) + gv872: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape401: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape400, gv872, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape400) + gv873: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc546: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv873, R.dtype("float16")) + _545: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), reshape401, alloc546) + R.vm.kill_object(reshape401) + gv874: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape402: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc546, gv874, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc546) + gv875: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape403: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape402, gv875, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape402) + gpt_neox_layers_4_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[52] + gv876: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc547: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv876, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc543, gpt_neox_layers_4_post_attention_layernorm_weight5, gpt_neox_layers_4_post_attention_layernorm_bias5, alloc547) + R.vm.kill_object(gpt_neox_layers_4_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_4_post_attention_layernorm_bias5) + gpt_neox_layers_4_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[58] + gv877: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc548: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv877, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc547, gpt_neox_layers_4_mlp_dense_h_to_4h_weight5, gpt_neox_layers_4_mlp_dense_h_to_4h_bias5, alloc548) + R.vm.kill_object(alloc547) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_4_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[60] + gv878: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc549: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv878, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc548, gpt_neox_layers_4_mlp_dense_4h_to_h_weight5, gpt_neox_layers_4_mlp_dense_4h_to_h_bias5, alloc549) + R.vm.kill_object(alloc548) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_4_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[56] + gv879: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc550: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv879, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape403, gpt_neox_layers_4_attention_dense_weight5, gpt_neox_layers_4_attention_dense_bias5, alloc549, alloc543, alloc550) + R.vm.kill_object(alloc543) + R.vm.kill_object(reshape403) + R.vm.kill_object(alloc549) + R.vm.kill_object(gpt_neox_layers_4_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_4_attention_dense_bias5) + gpt_neox_layers_5_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[62] + gv880: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc551: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv880, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc550, gpt_neox_layers_5_input_layernorm_weight5, gpt_neox_layers_5_input_layernorm_bias5, alloc551) + R.vm.kill_object(gpt_neox_layers_5_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_5_input_layernorm_bias5) + gpt_neox_layers_5_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[66] + gv881: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc552: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv881, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc551, gpt_neox_layers_5_attention_query_key_value_weight5, gpt_neox_layers_5_attention_query_key_value_bias5, alloc552) + R.vm.kill_object(alloc551) + R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_bias5) + gv882: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape404: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc552, gv882, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc552) + gv883: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape405: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape404, gv883, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape404) + gv884: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc553: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv884, R.dtype("float16")) + _552: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), reshape405, alloc553) + R.vm.kill_object(reshape405) + gv885: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape406: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc553, gv885, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc553) + gv886: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape407: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape406, gv886, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape406) + gpt_neox_layers_5_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[64] + gv887: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc554: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv887, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc550, gpt_neox_layers_5_post_attention_layernorm_weight5, gpt_neox_layers_5_post_attention_layernorm_bias5, alloc554) + R.vm.kill_object(gpt_neox_layers_5_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_5_post_attention_layernorm_bias5) + gpt_neox_layers_5_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[70] + gv888: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc555: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv888, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc554, gpt_neox_layers_5_mlp_dense_h_to_4h_weight5, gpt_neox_layers_5_mlp_dense_h_to_4h_bias5, alloc555) + R.vm.kill_object(alloc554) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_5_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[72] + gv889: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc556: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv889, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc555, gpt_neox_layers_5_mlp_dense_4h_to_h_weight5, gpt_neox_layers_5_mlp_dense_4h_to_h_bias5, alloc556) + R.vm.kill_object(alloc555) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_5_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[68] + gv890: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc557: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv890, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape407, gpt_neox_layers_5_attention_dense_weight5, gpt_neox_layers_5_attention_dense_bias5, alloc556, alloc550, alloc557) + R.vm.kill_object(alloc550) + R.vm.kill_object(reshape407) + R.vm.kill_object(alloc556) + R.vm.kill_object(gpt_neox_layers_5_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_5_attention_dense_bias5) + gpt_neox_layers_6_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[74] + gv891: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc558: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv891, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc557, gpt_neox_layers_6_input_layernorm_weight5, gpt_neox_layers_6_input_layernorm_bias5, alloc558) + R.vm.kill_object(gpt_neox_layers_6_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_6_input_layernorm_bias5) + gpt_neox_layers_6_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[78] + gv892: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc559: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv892, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc558, gpt_neox_layers_6_attention_query_key_value_weight5, gpt_neox_layers_6_attention_query_key_value_bias5, alloc559) + R.vm.kill_object(alloc558) + R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_bias5) + gv893: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape408: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc559, gv893, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc559) + gv894: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape409: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape408, gv894, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape408) + gv895: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc560: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv895, R.dtype("float16")) + _559: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), reshape409, alloc560) + R.vm.kill_object(reshape409) + gv896: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape410: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc560, gv896, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc560) + gv897: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape411: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape410, gv897, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape410) + gpt_neox_layers_6_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[76] + gv898: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc561: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv898, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc557, gpt_neox_layers_6_post_attention_layernorm_weight5, gpt_neox_layers_6_post_attention_layernorm_bias5, alloc561) + R.vm.kill_object(gpt_neox_layers_6_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_6_post_attention_layernorm_bias5) + gpt_neox_layers_6_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[82] + gv899: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc562: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv899, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc561, gpt_neox_layers_6_mlp_dense_h_to_4h_weight5, gpt_neox_layers_6_mlp_dense_h_to_4h_bias5, alloc562) + R.vm.kill_object(alloc561) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_6_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[84] + gv900: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc563: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv900, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc562, gpt_neox_layers_6_mlp_dense_4h_to_h_weight5, gpt_neox_layers_6_mlp_dense_4h_to_h_bias5, alloc563) + R.vm.kill_object(alloc562) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_6_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[80] + gv901: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc564: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv901, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape411, gpt_neox_layers_6_attention_dense_weight5, gpt_neox_layers_6_attention_dense_bias5, alloc563, alloc557, alloc564) + R.vm.kill_object(alloc557) + R.vm.kill_object(reshape411) + R.vm.kill_object(alloc563) + R.vm.kill_object(gpt_neox_layers_6_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_6_attention_dense_bias5) + gpt_neox_layers_7_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[86] + gv902: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc565: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv902, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc564, gpt_neox_layers_7_input_layernorm_weight5, gpt_neox_layers_7_input_layernorm_bias5, alloc565) + R.vm.kill_object(gpt_neox_layers_7_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_7_input_layernorm_bias5) + gpt_neox_layers_7_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[90] + gv903: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc566: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv903, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc565, gpt_neox_layers_7_attention_query_key_value_weight5, gpt_neox_layers_7_attention_query_key_value_bias5, alloc566) + R.vm.kill_object(alloc565) + R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_bias5) + gv904: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape412: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc566, gv904, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc566) + gv905: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape413: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape412, gv905, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape412) + gv906: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc567: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv906, R.dtype("float16")) + _566: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), reshape413, alloc567) + R.vm.kill_object(reshape413) + gv907: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape414: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc567, gv907, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc567) + gv908: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape415: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape414, gv908, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape414) + gpt_neox_layers_7_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[88] + gv909: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc568: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv909, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc564, gpt_neox_layers_7_post_attention_layernorm_weight5, gpt_neox_layers_7_post_attention_layernorm_bias5, alloc568) + R.vm.kill_object(gpt_neox_layers_7_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_7_post_attention_layernorm_bias5) + gpt_neox_layers_7_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[94] + gv910: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc569: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv910, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc568, gpt_neox_layers_7_mlp_dense_h_to_4h_weight5, gpt_neox_layers_7_mlp_dense_h_to_4h_bias5, alloc569) + R.vm.kill_object(alloc568) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_7_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[96] + gv911: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc570: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv911, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc569, gpt_neox_layers_7_mlp_dense_4h_to_h_weight5, gpt_neox_layers_7_mlp_dense_4h_to_h_bias5, alloc570) + R.vm.kill_object(alloc569) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_7_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[92] + gv912: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc571: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv912, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape415, gpt_neox_layers_7_attention_dense_weight5, gpt_neox_layers_7_attention_dense_bias5, alloc570, alloc564, alloc571) + R.vm.kill_object(alloc564) + R.vm.kill_object(reshape415) + R.vm.kill_object(alloc570) + R.vm.kill_object(gpt_neox_layers_7_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_7_attention_dense_bias5) + gpt_neox_layers_8_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[98] + gv913: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc572: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv913, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc571, gpt_neox_layers_8_input_layernorm_weight5, gpt_neox_layers_8_input_layernorm_bias5, alloc572) + R.vm.kill_object(gpt_neox_layers_8_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_8_input_layernorm_bias5) + gpt_neox_layers_8_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[102] + gv914: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc573: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv914, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc572, gpt_neox_layers_8_attention_query_key_value_weight5, gpt_neox_layers_8_attention_query_key_value_bias5, alloc573) + R.vm.kill_object(alloc572) + R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_bias5) + gv915: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape416: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc573, gv915, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc573) + gv916: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape417: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape416, gv916, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape416) + gv917: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc574: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv917, R.dtype("float16")) + _573: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), reshape417, alloc574) + R.vm.kill_object(reshape417) + gv918: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape418: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc574, gv918, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc574) + gv919: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape419: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape418, gv919, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape418) + gpt_neox_layers_8_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[100] + gv920: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc575: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv920, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc571, gpt_neox_layers_8_post_attention_layernorm_weight5, gpt_neox_layers_8_post_attention_layernorm_bias5, alloc575) + R.vm.kill_object(gpt_neox_layers_8_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_8_post_attention_layernorm_bias5) + gpt_neox_layers_8_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[106] + gv921: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc576: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv921, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc575, gpt_neox_layers_8_mlp_dense_h_to_4h_weight5, gpt_neox_layers_8_mlp_dense_h_to_4h_bias5, alloc576) + R.vm.kill_object(alloc575) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_8_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[108] + gv922: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc577: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv922, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc576, gpt_neox_layers_8_mlp_dense_4h_to_h_weight5, gpt_neox_layers_8_mlp_dense_4h_to_h_bias5, alloc577) + R.vm.kill_object(alloc576) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_8_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[104] + gv923: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc578: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv923, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape419, gpt_neox_layers_8_attention_dense_weight5, gpt_neox_layers_8_attention_dense_bias5, alloc577, alloc571, alloc578) + R.vm.kill_object(alloc571) + R.vm.kill_object(reshape419) + R.vm.kill_object(alloc577) + R.vm.kill_object(gpt_neox_layers_8_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_8_attention_dense_bias5) + gpt_neox_layers_9_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[110] + gv924: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc579: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv924, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc578, gpt_neox_layers_9_input_layernorm_weight5, gpt_neox_layers_9_input_layernorm_bias5, alloc579) + R.vm.kill_object(gpt_neox_layers_9_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_9_input_layernorm_bias5) + gpt_neox_layers_9_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[114] + gv925: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc580: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv925, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc579, gpt_neox_layers_9_attention_query_key_value_weight5, gpt_neox_layers_9_attention_query_key_value_bias5, alloc580) + R.vm.kill_object(alloc579) + R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_bias5) + gv926: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape420: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc580, gv926, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc580) + gv927: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape421: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape420, gv927, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape420) + gv928: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc581: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv928, R.dtype("float16")) + _580: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), reshape421, alloc581) + R.vm.kill_object(reshape421) + gv929: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape422: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc581, gv929, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc581) + gv930: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape423: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape422, gv930, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape422) + gpt_neox_layers_9_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[112] + gv931: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc582: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv931, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc578, gpt_neox_layers_9_post_attention_layernorm_weight5, gpt_neox_layers_9_post_attention_layernorm_bias5, alloc582) + R.vm.kill_object(gpt_neox_layers_9_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_9_post_attention_layernorm_bias5) + gpt_neox_layers_9_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[118] + gv932: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc583: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv932, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc582, gpt_neox_layers_9_mlp_dense_h_to_4h_weight5, gpt_neox_layers_9_mlp_dense_h_to_4h_bias5, alloc583) + R.vm.kill_object(alloc582) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_9_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[120] + gv933: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc584: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv933, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc583, gpt_neox_layers_9_mlp_dense_4h_to_h_weight5, gpt_neox_layers_9_mlp_dense_4h_to_h_bias5, alloc584) + R.vm.kill_object(alloc583) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_9_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[116] + gv934: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc585: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv934, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape423, gpt_neox_layers_9_attention_dense_weight5, gpt_neox_layers_9_attention_dense_bias5, alloc584, alloc578, alloc585) + R.vm.kill_object(alloc578) + R.vm.kill_object(reshape423) + R.vm.kill_object(alloc584) + R.vm.kill_object(gpt_neox_layers_9_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_9_attention_dense_bias5) + gpt_neox_layers_10_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[122] + gv935: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc586: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv935, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc585, gpt_neox_layers_10_input_layernorm_weight5, gpt_neox_layers_10_input_layernorm_bias5, alloc586) + R.vm.kill_object(gpt_neox_layers_10_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_10_input_layernorm_bias5) + gpt_neox_layers_10_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[126] + gv936: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc587: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv936, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc586, gpt_neox_layers_10_attention_query_key_value_weight5, gpt_neox_layers_10_attention_query_key_value_bias5, alloc587) + R.vm.kill_object(alloc586) + R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_bias5) + gv937: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape424: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc587, gv937, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc587) + gv938: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape425: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape424, gv938, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape424) + gv939: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc588: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv939, R.dtype("float16")) + _587: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), reshape425, alloc588) + R.vm.kill_object(reshape425) + gv940: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape426: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc588, gv940, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc588) + gv941: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape427: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape426, gv941, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape426) + gpt_neox_layers_10_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[124] + gv942: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc589: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv942, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc585, gpt_neox_layers_10_post_attention_layernorm_weight5, gpt_neox_layers_10_post_attention_layernorm_bias5, alloc589) + R.vm.kill_object(gpt_neox_layers_10_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_10_post_attention_layernorm_bias5) + gpt_neox_layers_10_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[130] + gv943: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc590: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv943, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc589, gpt_neox_layers_10_mlp_dense_h_to_4h_weight5, gpt_neox_layers_10_mlp_dense_h_to_4h_bias5, alloc590) + R.vm.kill_object(alloc589) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_10_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[132] + gv944: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc591: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv944, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc590, gpt_neox_layers_10_mlp_dense_4h_to_h_weight5, gpt_neox_layers_10_mlp_dense_4h_to_h_bias5, alloc591) + R.vm.kill_object(alloc590) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_10_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[128] + gv945: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc592: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv945, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape427, gpt_neox_layers_10_attention_dense_weight5, gpt_neox_layers_10_attention_dense_bias5, alloc591, alloc585, alloc592) + R.vm.kill_object(alloc585) + R.vm.kill_object(reshape427) + R.vm.kill_object(alloc591) + R.vm.kill_object(gpt_neox_layers_10_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_10_attention_dense_bias5) + gpt_neox_layers_11_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[134] + gv946: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc593: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv946, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc592, gpt_neox_layers_11_input_layernorm_weight5, gpt_neox_layers_11_input_layernorm_bias5, alloc593) + R.vm.kill_object(gpt_neox_layers_11_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_11_input_layernorm_bias5) + gpt_neox_layers_11_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[138] + gv947: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc594: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv947, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc593, gpt_neox_layers_11_attention_query_key_value_weight5, gpt_neox_layers_11_attention_query_key_value_bias5, alloc594) + R.vm.kill_object(alloc593) + R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_bias5) + gv948: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape428: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc594, gv948, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc594) + gv949: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape429: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape428, gv949, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape428) + gv950: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc595: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv950, R.dtype("float16")) + _594: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), reshape429, alloc595) + R.vm.kill_object(reshape429) + gv951: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape430: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc595, gv951, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc595) + gv952: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape431: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape430, gv952, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape430) + gpt_neox_layers_11_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[136] + gv953: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc596: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv953, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc592, gpt_neox_layers_11_post_attention_layernorm_weight5, gpt_neox_layers_11_post_attention_layernorm_bias5, alloc596) + R.vm.kill_object(gpt_neox_layers_11_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_11_post_attention_layernorm_bias5) + gpt_neox_layers_11_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[142] + gv954: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc597: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv954, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc596, gpt_neox_layers_11_mlp_dense_h_to_4h_weight5, gpt_neox_layers_11_mlp_dense_h_to_4h_bias5, alloc597) + R.vm.kill_object(alloc596) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_11_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[144] + gv955: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc598: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv955, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc597, gpt_neox_layers_11_mlp_dense_4h_to_h_weight5, gpt_neox_layers_11_mlp_dense_4h_to_h_bias5, alloc598) + R.vm.kill_object(alloc597) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_11_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[140] + gv956: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc599: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv956, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape431, gpt_neox_layers_11_attention_dense_weight5, gpt_neox_layers_11_attention_dense_bias5, alloc598, alloc592, alloc599) + R.vm.kill_object(alloc592) + R.vm.kill_object(reshape431) + R.vm.kill_object(alloc598) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias5) + gpt_neox_layers_12_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[145] + gpt_neox_layers_12_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[146] + gv957: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc600: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv957, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc599, gpt_neox_layers_12_input_layernorm_weight5, gpt_neox_layers_12_input_layernorm_bias5, alloc600) + R.vm.kill_object(gpt_neox_layers_12_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_12_input_layernorm_bias5) + gpt_neox_layers_12_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[149] + gpt_neox_layers_12_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[150] + gv958: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc601: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv958, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc600, gpt_neox_layers_12_attention_query_key_value_weight5, gpt_neox_layers_12_attention_query_key_value_bias5, alloc601) + R.vm.kill_object(alloc600) + R.vm.kill_object(gpt_neox_layers_12_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_12_attention_query_key_value_bias5) + gv959: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape432: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc601, gv959, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc601) + gv960: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape433: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape432, gv960, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape432) + gv961: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc602: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv961, R.dtype("float16")) + _601: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape433, alloc602) + R.vm.kill_object(reshape433) + gv962: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape434: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc602, gv962, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc602) + gv963: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape435: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape434, gv963, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape434) + gpt_neox_layers_12_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[147] + gpt_neox_layers_12_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[148] + gv964: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc603: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv964, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc599, gpt_neox_layers_12_post_attention_layernorm_weight5, gpt_neox_layers_12_post_attention_layernorm_bias5, alloc603) + R.vm.kill_object(gpt_neox_layers_12_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_12_post_attention_layernorm_bias5) + gpt_neox_layers_12_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[153] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[154] + gv965: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc604: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv965, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc603, gpt_neox_layers_12_mlp_dense_h_to_4h_weight5, gpt_neox_layers_12_mlp_dense_h_to_4h_bias5, alloc604) + R.vm.kill_object(alloc603) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_12_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[155] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[156] + gv966: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc605: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv966, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc604, gpt_neox_layers_12_mlp_dense_4h_to_h_weight5, gpt_neox_layers_12_mlp_dense_4h_to_h_bias5, alloc605) + R.vm.kill_object(alloc604) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_12_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[151] + gpt_neox_layers_12_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[152] + gv967: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc606: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv967, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape435, gpt_neox_layers_12_attention_dense_weight5, gpt_neox_layers_12_attention_dense_bias5, alloc605, alloc599, alloc606) + R.vm.kill_object(alloc599) + R.vm.kill_object(reshape435) + R.vm.kill_object(alloc605) + R.vm.kill_object(gpt_neox_layers_12_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_12_attention_dense_bias5) + gpt_neox_layers_13_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[157] + gpt_neox_layers_13_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[158] + gv968: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc607: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv968, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc606, gpt_neox_layers_13_input_layernorm_weight5, gpt_neox_layers_13_input_layernorm_bias5, alloc607) + R.vm.kill_object(gpt_neox_layers_13_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_13_input_layernorm_bias5) + gpt_neox_layers_13_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[161] + gpt_neox_layers_13_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[162] + gv969: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc608: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv969, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc607, gpt_neox_layers_13_attention_query_key_value_weight5, gpt_neox_layers_13_attention_query_key_value_bias5, alloc608) + R.vm.kill_object(alloc607) + R.vm.kill_object(gpt_neox_layers_13_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_13_attention_query_key_value_bias5) + gv970: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape436: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc608, gv970, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc608) + gv971: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape437: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape436, gv971, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape436) + gv972: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc609: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv972, R.dtype("float16")) + _608: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape437, alloc609) + R.vm.kill_object(reshape437) + gv973: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape438: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc609, gv973, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc609) + gv974: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape439: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape438, gv974, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape438) + gpt_neox_layers_13_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[159] + gpt_neox_layers_13_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[160] + gv975: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc610: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv975, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc606, gpt_neox_layers_13_post_attention_layernorm_weight5, gpt_neox_layers_13_post_attention_layernorm_bias5, alloc610) + R.vm.kill_object(gpt_neox_layers_13_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_13_post_attention_layernorm_bias5) + gpt_neox_layers_13_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[165] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[166] + gv976: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc611: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv976, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc610, gpt_neox_layers_13_mlp_dense_h_to_4h_weight5, gpt_neox_layers_13_mlp_dense_h_to_4h_bias5, alloc611) + R.vm.kill_object(alloc610) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_13_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[167] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[168] + gv977: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc612: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv977, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc611, gpt_neox_layers_13_mlp_dense_4h_to_h_weight5, gpt_neox_layers_13_mlp_dense_4h_to_h_bias5, alloc612) + R.vm.kill_object(alloc611) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_13_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[163] + gpt_neox_layers_13_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[164] + gv978: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc613: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv978, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape439, gpt_neox_layers_13_attention_dense_weight5, gpt_neox_layers_13_attention_dense_bias5, alloc612, alloc606, alloc613) + R.vm.kill_object(alloc606) + R.vm.kill_object(reshape439) + R.vm.kill_object(alloc612) + R.vm.kill_object(gpt_neox_layers_13_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_13_attention_dense_bias5) + gpt_neox_layers_14_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[169] + gpt_neox_layers_14_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[170] + gv979: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc614: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv979, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc613, gpt_neox_layers_14_input_layernorm_weight5, gpt_neox_layers_14_input_layernorm_bias5, alloc614) + R.vm.kill_object(gpt_neox_layers_14_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_14_input_layernorm_bias5) + gpt_neox_layers_14_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[173] + gpt_neox_layers_14_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[174] + gv980: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc615: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv980, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc614, gpt_neox_layers_14_attention_query_key_value_weight5, gpt_neox_layers_14_attention_query_key_value_bias5, alloc615) + R.vm.kill_object(alloc614) + R.vm.kill_object(gpt_neox_layers_14_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_14_attention_query_key_value_bias5) + gv981: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape440: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc615, gv981, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc615) + gv982: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape441: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape440, gv982, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape440) + gv983: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc616: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv983, R.dtype("float16")) + _615: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape441, alloc616) + R.vm.kill_object(reshape441) + gv984: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape442: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc616, gv984, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc616) + gv985: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape443: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape442, gv985, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape442) + gpt_neox_layers_14_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[171] + gpt_neox_layers_14_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[172] + gv986: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc617: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv986, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc613, gpt_neox_layers_14_post_attention_layernorm_weight5, gpt_neox_layers_14_post_attention_layernorm_bias5, alloc617) + R.vm.kill_object(gpt_neox_layers_14_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_14_post_attention_layernorm_bias5) + gpt_neox_layers_14_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[177] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[178] + gv987: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc618: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv987, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc617, gpt_neox_layers_14_mlp_dense_h_to_4h_weight5, gpt_neox_layers_14_mlp_dense_h_to_4h_bias5, alloc618) + R.vm.kill_object(alloc617) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_14_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[179] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[180] + gv988: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc619: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv988, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc618, gpt_neox_layers_14_mlp_dense_4h_to_h_weight5, gpt_neox_layers_14_mlp_dense_4h_to_h_bias5, alloc619) + R.vm.kill_object(alloc618) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_14_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[175] + gpt_neox_layers_14_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[176] + gv989: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc620: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv989, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape443, gpt_neox_layers_14_attention_dense_weight5, gpt_neox_layers_14_attention_dense_bias5, alloc619, alloc613, alloc620) + R.vm.kill_object(alloc613) + R.vm.kill_object(reshape443) + R.vm.kill_object(alloc619) + R.vm.kill_object(gpt_neox_layers_14_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_14_attention_dense_bias5) + gpt_neox_layers_15_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[181] + gpt_neox_layers_15_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[182] + gv990: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc621: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv990, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc620, gpt_neox_layers_15_input_layernorm_weight5, gpt_neox_layers_15_input_layernorm_bias5, alloc621) + R.vm.kill_object(gpt_neox_layers_15_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_15_input_layernorm_bias5) + gpt_neox_layers_15_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[185] + gpt_neox_layers_15_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[186] + gv991: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc622: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv991, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc621, gpt_neox_layers_15_attention_query_key_value_weight5, gpt_neox_layers_15_attention_query_key_value_bias5, alloc622) + R.vm.kill_object(alloc621) + R.vm.kill_object(gpt_neox_layers_15_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_15_attention_query_key_value_bias5) + gv992: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape444: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc622, gv992, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc622) + gv993: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape445: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape444, gv993, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape444) + gv994: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc623: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv994, R.dtype("float16")) + _622: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape445, alloc623) + R.vm.kill_object(reshape445) + gv995: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape446: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc623, gv995, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc623) + gv996: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape447: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape446, gv996, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape446) + gpt_neox_layers_15_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[183] + gpt_neox_layers_15_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[184] + gv997: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc624: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv997, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc620, gpt_neox_layers_15_post_attention_layernorm_weight5, gpt_neox_layers_15_post_attention_layernorm_bias5, alloc624) + R.vm.kill_object(gpt_neox_layers_15_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_15_post_attention_layernorm_bias5) + gpt_neox_layers_15_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[189] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[190] + gv998: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc625: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv998, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc624, gpt_neox_layers_15_mlp_dense_h_to_4h_weight5, gpt_neox_layers_15_mlp_dense_h_to_4h_bias5, alloc625) + R.vm.kill_object(alloc624) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_15_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[191] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[192] + gv999: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc626: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv999, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc625, gpt_neox_layers_15_mlp_dense_4h_to_h_weight5, gpt_neox_layers_15_mlp_dense_4h_to_h_bias5, alloc626) + R.vm.kill_object(alloc625) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_15_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[187] + gpt_neox_layers_15_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[188] + gv1000: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc627: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv1000, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape447, gpt_neox_layers_15_attention_dense_weight5, gpt_neox_layers_15_attention_dense_bias5, alloc626, alloc620, alloc627) + R.vm.kill_object(alloc620) + R.vm.kill_object(reshape447) + R.vm.kill_object(alloc626) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_bias5) + gpt_neox_layers_16_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[193] + gpt_neox_layers_16_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[194] + gv1001: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc628: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1001, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc627, gpt_neox_layers_16_input_layernorm_weight5, gpt_neox_layers_16_input_layernorm_bias5, alloc628) + R.vm.kill_object(gpt_neox_layers_16_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_16_input_layernorm_bias5) + gpt_neox_layers_16_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[197] + gpt_neox_layers_16_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[198] + gv1002: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc629: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1002, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc628, gpt_neox_layers_16_attention_query_key_value_weight5, gpt_neox_layers_16_attention_query_key_value_bias5, alloc629) + R.vm.kill_object(alloc628) + R.vm.kill_object(gpt_neox_layers_16_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_16_attention_query_key_value_bias5) + gv1003: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape448: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc629, gv1003, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc629) + gv1004: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape449: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape448, gv1004, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape448) + gv1005: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc630: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1005, R.dtype("float16")) + _629: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), reshape449, alloc630) + R.vm.kill_object(reshape449) + gv1006: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape450: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc630, gv1006, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc630) + gv1007: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape451: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape450, gv1007, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape450) + gpt_neox_layers_16_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[195] + gpt_neox_layers_16_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[196] + gv1008: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc631: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1008, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc627, gpt_neox_layers_16_post_attention_layernorm_weight5, gpt_neox_layers_16_post_attention_layernorm_bias5, alloc631) + R.vm.kill_object(gpt_neox_layers_16_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_16_post_attention_layernorm_bias5) + gpt_neox_layers_16_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[201] + gpt_neox_layers_16_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[202] + gv1009: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc632: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv1009, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc631, gpt_neox_layers_16_mlp_dense_h_to_4h_weight5, gpt_neox_layers_16_mlp_dense_h_to_4h_bias5, alloc632) + R.vm.kill_object(alloc631) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_16_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[203] + gpt_neox_layers_16_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[204] + gv1010: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc633: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1010, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc632, gpt_neox_layers_16_mlp_dense_4h_to_h_weight5, gpt_neox_layers_16_mlp_dense_4h_to_h_bias5, alloc633) + R.vm.kill_object(alloc632) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_16_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[199] + gpt_neox_layers_16_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[200] + gv1011: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc634: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv1011, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape451, gpt_neox_layers_16_attention_dense_weight5, gpt_neox_layers_16_attention_dense_bias5, alloc633, alloc627, alloc634) + R.vm.kill_object(alloc627) + R.vm.kill_object(reshape451) + R.vm.kill_object(alloc633) + R.vm.kill_object(gpt_neox_layers_16_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_16_attention_dense_bias5) + gpt_neox_layers_17_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[205] + gpt_neox_layers_17_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[206] + gv1012: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc635: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1012, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc634, gpt_neox_layers_17_input_layernorm_weight5, gpt_neox_layers_17_input_layernorm_bias5, alloc635) + R.vm.kill_object(gpt_neox_layers_17_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_17_input_layernorm_bias5) + gpt_neox_layers_17_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[209] + gpt_neox_layers_17_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[210] + gv1013: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc636: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1013, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc635, gpt_neox_layers_17_attention_query_key_value_weight5, gpt_neox_layers_17_attention_query_key_value_bias5, alloc636) + R.vm.kill_object(alloc635) + R.vm.kill_object(gpt_neox_layers_17_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_17_attention_query_key_value_bias5) + gv1014: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape452: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc636, gv1014, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc636) + gv1015: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape453: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape452, gv1015, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape452) + gv1016: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc637: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1016, R.dtype("float16")) + _636: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), reshape453, alloc637) + R.vm.kill_object(reshape453) + gv1017: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape454: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc637, gv1017, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc637) + gv1018: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape455: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape454, gv1018, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape454) + gpt_neox_layers_17_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[207] + gpt_neox_layers_17_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[208] + gv1019: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc638: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1019, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc634, gpt_neox_layers_17_post_attention_layernorm_weight5, gpt_neox_layers_17_post_attention_layernorm_bias5, alloc638) + R.vm.kill_object(gpt_neox_layers_17_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_17_post_attention_layernorm_bias5) + gpt_neox_layers_17_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[213] + gpt_neox_layers_17_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[214] + gv1020: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc639: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv1020, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc638, gpt_neox_layers_17_mlp_dense_h_to_4h_weight5, gpt_neox_layers_17_mlp_dense_h_to_4h_bias5, alloc639) + R.vm.kill_object(alloc638) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_17_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[215] + gpt_neox_layers_17_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[216] + gv1021: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc640: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1021, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc639, gpt_neox_layers_17_mlp_dense_4h_to_h_weight5, gpt_neox_layers_17_mlp_dense_4h_to_h_bias5, alloc640) + R.vm.kill_object(alloc639) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_17_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[211] + gpt_neox_layers_17_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[212] + gv1022: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc641: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv1022, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape455, gpt_neox_layers_17_attention_dense_weight5, gpt_neox_layers_17_attention_dense_bias5, alloc640, alloc634, alloc641) + R.vm.kill_object(alloc634) + R.vm.kill_object(reshape455) + R.vm.kill_object(alloc640) + R.vm.kill_object(gpt_neox_layers_17_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_17_attention_dense_bias5) + gpt_neox_layers_18_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[217] + gpt_neox_layers_18_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[218] + gv1023: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc642: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1023, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc641, gpt_neox_layers_18_input_layernorm_weight5, gpt_neox_layers_18_input_layernorm_bias5, alloc642) + R.vm.kill_object(gpt_neox_layers_18_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_18_input_layernorm_bias5) + gpt_neox_layers_18_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[221] + gpt_neox_layers_18_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[222] + gv1024: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc643: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1024, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc642, gpt_neox_layers_18_attention_query_key_value_weight5, gpt_neox_layers_18_attention_query_key_value_bias5, alloc643) + R.vm.kill_object(alloc642) + R.vm.kill_object(gpt_neox_layers_18_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_18_attention_query_key_value_bias5) + gv1025: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape456: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc643, gv1025, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc643) + gv1026: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape457: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape456, gv1026, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape456) + gv1027: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc644: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1027, R.dtype("float16")) + _643: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), reshape457, alloc644) + R.vm.kill_object(reshape457) + gv1028: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape458: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc644, gv1028, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc644) + gv1029: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape459: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape458, gv1029, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape458) + gpt_neox_layers_18_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[219] + gpt_neox_layers_18_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[220] + gv1030: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc645: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1030, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc641, gpt_neox_layers_18_post_attention_layernorm_weight5, gpt_neox_layers_18_post_attention_layernorm_bias5, alloc645) + R.vm.kill_object(gpt_neox_layers_18_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_18_post_attention_layernorm_bias5) + gpt_neox_layers_18_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[225] + gpt_neox_layers_18_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[226] + gv1031: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc646: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv1031, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc645, gpt_neox_layers_18_mlp_dense_h_to_4h_weight5, gpt_neox_layers_18_mlp_dense_h_to_4h_bias5, alloc646) + R.vm.kill_object(alloc645) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_18_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[227] + gpt_neox_layers_18_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[228] + gv1032: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc647: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1032, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc646, gpt_neox_layers_18_mlp_dense_4h_to_h_weight5, gpt_neox_layers_18_mlp_dense_4h_to_h_bias5, alloc647) + R.vm.kill_object(alloc646) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_18_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[223] + gpt_neox_layers_18_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[224] + gv1033: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc648: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv1033, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape459, gpt_neox_layers_18_attention_dense_weight5, gpt_neox_layers_18_attention_dense_bias5, alloc647, alloc641, alloc648) + R.vm.kill_object(alloc641) + R.vm.kill_object(reshape459) + R.vm.kill_object(alloc647) + R.vm.kill_object(gpt_neox_layers_18_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_18_attention_dense_bias5) + gpt_neox_layers_19_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[229] + gpt_neox_layers_19_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[230] + gv1034: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc649: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1034, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc648, gpt_neox_layers_19_input_layernorm_weight5, gpt_neox_layers_19_input_layernorm_bias5, alloc649) + R.vm.kill_object(gpt_neox_layers_19_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_19_input_layernorm_bias5) + gpt_neox_layers_19_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[233] + gpt_neox_layers_19_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[234] + gv1035: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc650: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1035, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc649, gpt_neox_layers_19_attention_query_key_value_weight5, gpt_neox_layers_19_attention_query_key_value_bias5, alloc650) + R.vm.kill_object(alloc649) + R.vm.kill_object(gpt_neox_layers_19_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_19_attention_query_key_value_bias5) + gv1036: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape460: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc650, gv1036, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc650) + gv1037: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape461: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape460, gv1037, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape460) + gv1038: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc651: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1038, R.dtype("float16")) + _650: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), reshape461, alloc651) + R.vm.kill_object(reshape461) + gv1039: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape462: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc651, gv1039, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc651) + gv1040: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape463: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape462, gv1040, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape462) + gpt_neox_layers_19_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[231] + gpt_neox_layers_19_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[232] + gv1041: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc652: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1041, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc648, gpt_neox_layers_19_post_attention_layernorm_weight5, gpt_neox_layers_19_post_attention_layernorm_bias5, alloc652) + R.vm.kill_object(gpt_neox_layers_19_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_19_post_attention_layernorm_bias5) + gpt_neox_layers_19_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[237] + gpt_neox_layers_19_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[238] + gv1042: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc653: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv1042, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc652, gpt_neox_layers_19_mlp_dense_h_to_4h_weight5, gpt_neox_layers_19_mlp_dense_h_to_4h_bias5, alloc653) + R.vm.kill_object(alloc652) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_19_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[239] + gpt_neox_layers_19_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[240] + gv1043: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc654: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1043, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc653, gpt_neox_layers_19_mlp_dense_4h_to_h_weight5, gpt_neox_layers_19_mlp_dense_4h_to_h_bias5, alloc654) + R.vm.kill_object(alloc653) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_19_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[235] + gpt_neox_layers_19_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[236] + gv1044: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc655: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv1044, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape463, gpt_neox_layers_19_attention_dense_weight5, gpt_neox_layers_19_attention_dense_bias5, alloc654, alloc648, alloc655) + R.vm.kill_object(alloc648) + R.vm.kill_object(reshape463) + R.vm.kill_object(alloc654) + R.vm.kill_object(gpt_neox_layers_19_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_19_attention_dense_bias5) + gpt_neox_layers_20_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[241] + gpt_neox_layers_20_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[242] + gv1045: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc656: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1045, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc655, gpt_neox_layers_20_input_layernorm_weight5, gpt_neox_layers_20_input_layernorm_bias5, alloc656) + R.vm.kill_object(gpt_neox_layers_20_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_20_input_layernorm_bias5) + gpt_neox_layers_20_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[245] + gpt_neox_layers_20_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[246] + gv1046: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc657: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1046, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc656, gpt_neox_layers_20_attention_query_key_value_weight5, gpt_neox_layers_20_attention_query_key_value_bias5, alloc657) + R.vm.kill_object(alloc656) + R.vm.kill_object(gpt_neox_layers_20_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_20_attention_query_key_value_bias5) + gv1047: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape464: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc657, gv1047, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc657) + gv1048: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape465: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape464, gv1048, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape464) + gv1049: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc658: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1049, R.dtype("float16")) + _657: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), reshape465, alloc658) + R.vm.kill_object(reshape465) + gv1050: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape466: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc658, gv1050, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc658) + gv1051: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape467: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape466, gv1051, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape466) + gpt_neox_layers_20_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[243] + gpt_neox_layers_20_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[244] + gv1052: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc659: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1052, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc655, gpt_neox_layers_20_post_attention_layernorm_weight5, gpt_neox_layers_20_post_attention_layernorm_bias5, alloc659) + R.vm.kill_object(gpt_neox_layers_20_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_20_post_attention_layernorm_bias5) + gpt_neox_layers_20_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[249] + gpt_neox_layers_20_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[250] + gv1053: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc660: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv1053, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc659, gpt_neox_layers_20_mlp_dense_h_to_4h_weight5, gpt_neox_layers_20_mlp_dense_h_to_4h_bias5, alloc660) + R.vm.kill_object(alloc659) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_20_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[251] + gpt_neox_layers_20_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[252] + gv1054: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc661: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1054, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc660, gpt_neox_layers_20_mlp_dense_4h_to_h_weight5, gpt_neox_layers_20_mlp_dense_4h_to_h_bias5, alloc661) + R.vm.kill_object(alloc660) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_20_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[247] + gpt_neox_layers_20_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[248] + gv1055: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc662: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv1055, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape467, gpt_neox_layers_20_attention_dense_weight5, gpt_neox_layers_20_attention_dense_bias5, alloc661, alloc655, alloc662) + R.vm.kill_object(alloc655) + R.vm.kill_object(reshape467) + R.vm.kill_object(alloc661) + R.vm.kill_object(gpt_neox_layers_20_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_20_attention_dense_bias5) + gpt_neox_layers_21_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[253] + gpt_neox_layers_21_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[254] + gv1056: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc663: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1056, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc662, gpt_neox_layers_21_input_layernorm_weight5, gpt_neox_layers_21_input_layernorm_bias5, alloc663) + R.vm.kill_object(gpt_neox_layers_21_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_21_input_layernorm_bias5) + gpt_neox_layers_21_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[257] + gpt_neox_layers_21_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[258] + gv1057: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc664: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1057, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc663, gpt_neox_layers_21_attention_query_key_value_weight5, gpt_neox_layers_21_attention_query_key_value_bias5, alloc664) + R.vm.kill_object(alloc663) + R.vm.kill_object(gpt_neox_layers_21_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_21_attention_query_key_value_bias5) + gv1058: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape468: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc664, gv1058, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc664) + gv1059: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape469: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape468, gv1059, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape468) + gv1060: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc665: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1060, R.dtype("float16")) + _664: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), reshape469, alloc665) + R.vm.kill_object(reshape469) + gv1061: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape470: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc665, gv1061, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc665) + gv1062: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape471: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape470, gv1062, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape470) + gpt_neox_layers_21_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[255] + gpt_neox_layers_21_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[256] + gv1063: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc666: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1063, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc662, gpt_neox_layers_21_post_attention_layernorm_weight5, gpt_neox_layers_21_post_attention_layernorm_bias5, alloc666) + R.vm.kill_object(gpt_neox_layers_21_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_21_post_attention_layernorm_bias5) + gpt_neox_layers_21_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[261] + gpt_neox_layers_21_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[262] + gv1064: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc667: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv1064, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc666, gpt_neox_layers_21_mlp_dense_h_to_4h_weight5, gpt_neox_layers_21_mlp_dense_h_to_4h_bias5, alloc667) + R.vm.kill_object(alloc666) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_21_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[263] + gpt_neox_layers_21_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[264] + gv1065: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc668: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1065, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc667, gpt_neox_layers_21_mlp_dense_4h_to_h_weight5, gpt_neox_layers_21_mlp_dense_4h_to_h_bias5, alloc668) + R.vm.kill_object(alloc667) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_21_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[259] + gpt_neox_layers_21_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[260] + gv1066: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc669: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv1066, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape471, gpt_neox_layers_21_attention_dense_weight5, gpt_neox_layers_21_attention_dense_bias5, alloc668, alloc662, alloc669) + R.vm.kill_object(alloc662) + R.vm.kill_object(reshape471) + R.vm.kill_object(alloc668) + R.vm.kill_object(gpt_neox_layers_21_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_21_attention_dense_bias5) + gpt_neox_layers_22_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[265] + gpt_neox_layers_22_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[266] + gv1067: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc670: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1067, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc669, gpt_neox_layers_22_input_layernorm_weight5, gpt_neox_layers_22_input_layernorm_bias5, alloc670) + R.vm.kill_object(gpt_neox_layers_22_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_22_input_layernorm_bias5) + gpt_neox_layers_22_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[269] + gpt_neox_layers_22_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[270] + gv1068: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc671: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1068, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc670, gpt_neox_layers_22_attention_query_key_value_weight5, gpt_neox_layers_22_attention_query_key_value_bias5, alloc671) + R.vm.kill_object(alloc670) + R.vm.kill_object(gpt_neox_layers_22_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_22_attention_query_key_value_bias5) + gv1069: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape472: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc671, gv1069, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc671) + gv1070: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape473: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape472, gv1070, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape472) + gv1071: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc672: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1071, R.dtype("float16")) + _671: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), reshape473, alloc672) + R.vm.kill_object(reshape473) + gv1072: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape474: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc672, gv1072, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc672) + gv1073: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape475: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape474, gv1073, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape474) + gpt_neox_layers_22_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[267] + gpt_neox_layers_22_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[268] + gv1074: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc673: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1074, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc669, gpt_neox_layers_22_post_attention_layernorm_weight5, gpt_neox_layers_22_post_attention_layernorm_bias5, alloc673) + R.vm.kill_object(gpt_neox_layers_22_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_22_post_attention_layernorm_bias5) + gpt_neox_layers_22_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[273] + gpt_neox_layers_22_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[274] + gv1075: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc674: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv1075, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc673, gpt_neox_layers_22_mlp_dense_h_to_4h_weight5, gpt_neox_layers_22_mlp_dense_h_to_4h_bias5, alloc674) + R.vm.kill_object(alloc673) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_22_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[275] + gpt_neox_layers_22_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[276] + gv1076: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc675: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1076, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc674, gpt_neox_layers_22_mlp_dense_4h_to_h_weight5, gpt_neox_layers_22_mlp_dense_4h_to_h_bias5, alloc675) + R.vm.kill_object(alloc674) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_22_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[271] + gpt_neox_layers_22_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[272] + gv1077: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc676: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv1077, R.dtype("float16")) + R.vm.kill_object(storage23) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape475, gpt_neox_layers_22_attention_dense_weight5, gpt_neox_layers_22_attention_dense_bias5, alloc675, alloc669, alloc676) + R.vm.kill_object(alloc669) + R.vm.kill_object(reshape475) + R.vm.kill_object(alloc675) + R.vm.kill_object(gpt_neox_layers_22_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_22_attention_dense_bias5) + gpt_neox_layers_23_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[277] + gpt_neox_layers_23_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[278] + gv1078: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc677: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1078, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc676, gpt_neox_layers_23_input_layernorm_weight5, gpt_neox_layers_23_input_layernorm_bias5, alloc677) + R.vm.kill_object(gpt_neox_layers_23_input_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_23_input_layernorm_bias5) + gpt_neox_layers_23_attention_query_key_value_weight5: R.Tensor((3072, 1024), dtype="float16") = packed_params[281] + gpt_neox_layers_23_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[282] + gv1079: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc678: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1079, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc677, gpt_neox_layers_23_attention_query_key_value_weight5, gpt_neox_layers_23_attention_query_key_value_bias5, alloc678) + R.vm.kill_object(alloc677) + R.vm.kill_object(gpt_neox_layers_23_attention_query_key_value_weight5) + R.vm.kill_object(gpt_neox_layers_23_attention_query_key_value_bias5) + gv1080: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape476: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc678, gv1080, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc678) + gv1081: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape477: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape476, gv1081, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape476) + gv1082: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc679: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1082, R.dtype("float16")) + _678: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), reshape477, alloc679) + R.vm.kill_object(reshape477) + gv1083: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape478: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc679, gv1083, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc679) + gv1084: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape479: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape478, gv1084, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape478) + gpt_neox_layers_23_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[279] + gpt_neox_layers_23_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[280] + gv1085: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc680: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1085, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc676, gpt_neox_layers_23_post_attention_layernorm_weight5, gpt_neox_layers_23_post_attention_layernorm_bias5, alloc680) + R.vm.kill_object(gpt_neox_layers_23_post_attention_layernorm_weight5) + R.vm.kill_object(gpt_neox_layers_23_post_attention_layernorm_bias5) + gpt_neox_layers_23_mlp_dense_h_to_4h_weight5: R.Tensor((4096, 1024), dtype="float16") = packed_params[285] + gpt_neox_layers_23_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[286] + gv1086: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc681: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv1086, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc680, gpt_neox_layers_23_mlp_dense_h_to_4h_weight5, gpt_neox_layers_23_mlp_dense_h_to_4h_bias5, alloc681) + R.vm.kill_object(alloc680) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_h_to_4h_weight5) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_h_to_4h_bias5) + gpt_neox_layers_23_mlp_dense_4h_to_h_weight5: R.Tensor((1024, 4096), dtype="float16") = packed_params[287] + gpt_neox_layers_23_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[288] + gv1087: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc682: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv1087, R.dtype("float16")) + R.vm.kill_object(storage22) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc681, gpt_neox_layers_23_mlp_dense_4h_to_h_weight5, gpt_neox_layers_23_mlp_dense_4h_to_h_bias5, alloc682) + R.vm.kill_object(alloc681) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_4h_to_h_weight5) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_4h_to_h_bias5) + gpt_neox_layers_23_attention_dense_weight5: R.Tensor((1024, 1024), dtype="float16") = packed_params[283] + gpt_neox_layers_23_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[284] + gv1088: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc683: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv1088, R.dtype("float16")) + R.vm.kill_object(storage24) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape479, gpt_neox_layers_23_attention_dense_weight5, gpt_neox_layers_23_attention_dense_bias5, alloc682, alloc676, alloc683) + R.vm.kill_object(alloc676) + R.vm.kill_object(reshape479) + R.vm.kill_object(alloc682) + R.vm.kill_object(gpt_neox_layers_23_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_23_attention_dense_bias5) + gpt_neox_final_layer_norm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[289] + gpt_neox_final_layer_norm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[290] + gv1089: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc684: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv1089, R.dtype("float16")) + R.vm.kill_object(storage21) + cls.gpt_neox_q0f16_layer_norm1(alloc683, gpt_neox_final_layer_norm_weight5, gpt_neox_final_layer_norm_bias5, alloc684) + R.vm.kill_object(alloc683) + R.vm.kill_object(gpt_neox_final_layer_norm_weight5) + R.vm.kill_object(gpt_neox_final_layer_norm_bias5) + embed_out_weight5: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[291] + gv1090: R.Shape(ndim=1) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(3), sinfo_args=(R.Shape(ndim=1),)) + storage25: R.Object = R.vm.alloc_storage(gv1090, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv1091: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(1), R.prim_value(1), sinfo_args=(R.Shape(ndim=3),)) + alloc685: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv1091, R.dtype("float32")) + R.vm.kill_object(storage25) + cls.gpt_neox_q0f16_fused_NT_matmul9_cast5(alloc684, embed_out_weight5, alloc685) + R.vm.kill_object(alloc684) + R.vm.kill_object(embed_out_weight5) + gv5: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc685, paged_kv_cache + R.vm.kill_object(alloc685) + gv1092: R.Tensor(dtype="float32", ndim=3) = gv5[0] + R.call_packed("vm.builtin.match_shape", gv1092, shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(3), R.prim_value(0), R.prim_value(3), R.prim_value(1), R.str("ErrorContext(fn=batch_verify, loc=return, annotation=R.Tuple(R.Tensor((1, seq_len, vocab_size), dtype=\"float32\"), R.Object)) "), sinfo_args=(R.Tuple,)) + return gv5 + + @R.function + def create_tir_paged_kv_cache(max_batch_size_: R.Shape(["max_batch_size"]), max_total_seq_len_: R.Shape(["max_total_seq_len"]), prefill_chunk_size_: R.Shape(["prefill_chunk_size"]), page_size_: R.Shape(["page_size"]), support_sliding_window_: R.Shape(["support_sliding_window"])) -> R.Object: + max_batch_size = T.int64() + max_total_seq_len = T.int64() + prefill_chunk_size = T.int64() + page_size = T.int64() + support_sliding_window = T.int64() + R.func_attr({"relax.force_pure": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) + cls = Module + shape_heap: R.Tensor(dtype="int64", ndim=1) = R.call_builtin_with_ctx("vm.builtin.alloc_shape_heap", (R.prim_value(5),), sinfo_args=(R.Tensor(dtype="int64", ndim=1),)) + R.call_packed("vm.builtin.check_shape_info", max_batch_size_, R.prim_value(1), R.str("ErrorContext(fn=create_tir_paged_kv_cache, loc=param[0], param=max_batch_size_, annotation=R.Shape([max_batch_size])) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_shape_info", max_total_seq_len_, R.prim_value(1), R.str("ErrorContext(fn=create_tir_paged_kv_cache, loc=param[1], param=max_total_seq_len_, annotation=R.Shape([max_total_seq_len])) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_shape_info", prefill_chunk_size_, R.prim_value(1), R.str("ErrorContext(fn=create_tir_paged_kv_cache, loc=param[2], param=prefill_chunk_size_, annotation=R.Shape([prefill_chunk_size])) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_shape_info", page_size_, R.prim_value(1), R.str("ErrorContext(fn=create_tir_paged_kv_cache, loc=param[3], param=page_size_, annotation=R.Shape([page_size])) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_shape_info", support_sliding_window_, R.prim_value(1), R.str("ErrorContext(fn=create_tir_paged_kv_cache, loc=param[4], param=support_sliding_window_, annotation=R.Shape([support_sliding_window])) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", max_batch_size_, shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(0), R.str("ErrorContext(fn=create_tir_paged_kv_cache, loc=param[0], param=max_batch_size_, annotation=R.Shape([max_batch_size])) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", max_total_seq_len_, shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(1), R.str("ErrorContext(fn=create_tir_paged_kv_cache, loc=param[1], param=max_total_seq_len_, annotation=R.Shape([max_total_seq_len])) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", prefill_chunk_size_, shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(2), R.str("ErrorContext(fn=create_tir_paged_kv_cache, loc=param[2], param=prefill_chunk_size_, annotation=R.Shape([prefill_chunk_size])) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", page_size_, shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(3), R.str("ErrorContext(fn=create_tir_paged_kv_cache, loc=param[3], param=page_size_, annotation=R.Shape([page_size])) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", support_sliding_window_, shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(4), R.str("ErrorContext(fn=create_tir_paged_kv_cache, loc=param[4], param=support_sliding_window_, annotation=R.Shape([support_sliding_window])) "), sinfo_args=(R.Tuple,)) + gv8: R.Shape(ndim=5) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(5), R.prim_value(1), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(1), R.prim_value(2), R.prim_value(1), R.prim_value(3), R.prim_value(1), R.prim_value(4), sinfo_args=(R.Shape(ndim=5),)) + paged_kv_cache: R.Object = R.call_packed("vm.builtin.paged_attention_kv_cache_create_reduced", gv8, R.shape([0, 24]), R.prim_value(16), R.prim_value(16), R.prim_value(64), R.prim_value(1), R.prim_value(1), R.prim_value(10000), R.const(0.0, "float16"), cls.gpt_neox_q0f16_tir_kv_cache_transpose_append, cls.gpt_neox_q0f16_batch_prefill_paged_kv, cls.gpt_neox_q0f16_batch_decode_paged_kv, cls.gpt_neox_q0f16_batch_prefill_paged_kv_sliding_window, cls.gpt_neox_q0f16_batch_decode_paged_kv_sliding_window, cls.gpt_neox_q0f16_batch_prefill_ragged_kv, cls.gpt_neox_q0f16_merge_state_inplace, cls.gpt_neox_q0f16_fused_rope, cls.gpt_neox_q0f16_copy_single_page, cls.gpt_neox_q0f16_tir_kv_cache_debug_get_kv, cls.gpt_neox_q0f16_compact_kv_copy, cls.gpt_neox_q0f16_batch_tree_attn, cls.gpt_neox_q0f16_tree_attn_paged_kv, R.prim_value(0), R.prim_value(0), sinfo_args=(R.Object,)) + return paged_kv_cache + + @R.function + def decode(input_embed: R.Tensor((1, 1, 1024), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor(("vocab_size", 1024), dtype="float16"))) -> R.Tuple(R.Tensor((1, 1, "vocab_size"), dtype="float32"), R.Object): + vocab_size = T.int64() + R.func_attr({"num_input": 2, "pipeline_parallel_stages": 1, "relax.force_pure": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) + cls = Module + shape_heap: R.Tensor(dtype="int64", ndim=1) = R.call_builtin_with_ctx("vm.builtin.alloc_shape_heap", (R.prim_value(3),), sinfo_args=(R.Tensor(dtype="int64", ndim=1),)) + R.call_packed("vm.builtin.check_tensor_info", input_embed, R.prim_value(3), R.dtype("float16"), R.str("ErrorContext(fn=decode, loc=param[0], param=input_embed, annotation=R.Tensor((1, 1, 1024), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(292), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv1093: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv1093, R.prim_value(2), R.dtype("float16"), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv1094: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[291] + R.call_packed("vm.builtin.check_tensor_info", gv1094, R.prim_value(2), R.dtype("float16"), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", input_embed, shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=decode, loc=param[0], param=input_embed, annotation=R.Tensor((1, 1, 1024), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv1093, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv1094, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(0), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func5(shape_heap) + gpt_neox_layers_0_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[2] + storage26: R.Object = R.vm.alloc_storage(R.shape([2048]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc686: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(input_embed, gpt_neox_layers_0_input_layernorm_weight2, gpt_neox_layers_0_input_layernorm_bias2, alloc686) + R.vm.kill_object(gpt_neox_layers_0_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_0_input_layernorm_bias2) + gpt_neox_layers_0_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[6] + storage27: R.Object = R.vm.alloc_storage(R.shape([6144]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc687: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc686, gpt_neox_layers_0_attention_query_key_value_weight2, gpt_neox_layers_0_attention_query_key_value_bias2, alloc687) + R.vm.kill_object(alloc686) + R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_bias2) + lv292: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc687, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc687) + alloc688: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _687: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), lv292, alloc688) + R.vm.kill_object(lv292) + lv293: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc688, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc688) + gpt_neox_layers_0_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[4] + alloc689: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(input_embed, gpt_neox_layers_0_post_attention_layernorm_weight2, gpt_neox_layers_0_post_attention_layernorm_bias2, alloc689) + R.vm.kill_object(gpt_neox_layers_0_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_0_post_attention_layernorm_bias2) + gpt_neox_layers_0_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[10] + storage28: R.Object = R.vm.alloc_storage(R.shape([8192]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc690: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc689, gpt_neox_layers_0_mlp_dense_h_to_4h_weight2, gpt_neox_layers_0_mlp_dense_h_to_4h_bias2, alloc690) + R.vm.kill_object(alloc689) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_0_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[12] + alloc691: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc690, gpt_neox_layers_0_mlp_dense_4h_to_h_weight2, gpt_neox_layers_0_mlp_dense_4h_to_h_bias2, alloc691) + R.vm.kill_object(alloc690) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_0_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[8] + alloc692: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv293, gpt_neox_layers_0_attention_dense_weight2, gpt_neox_layers_0_attention_dense_bias2, alloc691, input_embed, alloc692) + R.vm.kill_object(lv293) + R.vm.kill_object(alloc691) + R.vm.kill_object(gpt_neox_layers_0_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_0_attention_dense_bias2) + gpt_neox_layers_1_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[14] + alloc693: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc692, gpt_neox_layers_1_input_layernorm_weight2, gpt_neox_layers_1_input_layernorm_bias2, alloc693) + R.vm.kill_object(gpt_neox_layers_1_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_1_input_layernorm_bias2) + gpt_neox_layers_1_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[18] + alloc694: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc693, gpt_neox_layers_1_attention_query_key_value_weight2, gpt_neox_layers_1_attention_query_key_value_bias2, alloc694) + R.vm.kill_object(alloc693) + R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_bias2) + lv298: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc694, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc694) + alloc695: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _694: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), lv298, alloc695) + R.vm.kill_object(lv298) + lv299: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc695, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc695) + gpt_neox_layers_1_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[16] + alloc696: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc692, gpt_neox_layers_1_post_attention_layernorm_weight2, gpt_neox_layers_1_post_attention_layernorm_bias2, alloc696) + R.vm.kill_object(gpt_neox_layers_1_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_1_post_attention_layernorm_bias2) + gpt_neox_layers_1_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[22] + storage29: R.Object = R.vm.alloc_storage(R.shape([8192]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc697: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc696, gpt_neox_layers_1_mlp_dense_h_to_4h_weight2, gpt_neox_layers_1_mlp_dense_h_to_4h_bias2, alloc697) + R.vm.kill_object(alloc696) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_1_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[24] + alloc698: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc697, gpt_neox_layers_1_mlp_dense_4h_to_h_weight2, gpt_neox_layers_1_mlp_dense_4h_to_h_bias2, alloc698) + R.vm.kill_object(alloc697) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_1_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[20] + alloc699: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv299, gpt_neox_layers_1_attention_dense_weight2, gpt_neox_layers_1_attention_dense_bias2, alloc698, alloc692, alloc699) + R.vm.kill_object(alloc692) + R.vm.kill_object(lv299) + R.vm.kill_object(alloc698) + R.vm.kill_object(gpt_neox_layers_1_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_1_attention_dense_bias2) + gpt_neox_layers_2_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[26] + alloc700: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc699, gpt_neox_layers_2_input_layernorm_weight2, gpt_neox_layers_2_input_layernorm_bias2, alloc700) + R.vm.kill_object(gpt_neox_layers_2_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_2_input_layernorm_bias2) + gpt_neox_layers_2_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[30] + alloc701: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc700, gpt_neox_layers_2_attention_query_key_value_weight2, gpt_neox_layers_2_attention_query_key_value_bias2, alloc701) + R.vm.kill_object(alloc700) + R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_bias2) + lv304: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc701, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc701) + alloc702: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _701: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), lv304, alloc702) + R.vm.kill_object(lv304) + lv305: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc702, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc702) + gpt_neox_layers_2_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[28] + alloc703: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc699, gpt_neox_layers_2_post_attention_layernorm_weight2, gpt_neox_layers_2_post_attention_layernorm_bias2, alloc703) + R.vm.kill_object(gpt_neox_layers_2_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_2_post_attention_layernorm_bias2) + gpt_neox_layers_2_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[34] + alloc704: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc703, gpt_neox_layers_2_mlp_dense_h_to_4h_weight2, gpt_neox_layers_2_mlp_dense_h_to_4h_bias2, alloc704) + R.vm.kill_object(alloc703) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_2_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[36] + alloc705: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc704, gpt_neox_layers_2_mlp_dense_4h_to_h_weight2, gpt_neox_layers_2_mlp_dense_4h_to_h_bias2, alloc705) + R.vm.kill_object(alloc704) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_2_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[32] + alloc706: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv305, gpt_neox_layers_2_attention_dense_weight2, gpt_neox_layers_2_attention_dense_bias2, alloc705, alloc699, alloc706) + R.vm.kill_object(alloc699) + R.vm.kill_object(lv305) + R.vm.kill_object(alloc705) + R.vm.kill_object(gpt_neox_layers_2_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_2_attention_dense_bias2) + gpt_neox_layers_3_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[38] + alloc707: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc706, gpt_neox_layers_3_input_layernorm_weight2, gpt_neox_layers_3_input_layernorm_bias2, alloc707) + R.vm.kill_object(gpt_neox_layers_3_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_3_input_layernorm_bias2) + gpt_neox_layers_3_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[42] + alloc708: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc707, gpt_neox_layers_3_attention_query_key_value_weight2, gpt_neox_layers_3_attention_query_key_value_bias2, alloc708) + R.vm.kill_object(alloc707) + R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_bias2) + lv310: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc708, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc708) + alloc709: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _708: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), lv310, alloc709) + R.vm.kill_object(lv310) + lv311: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc709, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc709) + gpt_neox_layers_3_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[40] + alloc710: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc706, gpt_neox_layers_3_post_attention_layernorm_weight2, gpt_neox_layers_3_post_attention_layernorm_bias2, alloc710) + R.vm.kill_object(gpt_neox_layers_3_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_3_post_attention_layernorm_bias2) + gpt_neox_layers_3_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[46] + alloc711: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc710, gpt_neox_layers_3_mlp_dense_h_to_4h_weight2, gpt_neox_layers_3_mlp_dense_h_to_4h_bias2, alloc711) + R.vm.kill_object(alloc710) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_3_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[48] + alloc712: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc711, gpt_neox_layers_3_mlp_dense_4h_to_h_weight2, gpt_neox_layers_3_mlp_dense_4h_to_h_bias2, alloc712) + R.vm.kill_object(alloc711) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_3_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[44] + alloc713: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv311, gpt_neox_layers_3_attention_dense_weight2, gpt_neox_layers_3_attention_dense_bias2, alloc712, alloc706, alloc713) + R.vm.kill_object(alloc706) + R.vm.kill_object(lv311) + R.vm.kill_object(alloc712) + R.vm.kill_object(gpt_neox_layers_3_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_3_attention_dense_bias2) + gpt_neox_layers_4_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[50] + alloc714: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc713, gpt_neox_layers_4_input_layernorm_weight2, gpt_neox_layers_4_input_layernorm_bias2, alloc714) + R.vm.kill_object(gpt_neox_layers_4_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_4_input_layernorm_bias2) + gpt_neox_layers_4_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[54] + alloc715: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc714, gpt_neox_layers_4_attention_query_key_value_weight2, gpt_neox_layers_4_attention_query_key_value_bias2, alloc715) + R.vm.kill_object(alloc714) + R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_bias2) + lv316: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc715, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc715) + alloc716: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _715: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), lv316, alloc716) + R.vm.kill_object(lv316) + lv317: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc716, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc716) + gpt_neox_layers_4_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[52] + alloc717: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc713, gpt_neox_layers_4_post_attention_layernorm_weight2, gpt_neox_layers_4_post_attention_layernorm_bias2, alloc717) + R.vm.kill_object(gpt_neox_layers_4_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_4_post_attention_layernorm_bias2) + gpt_neox_layers_4_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[58] + alloc718: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc717, gpt_neox_layers_4_mlp_dense_h_to_4h_weight2, gpt_neox_layers_4_mlp_dense_h_to_4h_bias2, alloc718) + R.vm.kill_object(alloc717) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_4_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[60] + alloc719: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc718, gpt_neox_layers_4_mlp_dense_4h_to_h_weight2, gpt_neox_layers_4_mlp_dense_4h_to_h_bias2, alloc719) + R.vm.kill_object(alloc718) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_4_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[56] + alloc720: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv317, gpt_neox_layers_4_attention_dense_weight2, gpt_neox_layers_4_attention_dense_bias2, alloc719, alloc713, alloc720) + R.vm.kill_object(alloc713) + R.vm.kill_object(lv317) + R.vm.kill_object(alloc719) + R.vm.kill_object(gpt_neox_layers_4_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_4_attention_dense_bias2) + gpt_neox_layers_5_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[62] + alloc721: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc720, gpt_neox_layers_5_input_layernorm_weight2, gpt_neox_layers_5_input_layernorm_bias2, alloc721) + R.vm.kill_object(gpt_neox_layers_5_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_5_input_layernorm_bias2) + gpt_neox_layers_5_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[66] + alloc722: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc721, gpt_neox_layers_5_attention_query_key_value_weight2, gpt_neox_layers_5_attention_query_key_value_bias2, alloc722) + R.vm.kill_object(alloc721) + R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_bias2) + lv322: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc722, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc722) + alloc723: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _722: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), lv322, alloc723) + R.vm.kill_object(lv322) + lv323: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc723, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc723) + gpt_neox_layers_5_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[64] + alloc724: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc720, gpt_neox_layers_5_post_attention_layernorm_weight2, gpt_neox_layers_5_post_attention_layernorm_bias2, alloc724) + R.vm.kill_object(gpt_neox_layers_5_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_5_post_attention_layernorm_bias2) + gpt_neox_layers_5_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[70] + alloc725: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc724, gpt_neox_layers_5_mlp_dense_h_to_4h_weight2, gpt_neox_layers_5_mlp_dense_h_to_4h_bias2, alloc725) + R.vm.kill_object(alloc724) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_5_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[72] + alloc726: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc725, gpt_neox_layers_5_mlp_dense_4h_to_h_weight2, gpt_neox_layers_5_mlp_dense_4h_to_h_bias2, alloc726) + R.vm.kill_object(alloc725) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_5_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[68] + alloc727: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv323, gpt_neox_layers_5_attention_dense_weight2, gpt_neox_layers_5_attention_dense_bias2, alloc726, alloc720, alloc727) + R.vm.kill_object(alloc720) + R.vm.kill_object(lv323) + R.vm.kill_object(alloc726) + R.vm.kill_object(gpt_neox_layers_5_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_5_attention_dense_bias2) + gpt_neox_layers_6_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[74] + alloc728: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc727, gpt_neox_layers_6_input_layernorm_weight2, gpt_neox_layers_6_input_layernorm_bias2, alloc728) + R.vm.kill_object(gpt_neox_layers_6_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_6_input_layernorm_bias2) + gpt_neox_layers_6_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[78] + alloc729: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc728, gpt_neox_layers_6_attention_query_key_value_weight2, gpt_neox_layers_6_attention_query_key_value_bias2, alloc729) + R.vm.kill_object(alloc728) + R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_bias2) + lv328: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc729, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc729) + alloc730: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _729: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), lv328, alloc730) + R.vm.kill_object(lv328) + lv329: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc730, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc730) + gpt_neox_layers_6_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[76] + alloc731: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc727, gpt_neox_layers_6_post_attention_layernorm_weight2, gpt_neox_layers_6_post_attention_layernorm_bias2, alloc731) + R.vm.kill_object(gpt_neox_layers_6_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_6_post_attention_layernorm_bias2) + gpt_neox_layers_6_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[82] + alloc732: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc731, gpt_neox_layers_6_mlp_dense_h_to_4h_weight2, gpt_neox_layers_6_mlp_dense_h_to_4h_bias2, alloc732) + R.vm.kill_object(alloc731) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_6_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[84] + alloc733: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc732, gpt_neox_layers_6_mlp_dense_4h_to_h_weight2, gpt_neox_layers_6_mlp_dense_4h_to_h_bias2, alloc733) + R.vm.kill_object(alloc732) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_6_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[80] + alloc734: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv329, gpt_neox_layers_6_attention_dense_weight2, gpt_neox_layers_6_attention_dense_bias2, alloc733, alloc727, alloc734) + R.vm.kill_object(alloc727) + R.vm.kill_object(lv329) + R.vm.kill_object(alloc733) + R.vm.kill_object(gpt_neox_layers_6_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_6_attention_dense_bias2) + gpt_neox_layers_7_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[86] + alloc735: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc734, gpt_neox_layers_7_input_layernorm_weight2, gpt_neox_layers_7_input_layernorm_bias2, alloc735) + R.vm.kill_object(gpt_neox_layers_7_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_7_input_layernorm_bias2) + gpt_neox_layers_7_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[90] + alloc736: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc735, gpt_neox_layers_7_attention_query_key_value_weight2, gpt_neox_layers_7_attention_query_key_value_bias2, alloc736) + R.vm.kill_object(alloc735) + R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_bias2) + lv334: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc736, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc736) + alloc737: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _736: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), lv334, alloc737) + R.vm.kill_object(lv334) + lv335: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc737, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc737) + gpt_neox_layers_7_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[88] + alloc738: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc734, gpt_neox_layers_7_post_attention_layernorm_weight2, gpt_neox_layers_7_post_attention_layernorm_bias2, alloc738) + R.vm.kill_object(gpt_neox_layers_7_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_7_post_attention_layernorm_bias2) + gpt_neox_layers_7_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[94] + alloc739: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc738, gpt_neox_layers_7_mlp_dense_h_to_4h_weight2, gpt_neox_layers_7_mlp_dense_h_to_4h_bias2, alloc739) + R.vm.kill_object(alloc738) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_7_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[96] + alloc740: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc739, gpt_neox_layers_7_mlp_dense_4h_to_h_weight2, gpt_neox_layers_7_mlp_dense_4h_to_h_bias2, alloc740) + R.vm.kill_object(alloc739) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_7_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[92] + alloc741: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv335, gpt_neox_layers_7_attention_dense_weight2, gpt_neox_layers_7_attention_dense_bias2, alloc740, alloc734, alloc741) + R.vm.kill_object(alloc734) + R.vm.kill_object(lv335) + R.vm.kill_object(alloc740) + R.vm.kill_object(gpt_neox_layers_7_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_7_attention_dense_bias2) + gpt_neox_layers_8_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[98] + alloc742: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc741, gpt_neox_layers_8_input_layernorm_weight2, gpt_neox_layers_8_input_layernorm_bias2, alloc742) + R.vm.kill_object(gpt_neox_layers_8_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_8_input_layernorm_bias2) + gpt_neox_layers_8_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[102] + alloc743: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc742, gpt_neox_layers_8_attention_query_key_value_weight2, gpt_neox_layers_8_attention_query_key_value_bias2, alloc743) + R.vm.kill_object(alloc742) + R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_bias2) + lv340: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc743, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc743) + alloc744: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _743: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), lv340, alloc744) + R.vm.kill_object(lv340) + lv341: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc744, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc744) + gpt_neox_layers_8_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[100] + alloc745: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc741, gpt_neox_layers_8_post_attention_layernorm_weight2, gpt_neox_layers_8_post_attention_layernorm_bias2, alloc745) + R.vm.kill_object(gpt_neox_layers_8_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_8_post_attention_layernorm_bias2) + gpt_neox_layers_8_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[106] + alloc746: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc745, gpt_neox_layers_8_mlp_dense_h_to_4h_weight2, gpt_neox_layers_8_mlp_dense_h_to_4h_bias2, alloc746) + R.vm.kill_object(alloc745) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_8_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[108] + alloc747: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc746, gpt_neox_layers_8_mlp_dense_4h_to_h_weight2, gpt_neox_layers_8_mlp_dense_4h_to_h_bias2, alloc747) + R.vm.kill_object(alloc746) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_8_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[104] + alloc748: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv341, gpt_neox_layers_8_attention_dense_weight2, gpt_neox_layers_8_attention_dense_bias2, alloc747, alloc741, alloc748) + R.vm.kill_object(alloc741) + R.vm.kill_object(lv341) + R.vm.kill_object(alloc747) + R.vm.kill_object(gpt_neox_layers_8_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_8_attention_dense_bias2) + gpt_neox_layers_9_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[110] + alloc749: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc748, gpt_neox_layers_9_input_layernorm_weight2, gpt_neox_layers_9_input_layernorm_bias2, alloc749) + R.vm.kill_object(gpt_neox_layers_9_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_9_input_layernorm_bias2) + gpt_neox_layers_9_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[114] + alloc750: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc749, gpt_neox_layers_9_attention_query_key_value_weight2, gpt_neox_layers_9_attention_query_key_value_bias2, alloc750) + R.vm.kill_object(alloc749) + R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_bias2) + lv346: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc750, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc750) + alloc751: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _750: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), lv346, alloc751) + R.vm.kill_object(lv346) + lv347: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc751, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc751) + gpt_neox_layers_9_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[112] + alloc752: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc748, gpt_neox_layers_9_post_attention_layernorm_weight2, gpt_neox_layers_9_post_attention_layernorm_bias2, alloc752) + R.vm.kill_object(gpt_neox_layers_9_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_9_post_attention_layernorm_bias2) + gpt_neox_layers_9_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[118] + alloc753: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc752, gpt_neox_layers_9_mlp_dense_h_to_4h_weight2, gpt_neox_layers_9_mlp_dense_h_to_4h_bias2, alloc753) + R.vm.kill_object(alloc752) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_9_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[120] + alloc754: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc753, gpt_neox_layers_9_mlp_dense_4h_to_h_weight2, gpt_neox_layers_9_mlp_dense_4h_to_h_bias2, alloc754) + R.vm.kill_object(alloc753) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_9_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[116] + alloc755: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv347, gpt_neox_layers_9_attention_dense_weight2, gpt_neox_layers_9_attention_dense_bias2, alloc754, alloc748, alloc755) + R.vm.kill_object(alloc748) + R.vm.kill_object(lv347) + R.vm.kill_object(alloc754) + R.vm.kill_object(gpt_neox_layers_9_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_9_attention_dense_bias2) + gpt_neox_layers_10_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[122] + alloc756: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc755, gpt_neox_layers_10_input_layernorm_weight2, gpt_neox_layers_10_input_layernorm_bias2, alloc756) + R.vm.kill_object(gpt_neox_layers_10_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_10_input_layernorm_bias2) + gpt_neox_layers_10_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[126] + alloc757: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc756, gpt_neox_layers_10_attention_query_key_value_weight2, gpt_neox_layers_10_attention_query_key_value_bias2, alloc757) + R.vm.kill_object(alloc756) + R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_bias2) + lv352: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc757, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc757) + alloc758: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _757: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), lv352, alloc758) + R.vm.kill_object(lv352) + lv353: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc758, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc758) + gpt_neox_layers_10_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[124] + alloc759: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc755, gpt_neox_layers_10_post_attention_layernorm_weight2, gpt_neox_layers_10_post_attention_layernorm_bias2, alloc759) + R.vm.kill_object(gpt_neox_layers_10_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_10_post_attention_layernorm_bias2) + gpt_neox_layers_10_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[130] + alloc760: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc759, gpt_neox_layers_10_mlp_dense_h_to_4h_weight2, gpt_neox_layers_10_mlp_dense_h_to_4h_bias2, alloc760) + R.vm.kill_object(alloc759) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_10_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[132] + alloc761: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc760, gpt_neox_layers_10_mlp_dense_4h_to_h_weight2, gpt_neox_layers_10_mlp_dense_4h_to_h_bias2, alloc761) + R.vm.kill_object(alloc760) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_10_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[128] + alloc762: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv353, gpt_neox_layers_10_attention_dense_weight2, gpt_neox_layers_10_attention_dense_bias2, alloc761, alloc755, alloc762) + R.vm.kill_object(alloc755) + R.vm.kill_object(lv353) + R.vm.kill_object(alloc761) + R.vm.kill_object(gpt_neox_layers_10_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_10_attention_dense_bias2) + gpt_neox_layers_11_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[134] + alloc763: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc762, gpt_neox_layers_11_input_layernorm_weight2, gpt_neox_layers_11_input_layernorm_bias2, alloc763) + R.vm.kill_object(gpt_neox_layers_11_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_11_input_layernorm_bias2) + gpt_neox_layers_11_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[138] + alloc764: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc763, gpt_neox_layers_11_attention_query_key_value_weight2, gpt_neox_layers_11_attention_query_key_value_bias2, alloc764) + R.vm.kill_object(alloc763) + R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_bias2) + lv358: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc764, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc764) + alloc765: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _764: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), lv358, alloc765) + R.vm.kill_object(lv358) + lv359: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc765, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc765) + gpt_neox_layers_11_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[136] + alloc766: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc762, gpt_neox_layers_11_post_attention_layernorm_weight2, gpt_neox_layers_11_post_attention_layernorm_bias2, alloc766) + R.vm.kill_object(gpt_neox_layers_11_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_11_post_attention_layernorm_bias2) + gpt_neox_layers_11_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[142] + alloc767: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc766, gpt_neox_layers_11_mlp_dense_h_to_4h_weight2, gpt_neox_layers_11_mlp_dense_h_to_4h_bias2, alloc767) + R.vm.kill_object(alloc766) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_11_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[144] + alloc768: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc767, gpt_neox_layers_11_mlp_dense_4h_to_h_weight2, gpt_neox_layers_11_mlp_dense_4h_to_h_bias2, alloc768) + R.vm.kill_object(alloc767) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_11_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[140] + alloc769: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv359, gpt_neox_layers_11_attention_dense_weight2, gpt_neox_layers_11_attention_dense_bias2, alloc768, alloc762, alloc769) + R.vm.kill_object(alloc762) + R.vm.kill_object(lv359) + R.vm.kill_object(alloc768) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias2) + gpt_neox_layers_12_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[145] + gpt_neox_layers_12_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[146] + alloc770: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc769, gpt_neox_layers_12_input_layernorm_weight2, gpt_neox_layers_12_input_layernorm_bias2, alloc770) + R.vm.kill_object(gpt_neox_layers_12_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_12_input_layernorm_bias2) + gpt_neox_layers_12_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[149] + gpt_neox_layers_12_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[150] + alloc771: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc770, gpt_neox_layers_12_attention_query_key_value_weight2, gpt_neox_layers_12_attention_query_key_value_bias2, alloc771) + R.vm.kill_object(alloc770) + R.vm.kill_object(gpt_neox_layers_12_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_12_attention_query_key_value_bias2) + lv364: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc771, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc771) + alloc772: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _771: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), lv364, alloc772) + R.vm.kill_object(lv364) + lv365: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc772, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc772) + gpt_neox_layers_12_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[147] + gpt_neox_layers_12_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[148] + alloc773: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc769, gpt_neox_layers_12_post_attention_layernorm_weight2, gpt_neox_layers_12_post_attention_layernorm_bias2, alloc773) + R.vm.kill_object(gpt_neox_layers_12_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_12_post_attention_layernorm_bias2) + gpt_neox_layers_12_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[153] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[154] + alloc774: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc773, gpt_neox_layers_12_mlp_dense_h_to_4h_weight2, gpt_neox_layers_12_mlp_dense_h_to_4h_bias2, alloc774) + R.vm.kill_object(alloc773) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_12_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[155] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[156] + alloc775: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc774, gpt_neox_layers_12_mlp_dense_4h_to_h_weight2, gpt_neox_layers_12_mlp_dense_4h_to_h_bias2, alloc775) + R.vm.kill_object(alloc774) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_12_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[151] + gpt_neox_layers_12_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[152] + alloc776: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv365, gpt_neox_layers_12_attention_dense_weight2, gpt_neox_layers_12_attention_dense_bias2, alloc775, alloc769, alloc776) + R.vm.kill_object(alloc769) + R.vm.kill_object(lv365) + R.vm.kill_object(alloc775) + R.vm.kill_object(gpt_neox_layers_12_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_12_attention_dense_bias2) + gpt_neox_layers_13_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[157] + gpt_neox_layers_13_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[158] + alloc777: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc776, gpt_neox_layers_13_input_layernorm_weight2, gpt_neox_layers_13_input_layernorm_bias2, alloc777) + R.vm.kill_object(gpt_neox_layers_13_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_13_input_layernorm_bias2) + gpt_neox_layers_13_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[161] + gpt_neox_layers_13_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[162] + alloc778: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc777, gpt_neox_layers_13_attention_query_key_value_weight2, gpt_neox_layers_13_attention_query_key_value_bias2, alloc778) + R.vm.kill_object(alloc777) + R.vm.kill_object(gpt_neox_layers_13_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_13_attention_query_key_value_bias2) + lv370: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc778, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc778) + alloc779: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _778: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), lv370, alloc779) + R.vm.kill_object(lv370) + lv371: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc779, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc779) + gpt_neox_layers_13_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[159] + gpt_neox_layers_13_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[160] + alloc780: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc776, gpt_neox_layers_13_post_attention_layernorm_weight2, gpt_neox_layers_13_post_attention_layernorm_bias2, alloc780) + R.vm.kill_object(gpt_neox_layers_13_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_13_post_attention_layernorm_bias2) + gpt_neox_layers_13_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[165] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[166] + alloc781: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc780, gpt_neox_layers_13_mlp_dense_h_to_4h_weight2, gpt_neox_layers_13_mlp_dense_h_to_4h_bias2, alloc781) + R.vm.kill_object(alloc780) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_13_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[167] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[168] + alloc782: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc781, gpt_neox_layers_13_mlp_dense_4h_to_h_weight2, gpt_neox_layers_13_mlp_dense_4h_to_h_bias2, alloc782) + R.vm.kill_object(alloc781) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_13_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[163] + gpt_neox_layers_13_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[164] + alloc783: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv371, gpt_neox_layers_13_attention_dense_weight2, gpt_neox_layers_13_attention_dense_bias2, alloc782, alloc776, alloc783) + R.vm.kill_object(alloc776) + R.vm.kill_object(lv371) + R.vm.kill_object(alloc782) + R.vm.kill_object(gpt_neox_layers_13_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_13_attention_dense_bias2) + gpt_neox_layers_14_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[169] + gpt_neox_layers_14_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[170] + alloc784: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc783, gpt_neox_layers_14_input_layernorm_weight2, gpt_neox_layers_14_input_layernorm_bias2, alloc784) + R.vm.kill_object(gpt_neox_layers_14_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_14_input_layernorm_bias2) + gpt_neox_layers_14_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[173] + gpt_neox_layers_14_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[174] + alloc785: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc784, gpt_neox_layers_14_attention_query_key_value_weight2, gpt_neox_layers_14_attention_query_key_value_bias2, alloc785) + R.vm.kill_object(alloc784) + R.vm.kill_object(gpt_neox_layers_14_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_14_attention_query_key_value_bias2) + lv376: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc785, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc785) + alloc786: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _785: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), lv376, alloc786) + R.vm.kill_object(lv376) + lv377: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc786, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc786) + gpt_neox_layers_14_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[171] + gpt_neox_layers_14_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[172] + alloc787: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc783, gpt_neox_layers_14_post_attention_layernorm_weight2, gpt_neox_layers_14_post_attention_layernorm_bias2, alloc787) + R.vm.kill_object(gpt_neox_layers_14_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_14_post_attention_layernorm_bias2) + gpt_neox_layers_14_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[177] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[178] + alloc788: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc787, gpt_neox_layers_14_mlp_dense_h_to_4h_weight2, gpt_neox_layers_14_mlp_dense_h_to_4h_bias2, alloc788) + R.vm.kill_object(alloc787) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_14_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[179] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[180] + alloc789: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc788, gpt_neox_layers_14_mlp_dense_4h_to_h_weight2, gpt_neox_layers_14_mlp_dense_4h_to_h_bias2, alloc789) + R.vm.kill_object(alloc788) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_14_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[175] + gpt_neox_layers_14_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[176] + alloc790: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv377, gpt_neox_layers_14_attention_dense_weight2, gpt_neox_layers_14_attention_dense_bias2, alloc789, alloc783, alloc790) + R.vm.kill_object(alloc783) + R.vm.kill_object(lv377) + R.vm.kill_object(alloc789) + R.vm.kill_object(gpt_neox_layers_14_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_14_attention_dense_bias2) + gpt_neox_layers_15_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[181] + gpt_neox_layers_15_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[182] + alloc791: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc790, gpt_neox_layers_15_input_layernorm_weight2, gpt_neox_layers_15_input_layernorm_bias2, alloc791) + R.vm.kill_object(gpt_neox_layers_15_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_15_input_layernorm_bias2) + gpt_neox_layers_15_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[185] + gpt_neox_layers_15_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[186] + alloc792: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc791, gpt_neox_layers_15_attention_query_key_value_weight2, gpt_neox_layers_15_attention_query_key_value_bias2, alloc792) + R.vm.kill_object(alloc791) + R.vm.kill_object(gpt_neox_layers_15_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_15_attention_query_key_value_bias2) + lv382: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc792, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc792) + alloc793: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _792: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), lv382, alloc793) + R.vm.kill_object(lv382) + lv383: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc793, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc793) + gpt_neox_layers_15_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[183] + gpt_neox_layers_15_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[184] + alloc794: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc790, gpt_neox_layers_15_post_attention_layernorm_weight2, gpt_neox_layers_15_post_attention_layernorm_bias2, alloc794) + R.vm.kill_object(gpt_neox_layers_15_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_15_post_attention_layernorm_bias2) + gpt_neox_layers_15_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[189] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[190] + alloc795: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc794, gpt_neox_layers_15_mlp_dense_h_to_4h_weight2, gpt_neox_layers_15_mlp_dense_h_to_4h_bias2, alloc795) + R.vm.kill_object(alloc794) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_15_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[191] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[192] + alloc796: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc795, gpt_neox_layers_15_mlp_dense_4h_to_h_weight2, gpt_neox_layers_15_mlp_dense_4h_to_h_bias2, alloc796) + R.vm.kill_object(alloc795) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_15_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[187] + gpt_neox_layers_15_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[188] + alloc797: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv383, gpt_neox_layers_15_attention_dense_weight2, gpt_neox_layers_15_attention_dense_bias2, alloc796, alloc790, alloc797) + R.vm.kill_object(alloc790) + R.vm.kill_object(lv383) + R.vm.kill_object(alloc796) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_bias2) + gpt_neox_layers_16_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[193] + gpt_neox_layers_16_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[194] + alloc798: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc797, gpt_neox_layers_16_input_layernorm_weight2, gpt_neox_layers_16_input_layernorm_bias2, alloc798) + R.vm.kill_object(gpt_neox_layers_16_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_16_input_layernorm_bias2) + gpt_neox_layers_16_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[197] + gpt_neox_layers_16_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[198] + alloc799: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc798, gpt_neox_layers_16_attention_query_key_value_weight2, gpt_neox_layers_16_attention_query_key_value_bias2, alloc799) + R.vm.kill_object(alloc798) + R.vm.kill_object(gpt_neox_layers_16_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_16_attention_query_key_value_bias2) + lv388: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc799, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc799) + alloc800: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _799: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), lv388, alloc800) + R.vm.kill_object(lv388) + lv389: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc800, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc800) + gpt_neox_layers_16_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[195] + gpt_neox_layers_16_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[196] + alloc801: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc797, gpt_neox_layers_16_post_attention_layernorm_weight2, gpt_neox_layers_16_post_attention_layernorm_bias2, alloc801) + R.vm.kill_object(gpt_neox_layers_16_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_16_post_attention_layernorm_bias2) + gpt_neox_layers_16_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[201] + gpt_neox_layers_16_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[202] + alloc802: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc801, gpt_neox_layers_16_mlp_dense_h_to_4h_weight2, gpt_neox_layers_16_mlp_dense_h_to_4h_bias2, alloc802) + R.vm.kill_object(alloc801) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_16_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[203] + gpt_neox_layers_16_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[204] + alloc803: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc802, gpt_neox_layers_16_mlp_dense_4h_to_h_weight2, gpt_neox_layers_16_mlp_dense_4h_to_h_bias2, alloc803) + R.vm.kill_object(alloc802) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_16_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[199] + gpt_neox_layers_16_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[200] + alloc804: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv389, gpt_neox_layers_16_attention_dense_weight2, gpt_neox_layers_16_attention_dense_bias2, alloc803, alloc797, alloc804) + R.vm.kill_object(alloc797) + R.vm.kill_object(lv389) + R.vm.kill_object(alloc803) + R.vm.kill_object(gpt_neox_layers_16_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_16_attention_dense_bias2) + gpt_neox_layers_17_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[205] + gpt_neox_layers_17_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[206] + alloc805: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc804, gpt_neox_layers_17_input_layernorm_weight2, gpt_neox_layers_17_input_layernorm_bias2, alloc805) + R.vm.kill_object(gpt_neox_layers_17_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_17_input_layernorm_bias2) + gpt_neox_layers_17_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[209] + gpt_neox_layers_17_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[210] + alloc806: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc805, gpt_neox_layers_17_attention_query_key_value_weight2, gpt_neox_layers_17_attention_query_key_value_bias2, alloc806) + R.vm.kill_object(alloc805) + R.vm.kill_object(gpt_neox_layers_17_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_17_attention_query_key_value_bias2) + lv394: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc806, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc806) + alloc807: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _806: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), lv394, alloc807) + R.vm.kill_object(lv394) + lv395: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc807, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc807) + gpt_neox_layers_17_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[207] + gpt_neox_layers_17_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[208] + alloc808: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc804, gpt_neox_layers_17_post_attention_layernorm_weight2, gpt_neox_layers_17_post_attention_layernorm_bias2, alloc808) + R.vm.kill_object(gpt_neox_layers_17_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_17_post_attention_layernorm_bias2) + gpt_neox_layers_17_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[213] + gpt_neox_layers_17_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[214] + alloc809: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc808, gpt_neox_layers_17_mlp_dense_h_to_4h_weight2, gpt_neox_layers_17_mlp_dense_h_to_4h_bias2, alloc809) + R.vm.kill_object(alloc808) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_17_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[215] + gpt_neox_layers_17_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[216] + alloc810: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc809, gpt_neox_layers_17_mlp_dense_4h_to_h_weight2, gpt_neox_layers_17_mlp_dense_4h_to_h_bias2, alloc810) + R.vm.kill_object(alloc809) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_17_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[211] + gpt_neox_layers_17_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[212] + alloc811: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv395, gpt_neox_layers_17_attention_dense_weight2, gpt_neox_layers_17_attention_dense_bias2, alloc810, alloc804, alloc811) + R.vm.kill_object(alloc804) + R.vm.kill_object(lv395) + R.vm.kill_object(alloc810) + R.vm.kill_object(gpt_neox_layers_17_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_17_attention_dense_bias2) + gpt_neox_layers_18_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[217] + gpt_neox_layers_18_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[218] + alloc812: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc811, gpt_neox_layers_18_input_layernorm_weight2, gpt_neox_layers_18_input_layernorm_bias2, alloc812) + R.vm.kill_object(gpt_neox_layers_18_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_18_input_layernorm_bias2) + gpt_neox_layers_18_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[221] + gpt_neox_layers_18_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[222] + alloc813: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc812, gpt_neox_layers_18_attention_query_key_value_weight2, gpt_neox_layers_18_attention_query_key_value_bias2, alloc813) + R.vm.kill_object(alloc812) + R.vm.kill_object(gpt_neox_layers_18_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_18_attention_query_key_value_bias2) + lv400: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc813, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc813) + alloc814: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _813: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), lv400, alloc814) + R.vm.kill_object(lv400) + lv401: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc814, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc814) + gpt_neox_layers_18_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[219] + gpt_neox_layers_18_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[220] + alloc815: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc811, gpt_neox_layers_18_post_attention_layernorm_weight2, gpt_neox_layers_18_post_attention_layernorm_bias2, alloc815) + R.vm.kill_object(gpt_neox_layers_18_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_18_post_attention_layernorm_bias2) + gpt_neox_layers_18_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[225] + gpt_neox_layers_18_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[226] + alloc816: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc815, gpt_neox_layers_18_mlp_dense_h_to_4h_weight2, gpt_neox_layers_18_mlp_dense_h_to_4h_bias2, alloc816) + R.vm.kill_object(alloc815) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_18_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[227] + gpt_neox_layers_18_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[228] + alloc817: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc816, gpt_neox_layers_18_mlp_dense_4h_to_h_weight2, gpt_neox_layers_18_mlp_dense_4h_to_h_bias2, alloc817) + R.vm.kill_object(alloc816) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_18_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[223] + gpt_neox_layers_18_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[224] + alloc818: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv401, gpt_neox_layers_18_attention_dense_weight2, gpt_neox_layers_18_attention_dense_bias2, alloc817, alloc811, alloc818) + R.vm.kill_object(alloc811) + R.vm.kill_object(lv401) + R.vm.kill_object(alloc817) + R.vm.kill_object(gpt_neox_layers_18_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_18_attention_dense_bias2) + gpt_neox_layers_19_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[229] + gpt_neox_layers_19_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[230] + alloc819: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc818, gpt_neox_layers_19_input_layernorm_weight2, gpt_neox_layers_19_input_layernorm_bias2, alloc819) + R.vm.kill_object(gpt_neox_layers_19_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_19_input_layernorm_bias2) + gpt_neox_layers_19_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[233] + gpt_neox_layers_19_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[234] + alloc820: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc819, gpt_neox_layers_19_attention_query_key_value_weight2, gpt_neox_layers_19_attention_query_key_value_bias2, alloc820) + R.vm.kill_object(alloc819) + R.vm.kill_object(gpt_neox_layers_19_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_19_attention_query_key_value_bias2) + lv406: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc820, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc820) + alloc821: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _820: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), lv406, alloc821) + R.vm.kill_object(lv406) + lv407: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc821, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc821) + gpt_neox_layers_19_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[231] + gpt_neox_layers_19_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[232] + alloc822: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc818, gpt_neox_layers_19_post_attention_layernorm_weight2, gpt_neox_layers_19_post_attention_layernorm_bias2, alloc822) + R.vm.kill_object(gpt_neox_layers_19_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_19_post_attention_layernorm_bias2) + gpt_neox_layers_19_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[237] + gpt_neox_layers_19_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[238] + alloc823: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc822, gpt_neox_layers_19_mlp_dense_h_to_4h_weight2, gpt_neox_layers_19_mlp_dense_h_to_4h_bias2, alloc823) + R.vm.kill_object(alloc822) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_19_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[239] + gpt_neox_layers_19_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[240] + alloc824: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc823, gpt_neox_layers_19_mlp_dense_4h_to_h_weight2, gpt_neox_layers_19_mlp_dense_4h_to_h_bias2, alloc824) + R.vm.kill_object(alloc823) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_19_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[235] + gpt_neox_layers_19_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[236] + alloc825: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv407, gpt_neox_layers_19_attention_dense_weight2, gpt_neox_layers_19_attention_dense_bias2, alloc824, alloc818, alloc825) + R.vm.kill_object(alloc818) + R.vm.kill_object(lv407) + R.vm.kill_object(alloc824) + R.vm.kill_object(gpt_neox_layers_19_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_19_attention_dense_bias2) + gpt_neox_layers_20_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[241] + gpt_neox_layers_20_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[242] + alloc826: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc825, gpt_neox_layers_20_input_layernorm_weight2, gpt_neox_layers_20_input_layernorm_bias2, alloc826) + R.vm.kill_object(gpt_neox_layers_20_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_20_input_layernorm_bias2) + gpt_neox_layers_20_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[245] + gpt_neox_layers_20_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[246] + alloc827: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc826, gpt_neox_layers_20_attention_query_key_value_weight2, gpt_neox_layers_20_attention_query_key_value_bias2, alloc827) + R.vm.kill_object(alloc826) + R.vm.kill_object(gpt_neox_layers_20_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_20_attention_query_key_value_bias2) + lv412: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc827, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc827) + alloc828: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _827: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), lv412, alloc828) + R.vm.kill_object(lv412) + lv413: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc828, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc828) + gpt_neox_layers_20_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[243] + gpt_neox_layers_20_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[244] + alloc829: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc825, gpt_neox_layers_20_post_attention_layernorm_weight2, gpt_neox_layers_20_post_attention_layernorm_bias2, alloc829) + R.vm.kill_object(gpt_neox_layers_20_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_20_post_attention_layernorm_bias2) + gpt_neox_layers_20_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[249] + gpt_neox_layers_20_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[250] + alloc830: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc829, gpt_neox_layers_20_mlp_dense_h_to_4h_weight2, gpt_neox_layers_20_mlp_dense_h_to_4h_bias2, alloc830) + R.vm.kill_object(alloc829) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_20_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[251] + gpt_neox_layers_20_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[252] + alloc831: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc830, gpt_neox_layers_20_mlp_dense_4h_to_h_weight2, gpt_neox_layers_20_mlp_dense_4h_to_h_bias2, alloc831) + R.vm.kill_object(alloc830) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_20_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[247] + gpt_neox_layers_20_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[248] + alloc832: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv413, gpt_neox_layers_20_attention_dense_weight2, gpt_neox_layers_20_attention_dense_bias2, alloc831, alloc825, alloc832) + R.vm.kill_object(alloc825) + R.vm.kill_object(lv413) + R.vm.kill_object(alloc831) + R.vm.kill_object(gpt_neox_layers_20_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_20_attention_dense_bias2) + gpt_neox_layers_21_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[253] + gpt_neox_layers_21_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[254] + alloc833: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc832, gpt_neox_layers_21_input_layernorm_weight2, gpt_neox_layers_21_input_layernorm_bias2, alloc833) + R.vm.kill_object(gpt_neox_layers_21_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_21_input_layernorm_bias2) + gpt_neox_layers_21_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[257] + gpt_neox_layers_21_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[258] + alloc834: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc833, gpt_neox_layers_21_attention_query_key_value_weight2, gpt_neox_layers_21_attention_query_key_value_bias2, alloc834) + R.vm.kill_object(alloc833) + R.vm.kill_object(gpt_neox_layers_21_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_21_attention_query_key_value_bias2) + lv418: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc834, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc834) + alloc835: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _834: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), lv418, alloc835) + R.vm.kill_object(lv418) + lv419: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc835, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc835) + gpt_neox_layers_21_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[255] + gpt_neox_layers_21_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[256] + alloc836: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc832, gpt_neox_layers_21_post_attention_layernorm_weight2, gpt_neox_layers_21_post_attention_layernorm_bias2, alloc836) + R.vm.kill_object(gpt_neox_layers_21_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_21_post_attention_layernorm_bias2) + gpt_neox_layers_21_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[261] + gpt_neox_layers_21_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[262] + alloc837: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc836, gpt_neox_layers_21_mlp_dense_h_to_4h_weight2, gpt_neox_layers_21_mlp_dense_h_to_4h_bias2, alloc837) + R.vm.kill_object(alloc836) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_21_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[263] + gpt_neox_layers_21_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[264] + alloc838: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc837, gpt_neox_layers_21_mlp_dense_4h_to_h_weight2, gpt_neox_layers_21_mlp_dense_4h_to_h_bias2, alloc838) + R.vm.kill_object(alloc837) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_21_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[259] + gpt_neox_layers_21_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[260] + alloc839: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv419, gpt_neox_layers_21_attention_dense_weight2, gpt_neox_layers_21_attention_dense_bias2, alloc838, alloc832, alloc839) + R.vm.kill_object(alloc832) + R.vm.kill_object(lv419) + R.vm.kill_object(alloc838) + R.vm.kill_object(gpt_neox_layers_21_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_21_attention_dense_bias2) + gpt_neox_layers_22_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[265] + gpt_neox_layers_22_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[266] + alloc840: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc839, gpt_neox_layers_22_input_layernorm_weight2, gpt_neox_layers_22_input_layernorm_bias2, alloc840) + R.vm.kill_object(gpt_neox_layers_22_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_22_input_layernorm_bias2) + gpt_neox_layers_22_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[269] + gpt_neox_layers_22_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[270] + alloc841: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc840, gpt_neox_layers_22_attention_query_key_value_weight2, gpt_neox_layers_22_attention_query_key_value_bias2, alloc841) + R.vm.kill_object(alloc840) + R.vm.kill_object(gpt_neox_layers_22_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_22_attention_query_key_value_bias2) + lv424: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc841, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc841) + alloc842: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _841: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), lv424, alloc842) + R.vm.kill_object(lv424) + lv425: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc842, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc842) + gpt_neox_layers_22_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[267] + gpt_neox_layers_22_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[268] + alloc843: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc839, gpt_neox_layers_22_post_attention_layernorm_weight2, gpt_neox_layers_22_post_attention_layernorm_bias2, alloc843) + R.vm.kill_object(gpt_neox_layers_22_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_22_post_attention_layernorm_bias2) + gpt_neox_layers_22_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[273] + gpt_neox_layers_22_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[274] + alloc844: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc843, gpt_neox_layers_22_mlp_dense_h_to_4h_weight2, gpt_neox_layers_22_mlp_dense_h_to_4h_bias2, alloc844) + R.vm.kill_object(alloc843) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_22_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[275] + gpt_neox_layers_22_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[276] + alloc845: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc844, gpt_neox_layers_22_mlp_dense_4h_to_h_weight2, gpt_neox_layers_22_mlp_dense_4h_to_h_bias2, alloc845) + R.vm.kill_object(alloc844) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_22_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[271] + gpt_neox_layers_22_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[272] + alloc846: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage28, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + R.vm.kill_object(storage28) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv425, gpt_neox_layers_22_attention_dense_weight2, gpt_neox_layers_22_attention_dense_bias2, alloc845, alloc839, alloc846) + R.vm.kill_object(alloc839) + R.vm.kill_object(lv425) + R.vm.kill_object(alloc845) + R.vm.kill_object(gpt_neox_layers_22_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_22_attention_dense_bias2) + gpt_neox_layers_23_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[277] + gpt_neox_layers_23_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[278] + alloc847: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc846, gpt_neox_layers_23_input_layernorm_weight2, gpt_neox_layers_23_input_layernorm_bias2, alloc847) + R.vm.kill_object(gpt_neox_layers_23_input_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_23_input_layernorm_bias2) + gpt_neox_layers_23_attention_query_key_value_weight2: R.Tensor((3072, 1024), dtype="float16") = packed_params[281] + gpt_neox_layers_23_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[282] + alloc848: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc847, gpt_neox_layers_23_attention_query_key_value_weight2, gpt_neox_layers_23_attention_query_key_value_bias2, alloc848) + R.vm.kill_object(alloc847) + R.vm.kill_object(gpt_neox_layers_23_attention_query_key_value_weight2) + R.vm.kill_object(gpt_neox_layers_23_attention_query_key_value_bias2) + lv430: R.Tensor((1, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc848, R.shape([1, 48, 64]), sinfo_args=(R.Tensor((1, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc848) + alloc849: R.Tensor((1, 16, 64), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 16, 64]), R.dtype("float16")) + _848: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), lv430, alloc849) + R.vm.kill_object(lv430) + lv431: R.Tensor((1, 1, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc849, R.shape([1, 1, 1024]), sinfo_args=(R.Tensor((1, 1, 1024), dtype="float16"),)) + R.vm.kill_object(alloc849) + gpt_neox_layers_23_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[279] + gpt_neox_layers_23_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[280] + alloc850: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc846, gpt_neox_layers_23_post_attention_layernorm_weight2, gpt_neox_layers_23_post_attention_layernorm_bias2, alloc850) + R.vm.kill_object(gpt_neox_layers_23_post_attention_layernorm_weight2) + R.vm.kill_object(gpt_neox_layers_23_post_attention_layernorm_bias2) + gpt_neox_layers_23_mlp_dense_h_to_4h_weight2: R.Tensor((4096, 1024), dtype="float16") = packed_params[285] + gpt_neox_layers_23_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[286] + alloc851: R.Tensor((1, 1, 4096), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 4096]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc850, gpt_neox_layers_23_mlp_dense_h_to_4h_weight2, gpt_neox_layers_23_mlp_dense_h_to_4h_bias2, alloc851) + R.vm.kill_object(alloc850) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_h_to_4h_weight2) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_h_to_4h_bias2) + gpt_neox_layers_23_mlp_dense_4h_to_h_weight2: R.Tensor((1024, 4096), dtype="float16") = packed_params[287] + gpt_neox_layers_23_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[288] + alloc852: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage27, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + R.vm.kill_object(storage27) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc851, gpt_neox_layers_23_mlp_dense_4h_to_h_weight2, gpt_neox_layers_23_mlp_dense_4h_to_h_bias2, alloc852) + R.vm.kill_object(alloc851) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_4h_to_h_weight2) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_4h_to_h_bias2) + gpt_neox_layers_23_attention_dense_weight2: R.Tensor((1024, 1024), dtype="float16") = packed_params[283] + gpt_neox_layers_23_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[284] + alloc853: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + R.vm.kill_object(storage29) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv431, gpt_neox_layers_23_attention_dense_weight2, gpt_neox_layers_23_attention_dense_bias2, alloc852, alloc846, alloc853) + R.vm.kill_object(alloc846) + R.vm.kill_object(lv431) + R.vm.kill_object(alloc852) + R.vm.kill_object(gpt_neox_layers_23_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_23_attention_dense_bias2) + gpt_neox_final_layer_norm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[289] + gpt_neox_final_layer_norm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[290] + alloc854: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage26, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + R.vm.kill_object(storage26) + cls.gpt_neox_q0f16_layer_norm2(alloc853, gpt_neox_final_layer_norm_weight2, gpt_neox_final_layer_norm_bias2, alloc854) + R.vm.kill_object(alloc853) + R.vm.kill_object(gpt_neox_final_layer_norm_weight2) + R.vm.kill_object(gpt_neox_final_layer_norm_bias2) + embed_out_weight2: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[291] + gv1095: R.Shape(ndim=1) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(2), sinfo_args=(R.Shape(ndim=1),)) + storage30: R.Object = R.vm.alloc_storage(gv1095, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv1096: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), sinfo_args=(R.Shape(ndim=3),)) + alloc855: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage30, R.prim_value(0), gv1096, R.dtype("float32")) + R.vm.kill_object(storage30) + cls.gpt_neox_q0f16_fused_NT_matmul14_cast8(alloc854, embed_out_weight2, alloc855) + R.vm.kill_object(alloc854) + R.vm.kill_object(embed_out_weight2) + gv2: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc855, paged_kv_cache + R.vm.kill_object(alloc855) + gv1097: R.Tensor(dtype="float32", ndim=3) = gv2[0] + R.call_packed("vm.builtin.match_shape", gv1097, shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1), R.prim_value(3), R.prim_value(0), R.str("ErrorContext(fn=decode, loc=return, annotation=R.Tuple(R.Tensor((1, 1, vocab_size), dtype=\"float32\"), R.Object)) "), sinfo_args=(R.Tuple,)) + return gv2 + + @R.function + def embed(input_ids: R.Tensor(("seq_len",), dtype="int32"), packed_params: R.Tuple(R.Tensor(("vocab_size", 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor(("vocab_size", 1024), dtype="float16"))) -> R.Tensor(("seq_len", 1024), dtype="float16"): + seq_len = T.int64() + vocab_size = T.int64() + R.func_attr({"num_input": 1, "relax.force_pure": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) + cls = Module + shape_heap: R.Tensor(dtype="int64", ndim=1) = R.call_builtin_with_ctx("vm.builtin.alloc_shape_heap", (R.prim_value(2),), sinfo_args=(R.Tensor(dtype="int64", ndim=1),)) + R.call_packed("vm.builtin.check_tensor_info", input_ids, R.prim_value(1), R.dtype("int32"), R.str("ErrorContext(fn=embed, loc=param[0], param=input_ids, annotation=R.Tensor((seq_len,), dtype=\"int32\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(292), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv820: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv820, R.prim_value(2), R.dtype("float16"), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv821: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[291] + R.call_packed("vm.builtin.check_tensor_info", gv821, R.prim_value(2), R.dtype("float16"), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", input_ids, shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(0), R.str("ErrorContext(fn=embed, loc=param[0], param=input_ids, annotation=R.Tensor((seq_len,), dtype=\"int32\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv820, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv821, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gpt_neox_embed_in_weight: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[0] + storage20: R.Object = R.vm.alloc_storage(R.shape([4194304]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv822: R.Shape(ndim=2) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=2),)) + alloc515: R.Tensor(dtype="float16", ndim=2) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv822, R.dtype("float16")) + R.vm.kill_object(storage20) + cls.gpt_neox_q0f16_take1(gpt_neox_embed_in_weight, input_ids, alloc515) + R.vm.kill_object(gpt_neox_embed_in_weight) + R.call_packed("vm.builtin.match_shape", alloc515, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(0), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=embed, loc=return, annotation=R.Tensor((seq_len, 1024), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + return alloc515 + + @R.function + def prefill(input_embed: R.Tensor((1, "seq_len", 1024), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 1024), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 1024), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 1024), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 4096), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor(("vocab_size", 1024), dtype="float16"))) -> R.Tuple(R.Tensor((1, 1, "vocab_size"), dtype="float32"), R.Object): + vocab_size = T.int64() + seq_len = T.int64() + R.func_attr({"num_input": 2, "pipeline_parallel_stages": 1, "relax.force_pure": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) + cls = Module + shape_heap: R.Tensor(dtype="int64", ndim=1) = R.call_builtin_with_ctx("vm.builtin.alloc_shape_heap", (R.prim_value(4),), sinfo_args=(R.Tensor(dtype="int64", ndim=1),)) + R.call_packed("vm.builtin.check_tensor_info", input_embed, R.prim_value(3), R.dtype("float16"), R.str("ErrorContext(fn=prefill, loc=param[0], param=input_embed, annotation=R.Tensor((1, seq_len, 1024), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(292), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv279: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv279, R.prim_value(2), R.dtype("float16"), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv280: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[291] + R.call_packed("vm.builtin.check_tensor_info", gv280, R.prim_value(2), R.dtype("float16"), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", input_embed, shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=prefill, loc=param[0], param=input_embed, annotation=R.Tensor((1, seq_len, 1024), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv279, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv280, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(1024), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((3072, 1024), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float16\"), R.Tensor((1024, 1024), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((4096, 1024), dtype=\"float16\"), R.Tensor((4096,), dtype=\"float32\"), R.Tensor((1024, 4096), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float32\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((1024,), dtype=\"float16\"), R.Tensor((vocab_size, 1024), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func2(shape_heap) + gpt_neox_layers_0_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[2] + storage8: R.Object = R.vm.alloc_storage(R.shape([4194304]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv281: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc173: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv281, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(input_embed, gpt_neox_layers_0_input_layernorm_weight1, gpt_neox_layers_0_input_layernorm_bias1, alloc173) + R.vm.kill_object(gpt_neox_layers_0_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_0_input_layernorm_bias1) + gpt_neox_layers_0_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[6] + storage9: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv282: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc174: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv282, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc173, gpt_neox_layers_0_attention_query_key_value_weight1, gpt_neox_layers_0_attention_query_key_value_bias1, alloc174) + R.vm.kill_object(alloc173) + R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_bias1) + gv283: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc174, gv283, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc174) + gv284: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape1: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape, gv284, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape) + gv285: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc175: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv285, R.dtype("float16")) + _174: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), reshape1, alloc175) + R.vm.kill_object(reshape1) + gv286: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape2: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc175, gv286, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc175) + gv287: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape3: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape2, gv287, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape2) + gpt_neox_layers_0_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[4] + gv288: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc176: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv288, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(input_embed, gpt_neox_layers_0_post_attention_layernorm_weight1, gpt_neox_layers_0_post_attention_layernorm_bias1, alloc176) + R.vm.kill_object(gpt_neox_layers_0_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_0_post_attention_layernorm_bias1) + gpt_neox_layers_0_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[10] + storage10: R.Object = R.vm.alloc_storage(R.shape([16777216]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv289: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc177: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv289, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc176, gpt_neox_layers_0_mlp_dense_h_to_4h_weight1, gpt_neox_layers_0_mlp_dense_h_to_4h_bias1, alloc177) + R.vm.kill_object(alloc176) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_0_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[12] + gv290: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc178: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv290, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc177, gpt_neox_layers_0_mlp_dense_4h_to_h_weight1, gpt_neox_layers_0_mlp_dense_4h_to_h_bias1, alloc178) + R.vm.kill_object(alloc177) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_0_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[8] + gv291: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc179: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv291, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape3, gpt_neox_layers_0_attention_dense_weight1, gpt_neox_layers_0_attention_dense_bias1, alloc178, input_embed, alloc179) + R.vm.kill_object(reshape3) + R.vm.kill_object(alloc178) + R.vm.kill_object(gpt_neox_layers_0_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_0_attention_dense_bias1) + gpt_neox_layers_1_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[14] + gv292: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc180: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv292, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc179, gpt_neox_layers_1_input_layernorm_weight1, gpt_neox_layers_1_input_layernorm_bias1, alloc180) + R.vm.kill_object(gpt_neox_layers_1_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_1_input_layernorm_bias1) + gpt_neox_layers_1_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[18] + gv293: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc181: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv293, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc180, gpt_neox_layers_1_attention_query_key_value_weight1, gpt_neox_layers_1_attention_query_key_value_bias1, alloc181) + R.vm.kill_object(alloc180) + R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_bias1) + gv294: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape4: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc181, gv294, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc181) + gv295: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape5: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape4, gv295, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape4) + gv296: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc182: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv296, R.dtype("float16")) + _181: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), reshape5, alloc182) + R.vm.kill_object(reshape5) + gv297: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape6: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc182, gv297, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc182) + gv298: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape7: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape6, gv298, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape6) + gpt_neox_layers_1_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[16] + gv299: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc183: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv299, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc179, gpt_neox_layers_1_post_attention_layernorm_weight1, gpt_neox_layers_1_post_attention_layernorm_bias1, alloc183) + R.vm.kill_object(gpt_neox_layers_1_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_1_post_attention_layernorm_bias1) + gpt_neox_layers_1_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[22] + storage11: R.Object = R.vm.alloc_storage(R.shape([16777216]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv300: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc184: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv300, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc183, gpt_neox_layers_1_mlp_dense_h_to_4h_weight1, gpt_neox_layers_1_mlp_dense_h_to_4h_bias1, alloc184) + R.vm.kill_object(alloc183) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_1_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[24] + gv301: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc185: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv301, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc184, gpt_neox_layers_1_mlp_dense_4h_to_h_weight1, gpt_neox_layers_1_mlp_dense_4h_to_h_bias1, alloc185) + R.vm.kill_object(alloc184) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_1_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[20] + gv302: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc186: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv302, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape7, gpt_neox_layers_1_attention_dense_weight1, gpt_neox_layers_1_attention_dense_bias1, alloc185, alloc179, alloc186) + R.vm.kill_object(alloc179) + R.vm.kill_object(reshape7) + R.vm.kill_object(alloc185) + R.vm.kill_object(gpt_neox_layers_1_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_1_attention_dense_bias1) + gpt_neox_layers_2_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[26] + gv303: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc187: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv303, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc186, gpt_neox_layers_2_input_layernorm_weight1, gpt_neox_layers_2_input_layernorm_bias1, alloc187) + R.vm.kill_object(gpt_neox_layers_2_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_2_input_layernorm_bias1) + gpt_neox_layers_2_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[30] + gv304: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc188: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv304, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc187, gpt_neox_layers_2_attention_query_key_value_weight1, gpt_neox_layers_2_attention_query_key_value_bias1, alloc188) + R.vm.kill_object(alloc187) + R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_bias1) + gv305: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape8: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc188, gv305, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc188) + gv306: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape9: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape8, gv306, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape8) + gv307: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc189: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv307, R.dtype("float16")) + _188: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), reshape9, alloc189) + R.vm.kill_object(reshape9) + gv308: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape10: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc189, gv308, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc189) + gv309: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape11: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape10, gv309, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape10) + gpt_neox_layers_2_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[28] + gv310: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc190: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv310, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc186, gpt_neox_layers_2_post_attention_layernorm_weight1, gpt_neox_layers_2_post_attention_layernorm_bias1, alloc190) + R.vm.kill_object(gpt_neox_layers_2_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_2_post_attention_layernorm_bias1) + gpt_neox_layers_2_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[34] + gv311: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc191: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv311, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc190, gpt_neox_layers_2_mlp_dense_h_to_4h_weight1, gpt_neox_layers_2_mlp_dense_h_to_4h_bias1, alloc191) + R.vm.kill_object(alloc190) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_2_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[36] + gv312: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc192: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv312, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc191, gpt_neox_layers_2_mlp_dense_4h_to_h_weight1, gpt_neox_layers_2_mlp_dense_4h_to_h_bias1, alloc192) + R.vm.kill_object(alloc191) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_2_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[32] + gv313: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc193: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv313, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape11, gpt_neox_layers_2_attention_dense_weight1, gpt_neox_layers_2_attention_dense_bias1, alloc192, alloc186, alloc193) + R.vm.kill_object(alloc186) + R.vm.kill_object(reshape11) + R.vm.kill_object(alloc192) + R.vm.kill_object(gpt_neox_layers_2_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_2_attention_dense_bias1) + gpt_neox_layers_3_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[38] + gv314: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc194: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv314, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc193, gpt_neox_layers_3_input_layernorm_weight1, gpt_neox_layers_3_input_layernorm_bias1, alloc194) + R.vm.kill_object(gpt_neox_layers_3_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_3_input_layernorm_bias1) + gpt_neox_layers_3_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[42] + gv315: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc195: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv315, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc194, gpt_neox_layers_3_attention_query_key_value_weight1, gpt_neox_layers_3_attention_query_key_value_bias1, alloc195) + R.vm.kill_object(alloc194) + R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_bias1) + gv316: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape12: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc195, gv316, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc195) + gv317: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape13: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape12, gv317, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape12) + gv318: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc196: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv318, R.dtype("float16")) + _195: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), reshape13, alloc196) + R.vm.kill_object(reshape13) + gv319: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape14: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc196, gv319, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc196) + gv320: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape15: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape14, gv320, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape14) + gpt_neox_layers_3_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[40] + gv321: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc197: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv321, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc193, gpt_neox_layers_3_post_attention_layernorm_weight1, gpt_neox_layers_3_post_attention_layernorm_bias1, alloc197) + R.vm.kill_object(gpt_neox_layers_3_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_3_post_attention_layernorm_bias1) + gpt_neox_layers_3_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[46] + gv322: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc198: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv322, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc197, gpt_neox_layers_3_mlp_dense_h_to_4h_weight1, gpt_neox_layers_3_mlp_dense_h_to_4h_bias1, alloc198) + R.vm.kill_object(alloc197) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_3_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[48] + gv323: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc199: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv323, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc198, gpt_neox_layers_3_mlp_dense_4h_to_h_weight1, gpt_neox_layers_3_mlp_dense_4h_to_h_bias1, alloc199) + R.vm.kill_object(alloc198) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_3_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[44] + gv324: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc200: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv324, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape15, gpt_neox_layers_3_attention_dense_weight1, gpt_neox_layers_3_attention_dense_bias1, alloc199, alloc193, alloc200) + R.vm.kill_object(alloc193) + R.vm.kill_object(reshape15) + R.vm.kill_object(alloc199) + R.vm.kill_object(gpt_neox_layers_3_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_3_attention_dense_bias1) + gpt_neox_layers_4_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[50] + gv325: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc201: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv325, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc200, gpt_neox_layers_4_input_layernorm_weight1, gpt_neox_layers_4_input_layernorm_bias1, alloc201) + R.vm.kill_object(gpt_neox_layers_4_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_4_input_layernorm_bias1) + gpt_neox_layers_4_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[54] + gv326: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc202: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv326, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc201, gpt_neox_layers_4_attention_query_key_value_weight1, gpt_neox_layers_4_attention_query_key_value_bias1, alloc202) + R.vm.kill_object(alloc201) + R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_bias1) + gv327: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape16: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc202, gv327, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc202) + gv328: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape17: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape16, gv328, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape16) + gv329: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc203: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv329, R.dtype("float16")) + _202: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), reshape17, alloc203) + R.vm.kill_object(reshape17) + gv330: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape18: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc203, gv330, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc203) + gv331: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape19: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape18, gv331, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape18) + gpt_neox_layers_4_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[52] + gv332: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc204: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv332, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc200, gpt_neox_layers_4_post_attention_layernorm_weight1, gpt_neox_layers_4_post_attention_layernorm_bias1, alloc204) + R.vm.kill_object(gpt_neox_layers_4_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_4_post_attention_layernorm_bias1) + gpt_neox_layers_4_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[58] + gv333: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc205: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv333, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc204, gpt_neox_layers_4_mlp_dense_h_to_4h_weight1, gpt_neox_layers_4_mlp_dense_h_to_4h_bias1, alloc205) + R.vm.kill_object(alloc204) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_4_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[60] + gv334: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc206: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv334, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc205, gpt_neox_layers_4_mlp_dense_4h_to_h_weight1, gpt_neox_layers_4_mlp_dense_4h_to_h_bias1, alloc206) + R.vm.kill_object(alloc205) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_4_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[56] + gv335: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc207: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv335, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape19, gpt_neox_layers_4_attention_dense_weight1, gpt_neox_layers_4_attention_dense_bias1, alloc206, alloc200, alloc207) + R.vm.kill_object(alloc200) + R.vm.kill_object(reshape19) + R.vm.kill_object(alloc206) + R.vm.kill_object(gpt_neox_layers_4_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_4_attention_dense_bias1) + gpt_neox_layers_5_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[62] + gv336: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc208: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv336, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc207, gpt_neox_layers_5_input_layernorm_weight1, gpt_neox_layers_5_input_layernorm_bias1, alloc208) + R.vm.kill_object(gpt_neox_layers_5_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_5_input_layernorm_bias1) + gpt_neox_layers_5_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[66] + gv337: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc209: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv337, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc208, gpt_neox_layers_5_attention_query_key_value_weight1, gpt_neox_layers_5_attention_query_key_value_bias1, alloc209) + R.vm.kill_object(alloc208) + R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_bias1) + gv338: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape20: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc209, gv338, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc209) + gv339: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape21: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape20, gv339, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape20) + gv340: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc210: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv340, R.dtype("float16")) + _209: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), reshape21, alloc210) + R.vm.kill_object(reshape21) + gv341: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape22: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc210, gv341, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc210) + gv342: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape23: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape22, gv342, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape22) + gpt_neox_layers_5_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[64] + gv343: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc211: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv343, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc207, gpt_neox_layers_5_post_attention_layernorm_weight1, gpt_neox_layers_5_post_attention_layernorm_bias1, alloc211) + R.vm.kill_object(gpt_neox_layers_5_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_5_post_attention_layernorm_bias1) + gpt_neox_layers_5_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[70] + gv344: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc212: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv344, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc211, gpt_neox_layers_5_mlp_dense_h_to_4h_weight1, gpt_neox_layers_5_mlp_dense_h_to_4h_bias1, alloc212) + R.vm.kill_object(alloc211) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_5_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[72] + gv345: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc213: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv345, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc212, gpt_neox_layers_5_mlp_dense_4h_to_h_weight1, gpt_neox_layers_5_mlp_dense_4h_to_h_bias1, alloc213) + R.vm.kill_object(alloc212) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_5_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[68] + gv346: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc214: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv346, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape23, gpt_neox_layers_5_attention_dense_weight1, gpt_neox_layers_5_attention_dense_bias1, alloc213, alloc207, alloc214) + R.vm.kill_object(alloc207) + R.vm.kill_object(reshape23) + R.vm.kill_object(alloc213) + R.vm.kill_object(gpt_neox_layers_5_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_5_attention_dense_bias1) + gpt_neox_layers_6_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[74] + gv347: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc215: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv347, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc214, gpt_neox_layers_6_input_layernorm_weight1, gpt_neox_layers_6_input_layernorm_bias1, alloc215) + R.vm.kill_object(gpt_neox_layers_6_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_6_input_layernorm_bias1) + gpt_neox_layers_6_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[78] + gv348: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc216: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv348, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc215, gpt_neox_layers_6_attention_query_key_value_weight1, gpt_neox_layers_6_attention_query_key_value_bias1, alloc216) + R.vm.kill_object(alloc215) + R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_bias1) + gv349: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape24: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc216, gv349, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc216) + gv350: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape25: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape24, gv350, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape24) + gv351: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc217: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv351, R.dtype("float16")) + _216: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), reshape25, alloc217) + R.vm.kill_object(reshape25) + gv352: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape26: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc217, gv352, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc217) + gv353: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape27: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape26, gv353, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape26) + gpt_neox_layers_6_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[76] + gv354: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc218: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv354, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc214, gpt_neox_layers_6_post_attention_layernorm_weight1, gpt_neox_layers_6_post_attention_layernorm_bias1, alloc218) + R.vm.kill_object(gpt_neox_layers_6_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_6_post_attention_layernorm_bias1) + gpt_neox_layers_6_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[82] + gv355: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc219: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv355, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc218, gpt_neox_layers_6_mlp_dense_h_to_4h_weight1, gpt_neox_layers_6_mlp_dense_h_to_4h_bias1, alloc219) + R.vm.kill_object(alloc218) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_6_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[84] + gv356: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc220: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv356, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc219, gpt_neox_layers_6_mlp_dense_4h_to_h_weight1, gpt_neox_layers_6_mlp_dense_4h_to_h_bias1, alloc220) + R.vm.kill_object(alloc219) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_6_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[80] + gv357: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc221: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv357, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape27, gpt_neox_layers_6_attention_dense_weight1, gpt_neox_layers_6_attention_dense_bias1, alloc220, alloc214, alloc221) + R.vm.kill_object(alloc214) + R.vm.kill_object(reshape27) + R.vm.kill_object(alloc220) + R.vm.kill_object(gpt_neox_layers_6_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_6_attention_dense_bias1) + gpt_neox_layers_7_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[86] + gv358: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc222: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv358, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc221, gpt_neox_layers_7_input_layernorm_weight1, gpt_neox_layers_7_input_layernorm_bias1, alloc222) + R.vm.kill_object(gpt_neox_layers_7_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_7_input_layernorm_bias1) + gpt_neox_layers_7_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[90] + gv359: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc223: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv359, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc222, gpt_neox_layers_7_attention_query_key_value_weight1, gpt_neox_layers_7_attention_query_key_value_bias1, alloc223) + R.vm.kill_object(alloc222) + R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_bias1) + gv360: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape28: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc223, gv360, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc223) + gv361: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape29: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape28, gv361, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape28) + gv362: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc224: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv362, R.dtype("float16")) + _223: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), reshape29, alloc224) + R.vm.kill_object(reshape29) + gv363: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape30: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc224, gv363, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc224) + gv364: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape31: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape30, gv364, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape30) + gpt_neox_layers_7_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[88] + gv365: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc225: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv365, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc221, gpt_neox_layers_7_post_attention_layernorm_weight1, gpt_neox_layers_7_post_attention_layernorm_bias1, alloc225) + R.vm.kill_object(gpt_neox_layers_7_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_7_post_attention_layernorm_bias1) + gpt_neox_layers_7_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[94] + gv366: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc226: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv366, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc225, gpt_neox_layers_7_mlp_dense_h_to_4h_weight1, gpt_neox_layers_7_mlp_dense_h_to_4h_bias1, alloc226) + R.vm.kill_object(alloc225) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_7_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[96] + gv367: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc227: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv367, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc226, gpt_neox_layers_7_mlp_dense_4h_to_h_weight1, gpt_neox_layers_7_mlp_dense_4h_to_h_bias1, alloc227) + R.vm.kill_object(alloc226) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_7_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[92] + gv368: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc228: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv368, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape31, gpt_neox_layers_7_attention_dense_weight1, gpt_neox_layers_7_attention_dense_bias1, alloc227, alloc221, alloc228) + R.vm.kill_object(alloc221) + R.vm.kill_object(reshape31) + R.vm.kill_object(alloc227) + R.vm.kill_object(gpt_neox_layers_7_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_7_attention_dense_bias1) + gpt_neox_layers_8_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[98] + gv369: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc229: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv369, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc228, gpt_neox_layers_8_input_layernorm_weight1, gpt_neox_layers_8_input_layernorm_bias1, alloc229) + R.vm.kill_object(gpt_neox_layers_8_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_8_input_layernorm_bias1) + gpt_neox_layers_8_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[102] + gv370: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc230: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv370, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc229, gpt_neox_layers_8_attention_query_key_value_weight1, gpt_neox_layers_8_attention_query_key_value_bias1, alloc230) + R.vm.kill_object(alloc229) + R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_bias1) + gv371: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape32: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc230, gv371, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc230) + gv372: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape33: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape32, gv372, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape32) + gv373: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc231: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv373, R.dtype("float16")) + _230: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), reshape33, alloc231) + R.vm.kill_object(reshape33) + gv374: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape34: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc231, gv374, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc231) + gv375: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape35: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape34, gv375, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape34) + gpt_neox_layers_8_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[100] + gv376: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc232: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv376, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc228, gpt_neox_layers_8_post_attention_layernorm_weight1, gpt_neox_layers_8_post_attention_layernorm_bias1, alloc232) + R.vm.kill_object(gpt_neox_layers_8_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_8_post_attention_layernorm_bias1) + gpt_neox_layers_8_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[106] + gv377: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc233: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv377, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc232, gpt_neox_layers_8_mlp_dense_h_to_4h_weight1, gpt_neox_layers_8_mlp_dense_h_to_4h_bias1, alloc233) + R.vm.kill_object(alloc232) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_8_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[108] + gv378: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc234: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv378, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc233, gpt_neox_layers_8_mlp_dense_4h_to_h_weight1, gpt_neox_layers_8_mlp_dense_4h_to_h_bias1, alloc234) + R.vm.kill_object(alloc233) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_8_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[104] + gv379: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc235: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv379, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape35, gpt_neox_layers_8_attention_dense_weight1, gpt_neox_layers_8_attention_dense_bias1, alloc234, alloc228, alloc235) + R.vm.kill_object(alloc228) + R.vm.kill_object(reshape35) + R.vm.kill_object(alloc234) + R.vm.kill_object(gpt_neox_layers_8_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_8_attention_dense_bias1) + gpt_neox_layers_9_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[110] + gv380: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc236: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv380, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc235, gpt_neox_layers_9_input_layernorm_weight1, gpt_neox_layers_9_input_layernorm_bias1, alloc236) + R.vm.kill_object(gpt_neox_layers_9_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_9_input_layernorm_bias1) + gpt_neox_layers_9_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[114] + gv381: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc237: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv381, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc236, gpt_neox_layers_9_attention_query_key_value_weight1, gpt_neox_layers_9_attention_query_key_value_bias1, alloc237) + R.vm.kill_object(alloc236) + R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_bias1) + gv382: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape36: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc237, gv382, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc237) + gv383: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape37: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape36, gv383, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape36) + gv384: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc238: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv384, R.dtype("float16")) + _237: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), reshape37, alloc238) + R.vm.kill_object(reshape37) + gv385: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape38: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc238, gv385, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc238) + gv386: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape39: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape38, gv386, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape38) + gpt_neox_layers_9_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[112] + gv387: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc239: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv387, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc235, gpt_neox_layers_9_post_attention_layernorm_weight1, gpt_neox_layers_9_post_attention_layernorm_bias1, alloc239) + R.vm.kill_object(gpt_neox_layers_9_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_9_post_attention_layernorm_bias1) + gpt_neox_layers_9_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[118] + gv388: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc240: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv388, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc239, gpt_neox_layers_9_mlp_dense_h_to_4h_weight1, gpt_neox_layers_9_mlp_dense_h_to_4h_bias1, alloc240) + R.vm.kill_object(alloc239) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_9_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[120] + gv389: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc241: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv389, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc240, gpt_neox_layers_9_mlp_dense_4h_to_h_weight1, gpt_neox_layers_9_mlp_dense_4h_to_h_bias1, alloc241) + R.vm.kill_object(alloc240) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_9_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[116] + gv390: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc242: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv390, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape39, gpt_neox_layers_9_attention_dense_weight1, gpt_neox_layers_9_attention_dense_bias1, alloc241, alloc235, alloc242) + R.vm.kill_object(alloc235) + R.vm.kill_object(reshape39) + R.vm.kill_object(alloc241) + R.vm.kill_object(gpt_neox_layers_9_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_9_attention_dense_bias1) + gpt_neox_layers_10_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[122] + gv391: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc243: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv391, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc242, gpt_neox_layers_10_input_layernorm_weight1, gpt_neox_layers_10_input_layernorm_bias1, alloc243) + R.vm.kill_object(gpt_neox_layers_10_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_10_input_layernorm_bias1) + gpt_neox_layers_10_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[126] + gv392: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc244: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv392, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc243, gpt_neox_layers_10_attention_query_key_value_weight1, gpt_neox_layers_10_attention_query_key_value_bias1, alloc244) + R.vm.kill_object(alloc243) + R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_bias1) + gv393: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape40: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc244, gv393, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc244) + gv394: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape41: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape40, gv394, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape40) + gv395: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc245: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv395, R.dtype("float16")) + _244: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), reshape41, alloc245) + R.vm.kill_object(reshape41) + gv396: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape42: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc245, gv396, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc245) + gv397: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape43: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape42, gv397, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape42) + gpt_neox_layers_10_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[124] + gv398: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc246: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv398, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc242, gpt_neox_layers_10_post_attention_layernorm_weight1, gpt_neox_layers_10_post_attention_layernorm_bias1, alloc246) + R.vm.kill_object(gpt_neox_layers_10_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_10_post_attention_layernorm_bias1) + gpt_neox_layers_10_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[130] + gv399: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc247: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv399, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc246, gpt_neox_layers_10_mlp_dense_h_to_4h_weight1, gpt_neox_layers_10_mlp_dense_h_to_4h_bias1, alloc247) + R.vm.kill_object(alloc246) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_10_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[132] + gv400: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc248: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv400, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc247, gpt_neox_layers_10_mlp_dense_4h_to_h_weight1, gpt_neox_layers_10_mlp_dense_4h_to_h_bias1, alloc248) + R.vm.kill_object(alloc247) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_10_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[128] + gv401: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc249: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv401, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape43, gpt_neox_layers_10_attention_dense_weight1, gpt_neox_layers_10_attention_dense_bias1, alloc248, alloc242, alloc249) + R.vm.kill_object(alloc242) + R.vm.kill_object(reshape43) + R.vm.kill_object(alloc248) + R.vm.kill_object(gpt_neox_layers_10_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_10_attention_dense_bias1) + gpt_neox_layers_11_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[134] + gv402: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc250: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv402, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc249, gpt_neox_layers_11_input_layernorm_weight1, gpt_neox_layers_11_input_layernorm_bias1, alloc250) + R.vm.kill_object(gpt_neox_layers_11_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_11_input_layernorm_bias1) + gpt_neox_layers_11_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[138] + gv403: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc251: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv403, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc250, gpt_neox_layers_11_attention_query_key_value_weight1, gpt_neox_layers_11_attention_query_key_value_bias1, alloc251) + R.vm.kill_object(alloc250) + R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_bias1) + gv404: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape44: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc251, gv404, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc251) + gv405: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape45: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape44, gv405, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape44) + gv406: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc252: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv406, R.dtype("float16")) + _251: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), reshape45, alloc252) + R.vm.kill_object(reshape45) + gv407: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape46: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc252, gv407, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc252) + gv408: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape47: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape46, gv408, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape46) + gpt_neox_layers_11_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[136] + gv409: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc253: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv409, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc249, gpt_neox_layers_11_post_attention_layernorm_weight1, gpt_neox_layers_11_post_attention_layernorm_bias1, alloc253) + R.vm.kill_object(gpt_neox_layers_11_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_11_post_attention_layernorm_bias1) + gpt_neox_layers_11_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[142] + gv410: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc254: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv410, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc253, gpt_neox_layers_11_mlp_dense_h_to_4h_weight1, gpt_neox_layers_11_mlp_dense_h_to_4h_bias1, alloc254) + R.vm.kill_object(alloc253) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_11_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[144] + gv411: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc255: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv411, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc254, gpt_neox_layers_11_mlp_dense_4h_to_h_weight1, gpt_neox_layers_11_mlp_dense_4h_to_h_bias1, alloc255) + R.vm.kill_object(alloc254) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_11_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[140] + gv412: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc256: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv412, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape47, gpt_neox_layers_11_attention_dense_weight1, gpt_neox_layers_11_attention_dense_bias1, alloc255, alloc249, alloc256) + R.vm.kill_object(alloc249) + R.vm.kill_object(reshape47) + R.vm.kill_object(alloc255) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias1) + gpt_neox_layers_12_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[145] + gpt_neox_layers_12_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[146] + gv413: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc257: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv413, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc256, gpt_neox_layers_12_input_layernorm_weight1, gpt_neox_layers_12_input_layernorm_bias1, alloc257) + R.vm.kill_object(gpt_neox_layers_12_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_12_input_layernorm_bias1) + gpt_neox_layers_12_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[149] + gpt_neox_layers_12_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[150] + gv414: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc258: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv414, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc257, gpt_neox_layers_12_attention_query_key_value_weight1, gpt_neox_layers_12_attention_query_key_value_bias1, alloc258) + R.vm.kill_object(alloc257) + R.vm.kill_object(gpt_neox_layers_12_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_12_attention_query_key_value_bias1) + gv415: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape48: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc258, gv415, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc258) + gv416: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape49: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape48, gv416, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape48) + gv417: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc259: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv417, R.dtype("float16")) + _258: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape49, alloc259) + R.vm.kill_object(reshape49) + gv418: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape50: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc259, gv418, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc259) + gv419: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape51: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape50, gv419, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape50) + gpt_neox_layers_12_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[147] + gpt_neox_layers_12_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[148] + gv420: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc260: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv420, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc256, gpt_neox_layers_12_post_attention_layernorm_weight1, gpt_neox_layers_12_post_attention_layernorm_bias1, alloc260) + R.vm.kill_object(gpt_neox_layers_12_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_12_post_attention_layernorm_bias1) + gpt_neox_layers_12_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[153] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[154] + gv421: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc261: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv421, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc260, gpt_neox_layers_12_mlp_dense_h_to_4h_weight1, gpt_neox_layers_12_mlp_dense_h_to_4h_bias1, alloc261) + R.vm.kill_object(alloc260) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_12_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[155] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[156] + gv422: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc262: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv422, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc261, gpt_neox_layers_12_mlp_dense_4h_to_h_weight1, gpt_neox_layers_12_mlp_dense_4h_to_h_bias1, alloc262) + R.vm.kill_object(alloc261) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_12_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_12_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[151] + gpt_neox_layers_12_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[152] + gv423: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc263: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv423, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape51, gpt_neox_layers_12_attention_dense_weight1, gpt_neox_layers_12_attention_dense_bias1, alloc262, alloc256, alloc263) + R.vm.kill_object(alloc256) + R.vm.kill_object(reshape51) + R.vm.kill_object(alloc262) + R.vm.kill_object(gpt_neox_layers_12_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_12_attention_dense_bias1) + gpt_neox_layers_13_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[157] + gpt_neox_layers_13_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[158] + gv424: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc264: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv424, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc263, gpt_neox_layers_13_input_layernorm_weight1, gpt_neox_layers_13_input_layernorm_bias1, alloc264) + R.vm.kill_object(gpt_neox_layers_13_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_13_input_layernorm_bias1) + gpt_neox_layers_13_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[161] + gpt_neox_layers_13_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[162] + gv425: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc265: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv425, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc264, gpt_neox_layers_13_attention_query_key_value_weight1, gpt_neox_layers_13_attention_query_key_value_bias1, alloc265) + R.vm.kill_object(alloc264) + R.vm.kill_object(gpt_neox_layers_13_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_13_attention_query_key_value_bias1) + gv426: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape52: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc265, gv426, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc265) + gv427: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape53: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape52, gv427, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape52) + gv428: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc266: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv428, R.dtype("float16")) + _265: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape53, alloc266) + R.vm.kill_object(reshape53) + gv429: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape54: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc266, gv429, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc266) + gv430: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape55: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape54, gv430, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape54) + gpt_neox_layers_13_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[159] + gpt_neox_layers_13_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[160] + gv431: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc267: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv431, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc263, gpt_neox_layers_13_post_attention_layernorm_weight1, gpt_neox_layers_13_post_attention_layernorm_bias1, alloc267) + R.vm.kill_object(gpt_neox_layers_13_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_13_post_attention_layernorm_bias1) + gpt_neox_layers_13_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[165] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[166] + gv432: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc268: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv432, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc267, gpt_neox_layers_13_mlp_dense_h_to_4h_weight1, gpt_neox_layers_13_mlp_dense_h_to_4h_bias1, alloc268) + R.vm.kill_object(alloc267) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_13_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[167] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[168] + gv433: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc269: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv433, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc268, gpt_neox_layers_13_mlp_dense_4h_to_h_weight1, gpt_neox_layers_13_mlp_dense_4h_to_h_bias1, alloc269) + R.vm.kill_object(alloc268) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_13_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_13_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[163] + gpt_neox_layers_13_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[164] + gv434: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc270: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv434, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape55, gpt_neox_layers_13_attention_dense_weight1, gpt_neox_layers_13_attention_dense_bias1, alloc269, alloc263, alloc270) + R.vm.kill_object(alloc263) + R.vm.kill_object(reshape55) + R.vm.kill_object(alloc269) + R.vm.kill_object(gpt_neox_layers_13_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_13_attention_dense_bias1) + gpt_neox_layers_14_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[169] + gpt_neox_layers_14_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[170] + gv435: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc271: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv435, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc270, gpt_neox_layers_14_input_layernorm_weight1, gpt_neox_layers_14_input_layernorm_bias1, alloc271) + R.vm.kill_object(gpt_neox_layers_14_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_14_input_layernorm_bias1) + gpt_neox_layers_14_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[173] + gpt_neox_layers_14_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[174] + gv436: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc272: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv436, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc271, gpt_neox_layers_14_attention_query_key_value_weight1, gpt_neox_layers_14_attention_query_key_value_bias1, alloc272) + R.vm.kill_object(alloc271) + R.vm.kill_object(gpt_neox_layers_14_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_14_attention_query_key_value_bias1) + gv437: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape56: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc272, gv437, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc272) + gv438: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape57: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape56, gv438, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape56) + gv439: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc273: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv439, R.dtype("float16")) + _272: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape57, alloc273) + R.vm.kill_object(reshape57) + gv440: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape58: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc273, gv440, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc273) + gv441: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape59: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape58, gv441, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape58) + gpt_neox_layers_14_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[171] + gpt_neox_layers_14_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[172] + gv442: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc274: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv442, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc270, gpt_neox_layers_14_post_attention_layernorm_weight1, gpt_neox_layers_14_post_attention_layernorm_bias1, alloc274) + R.vm.kill_object(gpt_neox_layers_14_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_14_post_attention_layernorm_bias1) + gpt_neox_layers_14_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[177] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[178] + gv443: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc275: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv443, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc274, gpt_neox_layers_14_mlp_dense_h_to_4h_weight1, gpt_neox_layers_14_mlp_dense_h_to_4h_bias1, alloc275) + R.vm.kill_object(alloc274) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_14_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[179] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[180] + gv444: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc276: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv444, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc275, gpt_neox_layers_14_mlp_dense_4h_to_h_weight1, gpt_neox_layers_14_mlp_dense_4h_to_h_bias1, alloc276) + R.vm.kill_object(alloc275) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_14_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_14_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[175] + gpt_neox_layers_14_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[176] + gv445: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc277: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv445, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape59, gpt_neox_layers_14_attention_dense_weight1, gpt_neox_layers_14_attention_dense_bias1, alloc276, alloc270, alloc277) + R.vm.kill_object(alloc270) + R.vm.kill_object(reshape59) + R.vm.kill_object(alloc276) + R.vm.kill_object(gpt_neox_layers_14_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_14_attention_dense_bias1) + gpt_neox_layers_15_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[181] + gpt_neox_layers_15_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[182] + gv446: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc278: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv446, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc277, gpt_neox_layers_15_input_layernorm_weight1, gpt_neox_layers_15_input_layernorm_bias1, alloc278) + R.vm.kill_object(gpt_neox_layers_15_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_15_input_layernorm_bias1) + gpt_neox_layers_15_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[185] + gpt_neox_layers_15_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[186] + gv447: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc279: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv447, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc278, gpt_neox_layers_15_attention_query_key_value_weight1, gpt_neox_layers_15_attention_query_key_value_bias1, alloc279) + R.vm.kill_object(alloc278) + R.vm.kill_object(gpt_neox_layers_15_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_15_attention_query_key_value_bias1) + gv448: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape60: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc279, gv448, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc279) + gv449: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape61: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape60, gv449, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape60) + gv450: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc280: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv450, R.dtype("float16")) + _279: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape61, alloc280) + R.vm.kill_object(reshape61) + gv451: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape62: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc280, gv451, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc280) + gv452: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape63: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape62, gv452, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape62) + gpt_neox_layers_15_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[183] + gpt_neox_layers_15_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[184] + gv453: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc281: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv453, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc277, gpt_neox_layers_15_post_attention_layernorm_weight1, gpt_neox_layers_15_post_attention_layernorm_bias1, alloc281) + R.vm.kill_object(gpt_neox_layers_15_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_15_post_attention_layernorm_bias1) + gpt_neox_layers_15_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[189] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[190] + gv454: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc282: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv454, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc281, gpt_neox_layers_15_mlp_dense_h_to_4h_weight1, gpt_neox_layers_15_mlp_dense_h_to_4h_bias1, alloc282) + R.vm.kill_object(alloc281) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_15_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[191] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[192] + gv455: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc283: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv455, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc282, gpt_neox_layers_15_mlp_dense_4h_to_h_weight1, gpt_neox_layers_15_mlp_dense_4h_to_h_bias1, alloc283) + R.vm.kill_object(alloc282) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_15_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_15_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[187] + gpt_neox_layers_15_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[188] + gv456: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc284: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv456, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape63, gpt_neox_layers_15_attention_dense_weight1, gpt_neox_layers_15_attention_dense_bias1, alloc283, alloc277, alloc284) + R.vm.kill_object(alloc277) + R.vm.kill_object(reshape63) + R.vm.kill_object(alloc283) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_bias1) + gpt_neox_layers_16_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[193] + gpt_neox_layers_16_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[194] + gv457: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc285: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv457, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc284, gpt_neox_layers_16_input_layernorm_weight1, gpt_neox_layers_16_input_layernorm_bias1, alloc285) + R.vm.kill_object(gpt_neox_layers_16_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_16_input_layernorm_bias1) + gpt_neox_layers_16_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[197] + gpt_neox_layers_16_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[198] + gv458: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc286: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv458, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc285, gpt_neox_layers_16_attention_query_key_value_weight1, gpt_neox_layers_16_attention_query_key_value_bias1, alloc286) + R.vm.kill_object(alloc285) + R.vm.kill_object(gpt_neox_layers_16_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_16_attention_query_key_value_bias1) + gv459: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape64: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc286, gv459, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc286) + gv460: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape65: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape64, gv460, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape64) + gv461: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc287: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv461, R.dtype("float16")) + _286: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), reshape65, alloc287) + R.vm.kill_object(reshape65) + gv462: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape66: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc287, gv462, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc287) + gv463: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape67: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape66, gv463, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape66) + gpt_neox_layers_16_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[195] + gpt_neox_layers_16_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[196] + gv464: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc288: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv464, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc284, gpt_neox_layers_16_post_attention_layernorm_weight1, gpt_neox_layers_16_post_attention_layernorm_bias1, alloc288) + R.vm.kill_object(gpt_neox_layers_16_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_16_post_attention_layernorm_bias1) + gpt_neox_layers_16_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[201] + gpt_neox_layers_16_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[202] + gv465: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc289: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv465, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc288, gpt_neox_layers_16_mlp_dense_h_to_4h_weight1, gpt_neox_layers_16_mlp_dense_h_to_4h_bias1, alloc289) + R.vm.kill_object(alloc288) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_16_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[203] + gpt_neox_layers_16_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[204] + gv466: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc290: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv466, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc289, gpt_neox_layers_16_mlp_dense_4h_to_h_weight1, gpt_neox_layers_16_mlp_dense_4h_to_h_bias1, alloc290) + R.vm.kill_object(alloc289) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_16_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_16_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[199] + gpt_neox_layers_16_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[200] + gv467: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc291: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv467, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape67, gpt_neox_layers_16_attention_dense_weight1, gpt_neox_layers_16_attention_dense_bias1, alloc290, alloc284, alloc291) + R.vm.kill_object(alloc284) + R.vm.kill_object(reshape67) + R.vm.kill_object(alloc290) + R.vm.kill_object(gpt_neox_layers_16_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_16_attention_dense_bias1) + gpt_neox_layers_17_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[205] + gpt_neox_layers_17_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[206] + gv468: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc292: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv468, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc291, gpt_neox_layers_17_input_layernorm_weight1, gpt_neox_layers_17_input_layernorm_bias1, alloc292) + R.vm.kill_object(gpt_neox_layers_17_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_17_input_layernorm_bias1) + gpt_neox_layers_17_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[209] + gpt_neox_layers_17_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[210] + gv469: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc293: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv469, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc292, gpt_neox_layers_17_attention_query_key_value_weight1, gpt_neox_layers_17_attention_query_key_value_bias1, alloc293) + R.vm.kill_object(alloc292) + R.vm.kill_object(gpt_neox_layers_17_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_17_attention_query_key_value_bias1) + gv470: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape68: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc293, gv470, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc293) + gv471: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape69: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape68, gv471, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape68) + gv472: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc294: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv472, R.dtype("float16")) + _293: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), reshape69, alloc294) + R.vm.kill_object(reshape69) + gv473: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape70: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc294, gv473, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc294) + gv474: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape71: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape70, gv474, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape70) + gpt_neox_layers_17_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[207] + gpt_neox_layers_17_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[208] + gv475: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc295: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv475, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc291, gpt_neox_layers_17_post_attention_layernorm_weight1, gpt_neox_layers_17_post_attention_layernorm_bias1, alloc295) + R.vm.kill_object(gpt_neox_layers_17_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_17_post_attention_layernorm_bias1) + gpt_neox_layers_17_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[213] + gpt_neox_layers_17_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[214] + gv476: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc296: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv476, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc295, gpt_neox_layers_17_mlp_dense_h_to_4h_weight1, gpt_neox_layers_17_mlp_dense_h_to_4h_bias1, alloc296) + R.vm.kill_object(alloc295) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_17_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[215] + gpt_neox_layers_17_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[216] + gv477: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc297: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv477, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc296, gpt_neox_layers_17_mlp_dense_4h_to_h_weight1, gpt_neox_layers_17_mlp_dense_4h_to_h_bias1, alloc297) + R.vm.kill_object(alloc296) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_17_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_17_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[211] + gpt_neox_layers_17_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[212] + gv478: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc298: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv478, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape71, gpt_neox_layers_17_attention_dense_weight1, gpt_neox_layers_17_attention_dense_bias1, alloc297, alloc291, alloc298) + R.vm.kill_object(alloc291) + R.vm.kill_object(reshape71) + R.vm.kill_object(alloc297) + R.vm.kill_object(gpt_neox_layers_17_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_17_attention_dense_bias1) + gpt_neox_layers_18_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[217] + gpt_neox_layers_18_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[218] + gv479: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc299: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv479, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc298, gpt_neox_layers_18_input_layernorm_weight1, gpt_neox_layers_18_input_layernorm_bias1, alloc299) + R.vm.kill_object(gpt_neox_layers_18_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_18_input_layernorm_bias1) + gpt_neox_layers_18_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[221] + gpt_neox_layers_18_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[222] + gv480: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc300: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv480, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc299, gpt_neox_layers_18_attention_query_key_value_weight1, gpt_neox_layers_18_attention_query_key_value_bias1, alloc300) + R.vm.kill_object(alloc299) + R.vm.kill_object(gpt_neox_layers_18_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_18_attention_query_key_value_bias1) + gv481: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape72: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc300, gv481, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc300) + gv482: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape73: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape72, gv482, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape72) + gv483: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc301: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv483, R.dtype("float16")) + _300: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), reshape73, alloc301) + R.vm.kill_object(reshape73) + gv484: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape74: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc301, gv484, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc301) + gv485: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape75: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape74, gv485, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape74) + gpt_neox_layers_18_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[219] + gpt_neox_layers_18_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[220] + gv486: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc302: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv486, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc298, gpt_neox_layers_18_post_attention_layernorm_weight1, gpt_neox_layers_18_post_attention_layernorm_bias1, alloc302) + R.vm.kill_object(gpt_neox_layers_18_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_18_post_attention_layernorm_bias1) + gpt_neox_layers_18_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[225] + gpt_neox_layers_18_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[226] + gv487: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc303: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv487, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc302, gpt_neox_layers_18_mlp_dense_h_to_4h_weight1, gpt_neox_layers_18_mlp_dense_h_to_4h_bias1, alloc303) + R.vm.kill_object(alloc302) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_18_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[227] + gpt_neox_layers_18_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[228] + gv488: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc304: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv488, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc303, gpt_neox_layers_18_mlp_dense_4h_to_h_weight1, gpt_neox_layers_18_mlp_dense_4h_to_h_bias1, alloc304) + R.vm.kill_object(alloc303) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_18_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_18_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[223] + gpt_neox_layers_18_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[224] + gv489: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc305: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv489, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape75, gpt_neox_layers_18_attention_dense_weight1, gpt_neox_layers_18_attention_dense_bias1, alloc304, alloc298, alloc305) + R.vm.kill_object(alloc298) + R.vm.kill_object(reshape75) + R.vm.kill_object(alloc304) + R.vm.kill_object(gpt_neox_layers_18_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_18_attention_dense_bias1) + gpt_neox_layers_19_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[229] + gpt_neox_layers_19_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[230] + gv490: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc306: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv490, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc305, gpt_neox_layers_19_input_layernorm_weight1, gpt_neox_layers_19_input_layernorm_bias1, alloc306) + R.vm.kill_object(gpt_neox_layers_19_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_19_input_layernorm_bias1) + gpt_neox_layers_19_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[233] + gpt_neox_layers_19_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[234] + gv491: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc307: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv491, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc306, gpt_neox_layers_19_attention_query_key_value_weight1, gpt_neox_layers_19_attention_query_key_value_bias1, alloc307) + R.vm.kill_object(alloc306) + R.vm.kill_object(gpt_neox_layers_19_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_19_attention_query_key_value_bias1) + gv492: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape76: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc307, gv492, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc307) + gv493: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape77: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape76, gv493, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape76) + gv494: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc308: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv494, R.dtype("float16")) + _307: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), reshape77, alloc308) + R.vm.kill_object(reshape77) + gv495: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape78: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc308, gv495, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc308) + gv496: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape79: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape78, gv496, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape78) + gpt_neox_layers_19_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[231] + gpt_neox_layers_19_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[232] + gv497: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc309: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv497, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc305, gpt_neox_layers_19_post_attention_layernorm_weight1, gpt_neox_layers_19_post_attention_layernorm_bias1, alloc309) + R.vm.kill_object(gpt_neox_layers_19_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_19_post_attention_layernorm_bias1) + gpt_neox_layers_19_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[237] + gpt_neox_layers_19_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[238] + gv498: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc310: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv498, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc309, gpt_neox_layers_19_mlp_dense_h_to_4h_weight1, gpt_neox_layers_19_mlp_dense_h_to_4h_bias1, alloc310) + R.vm.kill_object(alloc309) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_19_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[239] + gpt_neox_layers_19_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[240] + gv499: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc311: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv499, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc310, gpt_neox_layers_19_mlp_dense_4h_to_h_weight1, gpt_neox_layers_19_mlp_dense_4h_to_h_bias1, alloc311) + R.vm.kill_object(alloc310) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_19_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_19_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[235] + gpt_neox_layers_19_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[236] + gv500: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc312: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv500, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape79, gpt_neox_layers_19_attention_dense_weight1, gpt_neox_layers_19_attention_dense_bias1, alloc311, alloc305, alloc312) + R.vm.kill_object(alloc305) + R.vm.kill_object(reshape79) + R.vm.kill_object(alloc311) + R.vm.kill_object(gpt_neox_layers_19_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_19_attention_dense_bias1) + gpt_neox_layers_20_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[241] + gpt_neox_layers_20_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[242] + gv501: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc313: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv501, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc312, gpt_neox_layers_20_input_layernorm_weight1, gpt_neox_layers_20_input_layernorm_bias1, alloc313) + R.vm.kill_object(gpt_neox_layers_20_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_20_input_layernorm_bias1) + gpt_neox_layers_20_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[245] + gpt_neox_layers_20_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[246] + gv502: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc314: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv502, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc313, gpt_neox_layers_20_attention_query_key_value_weight1, gpt_neox_layers_20_attention_query_key_value_bias1, alloc314) + R.vm.kill_object(alloc313) + R.vm.kill_object(gpt_neox_layers_20_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_20_attention_query_key_value_bias1) + gv503: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape80: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc314, gv503, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc314) + gv504: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape81: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape80, gv504, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape80) + gv505: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc315: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv505, R.dtype("float16")) + _314: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), reshape81, alloc315) + R.vm.kill_object(reshape81) + gv506: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape82: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc315, gv506, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc315) + gv507: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape83: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape82, gv507, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape82) + gpt_neox_layers_20_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[243] + gpt_neox_layers_20_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[244] + gv508: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc316: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv508, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc312, gpt_neox_layers_20_post_attention_layernorm_weight1, gpt_neox_layers_20_post_attention_layernorm_bias1, alloc316) + R.vm.kill_object(gpt_neox_layers_20_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_20_post_attention_layernorm_bias1) + gpt_neox_layers_20_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[249] + gpt_neox_layers_20_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[250] + gv509: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc317: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv509, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc316, gpt_neox_layers_20_mlp_dense_h_to_4h_weight1, gpt_neox_layers_20_mlp_dense_h_to_4h_bias1, alloc317) + R.vm.kill_object(alloc316) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_20_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[251] + gpt_neox_layers_20_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[252] + gv510: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc318: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv510, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc317, gpt_neox_layers_20_mlp_dense_4h_to_h_weight1, gpt_neox_layers_20_mlp_dense_4h_to_h_bias1, alloc318) + R.vm.kill_object(alloc317) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_20_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_20_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[247] + gpt_neox_layers_20_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[248] + gv511: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc319: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv511, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape83, gpt_neox_layers_20_attention_dense_weight1, gpt_neox_layers_20_attention_dense_bias1, alloc318, alloc312, alloc319) + R.vm.kill_object(alloc312) + R.vm.kill_object(reshape83) + R.vm.kill_object(alloc318) + R.vm.kill_object(gpt_neox_layers_20_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_20_attention_dense_bias1) + gpt_neox_layers_21_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[253] + gpt_neox_layers_21_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[254] + gv512: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc320: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv512, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc319, gpt_neox_layers_21_input_layernorm_weight1, gpt_neox_layers_21_input_layernorm_bias1, alloc320) + R.vm.kill_object(gpt_neox_layers_21_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_21_input_layernorm_bias1) + gpt_neox_layers_21_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[257] + gpt_neox_layers_21_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[258] + gv513: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc321: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv513, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc320, gpt_neox_layers_21_attention_query_key_value_weight1, gpt_neox_layers_21_attention_query_key_value_bias1, alloc321) + R.vm.kill_object(alloc320) + R.vm.kill_object(gpt_neox_layers_21_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_21_attention_query_key_value_bias1) + gv514: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape84: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc321, gv514, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc321) + gv515: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape85: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape84, gv515, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape84) + gv516: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc322: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv516, R.dtype("float16")) + _321: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), reshape85, alloc322) + R.vm.kill_object(reshape85) + gv517: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape86: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc322, gv517, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc322) + gv518: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape87: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape86, gv518, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape86) + gpt_neox_layers_21_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[255] + gpt_neox_layers_21_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[256] + gv519: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc323: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv519, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc319, gpt_neox_layers_21_post_attention_layernorm_weight1, gpt_neox_layers_21_post_attention_layernorm_bias1, alloc323) + R.vm.kill_object(gpt_neox_layers_21_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_21_post_attention_layernorm_bias1) + gpt_neox_layers_21_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[261] + gpt_neox_layers_21_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[262] + gv520: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc324: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv520, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc323, gpt_neox_layers_21_mlp_dense_h_to_4h_weight1, gpt_neox_layers_21_mlp_dense_h_to_4h_bias1, alloc324) + R.vm.kill_object(alloc323) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_21_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[263] + gpt_neox_layers_21_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[264] + gv521: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc325: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv521, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc324, gpt_neox_layers_21_mlp_dense_4h_to_h_weight1, gpt_neox_layers_21_mlp_dense_4h_to_h_bias1, alloc325) + R.vm.kill_object(alloc324) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_21_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_21_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[259] + gpt_neox_layers_21_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[260] + gv522: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc326: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv522, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape87, gpt_neox_layers_21_attention_dense_weight1, gpt_neox_layers_21_attention_dense_bias1, alloc325, alloc319, alloc326) + R.vm.kill_object(alloc319) + R.vm.kill_object(reshape87) + R.vm.kill_object(alloc325) + R.vm.kill_object(gpt_neox_layers_21_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_21_attention_dense_bias1) + gpt_neox_layers_22_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[265] + gpt_neox_layers_22_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[266] + gv523: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc327: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv523, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc326, gpt_neox_layers_22_input_layernorm_weight1, gpt_neox_layers_22_input_layernorm_bias1, alloc327) + R.vm.kill_object(gpt_neox_layers_22_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_22_input_layernorm_bias1) + gpt_neox_layers_22_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[269] + gpt_neox_layers_22_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[270] + gv524: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc328: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv524, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc327, gpt_neox_layers_22_attention_query_key_value_weight1, gpt_neox_layers_22_attention_query_key_value_bias1, alloc328) + R.vm.kill_object(alloc327) + R.vm.kill_object(gpt_neox_layers_22_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_22_attention_query_key_value_bias1) + gv525: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape88: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc328, gv525, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc328) + gv526: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape89: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape88, gv526, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape88) + gv527: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc329: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv527, R.dtype("float16")) + _328: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), reshape89, alloc329) + R.vm.kill_object(reshape89) + gv528: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape90: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc329, gv528, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc329) + gv529: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape91: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape90, gv529, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape90) + gpt_neox_layers_22_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[267] + gpt_neox_layers_22_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[268] + gv530: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc330: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv530, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc326, gpt_neox_layers_22_post_attention_layernorm_weight1, gpt_neox_layers_22_post_attention_layernorm_bias1, alloc330) + R.vm.kill_object(gpt_neox_layers_22_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_22_post_attention_layernorm_bias1) + gpt_neox_layers_22_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[273] + gpt_neox_layers_22_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[274] + gv531: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc331: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv531, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc330, gpt_neox_layers_22_mlp_dense_h_to_4h_weight1, gpt_neox_layers_22_mlp_dense_h_to_4h_bias1, alloc331) + R.vm.kill_object(alloc330) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_22_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[275] + gpt_neox_layers_22_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[276] + gv532: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc332: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv532, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc331, gpt_neox_layers_22_mlp_dense_4h_to_h_weight1, gpt_neox_layers_22_mlp_dense_4h_to_h_bias1, alloc332) + R.vm.kill_object(alloc331) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_22_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_22_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[271] + gpt_neox_layers_22_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[272] + gv533: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc333: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv533, R.dtype("float16")) + R.vm.kill_object(storage10) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape91, gpt_neox_layers_22_attention_dense_weight1, gpt_neox_layers_22_attention_dense_bias1, alloc332, alloc326, alloc333) + R.vm.kill_object(alloc326) + R.vm.kill_object(reshape91) + R.vm.kill_object(alloc332) + R.vm.kill_object(gpt_neox_layers_22_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_22_attention_dense_bias1) + gpt_neox_layers_23_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[277] + gpt_neox_layers_23_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[278] + gv534: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc334: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv534, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc333, gpt_neox_layers_23_input_layernorm_weight1, gpt_neox_layers_23_input_layernorm_bias1, alloc334) + R.vm.kill_object(gpt_neox_layers_23_input_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_23_input_layernorm_bias1) + gpt_neox_layers_23_attention_query_key_value_weight1: R.Tensor((3072, 1024), dtype="float16") = packed_params[281] + gpt_neox_layers_23_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[282] + gv535: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc335: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv535, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc334, gpt_neox_layers_23_attention_query_key_value_weight1, gpt_neox_layers_23_attention_query_key_value_bias1, alloc335) + R.vm.kill_object(alloc334) + R.vm.kill_object(gpt_neox_layers_23_attention_query_key_value_weight1) + R.vm.kill_object(gpt_neox_layers_23_attention_query_key_value_bias1) + gv536: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape92: R.Tensor((1, seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc335, gv536, sinfo_args=(R.Tensor((1, seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(alloc335) + gv537: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(48), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape93: R.Tensor((seq_len, 48, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape92, gv537, sinfo_args=(R.Tensor((seq_len, 48, 64), dtype="float16"),)) + R.vm.kill_object(reshape92) + gv538: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc336: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv538, R.dtype("float16")) + _335: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), reshape93, alloc336) + R.vm.kill_object(reshape93) + gv539: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(16), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape94: R.Tensor((1, seq_len, 16, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc336, gv539, sinfo_args=(R.Tensor((1, seq_len, 16, 64), dtype="float16"),)) + R.vm.kill_object(alloc336) + gv540: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + reshape95: R.Tensor((1, seq_len, 1024), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape94, gv540, sinfo_args=(R.Tensor((1, seq_len, 1024), dtype="float16"),)) + R.vm.kill_object(reshape94) + gpt_neox_layers_23_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[279] + gpt_neox_layers_23_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[280] + gv541: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc337: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv541, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc333, gpt_neox_layers_23_post_attention_layernorm_weight1, gpt_neox_layers_23_post_attention_layernorm_bias1, alloc337) + R.vm.kill_object(gpt_neox_layers_23_post_attention_layernorm_weight1) + R.vm.kill_object(gpt_neox_layers_23_post_attention_layernorm_bias1) + gpt_neox_layers_23_mlp_dense_h_to_4h_weight1: R.Tensor((4096, 1024), dtype="float16") = packed_params[285] + gpt_neox_layers_23_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[286] + gv542: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(4096), sinfo_args=(R.Shape(ndim=3),)) + alloc338: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv542, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc337, gpt_neox_layers_23_mlp_dense_h_to_4h_weight1, gpt_neox_layers_23_mlp_dense_h_to_4h_bias1, alloc338) + R.vm.kill_object(alloc337) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_h_to_4h_weight1) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_h_to_4h_bias1) + gpt_neox_layers_23_mlp_dense_4h_to_h_weight1: R.Tensor((1024, 4096), dtype="float16") = packed_params[287] + gpt_neox_layers_23_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[288] + gv543: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc339: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv543, R.dtype("float16")) + R.vm.kill_object(storage9) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc338, gpt_neox_layers_23_mlp_dense_4h_to_h_weight1, gpt_neox_layers_23_mlp_dense_4h_to_h_bias1, alloc339) + R.vm.kill_object(alloc338) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_4h_to_h_weight1) + R.vm.kill_object(gpt_neox_layers_23_mlp_dense_4h_to_h_bias1) + gpt_neox_layers_23_attention_dense_weight1: R.Tensor((1024, 1024), dtype="float16") = packed_params[283] + gpt_neox_layers_23_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[284] + gv544: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc340: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv544, R.dtype("float16")) + R.vm.kill_object(storage11) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape95, gpt_neox_layers_23_attention_dense_weight1, gpt_neox_layers_23_attention_dense_bias1, alloc339, alloc333, alloc340) + R.vm.kill_object(alloc333) + R.vm.kill_object(reshape95) + R.vm.kill_object(alloc339) + R.vm.kill_object(gpt_neox_layers_23_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_23_attention_dense_bias1) + gpt_neox_final_layer_norm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[289] + gpt_neox_final_layer_norm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[290] + gv545: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1024), sinfo_args=(R.Shape(ndim=3),)) + alloc341: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv545, R.dtype("float16")) + R.vm.kill_object(storage8) + cls.gpt_neox_q0f16_layer_norm1(alloc340, gpt_neox_final_layer_norm_weight1, gpt_neox_final_layer_norm_bias1, alloc341) + R.vm.kill_object(alloc340) + R.vm.kill_object(gpt_neox_final_layer_norm_weight1) + R.vm.kill_object(gpt_neox_final_layer_norm_bias1) + storage12: R.Object = R.vm.alloc_storage(R.shape([2048]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc342: R.Tensor((1, 1, 1024), dtype="float16") = R.vm.alloc_tensor(storage12, R.prim_value(0), R.shape([1, 1, 1024]), R.dtype("float16")) + R.vm.kill_object(storage12) + cls.gpt_neox_q0f16_index(alloc341, alloc342) + R.vm.kill_object(alloc341) + embed_out_weight1: R.Tensor((vocab_size, 1024), dtype="float16") = packed_params[291] + gv546: R.Shape(ndim=1) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(3), sinfo_args=(R.Shape(ndim=1),)) + storage13: R.Object = R.vm.alloc_storage(gv546, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv547: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(1), sinfo_args=(R.Shape(ndim=3),)) + alloc343: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv547, R.dtype("float32")) + R.vm.kill_object(storage13) + cls.gpt_neox_q0f16_fused_NT_matmul14_cast8(alloc342, embed_out_weight1, alloc343) + R.vm.kill_object(alloc342) + R.vm.kill_object(embed_out_weight1) + gv1: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc343, paged_kv_cache + R.vm.kill_object(alloc343) + gv548: R.Tensor(dtype="float32", ndim=3) = gv1[0] + R.call_packed("vm.builtin.match_shape", gv548, shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1), R.prim_value(3), R.prim_value(1), R.str("ErrorContext(fn=prefill, loc=return, annotation=R.Tuple(R.Tensor((1, 1, vocab_size), dtype=\"float32\"), R.Object)) "), sinfo_args=(R.Tuple,)) + return gv1 + + @R.function + def softmax_with_temperature(logits: R.Tensor(("batch_size", 1, "vocab_size"), dtype="float32"), temperature: R.Tensor(("batch_size",), dtype="float32")) -> R.Tensor(("batch_size", 1, "vocab_size"), dtype="float32"): + batch_size = T.int64(is_size_var=True) + vocab_size = T.int64(is_size_var=True) + R.func_attr({"relax.force_pure": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) + cls = Module + shape_heap: R.Tensor(dtype="int64", ndim=1) = R.call_builtin_with_ctx("vm.builtin.alloc_shape_heap", (R.prim_value(5),), sinfo_args=(R.Tensor(dtype="int64", ndim=1),)) + R.call_packed("vm.builtin.check_tensor_info", logits, R.prim_value(3), R.dtype("float32"), R.str("ErrorContext(fn=softmax_with_temperature, loc=param[0], param=logits, annotation=R.Tensor((batch_size, 1, vocab_size), dtype=\"float32\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tensor_info", temperature, R.prim_value(1), R.dtype("float32"), R.str("ErrorContext(fn=softmax_with_temperature, loc=param[1], param=temperature, annotation=R.Tensor((batch_size,), dtype=\"float32\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", logits, shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(1), R.str("ErrorContext(fn=softmax_with_temperature, loc=param[0], param=logits, annotation=R.Tensor((batch_size, 1, vocab_size), dtype=\"float32\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", temperature, shape_heap, R.prim_value(1), R.prim_value(3), R.prim_value(0), R.str("ErrorContext(fn=softmax_with_temperature, loc=param[1], param=temperature, annotation=R.Tensor((batch_size,), dtype=\"float32\")) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func(shape_heap) + gv: R.Shape(ndim=2) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(0), R.prim_value(1), R.prim_value(1), sinfo_args=(R.Shape(ndim=2),)) + lv: R.Tensor((batch_size, vocab_size), dtype="float32") = R.call_packed("vm.builtin.reshape", logits, gv, sinfo_args=(R.Tensor((batch_size, vocab_size), dtype="float32"),)) + gv1: R.Shape(ndim=1) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(2), sinfo_args=(R.Shape(ndim=1),)) + storage: R.Object = R.vm.alloc_storage(gv1, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv2: R.Shape(ndim=2) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(0), R.prim_value(1), R.prim_value(3), sinfo_args=(R.Shape(ndim=2),)) + alloc: R.Tensor(dtype="float32", ndim=2) = R.vm.alloc_tensor(storage, R.prim_value(0), gv2, R.dtype("float32")) + R.vm.kill_object(storage) + gv3: R.Shape(ndim=1) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(2), sinfo_args=(R.Shape(ndim=1),)) + storage1: R.Object = R.vm.alloc_storage(gv3, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv4: R.Shape(ndim=2) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(0), R.prim_value(1), R.prim_value(3), sinfo_args=(R.Shape(ndim=2),)) + alloc1: R.Tensor(dtype="float32", ndim=2) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv4, R.dtype("float32")) + R.vm.kill_object(storage1) + cls.gpt_neox_q0f16_chunk_lse(lv, temperature, alloc, alloc1) + lv1: R.Tuple(R.Tensor(dtype="float32", ndim=2), R.Tensor(dtype="float32", ndim=2)) = alloc, alloc1 + gv5: R.Shape(ndim=1) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(4), sinfo_args=(R.Shape(ndim=1),)) + storage2: R.Object = R.vm.alloc_storage(gv5, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv6: R.Shape(ndim=2) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(0), R.prim_value(1), R.prim_value(1), sinfo_args=(R.Shape(ndim=2),)) + alloc2: R.Tensor(dtype="float32", ndim=2) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv6, R.dtype("float32")) + R.vm.kill_object(storage2) + cls.gpt_neox_q0f16_softmax_with_chunked_sum(lv, temperature, alloc, alloc1, alloc2) + R.vm.kill_object(lv) + R.vm.kill_object(alloc) + R.vm.kill_object(alloc1) + gv7: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(1), sinfo_args=(R.Shape(ndim=3),)) + gv_1: R.Tensor((batch_size, 1, vocab_size), dtype="float32") = R.call_packed("vm.builtin.reshape", alloc2, gv7, sinfo_args=(R.Tensor((batch_size, 1, vocab_size), dtype="float32"),)) + R.vm.kill_object(alloc2) + return gv_1 \ No newline at end of file