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 +1,9492 @@ +# 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, 8, 256), "float16") + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(pages_handle, (max_num_pages, 2, 8, 16, 256), "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, 8, 256), "float16") + lse = T.match_buffer(lse_handle, (B, 8)) + # with T.block("root"): + sm_scale: T.float32 = T.float32(0.090168440055560212) + for bx in T.thread_binding(B, thread="blockIdx.x"): + for fused_by_bz in T.thread_binding(8, thread="blockIdx.y"): + for ty in T.thread_binding(1, thread="threadIdx.y"): + for tx in T.thread_binding(64, thread="threadIdx.x"): + for tz in T.thread_binding(4, 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 // 8 + ty + fused_by_bz % 8, tx * 4 - 128:tx * 4 - 128 + 260]) + T.writes(output[bx, fused_by_bz % 8 + fused_by_bz // 8 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 8 + fused_by_bz // 8 + 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((4, 256), "float16", scope="shared") + V_smem = T.alloc_buffer((4, 256), "float16", scope="shared") + O_allreduce = T.alloc_buffer((4, 1, 256), scope="shared") + md_allreduce = T.alloc_buffer((4, 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 % 8 + bz: T.int32 = fused_by_bz // 8 + 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 < 128, Q[bx, by + bz + ty, tx * 4 + vec + 128] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 128]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 256) / T.float32(256.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) + for iterator in range((kv_chunk_len[0] + 3) // 4): + tile_start_s: T.int32 = tz + ty + tile_start_g: T.int32 = iterator * 4 + 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 < 128, pages[page_no, 0, by, page_offset, tx * 4 + vec + 128] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 128]))), 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 % 256) / T.float32(256.0))}), 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 * 4 + 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(4): + 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, 8, 256), "float16") + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(pages_handle, (max_num_pages, 2, 8, 16, 256), "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, 8, 256), "float16") + lse = T.match_buffer(lse_handle, (B, 8)) + # with T.block("root"): + sm_scale: T.float32 = T.float32(0.090168440055560212) + for bx in T.thread_binding(B, thread="blockIdx.x"): + for fused_by_bz in T.thread_binding(8, thread="blockIdx.y"): + for ty in T.thread_binding(1, thread="threadIdx.y"): + for tx in T.thread_binding(64, thread="threadIdx.x"): + for tz in T.thread_binding(4, 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 // 8 + ty + fused_by_bz % 8, tx * 4 - 128:tx * 4 - 128 + 260]) + T.writes(output[bx, fused_by_bz % 8 + fused_by_bz // 8 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 8 + fused_by_bz // 8 + 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((4, 256), "float16", scope="shared") + V_smem = T.alloc_buffer((4, 256), "float16", scope="shared") + O_allreduce = T.alloc_buffer((4, 1, 256), scope="shared") + md_allreduce = T.alloc_buffer((4, 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 % 8 + bz: T.int32 = fused_by_bz // 8 + 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 < 128, Q[bx, by + bz + ty, tx * 4 + vec + 128] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 128]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 256) / T.float32(256.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) + for iterator in range((kv_chunk_len[0] + 3) // 4): + tile_start_s: T.int32 = tz + ty + tile_start_g: T.int32 = iterator * 4 + 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 < 128, pages[page_no, 0, by, page_offset, tx * 4 + vec + 128] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 128]))), 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 % 256) / T.float32(256.0))}), 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 * 4 + 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(4): + 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, 8, 256), "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, 8, 16, 256), "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, 8, 256), "float16") + lse = T.match_buffer(var_lse, (total_len, 8)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(8, 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((16, 256), "float16", scope="shared") + K_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + S_smem = T.alloc_buffer((16, 16), scope="shared") + S_local = T.alloc_buffer((16, 16), scope="local") + O_local = T.alloc_buffer((16, 256), scope="local") + m_smem = T.alloc_buffer((16,), scope="shared") + m_prev_smem = T.alloc_buffer((16,), scope="shared") + d_smem = T.alloc_buffer((16,), 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] + 16 - 1) // 16 + 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] + 16 - 1) // 16 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 16 + 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 < 16: + 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, 8): + with T.block("O_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + 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(8): + 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(16, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 256) + j = T.axis.spatial(256, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 256) + 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 < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), 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(8): + 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) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + 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 < 128, pages[page_no, 0, by, page_offset, j + 128] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 128]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), 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(8): + 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) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + 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[0:16, 0:256], K_smem[0:16, 0:256]) + T.writes(S_local[0:16, 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(1, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 + 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(32, 1, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + 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(256, 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(0.090168440055560212) + 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(1, 2): + with T.block("S_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + 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 < 16: + 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 < 16: + 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 < 16: + 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[0:16], m_smem[0:16], S_smem[0:16, 0:16], V_smem[0:16, 0:256]) + T.writes(O_local[0:16, 0:256]) + 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, 8): + with T.block("O_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) + j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + 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, 8): + with T.block("O_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + 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, 8): + with T.block("O_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + 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(16, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) + 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, 8, 256), "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, 8, 16, 256), "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, 8, 256), "float16") + lse = T.match_buffer(var_lse, (total_len, 8)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(8, 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((16, 256), "float16", scope="shared") + K_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + S_smem = T.alloc_buffer((16, 16), scope="shared") + S_local = T.alloc_buffer((16, 16), scope="local") + O_local = T.alloc_buffer((16, 256), scope="local") + m_smem = T.alloc_buffer((16,), scope="shared") + m_prev_smem = T.alloc_buffer((16,), scope="shared") + d_smem = T.alloc_buffer((16,), 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] + 16 - 1) // 16 + 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] + 16 - 1) // 16 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 16 + 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 < 16: + 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, 8): + with T.block("O_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + 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(8): + 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(16, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 256) + j = T.axis.spatial(256, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 256) + 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 < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), 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(8): + 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) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + 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 < 128, pages[page_no, 0, by, page_offset, j + 128] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 128]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), 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(8): + 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) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + 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[0:16, 0:256], K_smem[0:16, 0:256]) + T.writes(S_local[0:16, 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(1, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 + 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(32, 1, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + 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(256, 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(0.090168440055560212) + 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(1, 2): + with T.block("S_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + 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 < 16: + 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 < 16: + 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 < 16: + 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[0:16], m_smem[0:16], S_smem[0:16, 0:16], V_smem[0:16, 0:256]) + T.writes(O_local[0:16, 0:256]) + 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, 8): + with T.block("O_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) + j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + 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, 8): + with T.block("O_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + 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, 8): + with T.block("O_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + 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(16, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) + 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, 8, 256), "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, 8, 256), "float16") + v = T.match_buffer(var_v, (kv_len, 8, 256), "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, 8, 256), "float16") + lse = T.match_buffer(var_lse, (qo_len, 8)) + # with T.block("root"): + for lbx in T.thread_binding(8, thread="blockIdx.x"): + for lby in T.thread_binding(8, 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((16, 256), "float16", scope="shared") + K_smem = T.alloc_buffer((256, 16), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + S_smem = T.alloc_buffer((16, 16), scope="shared") + S_local = T.alloc_buffer((16, 16), scope="local") + O_local = T.alloc_buffer((16, 256), scope="local") + m_smem = T.alloc_buffer((16,), scope="shared") + m_prev_smem = T.alloc_buffer((16,), scope="shared") + d_smem = T.alloc_buffer((16,), 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] + 16 - 1) // 16 + 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] + 16 - 1) // 16 + 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] * 16 + 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 < 16: + 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(8): + with T.block("O_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1_0 * 8 + 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(2, 2): + for lj_1 in T.vectorized(8): + with T.block("Q_load"): + i = T.axis.spatial(16, li_0 * 8 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) // 16) + j = T.axis.spatial(256, lj_0_0 * 128 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) % 16 * 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 < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), 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 + 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(2, 2): + for ly_1 in T.vectorized(8): + with T.block("K_load"): + i = T.axis.spatial(16, lz_0 * 8 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 16) + j = T.axis.spatial(256, ly_0_0 * 128 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 16 * 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 - 128:j - 128 + 257]) + 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 < 128, k[L_kv_base + cur_L, by, j + 128] * T.float16(-1.0), k[L_kv_base + cur_L, by, j - 128]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), 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(2, 2): + for ly_1 in T.vectorized(8): + with T.block("V_load"): + i = T.axis.spatial(16, lz_0 * 8 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 16) + j = T.axis.spatial(256, ly_0_0 * 128 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 16 * 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[0:16, 0:256], K_smem[0:256, 0:16]) + T.writes(S_local[0:16, 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 in T.unroll(1): + for lj_1_0_init in T.unroll(1): + for lj_1_1_init in T.vectorized(2): + with T.block("S_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 + 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_0_init * 2 + 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(16): + for li_1 in T.unroll(1): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(2): + for lk_1 in range(16): + with T.block("S_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1_0 * 2 + lj_1_1) + k_1 = T.axis.reduce(256, 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(0.090168440055560212) + 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(1): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(2): + with T.block("S_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1_0 * 2 + 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 < 16: + 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 < 16: + 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 < 16: + 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[0:16], m_smem[0:16], S_smem[0:16, 0:16], V_smem[0:16, 0:256]) + T.writes(O_local[0:16, 0:256]) + 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(8): + with T.block("O_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) + j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + lj_1_0_init * 8 + 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(1, 16): + for li_1 in T.unroll(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(8): + with T.block("O_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1_0 * 8 + lj_1_1) + k_1 = T.axis.reduce(16, 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(8): + with T.block("O_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1_0 * 8 + 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(16, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) + 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, 8, 256), "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, 8, 256), "float16") + v = T.match_buffer(var_v, (kv_len, 8, 256), "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, 8, 256), "float16") + lse = T.match_buffer(var_lse, (qo_len, 8)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(8, 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((16, 256), "float16", scope="shared") + K_smem = T.alloc_buffer((32, 256), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 256), "float16", scope="shared") + S_smem = T.alloc_buffer((16, 32), scope="shared") + S_local = T.alloc_buffer((16, 32), scope="local") + O_local = T.alloc_buffer((16, 256), scope="local") + m_smem = T.alloc_buffer((16,), scope="shared") + m_prev_smem = T.alloc_buffer((16,), scope="shared") + d_smem = T.alloc_buffer((16,), 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] + 16 - 1) // 16 + 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] + 16 - 1) // 16 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 16 + 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 < 16: + 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, 8): + with T.block("O_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + 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(8): + 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(16, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 256) + j = T.axis.spatial(256, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 256) + 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 < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), 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(16): + 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) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + 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 < 128, k[cur_L, by, j + 128] * T.float16(-1.0), k[cur_L, by, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), 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[0:16, 0:256], K_smem[0:32, 0:256]) + T.writes(S_local[0:16, 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, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 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(32, 2, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 2 + lj_1) + k_1 = T.axis.reduce(256, 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(0.090168440055560212) + 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(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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 < 16: + 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 < 16: + 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 < 16: + 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[0:16], m_smem[0:16], S_smem[0:16, 0:32], V_smem[0:32, 0:256]) + T.writes(O_local[0:16, 0:256]) + 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, 8): + with T.block("O_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) + j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + 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, 8): + with T.block("O_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + 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, 8): + with T.block("O_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + 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(16, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) + 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, 8, 16, 256), "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 * 8, thread="blockIdx.x"): + for bhd_i in T.thread_binding(256, thread="threadIdx.x"): + b: T.int32 = (bhd_o * 256 + bhd_i) // 2048 + h: T.int32 = (bhd_o * 256 + bhd_i) // 256 % 8 + d: T.int32 = (bhd_o * 256 + bhd_i) % 256 + if bhd_o * 256 + bhd_i < batch_size * 8 * 256: + 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, 8, page_size, 256), "float16", offset_factor=1) + # with T.block("root"): + for b in T.thread_binding(copy_length * 8, thread="blockIdx.x"): + for t in T.thread_binding(256, thread="threadIdx.x"): + with T.block("copy"): + vh = T.axis.spatial(8, (b * 256 + t) // (copy_length * 256)) + vp = T.axis.spatial(copy_length, (b * 256 + t) % (copy_length * 256) // 256) + vd = T.axis.spatial(256, (b * 256 + t) % 256) + T.where(b * 256 + t < copy_length * 8 * 256) + 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_norm33: T.Buffer((1, 1, 2048), "float16"), gpt_neox_layers_0_attention_query_key_value_weight2: T.Buffer((6144, 2048), "float16"), gpt_neox_layers_0_attention_query_key_value_bias2: T.Buffer((6144,), "float16"), T_add_intermediate: T.Buffer((1, 1, 6144), "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, 6144), "float16", scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((256, 1, 1, 6144), "float16", scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((32, 1, 1, 6144), "float16", scope="local") + gpt_neox_layers_0_attention_query_key_value_weight2_local = T.alloc_buffer((6144, 2048), "float16", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(3072, 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(6144, 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(8, 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(6144, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(2048, 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(6144, 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_norm33[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_norm33[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(6144, 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(6144, 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(6144, 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(6144, 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(reshape67: T.Buffer((1, 1, 2048), "float16"), gpt_neox_layers_0_attention_dense_weight2: T.Buffer((2048, 2048), "float16"), gpt_neox_layers_0_attention_dense_bias2: T.Buffer((2048,), "float16"), astype34: T.Buffer((1, 1, 2048), "float16"), input_embed: T.Buffer((1, 1, 2048), "float16"), T_add_intermediate_1_2: T.Buffer((1, 1, 2048), "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, 2048), "float16", scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((256, 1, 1, 2048), "float16", scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((32, 1, 1, 2048), "float16", scope="local") + gpt_neox_layers_0_attention_dense_weight2_local = T.alloc_buffer((2048, 2048), "float16", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(1024, 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(2048, 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(8, 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(2048, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(2048, 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(2048, 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], reshape67[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] + reshape67[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(2048, 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(2048, 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(2048, 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(2048, u_fused_ax0_fused_fused_0 * 2 + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) + T.reads(astype34[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] = astype34[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_norm34: T.Buffer((1, 1, 2048), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_weight2: T.Buffer((8192, 2048), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: T.Buffer((8192,), "float32"), compute_intermediate: T.Buffer((1, 1, 8192), "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, 8192), scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((256, 1, 1, 8192), scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((32, 1, 1, 8192), scope="local") + gpt_neox_layers_0_mlp_dense_h_to_4h_weight2_local = T.alloc_buffer((8192, 2048), "float16", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(4096, 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(8192, 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(8, 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(8192, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(2048, 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(8192, 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_norm34[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_norm34[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(8192, 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(8192, 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(8192, 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(8192, 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(astype33: T.Buffer((1, 1, 8192), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_weight2: T.Buffer((2048, 8192), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: T.Buffer((2048,), "float32"), compute_intermediate: T.Buffer((1, 1, 2048), "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, 2048), scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((256, 1, 1, 2048), scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((32, 1, 1, 2048), scope="local") + gpt_neox_layers_0_mlp_dense_4h_to_h_weight2_local = T.alloc_buffer((2048, 8192), "float16", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(1024, 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(2048, 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(32, 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(2048, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(8192, 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(2048, 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], astype33[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", astype33[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(2048, 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(2048, 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(2048, 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(2048, 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_norm65: T.Buffer((1, 1, 2048), "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, 2048), "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, 2048), "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(8, 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(2048, 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_norm65[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_norm65[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_reshape195: T.handle, gpt_neox_layers_0_attention_dense_weight4: T.Buffer((2048, 2048), "float16"), gpt_neox_layers_0_attention_dense_bias4: T.Buffer((2048,), "float16"), p_astype100: 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() + reshape195 = T.match_buffer(p_reshape195, (batch_size, 1, 2048), "float16") + astype100 = T.match_buffer(p_astype100, (batch_size, 1, 2048), "float16") + input_embeds = T.match_buffer(p_input_embeds, (batch_size, 1, 2048), "float16") + T_add_intermediate_1_2 = T.match_buffer(p_output0, (batch_size, 1, 2048), "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, 2048), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 1) // 2 * 2, 1, 2048), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 1) // 2 * 2, 1, 2048), "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(512, 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(2048, 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(8, 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(2048, 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], reshape195[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, reshape195[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(2048, 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(2048, 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(2048, 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(2048, 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(astype100[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] = astype100[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, 2048), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 3) // 4 * 4, 1, 2048), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 3) // 4 * 4, 1, 2048), "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(512, 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(2048, 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(8, 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(2048, 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], reshape195[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, reshape195[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(2048, 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(2048, 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(2048, 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(2048, 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(astype100[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] = astype100[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, 2048), "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(8, 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(2048, 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(128, 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(2048, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(2048, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], reshape195[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, reshape195[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(2048, 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(astype100[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] = astype100[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_norm100: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_weight4: T.Buffer((8192, 2048), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: T.Buffer((8192,), "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_norm100 = T.match_buffer(p_layer_norm100, (batch_size, 1, 2048), "float16") + compute_intermediate = T.match_buffer(p_output0, (batch_size, 1, 8192), "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, 8192), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 1) // 2 * 2, 1, 8192), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 1) // 2 * 2, 1, 8192), scope="local") + for ax0_0 in T.thread_binding((batch_size + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(2048, 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(8192, 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(8, 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(8192, 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_norm100[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_norm100[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(8192, 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(8192, 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(8192, 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(8192, 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, 8192), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 3) // 4 * 4, 1, 8192), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 3) // 4 * 4, 1, 8192), scope="local") + for ax0_0 in T.thread_binding((batch_size + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(2048, 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(8192, 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(8, 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(8192, 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_norm100[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_norm100[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(8192, 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(8192, 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(8192, 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(8192, 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, 8192), 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(32, 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(8192, 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(128, 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(8192, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(2048, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm100[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_norm100[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(8192, 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_astype99: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_weight4: T.Buffer((2048, 8192), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: T.Buffer((2048,), "float32"), p_output0: T.handle): + T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + astype99 = T.match_buffer(p_astype99, (batch_size, 1, 8192), "float16") + compute_intermediate = T.match_buffer(p_output0, (batch_size, 1, 2048), "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, 2048), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 1) // 2 * 2, 1, 2048), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 1) // 2 * 2, 1, 2048), scope="local") + for ax0_0 in T.thread_binding((batch_size + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(512, 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(2048, 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(32, 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(2048, 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], astype99[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, astype99[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(2048, 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(2048, 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(2048, 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(2048, 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, 2048), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 3) // 4 * 4, 1, 2048), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 3) // 4 * 4, 1, 2048), scope="local") + for ax0_0 in T.thread_binding((batch_size + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(512, 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(2048, 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(32, 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(2048, 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], astype99[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, astype99[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(2048, 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(2048, 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(2048, 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(2048, 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, 2048), 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(8, 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(2048, 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(512, 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(2048, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(8192, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], astype99[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, astype99[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(2048, 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_norm131: 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_norm131 = T.match_buffer(p_layer_norm131, (batch_size, 1, 2048), "float16") + vocab_size = T.int32() + embed_out_weight4 = T.match_buffer(p_embed_out_weight4, (vocab_size, 2048), "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(128, 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(2048, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm131[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_norm131[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_norm66: T.handle, gpt_neox_layers_0_attention_query_key_value_weight3: T.Buffer((6144, 2048), "float16"), gpt_neox_layers_0_attention_query_key_value_bias3: T.Buffer((6144,), "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_norm66 = T.match_buffer(p_layer_norm66, (1, seq_len, 2048), "float16") + T_add_intermediate = T.match_buffer(p_output0, (1, seq_len, 6144), "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, 6144), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 1) // 2 * 2, 6144), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 1) // 2 * 2, 6144), "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(1536, 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(6144, 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(8, 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(6144, 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_norm66[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_norm66[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(6144, 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(6144, 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(6144, 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(6144, 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, 6144), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 3) // 4 * 4, 6144), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 3) // 4 * 4, 6144), "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(1536, 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(6144, 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(8, 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(6144, 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_norm66[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_norm66[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(6144, 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(6144, 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(6144, 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(6144, 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, 6144), "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(24, 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(6144, 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(128, 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(6144, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(2048, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm66[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_norm66[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(6144, 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_reshape131: T.handle, gpt_neox_layers_0_attention_dense_weight3: T.Buffer((2048, 2048), "float16"), gpt_neox_layers_0_attention_dense_bias3: T.Buffer((2048,), "float16"), p_astype67: 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() + reshape131 = T.match_buffer(p_reshape131, (1, seq_len, 2048), "float16") + astype67 = T.match_buffer(p_astype67, (1, seq_len, 2048), "float16") + input_embeds = T.match_buffer(p_input_embeds, (1, seq_len, 2048), "float16") + T_add_intermediate_1_2 = T.match_buffer(p_output0, (1, seq_len, 2048), "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, 2048), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 1) // 2 * 2, 2048), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 1) // 2 * 2, 2048), "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(512, 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(2048, 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(8, 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(2048, 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], reshape131[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, reshape131[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(2048, 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(2048, 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(2048, 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(2048, 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(astype67[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] = astype67[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, 2048), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 3) // 4 * 4, 2048), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 3) // 4 * 4, 2048), "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(512, 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(2048, 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(8, 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(2048, 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], reshape131[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, reshape131[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(2048, 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(2048, 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(2048, 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(2048, 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(astype67[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] = astype67[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, 2048), "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(8, 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(2048, 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(128, 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(2048, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(2048, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], reshape131[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, reshape131[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(2048, 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(astype67[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] = astype67[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_norm67: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_weight3: T.Buffer((8192, 2048), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: T.Buffer((8192,), "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_norm67 = T.match_buffer(p_layer_norm67, (1, seq_len, 2048), "float16") + compute_intermediate = T.match_buffer(p_output0, (1, seq_len, 8192), "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, 8192), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 1) // 2 * 2, 8192), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 1) // 2 * 2, 8192), scope="local") + for ax0_0 in T.thread_binding((seq_len + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(2048, 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(8192, 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(8, 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(8192, 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_norm67[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_norm67[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(8192, 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(8192, 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(8192, 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(8192, 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, 8192), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 3) // 4 * 4, 8192), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 3) // 4 * 4, 8192), scope="local") + for ax0_0 in T.thread_binding((seq_len + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(2048, 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(8192, 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(8, 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(8192, 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_norm67[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_norm67[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(8192, 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(8192, 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(8192, 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(8192, 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, 8192), 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(32, 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(8192, 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(128, 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(8192, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(2048, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm67[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_norm67[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(8192, 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_astype66: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_weight3: T.Buffer((2048, 8192), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: T.Buffer((2048,), "float32"), p_output0: T.handle): + T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + astype66 = T.match_buffer(p_astype66, (1, seq_len, 8192), "float16") + compute_intermediate = T.match_buffer(p_output0, (1, seq_len, 2048), "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, 2048), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 1) // 2 * 2, 2048), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 1) // 2 * 2, 2048), scope="local") + for ax0_0 in T.thread_binding((seq_len + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(512, 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(2048, 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(32, 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(2048, 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], astype66[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, astype66[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(2048, 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(2048, 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(2048, 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(2048, 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, 2048), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 3) // 4 * 4, 2048), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 3) // 4 * 4, 2048), scope="local") + for ax0_0 in T.thread_binding((seq_len + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(512, 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(2048, 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(32, 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(2048, 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], astype66[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, astype66[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(2048, 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(2048, 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(2048, 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(2048, 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, 2048), 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(8, 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(2048, 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(512, 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(2048, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(8192, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], astype66[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, astype66[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(2048, 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, 2048), "float16") + vocab_size = T.int32() + embed_out_weight3 = T.match_buffer(p_embed_out_weight3, (vocab_size, 2048), "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(128, 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(2048, 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_norm99: T.handle, gpt_neox_layers_0_attention_query_key_value_weight4: T.Buffer((6144, 2048), "float16"), gpt_neox_layers_0_attention_query_key_value_bias4: T.Buffer((6144,), "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_norm99 = T.match_buffer(p_layer_norm99, (batch_size, 1, 2048), "float16") + T_add_intermediate = T.match_buffer(p_output0, (batch_size, 1, 6144), "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, 6144), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 1) // 2 * 2, 1, 6144), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 1) // 2 * 2, 1, 6144), "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(1536, 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(6144, 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(8, 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(6144, 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_norm99[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_norm99[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(6144, 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(6144, 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(6144, 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(6144, 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, 6144), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 3) // 4 * 4, 1, 6144), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 3) // 4 * 4, 1, 6144), "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(1536, 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(6144, 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(8, 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(6144, 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_norm99[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_norm99[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(6144, 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(6144, 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(6144, 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(6144, 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, 6144), "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(24, 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(6144, 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(128, 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(6144, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(2048, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm99[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_norm99[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(6144, 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(lv17: T.Buffer((1, 8, 256), "float16"), T_reshape_intermediate_1: T.Buffer((1, 1, 2048), "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(8, 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(2048, ax0_fused_0 * 256 + ax0_fused_1) + T.reads(lv17[0, v0 // 256, v0 % 256]) + T.writes(T_reshape_intermediate_1[0, 0, v0]) + T_reshape_intermediate_1[0, 0, v0] = lv17[0, v0 // 256, v0 % 256] + + @T.prim_func + def gpt_neox_q0f16_fused_reshape8_reshape9(add96: T.Buffer((1, 1, 6144), "float16"), T_reshape_intermediate_1: T.Buffer((1, 24, 256), "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(24, 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(24, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) // 256) + v1 = T.axis.spatial(256, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 256) + T.reads(add96[0, 0, v0 * 256 + v1]) + T.writes(T_reshape_intermediate_1[0, v0, v1]) + T_reshape_intermediate_1[0, v0, v1] = add96[0, 0, v0 * 256 + 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, 24, 256), "float16") + position_map = T.match_buffer(var_position_map, (seq_len,), "int32", offset_factor=1) + q = T.match_buffer(var_q, (seq_len, 8, 256), "float16") + k = T.match_buffer(var_k, (seq_len, 8, 256), "float16") + v = T.match_buffer(var_v, (seq_len, 8, 256), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 24, 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) // 6144) + v1 = T.axis.spatial(24, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 6144 // 256) + v2 = T.axis.spatial(256, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 256) + T.reads(position_map[v0], qkv[v0, v1, v2 + -32:v2 + -32 + 65]) + T.writes(q[v0, v1, v2], k[v0, v1 + -8, v2], v[v0, v1 + -16, v2]) + if v1 < 8: + freq = T.float32() + q[v0, v1, v2] = T.if_then_else(apply_rope > 0 and v2 < 64, 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 < 32, qkv[v0, v1, v2 + 32] * T.float16(-1.0), qkv[v0, v1, v2 + -32]))), where={freq: T.Cast("float32", position_map[v0]) / T.pow(T.float32(10000.0), T.Cast("float32", v2 * 2 % 64) / T.float32(64.0))}), qkv[v0, v1, v2]) + else: + if v1 < 16: + freq = T.float32() + k[v0, v1 + -8, v2] = T.if_then_else(apply_rope > 0 and v2 < 64, 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 < 32, qkv[v0, v1, v2 + 32] * T.float16(-1.0), qkv[v0, v1, v2 + -32]))), where={freq: T.Cast("float32", position_map[v0]) / T.pow(T.float32(10000.0), T.Cast("float32", v2 * 2 % 64) / T.float32(64.0))}), qkv[v0, v1, v2]) + else: + v[v0, v1 + -16, 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_norm32: T.handle, index: T.Buffer((1, 1, 2048), "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_norm32 = T.match_buffer(var_layer_norm32, (1, seq_len, 2048), "float16") + # with T.block("root"): + for ax0_fused_0 in T.thread_binding(8, thread="blockIdx.x"): + for ax0_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("index"): + v0 = T.axis.spatial(2048, ax0_fused_0 * 256 + ax0_fused_1) + T.reads(layer_norm32[0, seq_len - 1, v0]) + T.writes(index[0, 0, v0]) + index[0, 0, v0] = layer_norm32[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((2048,), "float16"), gpt_neox_layers_0_input_layernorm_bias4: T.Buffer((2048,), "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, 2048), "float16") + T_layer_norm = T.match_buffer(var_T_layer_norm, (batch_size, 1, 2048), "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(8, 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(2048, 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(8, 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(2048, 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(0.00048828125)) * T.rsqrt(input_embeds_red_temp_v1_shared[v0, 0] * T.float32(0.00048828125) - input_embeds_red_temp_v0_shared[v0, 0] * T.float32(0.00048828125) * (input_embeds_red_temp_v0_shared[v0, 0] * T.float32(0.00048828125)) + 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((2048,), "float16"), gpt_neox_layers_0_input_layernorm_bias3: T.Buffer((2048,), "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, 2048), "float16") + T_layer_norm = T.match_buffer(var_T_layer_norm, (1, seq_len, 2048), "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(8, 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(2048, 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(8, 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(2048, 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(0.00048828125)) * T.rsqrt(input_embeds_red_temp_v1_shared[0, v0] * T.float32(0.00048828125) - input_embeds_red_temp_v0_shared[0, v0] * T.float32(0.00048828125) * (input_embeds_red_temp_v0_shared[0, v0] * T.float32(0.00048828125)) + 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, 2048), "float16"), gpt_neox_layers_0_input_layernorm_weight2: T.Buffer((2048,), "float16"), gpt_neox_layers_0_input_layernorm_bias2: T.Buffer((2048,), "float16"), T_layer_norm: T.Buffer((1, 1, 2048), "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(8, 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(2048, 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(8, 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(2048, 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(0.00048828125)) * T.rsqrt(input_embed_red_temp_v1_shared[0, 0] * T.float32(0.00048828125) - input_embed_red_temp_v0_shared[0, 0] * T.float32(0.00048828125) * (input_embed_red_temp_v0_shared[0, 0] * T.float32(0.00048828125)) + 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(2, thread="blockIdx.y"): + for ty in T.thread_binding(4, thread="threadIdx.y"): + for tx in T.thread_binding(64, thread="threadIdx.x"): + with T.block("merge"): + T.reads(S[bx, ty + by * 4], S_other[bx, ty + by * 4], V[bx, ty + by * 4, tx * 4:tx * 4 + 4], V_other[bx, ty + by * 4, tx * 4:tx * 4 + 4]) + T.writes(V[bx, ty + by * 4, tx * 4:tx * 4 + 4], S[bx, ty + by * 4]) + 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 * 4] + s_other_val[0] = S_other[bx, ty + by * 4] + 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 * 4, tx * 4 + vec] + for vec in T.vectorized(4): + v_other_vec[vec] = V_other[bx, ty + by * 4, 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 * 4, tx * 4 + vec] = v_vec[vec] + S[bx, ty + by * 4] = T.log2(s_val[0] + s_other_val[0]) + s_max[0] + + @T.prim_func + def gpt_neox_q0f16_reshape(var_add288: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + add288 = T.match_buffer(var_add288, (batch_size, 1, 6144), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, 1, 24, 256), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * 24, 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) // 6144) + v1 = T.axis.spatial(24, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 6144 // 256) + v2 = T.axis.spatial(256, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 256) + T.reads(add288[v0, 0, v1 * 256 + v2]) + T.writes(T_reshape[v0, 0, v1, v2]) + T_reshape[v0, 0, v1, v2] = add288[v0, 0, v1 * 256 + v2] + + @T.prim_func + def gpt_neox_q0f16_reshape1(var_reshape192: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + reshape192 = T.match_buffer(var_reshape192, (batch_size, 1, 24, 256), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, 24, 256), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * 24, 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) // 6144) + v1 = T.axis.spatial(24, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 6144 // 256) + v2 = T.axis.spatial(256, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 256) + T.reads(reshape192[v0, 0, v1, v2]) + T.writes(T_reshape[v0, v1, v2]) + T_reshape[v0, v1, v2] = reshape192[v0, 0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_reshape2(var_lv49: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + lv49 = T.match_buffer(var_lv49, (batch_size, 8, 256), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, 1, 8, 256), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * 8, 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) // 2048) + v1 = T.axis.spatial(8, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 2048 // 256) + v2 = T.axis.spatial(256, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 256) + T.reads(lv49[v0, v1, v2]) + T.writes(T_reshape[v0, 0, v1, v2]) + T_reshape[v0, 0, v1, v2] = lv49[v0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_reshape3(var_reshape194: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + reshape194 = T.match_buffer(var_reshape194, (batch_size, 1, 8, 256), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, 1, 2048), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(batch_size * 8, 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) // 2048) + v1 = T.axis.spatial(2048, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 2048) + T.reads(reshape194[v0, 0, v1 // 256, v1 % 256]) + T.writes(T_reshape[v0, 0, v1]) + T_reshape[v0, 0, v1] = reshape194[v0, 0, v1 // 256, v1 % 256] + + @T.prim_func + def gpt_neox_q0f16_reshape4(var_add192: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + add192 = T.match_buffer(var_add192, (1, seq_len, 6144), "float16") + T_reshape = T.match_buffer(var_T_reshape, (1, seq_len, 24, 256), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 24, 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) // 6144) + v1 = T.axis.spatial(24, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 6144 // 256) + v2 = T.axis.spatial(256, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 256) + T.reads(add192[0, v0, v1 * 256 + v2]) + T.writes(T_reshape[0, v0, v1, v2]) + T_reshape[0, v0, v1, v2] = add192[0, v0, v1 * 256 + v2] + + @T.prim_func + def gpt_neox_q0f16_reshape5(var_reshape128: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + reshape128 = T.match_buffer(var_reshape128, (1, seq_len, 24, 256), "float16") + T_reshape = T.match_buffer(var_T_reshape, (seq_len, 24, 256), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 24, 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) // 6144) + v1 = T.axis.spatial(24, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 6144 // 256) + v2 = T.axis.spatial(256, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 256) + T.reads(reshape128[0, v0, v1, v2]) + T.writes(T_reshape[v0, v1, v2]) + T_reshape[v0, v1, v2] = reshape128[0, v0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_reshape6(var_lv33: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + lv33 = T.match_buffer(var_lv33, (seq_len, 8, 256), "float16") + T_reshape = T.match_buffer(var_T_reshape, (1, seq_len, 8, 256), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 8, 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) // 2048) + v1 = T.axis.spatial(8, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 2048 // 256) + v2 = T.axis.spatial(256, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 256) + T.reads(lv33[v0, v1, v2]) + T.writes(T_reshape[0, v0, v1, v2]) + T_reshape[0, v0, v1, v2] = lv33[v0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_reshape7(var_reshape130: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + reshape130 = T.match_buffer(var_reshape130, (1, seq_len, 8, 256), "float16") + T_reshape = T.match_buffer(var_T_reshape, (1, seq_len, 2048), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(seq_len * 8, 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) // 2048) + v1 = T.axis.spatial(2048, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 2048) + T.reads(reshape130[0, v0, v1 // 256, v1 % 256]) + T.writes(T_reshape[0, v0, v1]) + T_reshape[0, v0, v1] = reshape130[0, v0, v1 // 256, v1 % 256] + + @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(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_func1(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_func2(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_func3(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_shape_func4(H: T.Buffer((T.int64(5),), "int64")): + T.func_attr({"tir.is_host_func": 1}) + H[T.int64(4)] = T.int64(512) * H[T.int64(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)) + + @T.prim_func + def gpt_neox_q0f16_shape_func5(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_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_norm98: 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_norm98 = T.match_buffer(var_layer_norm98, (1, seq_len, 2048), "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, 2048), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(batch_size * 8, 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) // 2048) + v1 = T.axis.spatial(2048, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 2048) + T.reads(layer_norm98[0, logit_positions[v0], v1], logit_positions[v0]) + T.writes(T_take[0, v0, v1]) + T_take[0, v0, v1] = layer_norm98[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, 2048), "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, 2048), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(seq_len * 8, 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) // 2048) + v1 = T.axis.spatial(2048, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 2048) + 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, 8, page_size, 256), "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, (16, seqlen, 8, 256), "float16") + v_data = T.match_buffer(var_v_data, (16, seqlen, 8, 256), "float16") + # with T.block("root"): + for p_h_d_fused_0 in T.thread_binding(seqlen * 8, 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) // 2048) + vh = T.axis.spatial(8, (p_h_d_fused_0 * 256 + p_h_d_fused_1) % 2048 // 256) + vd = T.axis.spatial(256, (p_h_d_fused_0 * 256 + p_h_d_fused_1) % 256) + 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, 8, 16, 256), "float16", offset_factor=1) + ntoken = T.int32(is_size_var=True) + k_data = T.match_buffer(var_k_data, (ntoken, 8, 256), "float16") + v_data = T.match_buffer(var_v_data, (ntoken, 8, 256), "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 * 8, 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) // 2048] != -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) // 2048) + vh = T.axis.spatial(8, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) % 2048 // 256) + vf = T.axis.spatial(256, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) % 256) + 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) // 2048) + vh = T.axis.spatial(8, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) % 2048 // 256) + vf = T.axis.spatial(256, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) % 256) + 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, 8, 256), "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, 8, 32, 256), "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, 8, 256), "float16") + lse = T.match_buffer(var_lse, (total_len, 8)) + 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(8, 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((16, 256), "float16", scope="shared") + K_smem = T.alloc_buffer((32, 256), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 256), "float16", scope="shared") + S_smem = T.alloc_buffer((16, 32), scope="shared") + S_local = T.alloc_buffer((16, 32), scope="local") + O_local = T.alloc_buffer((16, 256), scope="local") + m_smem = T.alloc_buffer((16,), scope="shared") + m_prev_smem = T.alloc_buffer((16,), scope="shared") + d_smem = T.alloc_buffer((16,), 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] + 16 - 1) // 16 + 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] + 16 - 1) // 16 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 16 + 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 < 16: + 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, 8): + with T.block("O_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + 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(8): + 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(16, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 256) + j = T.axis.spatial(256, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 256) + 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 < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), 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(16): + 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) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + 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(16): + 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) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + 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[0:16, 0:256], K_smem[0:32, 0:256]) + T.writes(S_local[0:16, 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, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 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(32, 2, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 2 + lj_1) + k = T.axis.reduce(256, 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(0.090168440055560212) + 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(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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 < 16: + 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 < 16: + 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 < 16: + 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[0:16], m_smem[0:16], S_smem[0:16, 0:32], V_smem[0:32, 0:256]) + T.writes(O_local[0:16, 0:256]) + 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, 8): + with T.block("O_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) + j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + 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, 8): + with T.block("O_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + 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, 8): + with T.block("O_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + 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(16, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) + 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\", 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.12.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.13.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.14.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.input_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.input_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.post_attention_layernorm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.post_attention_layernorm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.attention.query_key_value.weight\", \"shape\": [6144, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.attention.query_key_value.bias\", \"shape\": [6144], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.attention.dense.weight\", \"shape\": [2048, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.attention.dense.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.mlp.dense_h_to_4h.weight\", \"shape\": [8192, 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.mlp.dense_h_to_4h.bias\", \"shape\": [8192], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.mlp.dense_4h_to_h.weight\", \"shape\": [2048, 8192], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.15.mlp.dense_4h_to_h.bias\", \"shape\": [2048], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.final_layer_norm.weight\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.final_layer_norm.bias\", \"shape\": [2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"embed_out.weight\", \"shape\": [\"vocab_size\", 2048], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}], \"kv_cache\": {\"num_hidden_layers\": 16, \"num_attention_heads\": 8, \"num_key_value_heads\": 8, \"head_dim\": 256}, \"memory_usage\": {\"alloc_embedding_tensor\": 8388608, \"batch_decode\": 6291456, \"batch_prefill\": 101187584, \"batch_verify\": 100663296, \"create_tir_paged_kv_cache\": 0, \"decode\": 49152, \"embed\": 8388608, \"prefill\": 100667392, \"softmax_with_temperature\": 0}}") + + @R.function + def alloc_embedding_tensor() -> R.Tensor((2048, 2048), 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([8388608]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv: R.Tensor((2048, 2048), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([2048, 2048]), R.dtype("float16")) + R.vm.kill_object(storage) + return gv + + @R.function + def batch_decode(input_embeds: R.Tensor(("batch_size", 1, 2048), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 2048), 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, 2048), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(196), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv564: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv564, 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, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv565: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[195] + R.call_packed("vm.builtin.check_tensor_info", gv565, 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, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), 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(2048), R.str("ErrorContext(fn=batch_decode, loc=param[0], param=input_embeds, annotation=R.Tensor((batch_size, 1, 2048), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv564, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(2048), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv565, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(2048), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func5(shape_heap) + gpt_neox_layers_0_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[2] + storage26: R.Object = R.vm.alloc_storage(R.shape([524288]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc462: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv566, 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, alloc462) + 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((6144, 2048), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[6] + storage27: R.Object = R.vm.alloc_storage(R.shape([1572864]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv567: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc463: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv567, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc462, gpt_neox_layers_0_attention_query_key_value_weight4, gpt_neox_layers_0_attention_query_key_value_bias4, alloc463) + R.vm.kill_object(alloc462) + 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) + gv568: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape192: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc463, gv568, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc463) + gv569: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape193: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape192, gv569, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape192) + gv570: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc464: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv570, R.dtype("float16")) + _463: 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, alloc464) + R.vm.kill_object(reshape193) + gv571: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape194: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc464, gv571, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc464) + gv572: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape195: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape194, gv572, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape194) + gpt_neox_layers_0_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[4] + gv573: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc465: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv573, 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, alloc465) + 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((8192, 2048), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[10] + storage28: R.Object = R.vm.alloc_storage(R.shape([2097152]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv574: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc466: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv574, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc465, gpt_neox_layers_0_mlp_dense_h_to_4h_weight4, gpt_neox_layers_0_mlp_dense_h_to_4h_bias4, alloc466) + R.vm.kill_object(alloc465) + 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((2048, 8192), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[12] + gv575: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc467: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv575, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc466, gpt_neox_layers_0_mlp_dense_4h_to_h_weight4, gpt_neox_layers_0_mlp_dense_4h_to_h_bias4, alloc467) + R.vm.kill_object(alloc466) + 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((2048, 2048), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[8] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc468: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv576, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape195, gpt_neox_layers_0_attention_dense_weight4, gpt_neox_layers_0_attention_dense_bias4, alloc467, input_embeds, alloc468) + R.vm.kill_object(reshape195) + R.vm.kill_object(alloc467) + 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((2048,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[14] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc469: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv577, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc468, gpt_neox_layers_1_input_layernorm_weight4, gpt_neox_layers_1_input_layernorm_bias4, alloc469) + 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((6144, 2048), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[18] + gv578: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc470: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv578, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc469, gpt_neox_layers_1_attention_query_key_value_weight4, gpt_neox_layers_1_attention_query_key_value_bias4, alloc470) + R.vm.kill_object(alloc469) + 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) + gv579: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape196: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc470, gv579, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc470) + gv580: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape197: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape196, gv580, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape196) + gv581: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc471: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv581, R.dtype("float16")) + _470: 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, alloc471) + R.vm.kill_object(reshape197) + gv582: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape198: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc471, gv582, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc471) + gv583: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape199: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape198, gv583, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape198) + gpt_neox_layers_1_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[16] + gv584: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc472: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv584, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc468, gpt_neox_layers_1_post_attention_layernorm_weight4, gpt_neox_layers_1_post_attention_layernorm_bias4, alloc472) + 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((8192, 2048), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[22] + storage29: R.Object = R.vm.alloc_storage(R.shape([2097152]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv585: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc473: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv585, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc472, gpt_neox_layers_1_mlp_dense_h_to_4h_weight4, gpt_neox_layers_1_mlp_dense_h_to_4h_bias4, alloc473) + R.vm.kill_object(alloc472) + 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((2048, 8192), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[24] + gv586: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc474: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv586, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc473, gpt_neox_layers_1_mlp_dense_4h_to_h_weight4, gpt_neox_layers_1_mlp_dense_4h_to_h_bias4, alloc474) + R.vm.kill_object(alloc473) + 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((2048, 2048), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[20] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc475: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv587, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape199, gpt_neox_layers_1_attention_dense_weight4, gpt_neox_layers_1_attention_dense_bias4, alloc474, alloc468, alloc475) + R.vm.kill_object(alloc468) + R.vm.kill_object(reshape199) + R.vm.kill_object(alloc474) + 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((2048,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[26] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc476: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv588, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc475, gpt_neox_layers_2_input_layernorm_weight4, gpt_neox_layers_2_input_layernorm_bias4, alloc476) + 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((6144, 2048), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[30] + gv589: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc477: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv589, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc476, gpt_neox_layers_2_attention_query_key_value_weight4, gpt_neox_layers_2_attention_query_key_value_bias4, alloc477) + R.vm.kill_object(alloc476) + 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) + gv590: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape200: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc477, gv590, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc477) + gv591: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape201: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape200, gv591, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape200) + gv592: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc478: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv592, R.dtype("float16")) + _477: 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, alloc478) + R.vm.kill_object(reshape201) + gv593: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape202: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc478, gv593, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc478) + gv594: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape203: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape202, gv594, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape202) + gpt_neox_layers_2_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[28] + gv595: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc479: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv595, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc475, gpt_neox_layers_2_post_attention_layernorm_weight4, gpt_neox_layers_2_post_attention_layernorm_bias4, alloc479) + 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((8192, 2048), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[34] + gv596: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc480: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv596, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc479, gpt_neox_layers_2_mlp_dense_h_to_4h_weight4, gpt_neox_layers_2_mlp_dense_h_to_4h_bias4, alloc480) + R.vm.kill_object(alloc479) + 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((2048, 8192), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[36] + gv597: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc481: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv597, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc480, gpt_neox_layers_2_mlp_dense_4h_to_h_weight4, gpt_neox_layers_2_mlp_dense_4h_to_h_bias4, alloc481) + R.vm.kill_object(alloc480) + 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((2048, 2048), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[32] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc482: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv598, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape203, gpt_neox_layers_2_attention_dense_weight4, gpt_neox_layers_2_attention_dense_bias4, alloc481, alloc475, alloc482) + R.vm.kill_object(alloc475) + R.vm.kill_object(reshape203) + R.vm.kill_object(alloc481) + 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((2048,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[38] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc483: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv599, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc482, gpt_neox_layers_3_input_layernorm_weight4, gpt_neox_layers_3_input_layernorm_bias4, alloc483) + 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((6144, 2048), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[42] + gv600: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc484: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv600, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc483, gpt_neox_layers_3_attention_query_key_value_weight4, gpt_neox_layers_3_attention_query_key_value_bias4, alloc484) + R.vm.kill_object(alloc483) + 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) + gv601: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape204: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc484, gv601, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc484) + gv602: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape205: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape204, gv602, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape204) + gv603: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc485: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv603, R.dtype("float16")) + _484: 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, alloc485) + R.vm.kill_object(reshape205) + gv604: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape206: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc485, gv604, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc485) + gv605: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape207: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape206, gv605, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape206) + gpt_neox_layers_3_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[40] + gv606: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc486: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv606, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc482, gpt_neox_layers_3_post_attention_layernorm_weight4, gpt_neox_layers_3_post_attention_layernorm_bias4, alloc486) + 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((8192, 2048), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[46] + gv607: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc487: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv607, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc486, gpt_neox_layers_3_mlp_dense_h_to_4h_weight4, gpt_neox_layers_3_mlp_dense_h_to_4h_bias4, alloc487) + R.vm.kill_object(alloc486) + 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((2048, 8192), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[48] + gv608: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc488: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv608, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc487, gpt_neox_layers_3_mlp_dense_4h_to_h_weight4, gpt_neox_layers_3_mlp_dense_4h_to_h_bias4, alloc488) + R.vm.kill_object(alloc487) + 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((2048, 2048), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[44] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc489: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv609, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape207, gpt_neox_layers_3_attention_dense_weight4, gpt_neox_layers_3_attention_dense_bias4, alloc488, alloc482, alloc489) + R.vm.kill_object(alloc482) + R.vm.kill_object(reshape207) + R.vm.kill_object(alloc488) + 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((2048,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[50] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc490: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv610, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc489, gpt_neox_layers_4_input_layernorm_weight4, gpt_neox_layers_4_input_layernorm_bias4, alloc490) + 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((6144, 2048), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[54] + gv611: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc491: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv611, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc490, gpt_neox_layers_4_attention_query_key_value_weight4, gpt_neox_layers_4_attention_query_key_value_bias4, alloc491) + R.vm.kill_object(alloc490) + 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) + gv612: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape208: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc491, gv612, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc491) + gv613: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape209: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape208, gv613, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape208) + gv614: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc492: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv614, R.dtype("float16")) + _491: 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, alloc492) + R.vm.kill_object(reshape209) + gv615: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape210: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc492, gv615, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc492) + gv616: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape211: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape210, gv616, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape210) + gpt_neox_layers_4_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[52] + gv617: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc493: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv617, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc489, gpt_neox_layers_4_post_attention_layernorm_weight4, gpt_neox_layers_4_post_attention_layernorm_bias4, alloc493) + 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((8192, 2048), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[58] + gv618: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc494: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv618, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc493, gpt_neox_layers_4_mlp_dense_h_to_4h_weight4, gpt_neox_layers_4_mlp_dense_h_to_4h_bias4, alloc494) + R.vm.kill_object(alloc493) + 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((2048, 8192), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[60] + gv619: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc495: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv619, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc494, gpt_neox_layers_4_mlp_dense_4h_to_h_weight4, gpt_neox_layers_4_mlp_dense_4h_to_h_bias4, alloc495) + R.vm.kill_object(alloc494) + 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((2048, 2048), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[56] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc496: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv620, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape211, gpt_neox_layers_4_attention_dense_weight4, gpt_neox_layers_4_attention_dense_bias4, alloc495, alloc489, alloc496) + R.vm.kill_object(alloc489) + R.vm.kill_object(reshape211) + R.vm.kill_object(alloc495) + 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((2048,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[62] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc497: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv621, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc496, gpt_neox_layers_5_input_layernorm_weight4, gpt_neox_layers_5_input_layernorm_bias4, alloc497) + 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((6144, 2048), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[66] + gv622: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc498: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv622, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc497, gpt_neox_layers_5_attention_query_key_value_weight4, gpt_neox_layers_5_attention_query_key_value_bias4, alloc498) + R.vm.kill_object(alloc497) + 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) + gv623: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape212: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc498, gv623, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc498) + gv624: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape213: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape212, gv624, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape212) + gv625: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc499: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv625, R.dtype("float16")) + _498: 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, alloc499) + R.vm.kill_object(reshape213) + gv626: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape214: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc499, gv626, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc499) + gv627: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape215: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape214, gv627, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape214) + gpt_neox_layers_5_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[64] + gv628: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc500: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv628, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc496, gpt_neox_layers_5_post_attention_layernorm_weight4, gpt_neox_layers_5_post_attention_layernorm_bias4, alloc500) + 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((8192, 2048), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[70] + gv629: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc501: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv629, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc500, gpt_neox_layers_5_mlp_dense_h_to_4h_weight4, gpt_neox_layers_5_mlp_dense_h_to_4h_bias4, alloc501) + R.vm.kill_object(alloc500) + 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((2048, 8192), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[72] + gv630: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc502: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv630, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc501, gpt_neox_layers_5_mlp_dense_4h_to_h_weight4, gpt_neox_layers_5_mlp_dense_4h_to_h_bias4, alloc502) + R.vm.kill_object(alloc501) + 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((2048, 2048), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[68] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc503: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv631, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape215, gpt_neox_layers_5_attention_dense_weight4, gpt_neox_layers_5_attention_dense_bias4, alloc502, alloc496, alloc503) + R.vm.kill_object(alloc496) + R.vm.kill_object(reshape215) + R.vm.kill_object(alloc502) + 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((2048,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[74] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc504: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv632, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc503, gpt_neox_layers_6_input_layernorm_weight4, gpt_neox_layers_6_input_layernorm_bias4, alloc504) + 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((6144, 2048), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[78] + gv633: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc505: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv633, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc504, gpt_neox_layers_6_attention_query_key_value_weight4, gpt_neox_layers_6_attention_query_key_value_bias4, alloc505) + R.vm.kill_object(alloc504) + 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) + gv634: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape216: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc505, gv634, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc505) + gv635: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape217: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape216, gv635, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape216) + gv636: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc506: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv636, R.dtype("float16")) + _505: 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, alloc506) + R.vm.kill_object(reshape217) + gv637: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape218: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc506, gv637, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc506) + gv638: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape219: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape218, gv638, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape218) + gpt_neox_layers_6_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[76] + gv639: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc507: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv639, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc503, gpt_neox_layers_6_post_attention_layernorm_weight4, gpt_neox_layers_6_post_attention_layernorm_bias4, alloc507) + 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((8192, 2048), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[82] + gv640: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc508: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv640, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc507, gpt_neox_layers_6_mlp_dense_h_to_4h_weight4, gpt_neox_layers_6_mlp_dense_h_to_4h_bias4, alloc508) + R.vm.kill_object(alloc507) + 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((2048, 8192), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[84] + gv641: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc509: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv641, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc508, gpt_neox_layers_6_mlp_dense_4h_to_h_weight4, gpt_neox_layers_6_mlp_dense_4h_to_h_bias4, alloc509) + R.vm.kill_object(alloc508) + 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((2048, 2048), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[80] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc510: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv642, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape219, gpt_neox_layers_6_attention_dense_weight4, gpt_neox_layers_6_attention_dense_bias4, alloc509, alloc503, alloc510) + R.vm.kill_object(alloc503) + R.vm.kill_object(reshape219) + R.vm.kill_object(alloc509) + 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((2048,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[86] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc511: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv643, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc510, gpt_neox_layers_7_input_layernorm_weight4, gpt_neox_layers_7_input_layernorm_bias4, alloc511) + 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((6144, 2048), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[90] + gv644: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc512: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv644, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc511, gpt_neox_layers_7_attention_query_key_value_weight4, gpt_neox_layers_7_attention_query_key_value_bias4, alloc512) + R.vm.kill_object(alloc511) + 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) + gv645: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape220: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc512, gv645, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc512) + gv646: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape221: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape220, gv646, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape220) + gv647: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc513: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv647, R.dtype("float16")) + _512: 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, alloc513) + R.vm.kill_object(reshape221) + gv648: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape222: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc513, gv648, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc513) + gv649: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape223: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape222, gv649, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape222) + gpt_neox_layers_7_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[88] + gv650: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc514: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv650, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc510, gpt_neox_layers_7_post_attention_layernorm_weight4, gpt_neox_layers_7_post_attention_layernorm_bias4, alloc514) + 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((8192, 2048), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[94] + gv651: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc515: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv651, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc514, gpt_neox_layers_7_mlp_dense_h_to_4h_weight4, gpt_neox_layers_7_mlp_dense_h_to_4h_bias4, alloc515) + R.vm.kill_object(alloc514) + 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((2048, 8192), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[96] + gv652: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc516: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv652, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc515, gpt_neox_layers_7_mlp_dense_4h_to_h_weight4, gpt_neox_layers_7_mlp_dense_4h_to_h_bias4, alloc516) + R.vm.kill_object(alloc515) + 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((2048, 2048), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[92] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc517: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv653, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape223, gpt_neox_layers_7_attention_dense_weight4, gpt_neox_layers_7_attention_dense_bias4, alloc516, alloc510, alloc517) + R.vm.kill_object(alloc510) + R.vm.kill_object(reshape223) + R.vm.kill_object(alloc516) + 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((2048,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[98] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc518: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv654, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc517, gpt_neox_layers_8_input_layernorm_weight4, gpt_neox_layers_8_input_layernorm_bias4, alloc518) + 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((6144, 2048), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[102] + gv655: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc519: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv655, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc518, gpt_neox_layers_8_attention_query_key_value_weight4, gpt_neox_layers_8_attention_query_key_value_bias4, alloc519) + R.vm.kill_object(alloc518) + 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) + gv656: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape224: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc519, gv656, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc519) + gv657: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape225: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape224, gv657, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape224) + gv658: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc520: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv658, R.dtype("float16")) + _519: 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, alloc520) + R.vm.kill_object(reshape225) + gv659: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape226: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc520, gv659, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc520) + gv660: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape227: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape226, gv660, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape226) + gpt_neox_layers_8_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[100] + gv661: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc521: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv661, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc517, gpt_neox_layers_8_post_attention_layernorm_weight4, gpt_neox_layers_8_post_attention_layernorm_bias4, alloc521) + 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((8192, 2048), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[106] + gv662: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc522: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv662, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc521, gpt_neox_layers_8_mlp_dense_h_to_4h_weight4, gpt_neox_layers_8_mlp_dense_h_to_4h_bias4, alloc522) + R.vm.kill_object(alloc521) + 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((2048, 8192), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[108] + gv663: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc523: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv663, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc522, gpt_neox_layers_8_mlp_dense_4h_to_h_weight4, gpt_neox_layers_8_mlp_dense_4h_to_h_bias4, alloc523) + R.vm.kill_object(alloc522) + 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((2048, 2048), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[104] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc524: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv664, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape227, gpt_neox_layers_8_attention_dense_weight4, gpt_neox_layers_8_attention_dense_bias4, alloc523, alloc517, alloc524) + R.vm.kill_object(alloc517) + R.vm.kill_object(reshape227) + R.vm.kill_object(alloc523) + 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((2048,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[110] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc525: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv665, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc524, gpt_neox_layers_9_input_layernorm_weight4, gpt_neox_layers_9_input_layernorm_bias4, alloc525) + 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((6144, 2048), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[114] + gv666: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc526: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv666, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc525, gpt_neox_layers_9_attention_query_key_value_weight4, gpt_neox_layers_9_attention_query_key_value_bias4, alloc526) + R.vm.kill_object(alloc525) + 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) + gv667: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape228: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc526, gv667, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc526) + gv668: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape229: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape228, gv668, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape228) + gv669: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc527: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv669, R.dtype("float16")) + _526: 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, alloc527) + R.vm.kill_object(reshape229) + gv670: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape230: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc527, gv670, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc527) + gv671: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape231: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape230, gv671, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape230) + gpt_neox_layers_9_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[112] + gv672: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc528: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv672, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc524, gpt_neox_layers_9_post_attention_layernorm_weight4, gpt_neox_layers_9_post_attention_layernorm_bias4, alloc528) + 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((8192, 2048), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[118] + gv673: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc529: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv673, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc528, gpt_neox_layers_9_mlp_dense_h_to_4h_weight4, gpt_neox_layers_9_mlp_dense_h_to_4h_bias4, alloc529) + R.vm.kill_object(alloc528) + 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((2048, 8192), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[120] + gv674: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc530: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv674, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc529, gpt_neox_layers_9_mlp_dense_4h_to_h_weight4, gpt_neox_layers_9_mlp_dense_4h_to_h_bias4, alloc530) + R.vm.kill_object(alloc529) + 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((2048, 2048), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[116] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc531: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv675, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape231, gpt_neox_layers_9_attention_dense_weight4, gpt_neox_layers_9_attention_dense_bias4, alloc530, alloc524, alloc531) + R.vm.kill_object(alloc524) + R.vm.kill_object(reshape231) + R.vm.kill_object(alloc530) + 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((2048,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[122] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc532: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv676, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc531, gpt_neox_layers_10_input_layernorm_weight4, gpt_neox_layers_10_input_layernorm_bias4, alloc532) + 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((6144, 2048), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[126] + gv677: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc533: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv677, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc532, gpt_neox_layers_10_attention_query_key_value_weight4, gpt_neox_layers_10_attention_query_key_value_bias4, alloc533) + R.vm.kill_object(alloc532) + 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) + gv678: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape232: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc533, gv678, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc533) + gv679: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape233: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape232, gv679, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape232) + gv680: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc534: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv680, R.dtype("float16")) + _533: 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, alloc534) + R.vm.kill_object(reshape233) + gv681: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape234: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc534, gv681, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc534) + gv682: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape235: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape234, gv682, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape234) + gpt_neox_layers_10_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[124] + gv683: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc535: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv683, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc531, gpt_neox_layers_10_post_attention_layernorm_weight4, gpt_neox_layers_10_post_attention_layernorm_bias4, alloc535) + 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((8192, 2048), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[130] + gv684: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc536: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv684, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc535, gpt_neox_layers_10_mlp_dense_h_to_4h_weight4, gpt_neox_layers_10_mlp_dense_h_to_4h_bias4, alloc536) + R.vm.kill_object(alloc535) + 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((2048, 8192), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[132] + gv685: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc537: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv685, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc536, gpt_neox_layers_10_mlp_dense_4h_to_h_weight4, gpt_neox_layers_10_mlp_dense_4h_to_h_bias4, alloc537) + R.vm.kill_object(alloc536) + 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((2048, 2048), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[128] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc538: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv686, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape235, gpt_neox_layers_10_attention_dense_weight4, gpt_neox_layers_10_attention_dense_bias4, alloc537, alloc531, alloc538) + R.vm.kill_object(alloc531) + R.vm.kill_object(reshape235) + R.vm.kill_object(alloc537) + 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((2048,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[134] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc539: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv687, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc538, gpt_neox_layers_11_input_layernorm_weight4, gpt_neox_layers_11_input_layernorm_bias4, alloc539) + 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((6144, 2048), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[138] + gv688: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc540: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv688, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc539, gpt_neox_layers_11_attention_query_key_value_weight4, gpt_neox_layers_11_attention_query_key_value_bias4, alloc540) + R.vm.kill_object(alloc539) + 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) + gv689: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape236: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc540, gv689, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc540) + gv690: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape237: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape236, gv690, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape236) + gv691: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc541: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv691, R.dtype("float16")) + _540: 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, alloc541) + R.vm.kill_object(reshape237) + gv692: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape238: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc541, gv692, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc541) + gv693: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape239: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape238, gv693, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape238) + gpt_neox_layers_11_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[136] + gv694: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc542: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv694, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc538, gpt_neox_layers_11_post_attention_layernorm_weight4, gpt_neox_layers_11_post_attention_layernorm_bias4, alloc542) + 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((8192, 2048), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[142] + gv695: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc543: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv695, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc542, gpt_neox_layers_11_mlp_dense_h_to_4h_weight4, gpt_neox_layers_11_mlp_dense_h_to_4h_bias4, alloc543) + R.vm.kill_object(alloc542) + 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((2048, 8192), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[144] + gv696: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc544: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv696, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc543, gpt_neox_layers_11_mlp_dense_4h_to_h_weight4, gpt_neox_layers_11_mlp_dense_4h_to_h_bias4, alloc544) + R.vm.kill_object(alloc543) + 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((2048, 2048), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[140] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc545: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv697, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape239, gpt_neox_layers_11_attention_dense_weight4, gpt_neox_layers_11_attention_dense_bias4, alloc544, alloc538, alloc545) + R.vm.kill_object(alloc538) + R.vm.kill_object(reshape239) + R.vm.kill_object(alloc544) + 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((2048,), dtype="float16") = packed_params[145] + gpt_neox_layers_12_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[146] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc546: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv698, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc545, gpt_neox_layers_12_input_layernorm_weight4, gpt_neox_layers_12_input_layernorm_bias4, alloc546) + 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((6144, 2048), dtype="float16") = packed_params[149] + gpt_neox_layers_12_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[150] + gv699: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc547: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv699, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc546, gpt_neox_layers_12_attention_query_key_value_weight4, gpt_neox_layers_12_attention_query_key_value_bias4, alloc547) + R.vm.kill_object(alloc546) + 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) + gv700: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape240: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc547, gv700, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc547) + gv701: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape241: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape240, gv701, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape240) + gv702: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc548: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv702, R.dtype("float16")) + _547: 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, alloc548) + R.vm.kill_object(reshape241) + gv703: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape242: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc548, gv703, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc548) + gv704: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape243: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape242, gv704, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape242) + gpt_neox_layers_12_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[147] + gpt_neox_layers_12_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[148] + gv705: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc549: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv705, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc545, gpt_neox_layers_12_post_attention_layernorm_weight4, gpt_neox_layers_12_post_attention_layernorm_bias4, alloc549) + 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((8192, 2048), dtype="float16") = packed_params[153] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[154] + gv706: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc550: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv706, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc549, gpt_neox_layers_12_mlp_dense_h_to_4h_weight4, gpt_neox_layers_12_mlp_dense_h_to_4h_bias4, alloc550) + R.vm.kill_object(alloc549) + 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((2048, 8192), dtype="float16") = packed_params[155] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[156] + gv707: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc551: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv707, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc550, gpt_neox_layers_12_mlp_dense_4h_to_h_weight4, gpt_neox_layers_12_mlp_dense_4h_to_h_bias4, alloc551) + R.vm.kill_object(alloc550) + 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((2048, 2048), dtype="float16") = packed_params[151] + gpt_neox_layers_12_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[152] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc552: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv708, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape243, gpt_neox_layers_12_attention_dense_weight4, gpt_neox_layers_12_attention_dense_bias4, alloc551, alloc545, alloc552) + R.vm.kill_object(alloc545) + R.vm.kill_object(reshape243) + R.vm.kill_object(alloc551) + 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((2048,), dtype="float16") = packed_params[157] + gpt_neox_layers_13_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[158] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc553: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv709, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc552, gpt_neox_layers_13_input_layernorm_weight4, gpt_neox_layers_13_input_layernorm_bias4, alloc553) + 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((6144, 2048), dtype="float16") = packed_params[161] + gpt_neox_layers_13_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[162] + gv710: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc554: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv710, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc553, gpt_neox_layers_13_attention_query_key_value_weight4, gpt_neox_layers_13_attention_query_key_value_bias4, alloc554) + R.vm.kill_object(alloc553) + 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) + gv711: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape244: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc554, gv711, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc554) + gv712: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape245: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape244, gv712, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape244) + gv713: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc555: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv713, R.dtype("float16")) + _554: 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, alloc555) + R.vm.kill_object(reshape245) + gv714: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape246: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc555, gv714, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc555) + gv715: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape247: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape246, gv715, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape246) + gpt_neox_layers_13_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[159] + gpt_neox_layers_13_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[160] + gv716: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc556: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv716, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc552, gpt_neox_layers_13_post_attention_layernorm_weight4, gpt_neox_layers_13_post_attention_layernorm_bias4, alloc556) + 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((8192, 2048), dtype="float16") = packed_params[165] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[166] + gv717: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc557: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv717, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc556, gpt_neox_layers_13_mlp_dense_h_to_4h_weight4, gpt_neox_layers_13_mlp_dense_h_to_4h_bias4, alloc557) + R.vm.kill_object(alloc556) + 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((2048, 8192), dtype="float16") = packed_params[167] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[168] + gv718: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc558: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv718, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc557, gpt_neox_layers_13_mlp_dense_4h_to_h_weight4, gpt_neox_layers_13_mlp_dense_4h_to_h_bias4, alloc558) + R.vm.kill_object(alloc557) + 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((2048, 2048), dtype="float16") = packed_params[163] + gpt_neox_layers_13_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[164] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc559: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv719, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape247, gpt_neox_layers_13_attention_dense_weight4, gpt_neox_layers_13_attention_dense_bias4, alloc558, alloc552, alloc559) + R.vm.kill_object(alloc552) + R.vm.kill_object(reshape247) + R.vm.kill_object(alloc558) + 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((2048,), dtype="float16") = packed_params[169] + gpt_neox_layers_14_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[170] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc560: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv720, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc559, gpt_neox_layers_14_input_layernorm_weight4, gpt_neox_layers_14_input_layernorm_bias4, alloc560) + 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((6144, 2048), dtype="float16") = packed_params[173] + gpt_neox_layers_14_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[174] + gv721: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc561: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv721, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc560, gpt_neox_layers_14_attention_query_key_value_weight4, gpt_neox_layers_14_attention_query_key_value_bias4, alloc561) + R.vm.kill_object(alloc560) + 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) + gv722: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape248: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc561, gv722, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc561) + gv723: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape249: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape248, gv723, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape248) + gv724: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc562: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv724, R.dtype("float16")) + _561: 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, alloc562) + R.vm.kill_object(reshape249) + gv725: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape250: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc562, gv725, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc562) + gv726: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape251: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape250, gv726, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape250) + gpt_neox_layers_14_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[171] + gpt_neox_layers_14_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[172] + gv727: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc563: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv727, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc559, gpt_neox_layers_14_post_attention_layernorm_weight4, gpt_neox_layers_14_post_attention_layernorm_bias4, alloc563) + 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((8192, 2048), dtype="float16") = packed_params[177] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[178] + gv728: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc564: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv728, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc563, gpt_neox_layers_14_mlp_dense_h_to_4h_weight4, gpt_neox_layers_14_mlp_dense_h_to_4h_bias4, alloc564) + R.vm.kill_object(alloc563) + 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((2048, 8192), dtype="float16") = packed_params[179] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[180] + gv729: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc565: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv729, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc564, gpt_neox_layers_14_mlp_dense_4h_to_h_weight4, gpt_neox_layers_14_mlp_dense_4h_to_h_bias4, alloc565) + R.vm.kill_object(alloc564) + 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((2048, 2048), dtype="float16") = packed_params[175] + gpt_neox_layers_14_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[176] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc566: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv730, R.dtype("float16")) + R.vm.kill_object(storage28) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape251, gpt_neox_layers_14_attention_dense_weight4, gpt_neox_layers_14_attention_dense_bias4, alloc565, alloc559, alloc566) + R.vm.kill_object(alloc559) + R.vm.kill_object(reshape251) + R.vm.kill_object(alloc565) + 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((2048,), dtype="float16") = packed_params[181] + gpt_neox_layers_15_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[182] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc567: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv731, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc566, gpt_neox_layers_15_input_layernorm_weight4, gpt_neox_layers_15_input_layernorm_bias4, alloc567) + 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((6144, 2048), dtype="float16") = packed_params[185] + gpt_neox_layers_15_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[186] + gv732: R.Shape(ndim=3) = R.call_packed("vm.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(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc568: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv732, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc567, gpt_neox_layers_15_attention_query_key_value_weight4, gpt_neox_layers_15_attention_query_key_value_bias4, alloc568) + R.vm.kill_object(alloc567) + 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) + gv733: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape252: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc568, gv733, sinfo_args=(R.Tensor((batch_size, 1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc568) + gv734: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape253: R.Tensor((batch_size, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape252, gv734, sinfo_args=(R.Tensor((batch_size, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape252) + gv735: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc569: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv735, R.dtype("float16")) + _568: 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, alloc569) + R.vm.kill_object(reshape253) + gv736: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape254: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc569, gv736, sinfo_args=(R.Tensor((batch_size, 1, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc569) + gv737: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape255: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape254, gv737, sinfo_args=(R.Tensor((batch_size, 1, 2048), dtype="float16"),)) + R.vm.kill_object(reshape254) + gpt_neox_layers_15_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[183] + gpt_neox_layers_15_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[184] + gv738: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc570: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv738, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc566, gpt_neox_layers_15_post_attention_layernorm_weight4, gpt_neox_layers_15_post_attention_layernorm_bias4, alloc570) + 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((8192, 2048), dtype="float16") = packed_params[189] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[190] + gv739: R.Shape(ndim=3) = R.call_packed("vm.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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc571: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv739, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc570, gpt_neox_layers_15_mlp_dense_h_to_4h_weight4, gpt_neox_layers_15_mlp_dense_h_to_4h_bias4, alloc571) + R.vm.kill_object(alloc570) + 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((2048, 8192), dtype="float16") = packed_params[191] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[192] + gv740: R.Shape(ndim=3) = R.call_packed("vm.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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc572: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv740, R.dtype("float16")) + R.vm.kill_object(storage27) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc571, gpt_neox_layers_15_mlp_dense_4h_to_h_weight4, gpt_neox_layers_15_mlp_dense_4h_to_h_bias4, alloc572) + R.vm.kill_object(alloc571) + 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((2048, 2048), dtype="float16") = packed_params[187] + gpt_neox_layers_15_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[188] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc573: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv741, R.dtype("float16")) + R.vm.kill_object(storage29) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape255, gpt_neox_layers_15_attention_dense_weight4, gpt_neox_layers_15_attention_dense_bias4, alloc572, alloc566, alloc573) + R.vm.kill_object(alloc566) + R.vm.kill_object(reshape255) + R.vm.kill_object(alloc572) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_bias4) + gpt_neox_final_layer_norm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[193] + gpt_neox_final_layer_norm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[194] + 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(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc574: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv742, R.dtype("float16")) + R.vm.kill_object(storage26) + cls.gpt_neox_q0f16_layer_norm(alloc573, gpt_neox_final_layer_norm_weight4, gpt_neox_final_layer_norm_bias4, alloc574) + R.vm.kill_object(alloc573) + 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, 2048), dtype="float16") = packed_params[195] + gv743: 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),)) + storage30: R.Object = R.vm.alloc_storage(gv743, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv744: R.Shape(ndim=3) = R.call_packed("vm.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),)) + alloc575: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage30, R.prim_value(0), gv744, R.dtype("float32")) + R.vm.kill_object(storage30) + cls.gpt_neox_q0f16_fused_NT_matmul4_cast2(alloc574, embed_out_weight4, alloc575) + R.vm.kill_object(alloc574) + R.vm.kill_object(embed_out_weight4) + gv4: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc575, paged_kv_cache + R.vm.kill_object(alloc575) + gv745: R.Tensor(dtype="float32", ndim=3) = gv4[0] + R.call_packed("vm.builtin.match_shape", gv745, 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", 2048), dtype="float16"), logit_positions: R.Tensor(("batch_size",), dtype="int32"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 2048), 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, 2048), 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(196), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv368: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv368, 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, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv369: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[195] + R.call_packed("vm.builtin.check_tensor_info", gv369, 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, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), 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(2048), R.str("ErrorContext(fn=batch_prefill, loc=param[0], param=input_embeds, annotation=R.Tensor((1, seq_len, 2048), 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", gv368, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(2), R.prim_value(0), R.prim_value(2048), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv369, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(2), R.prim_value(0), R.prim_value(2048), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func2(shape_heap) + gpt_neox_layers_0_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[2] + storage12: R.Object = R.vm.alloc_storage(R.shape([8388608]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc230: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv370, 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, alloc230) + 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((6144, 2048), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[6] + storage13: R.Object = R.vm.alloc_storage(R.shape([25165824]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv371: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc231: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv371, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc230, gpt_neox_layers_0_attention_query_key_value_weight3, gpt_neox_layers_0_attention_query_key_value_bias3, alloc231) + R.vm.kill_object(alloc230) + 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) + gv372: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape128: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc231, gv372, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc231) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape129: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape128, gv373, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape128) + gv374: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc232: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv374, R.dtype("float16")) + _232: 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)), reshape129, alloc232) + R.vm.kill_object(reshape129) + gv375: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape130: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc232, gv375, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc232) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape131: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape130, gv376, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape130) + gpt_neox_layers_0_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[4] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc233: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv377, 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, alloc233) + 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((8192, 2048), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[10] + storage14: R.Object = R.vm.alloc_storage(R.shape([33554432]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc234: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv378, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc233, gpt_neox_layers_0_mlp_dense_h_to_4h_weight3, gpt_neox_layers_0_mlp_dense_h_to_4h_bias3, alloc234) + R.vm.kill_object(alloc233) + 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((2048, 8192), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[12] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc235: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv379, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc234, gpt_neox_layers_0_mlp_dense_4h_to_h_weight3, gpt_neox_layers_0_mlp_dense_4h_to_h_bias3, alloc235) + R.vm.kill_object(alloc234) + 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((2048, 2048), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[8] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc236: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv380, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape131, gpt_neox_layers_0_attention_dense_weight3, gpt_neox_layers_0_attention_dense_bias3, alloc235, input_embeds, alloc236) + R.vm.kill_object(reshape131) + R.vm.kill_object(alloc235) + 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((2048,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[14] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc237: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv381, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc236, gpt_neox_layers_1_input_layernorm_weight3, gpt_neox_layers_1_input_layernorm_bias3, alloc237) + 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((6144, 2048), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[18] + gv382: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc238: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv382, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc237, gpt_neox_layers_1_attention_query_key_value_weight3, gpt_neox_layers_1_attention_query_key_value_bias3, alloc238) + R.vm.kill_object(alloc237) + 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) + gv383: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape132: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc238, gv383, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc238) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape133: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape132, gv384, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape132) + gv385: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc239: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv385, R.dtype("float16")) + _239: 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)), reshape133, alloc239) + R.vm.kill_object(reshape133) + gv386: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape134: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc239, gv386, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc239) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape135: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape134, gv387, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape134) + gpt_neox_layers_1_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[16] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc240: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv388, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc236, gpt_neox_layers_1_post_attention_layernorm_weight3, gpt_neox_layers_1_post_attention_layernorm_bias3, alloc240) + 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((8192, 2048), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[22] + storage15: R.Object = R.vm.alloc_storage(R.shape([33554432]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc241: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv389, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc240, gpt_neox_layers_1_mlp_dense_h_to_4h_weight3, gpt_neox_layers_1_mlp_dense_h_to_4h_bias3, alloc241) + R.vm.kill_object(alloc240) + 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((2048, 8192), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[24] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc242: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv390, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc241, gpt_neox_layers_1_mlp_dense_4h_to_h_weight3, gpt_neox_layers_1_mlp_dense_4h_to_h_bias3, alloc242) + R.vm.kill_object(alloc241) + 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((2048, 2048), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[20] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc243: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv391, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape135, gpt_neox_layers_1_attention_dense_weight3, gpt_neox_layers_1_attention_dense_bias3, alloc242, alloc236, alloc243) + R.vm.kill_object(alloc236) + R.vm.kill_object(reshape135) + R.vm.kill_object(alloc242) + 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((2048,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[26] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc244: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv392, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc243, gpt_neox_layers_2_input_layernorm_weight3, gpt_neox_layers_2_input_layernorm_bias3, alloc244) + 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((6144, 2048), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[30] + gv393: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc245: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv393, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc244, gpt_neox_layers_2_attention_query_key_value_weight3, gpt_neox_layers_2_attention_query_key_value_bias3, alloc245) + R.vm.kill_object(alloc244) + 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) + gv394: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape136: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc245, gv394, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc245) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape137: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape136, gv395, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape136) + gv396: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc246: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv396, R.dtype("float16")) + _246: 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)), reshape137, alloc246) + R.vm.kill_object(reshape137) + gv397: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape138: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc246, gv397, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc246) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape139: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape138, gv398, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape138) + gpt_neox_layers_2_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[28] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc247: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv399, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc243, gpt_neox_layers_2_post_attention_layernorm_weight3, gpt_neox_layers_2_post_attention_layernorm_bias3, alloc247) + 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((8192, 2048), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[34] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc248: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv400, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc247, gpt_neox_layers_2_mlp_dense_h_to_4h_weight3, gpt_neox_layers_2_mlp_dense_h_to_4h_bias3, alloc248) + R.vm.kill_object(alloc247) + 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((2048, 8192), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[36] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc249: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv401, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc248, gpt_neox_layers_2_mlp_dense_4h_to_h_weight3, gpt_neox_layers_2_mlp_dense_4h_to_h_bias3, alloc249) + R.vm.kill_object(alloc248) + 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((2048, 2048), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[32] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc250: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv402, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape139, gpt_neox_layers_2_attention_dense_weight3, gpt_neox_layers_2_attention_dense_bias3, alloc249, alloc243, alloc250) + R.vm.kill_object(alloc243) + R.vm.kill_object(reshape139) + R.vm.kill_object(alloc249) + 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((2048,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[38] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc251: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv403, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc250, gpt_neox_layers_3_input_layernorm_weight3, gpt_neox_layers_3_input_layernorm_bias3, alloc251) + 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((6144, 2048), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[42] + gv404: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc252: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv404, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc251, gpt_neox_layers_3_attention_query_key_value_weight3, gpt_neox_layers_3_attention_query_key_value_bias3, alloc252) + R.vm.kill_object(alloc251) + 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) + gv405: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape140: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc252, gv405, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc252) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape141: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape140, gv406, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape140) + gv407: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc253: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv407, R.dtype("float16")) + _253: 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)), reshape141, alloc253) + R.vm.kill_object(reshape141) + gv408: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape142: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc253, gv408, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc253) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape143: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape142, gv409, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape142) + gpt_neox_layers_3_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[40] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc254: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv410, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc250, gpt_neox_layers_3_post_attention_layernorm_weight3, gpt_neox_layers_3_post_attention_layernorm_bias3, alloc254) + 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((8192, 2048), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[46] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc255: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv411, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc254, gpt_neox_layers_3_mlp_dense_h_to_4h_weight3, gpt_neox_layers_3_mlp_dense_h_to_4h_bias3, alloc255) + R.vm.kill_object(alloc254) + 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((2048, 8192), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[48] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc256: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv412, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc255, gpt_neox_layers_3_mlp_dense_4h_to_h_weight3, gpt_neox_layers_3_mlp_dense_4h_to_h_bias3, alloc256) + R.vm.kill_object(alloc255) + 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((2048, 2048), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[44] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc257: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv413, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape143, gpt_neox_layers_3_attention_dense_weight3, gpt_neox_layers_3_attention_dense_bias3, alloc256, alloc250, alloc257) + R.vm.kill_object(alloc250) + R.vm.kill_object(reshape143) + R.vm.kill_object(alloc256) + 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((2048,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[50] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc258: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv414, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc257, gpt_neox_layers_4_input_layernorm_weight3, gpt_neox_layers_4_input_layernorm_bias3, alloc258) + 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((6144, 2048), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[54] + gv415: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc259: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv415, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc258, gpt_neox_layers_4_attention_query_key_value_weight3, gpt_neox_layers_4_attention_query_key_value_bias3, alloc259) + R.vm.kill_object(alloc258) + 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) + gv416: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape144: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc259, gv416, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc259) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape145: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape144, gv417, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape144) + gv418: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc260: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv418, R.dtype("float16")) + _260: 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)), reshape145, alloc260) + R.vm.kill_object(reshape145) + gv419: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape146: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc260, gv419, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc260) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape147: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape146, gv420, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape146) + gpt_neox_layers_4_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[52] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc261: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv421, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc257, gpt_neox_layers_4_post_attention_layernorm_weight3, gpt_neox_layers_4_post_attention_layernorm_bias3, alloc261) + 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((8192, 2048), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[58] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc262: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv422, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc261, gpt_neox_layers_4_mlp_dense_h_to_4h_weight3, gpt_neox_layers_4_mlp_dense_h_to_4h_bias3, alloc262) + R.vm.kill_object(alloc261) + 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((2048, 8192), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[60] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc263: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv423, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc262, gpt_neox_layers_4_mlp_dense_4h_to_h_weight3, gpt_neox_layers_4_mlp_dense_4h_to_h_bias3, alloc263) + R.vm.kill_object(alloc262) + 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((2048, 2048), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[56] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc264: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv424, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape147, gpt_neox_layers_4_attention_dense_weight3, gpt_neox_layers_4_attention_dense_bias3, alloc263, alloc257, alloc264) + R.vm.kill_object(alloc257) + R.vm.kill_object(reshape147) + R.vm.kill_object(alloc263) + 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((2048,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[62] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc265: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv425, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc264, gpt_neox_layers_5_input_layernorm_weight3, gpt_neox_layers_5_input_layernorm_bias3, alloc265) + 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((6144, 2048), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[66] + gv426: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc266: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv426, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc265, gpt_neox_layers_5_attention_query_key_value_weight3, gpt_neox_layers_5_attention_query_key_value_bias3, alloc266) + R.vm.kill_object(alloc265) + 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) + gv427: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape148: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc266, gv427, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc266) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape149: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape148, gv428, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape148) + gv429: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc267: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv429, R.dtype("float16")) + _267: 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)), reshape149, alloc267) + R.vm.kill_object(reshape149) + gv430: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape150: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc267, gv430, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc267) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape151: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape150, gv431, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape150) + gpt_neox_layers_5_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[64] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc268: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv432, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc264, gpt_neox_layers_5_post_attention_layernorm_weight3, gpt_neox_layers_5_post_attention_layernorm_bias3, alloc268) + 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((8192, 2048), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[70] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc269: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv433, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc268, gpt_neox_layers_5_mlp_dense_h_to_4h_weight3, gpt_neox_layers_5_mlp_dense_h_to_4h_bias3, alloc269) + R.vm.kill_object(alloc268) + 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((2048, 8192), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[72] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc270: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv434, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc269, gpt_neox_layers_5_mlp_dense_4h_to_h_weight3, gpt_neox_layers_5_mlp_dense_4h_to_h_bias3, alloc270) + R.vm.kill_object(alloc269) + 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((2048, 2048), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[68] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc271: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv435, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape151, gpt_neox_layers_5_attention_dense_weight3, gpt_neox_layers_5_attention_dense_bias3, alloc270, alloc264, alloc271) + R.vm.kill_object(alloc264) + R.vm.kill_object(reshape151) + R.vm.kill_object(alloc270) + 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((2048,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[74] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc272: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv436, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc271, gpt_neox_layers_6_input_layernorm_weight3, gpt_neox_layers_6_input_layernorm_bias3, alloc272) + 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((6144, 2048), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[78] + gv437: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc273: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv437, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc272, gpt_neox_layers_6_attention_query_key_value_weight3, gpt_neox_layers_6_attention_query_key_value_bias3, alloc273) + R.vm.kill_object(alloc272) + 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) + gv438: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape152: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc273, gv438, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc273) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape153: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape152, gv439, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape152) + gv440: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc274: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv440, R.dtype("float16")) + _274: 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)), reshape153, alloc274) + R.vm.kill_object(reshape153) + gv441: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape154: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc274, gv441, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc274) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape155: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape154, gv442, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape154) + gpt_neox_layers_6_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[76] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc275: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv443, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc271, gpt_neox_layers_6_post_attention_layernorm_weight3, gpt_neox_layers_6_post_attention_layernorm_bias3, alloc275) + 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((8192, 2048), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[82] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc276: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv444, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc275, gpt_neox_layers_6_mlp_dense_h_to_4h_weight3, gpt_neox_layers_6_mlp_dense_h_to_4h_bias3, alloc276) + R.vm.kill_object(alloc275) + 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((2048, 8192), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[84] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc277: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv445, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc276, gpt_neox_layers_6_mlp_dense_4h_to_h_weight3, gpt_neox_layers_6_mlp_dense_4h_to_h_bias3, alloc277) + R.vm.kill_object(alloc276) + 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((2048, 2048), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[80] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc278: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv446, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape155, gpt_neox_layers_6_attention_dense_weight3, gpt_neox_layers_6_attention_dense_bias3, alloc277, alloc271, alloc278) + R.vm.kill_object(alloc271) + R.vm.kill_object(reshape155) + R.vm.kill_object(alloc277) + 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((2048,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[86] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc279: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv447, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc278, gpt_neox_layers_7_input_layernorm_weight3, gpt_neox_layers_7_input_layernorm_bias3, alloc279) + 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((6144, 2048), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[90] + gv448: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc280: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv448, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc279, gpt_neox_layers_7_attention_query_key_value_weight3, gpt_neox_layers_7_attention_query_key_value_bias3, alloc280) + R.vm.kill_object(alloc279) + 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) + gv449: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape156: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc280, gv449, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc280) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape157: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape156, gv450, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape156) + gv451: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc281: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv451, R.dtype("float16")) + _281: 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)), reshape157, alloc281) + R.vm.kill_object(reshape157) + gv452: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape158: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc281, gv452, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc281) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape159: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape158, gv453, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape158) + gpt_neox_layers_7_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[88] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc282: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv454, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc278, gpt_neox_layers_7_post_attention_layernorm_weight3, gpt_neox_layers_7_post_attention_layernorm_bias3, alloc282) + 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((8192, 2048), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[94] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc283: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv455, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc282, gpt_neox_layers_7_mlp_dense_h_to_4h_weight3, gpt_neox_layers_7_mlp_dense_h_to_4h_bias3, alloc283) + R.vm.kill_object(alloc282) + 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((2048, 8192), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[96] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc284: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv456, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc283, gpt_neox_layers_7_mlp_dense_4h_to_h_weight3, gpt_neox_layers_7_mlp_dense_4h_to_h_bias3, alloc284) + R.vm.kill_object(alloc283) + 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((2048, 2048), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[92] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc285: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv457, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape159, gpt_neox_layers_7_attention_dense_weight3, gpt_neox_layers_7_attention_dense_bias3, alloc284, alloc278, alloc285) + R.vm.kill_object(alloc278) + R.vm.kill_object(reshape159) + R.vm.kill_object(alloc284) + 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((2048,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[98] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc286: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv458, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc285, gpt_neox_layers_8_input_layernorm_weight3, gpt_neox_layers_8_input_layernorm_bias3, alloc286) + 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((6144, 2048), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[102] + gv459: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc287: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv459, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc286, gpt_neox_layers_8_attention_query_key_value_weight3, gpt_neox_layers_8_attention_query_key_value_bias3, alloc287) + R.vm.kill_object(alloc286) + 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) + gv460: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape160: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc287, gv460, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc287) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape161: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape160, gv461, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape160) + gv462: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc288: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv462, R.dtype("float16")) + _288: 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)), reshape161, alloc288) + R.vm.kill_object(reshape161) + gv463: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape162: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc288, gv463, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc288) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape163: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape162, gv464, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape162) + gpt_neox_layers_8_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[100] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc289: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv465, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc285, gpt_neox_layers_8_post_attention_layernorm_weight3, gpt_neox_layers_8_post_attention_layernorm_bias3, alloc289) + 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((8192, 2048), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[106] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc290: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv466, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc289, gpt_neox_layers_8_mlp_dense_h_to_4h_weight3, gpt_neox_layers_8_mlp_dense_h_to_4h_bias3, alloc290) + R.vm.kill_object(alloc289) + 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((2048, 8192), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[108] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc291: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv467, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc290, gpt_neox_layers_8_mlp_dense_4h_to_h_weight3, gpt_neox_layers_8_mlp_dense_4h_to_h_bias3, alloc291) + R.vm.kill_object(alloc290) + 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((2048, 2048), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[104] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc292: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv468, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape163, gpt_neox_layers_8_attention_dense_weight3, gpt_neox_layers_8_attention_dense_bias3, alloc291, alloc285, alloc292) + R.vm.kill_object(alloc285) + R.vm.kill_object(reshape163) + R.vm.kill_object(alloc291) + 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((2048,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[110] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc293: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv469, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc292, gpt_neox_layers_9_input_layernorm_weight3, gpt_neox_layers_9_input_layernorm_bias3, alloc293) + 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((6144, 2048), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[114] + gv470: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc294: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv470, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc293, gpt_neox_layers_9_attention_query_key_value_weight3, gpt_neox_layers_9_attention_query_key_value_bias3, alloc294) + R.vm.kill_object(alloc293) + 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) + gv471: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape164: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc294, gv471, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc294) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape165: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape164, gv472, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape164) + gv473: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc295: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv473, R.dtype("float16")) + _295: 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)), reshape165, alloc295) + R.vm.kill_object(reshape165) + gv474: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape166: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc295, gv474, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc295) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape167: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape166, gv475, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape166) + gpt_neox_layers_9_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[112] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc296: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv476, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc292, gpt_neox_layers_9_post_attention_layernorm_weight3, gpt_neox_layers_9_post_attention_layernorm_bias3, alloc296) + 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((8192, 2048), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[118] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc297: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv477, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc296, gpt_neox_layers_9_mlp_dense_h_to_4h_weight3, gpt_neox_layers_9_mlp_dense_h_to_4h_bias3, alloc297) + R.vm.kill_object(alloc296) + 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((2048, 8192), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[120] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc298: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv478, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc297, gpt_neox_layers_9_mlp_dense_4h_to_h_weight3, gpt_neox_layers_9_mlp_dense_4h_to_h_bias3, alloc298) + R.vm.kill_object(alloc297) + 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((2048, 2048), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[116] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc299: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv479, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape167, gpt_neox_layers_9_attention_dense_weight3, gpt_neox_layers_9_attention_dense_bias3, alloc298, alloc292, alloc299) + R.vm.kill_object(alloc292) + R.vm.kill_object(reshape167) + R.vm.kill_object(alloc298) + 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((2048,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[122] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc300: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv480, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc299, gpt_neox_layers_10_input_layernorm_weight3, gpt_neox_layers_10_input_layernorm_bias3, alloc300) + 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((6144, 2048), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[126] + gv481: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc301: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv481, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc300, gpt_neox_layers_10_attention_query_key_value_weight3, gpt_neox_layers_10_attention_query_key_value_bias3, alloc301) + R.vm.kill_object(alloc300) + 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) + gv482: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape168: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc301, gv482, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc301) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape169: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape168, gv483, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape168) + gv484: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc302: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv484, R.dtype("float16")) + _302: 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)), reshape169, alloc302) + R.vm.kill_object(reshape169) + gv485: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape170: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc302, gv485, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc302) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape171: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape170, gv486, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape170) + gpt_neox_layers_10_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[124] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc303: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv487, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc299, gpt_neox_layers_10_post_attention_layernorm_weight3, gpt_neox_layers_10_post_attention_layernorm_bias3, alloc303) + 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((8192, 2048), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[130] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc304: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv488, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc303, gpt_neox_layers_10_mlp_dense_h_to_4h_weight3, gpt_neox_layers_10_mlp_dense_h_to_4h_bias3, alloc304) + R.vm.kill_object(alloc303) + 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((2048, 8192), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[132] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc305: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv489, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc304, gpt_neox_layers_10_mlp_dense_4h_to_h_weight3, gpt_neox_layers_10_mlp_dense_4h_to_h_bias3, alloc305) + R.vm.kill_object(alloc304) + 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((2048, 2048), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[128] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc306: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv490, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape171, gpt_neox_layers_10_attention_dense_weight3, gpt_neox_layers_10_attention_dense_bias3, alloc305, alloc299, alloc306) + R.vm.kill_object(alloc299) + R.vm.kill_object(reshape171) + R.vm.kill_object(alloc305) + 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((2048,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[134] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc307: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv491, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc306, gpt_neox_layers_11_input_layernorm_weight3, gpt_neox_layers_11_input_layernorm_bias3, alloc307) + 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((6144, 2048), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[138] + gv492: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc308: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv492, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc307, gpt_neox_layers_11_attention_query_key_value_weight3, gpt_neox_layers_11_attention_query_key_value_bias3, alloc308) + R.vm.kill_object(alloc307) + 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) + gv493: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape172: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc308, gv493, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc308) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape173: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape172, gv494, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape172) + gv495: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc309: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv495, R.dtype("float16")) + _309: 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)), reshape173, alloc309) + R.vm.kill_object(reshape173) + gv496: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape174: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc309, gv496, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc309) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape175: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape174, gv497, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape174) + gpt_neox_layers_11_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[136] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc310: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv498, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc306, gpt_neox_layers_11_post_attention_layernorm_weight3, gpt_neox_layers_11_post_attention_layernorm_bias3, alloc310) + 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((8192, 2048), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[142] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc311: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv499, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc310, gpt_neox_layers_11_mlp_dense_h_to_4h_weight3, gpt_neox_layers_11_mlp_dense_h_to_4h_bias3, alloc311) + R.vm.kill_object(alloc310) + 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((2048, 8192), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[144] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc312: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv500, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc311, gpt_neox_layers_11_mlp_dense_4h_to_h_weight3, gpt_neox_layers_11_mlp_dense_4h_to_h_bias3, alloc312) + R.vm.kill_object(alloc311) + 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((2048, 2048), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[140] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc313: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv501, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape175, gpt_neox_layers_11_attention_dense_weight3, gpt_neox_layers_11_attention_dense_bias3, alloc312, alloc306, alloc313) + R.vm.kill_object(alloc306) + R.vm.kill_object(reshape175) + R.vm.kill_object(alloc312) + 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((2048,), dtype="float16") = packed_params[145] + gpt_neox_layers_12_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[146] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc314: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv502, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc313, gpt_neox_layers_12_input_layernorm_weight3, gpt_neox_layers_12_input_layernorm_bias3, alloc314) + 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((6144, 2048), dtype="float16") = packed_params[149] + gpt_neox_layers_12_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[150] + gv503: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc315: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv503, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc314, gpt_neox_layers_12_attention_query_key_value_weight3, gpt_neox_layers_12_attention_query_key_value_bias3, alloc315) + R.vm.kill_object(alloc314) + 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) + gv504: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape176: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc315, gv504, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc315) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape177: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape176, gv505, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape176) + gv506: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc316: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv506, R.dtype("float16")) + _316: 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)), reshape177, alloc316) + R.vm.kill_object(reshape177) + gv507: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape178: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc316, gv507, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc316) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape179: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape178, gv508, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape178) + gpt_neox_layers_12_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[147] + gpt_neox_layers_12_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[148] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc317: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv509, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc313, gpt_neox_layers_12_post_attention_layernorm_weight3, gpt_neox_layers_12_post_attention_layernorm_bias3, alloc317) + 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((8192, 2048), dtype="float16") = packed_params[153] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[154] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc318: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv510, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc317, gpt_neox_layers_12_mlp_dense_h_to_4h_weight3, gpt_neox_layers_12_mlp_dense_h_to_4h_bias3, alloc318) + R.vm.kill_object(alloc317) + 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((2048, 8192), dtype="float16") = packed_params[155] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[156] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc319: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv511, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc318, gpt_neox_layers_12_mlp_dense_4h_to_h_weight3, gpt_neox_layers_12_mlp_dense_4h_to_h_bias3, alloc319) + R.vm.kill_object(alloc318) + 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((2048, 2048), dtype="float16") = packed_params[151] + gpt_neox_layers_12_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[152] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc320: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv512, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape179, gpt_neox_layers_12_attention_dense_weight3, gpt_neox_layers_12_attention_dense_bias3, alloc319, alloc313, alloc320) + R.vm.kill_object(alloc313) + R.vm.kill_object(reshape179) + R.vm.kill_object(alloc319) + 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((2048,), dtype="float16") = packed_params[157] + gpt_neox_layers_13_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[158] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc321: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv513, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc320, gpt_neox_layers_13_input_layernorm_weight3, gpt_neox_layers_13_input_layernorm_bias3, alloc321) + 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((6144, 2048), dtype="float16") = packed_params[161] + gpt_neox_layers_13_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[162] + gv514: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc322: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv514, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc321, gpt_neox_layers_13_attention_query_key_value_weight3, gpt_neox_layers_13_attention_query_key_value_bias3, alloc322) + R.vm.kill_object(alloc321) + 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) + gv515: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape180: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc322, gv515, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc322) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape181: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape180, gv516, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape180) + gv517: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc323: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv517, R.dtype("float16")) + _323: 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)), reshape181, alloc323) + R.vm.kill_object(reshape181) + gv518: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape182: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc323, gv518, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc323) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape183: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape182, gv519, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape182) + gpt_neox_layers_13_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[159] + gpt_neox_layers_13_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[160] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc324: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv520, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc320, gpt_neox_layers_13_post_attention_layernorm_weight3, gpt_neox_layers_13_post_attention_layernorm_bias3, alloc324) + 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((8192, 2048), dtype="float16") = packed_params[165] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[166] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc325: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv521, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc324, gpt_neox_layers_13_mlp_dense_h_to_4h_weight3, gpt_neox_layers_13_mlp_dense_h_to_4h_bias3, alloc325) + R.vm.kill_object(alloc324) + 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((2048, 8192), dtype="float16") = packed_params[167] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[168] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc326: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv522, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc325, gpt_neox_layers_13_mlp_dense_4h_to_h_weight3, gpt_neox_layers_13_mlp_dense_4h_to_h_bias3, alloc326) + R.vm.kill_object(alloc325) + 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((2048, 2048), dtype="float16") = packed_params[163] + gpt_neox_layers_13_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[164] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc327: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv523, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape183, gpt_neox_layers_13_attention_dense_weight3, gpt_neox_layers_13_attention_dense_bias3, alloc326, alloc320, alloc327) + R.vm.kill_object(alloc320) + R.vm.kill_object(reshape183) + R.vm.kill_object(alloc326) + 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((2048,), dtype="float16") = packed_params[169] + gpt_neox_layers_14_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[170] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc328: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv524, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc327, gpt_neox_layers_14_input_layernorm_weight3, gpt_neox_layers_14_input_layernorm_bias3, alloc328) + 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((6144, 2048), dtype="float16") = packed_params[173] + gpt_neox_layers_14_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[174] + gv525: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc329: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv525, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc328, gpt_neox_layers_14_attention_query_key_value_weight3, gpt_neox_layers_14_attention_query_key_value_bias3, alloc329) + R.vm.kill_object(alloc328) + 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) + gv526: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape184: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc329, gv526, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc329) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape185: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape184, gv527, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape184) + gv528: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc330: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv528, R.dtype("float16")) + _330: 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)), reshape185, alloc330) + R.vm.kill_object(reshape185) + gv529: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape186: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc330, gv529, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc330) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape187: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape186, gv530, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape186) + gpt_neox_layers_14_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[171] + gpt_neox_layers_14_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[172] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc331: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv531, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc327, gpt_neox_layers_14_post_attention_layernorm_weight3, gpt_neox_layers_14_post_attention_layernorm_bias3, alloc331) + 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((8192, 2048), dtype="float16") = packed_params[177] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[178] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc332: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv532, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc331, gpt_neox_layers_14_mlp_dense_h_to_4h_weight3, gpt_neox_layers_14_mlp_dense_h_to_4h_bias3, alloc332) + R.vm.kill_object(alloc331) + 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((2048, 8192), dtype="float16") = packed_params[179] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[180] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc333: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv533, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc332, gpt_neox_layers_14_mlp_dense_4h_to_h_weight3, gpt_neox_layers_14_mlp_dense_4h_to_h_bias3, alloc333) + R.vm.kill_object(alloc332) + 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((2048, 2048), dtype="float16") = packed_params[175] + gpt_neox_layers_14_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[176] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc334: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv534, R.dtype("float16")) + R.vm.kill_object(storage14) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape187, gpt_neox_layers_14_attention_dense_weight3, gpt_neox_layers_14_attention_dense_bias3, alloc333, alloc327, alloc334) + R.vm.kill_object(alloc327) + R.vm.kill_object(reshape187) + R.vm.kill_object(alloc333) + 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((2048,), dtype="float16") = packed_params[181] + gpt_neox_layers_15_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[182] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc335: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv535, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc334, gpt_neox_layers_15_input_layernorm_weight3, gpt_neox_layers_15_input_layernorm_bias3, alloc335) + 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((6144, 2048), dtype="float16") = packed_params[185] + gpt_neox_layers_15_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[186] + gv536: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc336: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv536, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc335, gpt_neox_layers_15_attention_query_key_value_weight3, gpt_neox_layers_15_attention_query_key_value_bias3, alloc336) + R.vm.kill_object(alloc335) + 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) + gv537: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape188: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc336, gv537, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc336) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape189: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape188, gv538, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape188) + gv539: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc337: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv539, R.dtype("float16")) + _337: 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)), reshape189, alloc337) + R.vm.kill_object(reshape189) + gv540: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape190: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc337, gv540, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc337) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape191: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape190, gv541, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape190) + gpt_neox_layers_15_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[183] + gpt_neox_layers_15_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[184] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc338: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv542, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc334, gpt_neox_layers_15_post_attention_layernorm_weight3, gpt_neox_layers_15_post_attention_layernorm_bias3, alloc338) + 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((8192, 2048), dtype="float16") = packed_params[189] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[190] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc339: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv543, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc338, gpt_neox_layers_15_mlp_dense_h_to_4h_weight3, gpt_neox_layers_15_mlp_dense_h_to_4h_bias3, alloc339) + R.vm.kill_object(alloc338) + 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((2048, 8192), dtype="float16") = packed_params[191] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[192] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc340: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv544, R.dtype("float16")) + R.vm.kill_object(storage13) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc339, gpt_neox_layers_15_mlp_dense_4h_to_h_weight3, gpt_neox_layers_15_mlp_dense_4h_to_h_bias3, alloc340) + R.vm.kill_object(alloc339) + 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((2048, 2048), dtype="float16") = packed_params[187] + gpt_neox_layers_15_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[188] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc341: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv545, R.dtype("float16")) + R.vm.kill_object(storage15) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape191, gpt_neox_layers_15_attention_dense_weight3, gpt_neox_layers_15_attention_dense_bias3, alloc340, alloc334, alloc341) + R.vm.kill_object(alloc334) + R.vm.kill_object(reshape191) + R.vm.kill_object(alloc340) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_bias3) + gpt_neox_final_layer_norm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[193] + gpt_neox_final_layer_norm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[194] + gv546: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc342: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv546, R.dtype("float16")) + R.vm.kill_object(storage12) + cls.gpt_neox_q0f16_layer_norm1(alloc341, gpt_neox_final_layer_norm_weight3, gpt_neox_final_layer_norm_bias3, alloc342) + R.vm.kill_object(alloc341) + R.vm.kill_object(gpt_neox_final_layer_norm_weight3) + R.vm.kill_object(gpt_neox_final_layer_norm_bias3) + storage16: R.Object = R.vm.alloc_storage(R.shape([524288]), 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(1), R.prim_value(1), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc343: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv547, R.dtype("float16")) + R.vm.kill_object(storage16) + cls.gpt_neox_q0f16_take(alloc342, logit_positions, alloc343) + R.vm.kill_object(alloc342) + embed_out_weight3: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[195] + gv548: 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),)) + storage17: R.Object = R.vm.alloc_storage(gv548, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv549: R.Shape(ndim=3) = R.call_packed("vm.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),)) + alloc344: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv549, R.dtype("float32")) + R.vm.kill_object(storage17) + cls.gpt_neox_q0f16_fused_NT_matmul9_cast5(alloc343, embed_out_weight3, alloc344) + R.vm.kill_object(alloc343) + R.vm.kill_object(embed_out_weight3) + gv3: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc344, paged_kv_cache + R.vm.kill_object(alloc344) + gv550: R.Tensor(dtype="float32", ndim=3) = gv3[0] + R.call_packed("vm.builtin.match_shape", gv550, 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", 2048), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 2048), 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, 2048), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(196), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv182: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv182, 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, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv183: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[195] + R.call_packed("vm.builtin.check_tensor_info", gv183, 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, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), 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(2048), R.str("ErrorContext(fn=batch_verify, loc=param[0], param=input_embeds, annotation=R.Tensor((1, seq_len, 2048), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv182, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(2048), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv183, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(2048), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func1(shape_heap) + gpt_neox_layers_0_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[2] + storage6: R.Object = R.vm.alloc_storage(R.shape([8388608]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv184: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc115: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv184, 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, alloc115) + 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((6144, 2048), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[6] + storage7: R.Object = R.vm.alloc_storage(R.shape([25165824]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv185: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc116: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv185, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc115, gpt_neox_layers_0_attention_query_key_value_weight5, gpt_neox_layers_0_attention_query_key_value_bias5, alloc116) + R.vm.kill_object(alloc115) + 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) + gv186: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape256: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc116, gv186, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc116) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape257: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape256, gv187, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape256) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc117: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv188, R.dtype("float16")) + _117: 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)), reshape257, alloc117) + R.vm.kill_object(reshape257) + gv189: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape258: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc117, gv189, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc117) + gv190: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape259: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape258, gv190, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape258) + gpt_neox_layers_0_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[4] + gv191: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc118: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv191, 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, alloc118) + 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((8192, 2048), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[10] + storage8: R.Object = R.vm.alloc_storage(R.shape([33554432]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv192: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc119: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv192, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc118, gpt_neox_layers_0_mlp_dense_h_to_4h_weight5, gpt_neox_layers_0_mlp_dense_h_to_4h_bias5, alloc119) + R.vm.kill_object(alloc118) + 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((2048, 8192), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[12] + gv193: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc120: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv193, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc119, gpt_neox_layers_0_mlp_dense_4h_to_h_weight5, gpt_neox_layers_0_mlp_dense_4h_to_h_bias5, alloc120) + R.vm.kill_object(alloc119) + 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((2048, 2048), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[8] + gv194: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc121: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv194, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape259, gpt_neox_layers_0_attention_dense_weight5, gpt_neox_layers_0_attention_dense_bias5, alloc120, input_embeds, alloc121) + R.vm.kill_object(reshape259) + R.vm.kill_object(alloc120) + 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((2048,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[14] + gv195: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc122: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv195, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc121, gpt_neox_layers_1_input_layernorm_weight5, gpt_neox_layers_1_input_layernorm_bias5, alloc122) + 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((6144, 2048), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[18] + gv196: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc123: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv196, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc122, gpt_neox_layers_1_attention_query_key_value_weight5, gpt_neox_layers_1_attention_query_key_value_bias5, alloc123) + R.vm.kill_object(alloc122) + 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) + gv197: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape260: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc123, gv197, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc123) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape261: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape260, gv198, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape260) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc124: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv199, R.dtype("float16")) + _124: 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)), reshape261, alloc124) + R.vm.kill_object(reshape261) + gv200: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape262: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc124, gv200, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc124) + gv201: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape263: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape262, gv201, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape262) + gpt_neox_layers_1_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[16] + gv202: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc125: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv202, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc121, gpt_neox_layers_1_post_attention_layernorm_weight5, gpt_neox_layers_1_post_attention_layernorm_bias5, alloc125) + 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((8192, 2048), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[22] + storage9: R.Object = R.vm.alloc_storage(R.shape([33554432]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv203: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc126: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv203, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc125, gpt_neox_layers_1_mlp_dense_h_to_4h_weight5, gpt_neox_layers_1_mlp_dense_h_to_4h_bias5, alloc126) + R.vm.kill_object(alloc125) + 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((2048, 8192), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[24] + gv204: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc127: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv204, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc126, gpt_neox_layers_1_mlp_dense_4h_to_h_weight5, gpt_neox_layers_1_mlp_dense_4h_to_h_bias5, alloc127) + R.vm.kill_object(alloc126) + 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((2048, 2048), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[20] + gv205: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc128: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv205, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape263, gpt_neox_layers_1_attention_dense_weight5, gpt_neox_layers_1_attention_dense_bias5, alloc127, alloc121, alloc128) + R.vm.kill_object(alloc121) + R.vm.kill_object(reshape263) + R.vm.kill_object(alloc127) + 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((2048,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[26] + gv206: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc129: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv206, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc128, gpt_neox_layers_2_input_layernorm_weight5, gpt_neox_layers_2_input_layernorm_bias5, alloc129) + 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((6144, 2048), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[30] + gv207: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc130: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv207, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc129, gpt_neox_layers_2_attention_query_key_value_weight5, gpt_neox_layers_2_attention_query_key_value_bias5, alloc130) + R.vm.kill_object(alloc129) + 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) + gv208: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape264: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc130, gv208, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc130) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape265: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape264, gv209, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape264) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc131: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv210, R.dtype("float16")) + _131: 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)), reshape265, alloc131) + R.vm.kill_object(reshape265) + gv211: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape266: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc131, gv211, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc131) + gv212: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape267: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape266, gv212, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape266) + gpt_neox_layers_2_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[28] + gv213: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc132: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv213, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc128, gpt_neox_layers_2_post_attention_layernorm_weight5, gpt_neox_layers_2_post_attention_layernorm_bias5, alloc132) + 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((8192, 2048), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[34] + gv214: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc133: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv214, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc132, gpt_neox_layers_2_mlp_dense_h_to_4h_weight5, gpt_neox_layers_2_mlp_dense_h_to_4h_bias5, alloc133) + R.vm.kill_object(alloc132) + 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((2048, 8192), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[36] + gv215: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc134: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv215, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc133, gpt_neox_layers_2_mlp_dense_4h_to_h_weight5, gpt_neox_layers_2_mlp_dense_4h_to_h_bias5, alloc134) + R.vm.kill_object(alloc133) + 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((2048, 2048), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[32] + gv216: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc135: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv216, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape267, gpt_neox_layers_2_attention_dense_weight5, gpt_neox_layers_2_attention_dense_bias5, alloc134, alloc128, alloc135) + R.vm.kill_object(alloc128) + R.vm.kill_object(reshape267) + R.vm.kill_object(alloc134) + 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((2048,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[38] + gv217: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc136: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv217, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc135, gpt_neox_layers_3_input_layernorm_weight5, gpt_neox_layers_3_input_layernorm_bias5, alloc136) + 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((6144, 2048), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[42] + gv218: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc137: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv218, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc136, gpt_neox_layers_3_attention_query_key_value_weight5, gpt_neox_layers_3_attention_query_key_value_bias5, alloc137) + R.vm.kill_object(alloc136) + 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) + gv219: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape268: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc137, gv219, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc137) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape269: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape268, gv220, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape268) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc138: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv221, R.dtype("float16")) + _138: 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)), reshape269, alloc138) + R.vm.kill_object(reshape269) + gv222: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape270: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc138, gv222, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc138) + gv223: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape271: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape270, gv223, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape270) + gpt_neox_layers_3_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[40] + gv224: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc139: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv224, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc135, gpt_neox_layers_3_post_attention_layernorm_weight5, gpt_neox_layers_3_post_attention_layernorm_bias5, alloc139) + 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((8192, 2048), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[46] + gv225: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc140: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv225, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc139, gpt_neox_layers_3_mlp_dense_h_to_4h_weight5, gpt_neox_layers_3_mlp_dense_h_to_4h_bias5, alloc140) + R.vm.kill_object(alloc139) + 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((2048, 8192), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[48] + gv226: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc141: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv226, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc140, gpt_neox_layers_3_mlp_dense_4h_to_h_weight5, gpt_neox_layers_3_mlp_dense_4h_to_h_bias5, alloc141) + R.vm.kill_object(alloc140) + 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((2048, 2048), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[44] + gv227: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc142: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv227, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape271, gpt_neox_layers_3_attention_dense_weight5, gpt_neox_layers_3_attention_dense_bias5, alloc141, alloc135, alloc142) + R.vm.kill_object(alloc135) + R.vm.kill_object(reshape271) + R.vm.kill_object(alloc141) + 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((2048,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[50] + gv228: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc143: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv228, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc142, gpt_neox_layers_4_input_layernorm_weight5, gpt_neox_layers_4_input_layernorm_bias5, alloc143) + 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((6144, 2048), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[54] + gv229: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc144: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv229, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc143, gpt_neox_layers_4_attention_query_key_value_weight5, gpt_neox_layers_4_attention_query_key_value_bias5, alloc144) + R.vm.kill_object(alloc143) + 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) + gv230: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape272: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc144, gv230, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc144) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape273: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape272, gv231, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape272) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc145: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv232, R.dtype("float16")) + _145: 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)), reshape273, alloc145) + R.vm.kill_object(reshape273) + gv233: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape274: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc145, gv233, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc145) + gv234: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape275: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape274, gv234, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape274) + gpt_neox_layers_4_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[52] + gv235: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc146: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv235, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc142, gpt_neox_layers_4_post_attention_layernorm_weight5, gpt_neox_layers_4_post_attention_layernorm_bias5, alloc146) + 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((8192, 2048), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[58] + gv236: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc147: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv236, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc146, gpt_neox_layers_4_mlp_dense_h_to_4h_weight5, gpt_neox_layers_4_mlp_dense_h_to_4h_bias5, alloc147) + R.vm.kill_object(alloc146) + 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((2048, 8192), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[60] + gv237: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc148: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv237, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc147, gpt_neox_layers_4_mlp_dense_4h_to_h_weight5, gpt_neox_layers_4_mlp_dense_4h_to_h_bias5, alloc148) + R.vm.kill_object(alloc147) + 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((2048, 2048), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[56] + gv238: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc149: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv238, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape275, gpt_neox_layers_4_attention_dense_weight5, gpt_neox_layers_4_attention_dense_bias5, alloc148, alloc142, alloc149) + R.vm.kill_object(alloc142) + R.vm.kill_object(reshape275) + R.vm.kill_object(alloc148) + 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((2048,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[62] + gv239: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc150: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv239, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc149, gpt_neox_layers_5_input_layernorm_weight5, gpt_neox_layers_5_input_layernorm_bias5, alloc150) + 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((6144, 2048), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[66] + gv240: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc151: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv240, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc150, gpt_neox_layers_5_attention_query_key_value_weight5, gpt_neox_layers_5_attention_query_key_value_bias5, alloc151) + R.vm.kill_object(alloc150) + 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) + gv241: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape276: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc151, gv241, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc151) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape277: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape276, gv242, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape276) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc152: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv243, R.dtype("float16")) + _152: 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)), reshape277, alloc152) + R.vm.kill_object(reshape277) + gv244: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape278: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc152, gv244, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc152) + gv245: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape279: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape278, gv245, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape278) + gpt_neox_layers_5_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[64] + gv246: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc153: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv246, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc149, gpt_neox_layers_5_post_attention_layernorm_weight5, gpt_neox_layers_5_post_attention_layernorm_bias5, alloc153) + 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((8192, 2048), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[70] + gv247: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc154: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv247, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc153, gpt_neox_layers_5_mlp_dense_h_to_4h_weight5, gpt_neox_layers_5_mlp_dense_h_to_4h_bias5, alloc154) + R.vm.kill_object(alloc153) + 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((2048, 8192), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[72] + gv248: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc155: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv248, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc154, gpt_neox_layers_5_mlp_dense_4h_to_h_weight5, gpt_neox_layers_5_mlp_dense_4h_to_h_bias5, alloc155) + R.vm.kill_object(alloc154) + 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((2048, 2048), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[68] + gv249: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc156: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv249, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape279, gpt_neox_layers_5_attention_dense_weight5, gpt_neox_layers_5_attention_dense_bias5, alloc155, alloc149, alloc156) + R.vm.kill_object(alloc149) + R.vm.kill_object(reshape279) + R.vm.kill_object(alloc155) + 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((2048,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[74] + gv250: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc157: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv250, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc156, gpt_neox_layers_6_input_layernorm_weight5, gpt_neox_layers_6_input_layernorm_bias5, alloc157) + 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((6144, 2048), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[78] + gv251: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc158: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv251, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc157, gpt_neox_layers_6_attention_query_key_value_weight5, gpt_neox_layers_6_attention_query_key_value_bias5, alloc158) + R.vm.kill_object(alloc157) + 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) + gv252: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape280: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc158, gv252, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc158) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape281: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape280, gv253, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape280) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc159: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv254, R.dtype("float16")) + _159: 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)), reshape281, alloc159) + R.vm.kill_object(reshape281) + gv255: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape282: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc159, gv255, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc159) + gv256: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape283: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape282, gv256, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape282) + gpt_neox_layers_6_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[76] + gv257: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc160: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv257, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc156, gpt_neox_layers_6_post_attention_layernorm_weight5, gpt_neox_layers_6_post_attention_layernorm_bias5, alloc160) + 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((8192, 2048), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[82] + gv258: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc161: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv258, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc160, gpt_neox_layers_6_mlp_dense_h_to_4h_weight5, gpt_neox_layers_6_mlp_dense_h_to_4h_bias5, alloc161) + R.vm.kill_object(alloc160) + 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((2048, 8192), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[84] + gv259: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc162: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv259, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc161, gpt_neox_layers_6_mlp_dense_4h_to_h_weight5, gpt_neox_layers_6_mlp_dense_4h_to_h_bias5, alloc162) + R.vm.kill_object(alloc161) + 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((2048, 2048), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[80] + gv260: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc163: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv260, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape283, gpt_neox_layers_6_attention_dense_weight5, gpt_neox_layers_6_attention_dense_bias5, alloc162, alloc156, alloc163) + R.vm.kill_object(alloc156) + R.vm.kill_object(reshape283) + R.vm.kill_object(alloc162) + 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((2048,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[86] + gv261: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc164: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv261, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc163, gpt_neox_layers_7_input_layernorm_weight5, gpt_neox_layers_7_input_layernorm_bias5, alloc164) + 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((6144, 2048), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[90] + gv262: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc165: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv262, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc164, gpt_neox_layers_7_attention_query_key_value_weight5, gpt_neox_layers_7_attention_query_key_value_bias5, alloc165) + R.vm.kill_object(alloc164) + 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) + gv263: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape284: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc165, gv263, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc165) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape285: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape284, gv264, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape284) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc166: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv265, R.dtype("float16")) + _166: 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)), reshape285, alloc166) + R.vm.kill_object(reshape285) + gv266: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape286: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc166, gv266, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc166) + gv267: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape287: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape286, gv267, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape286) + gpt_neox_layers_7_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[88] + gv268: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc167: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv268, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc163, gpt_neox_layers_7_post_attention_layernorm_weight5, gpt_neox_layers_7_post_attention_layernorm_bias5, alloc167) + 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((8192, 2048), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[94] + gv269: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc168: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv269, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc167, gpt_neox_layers_7_mlp_dense_h_to_4h_weight5, gpt_neox_layers_7_mlp_dense_h_to_4h_bias5, alloc168) + R.vm.kill_object(alloc167) + 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((2048, 8192), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[96] + gv270: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc169: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv270, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc168, gpt_neox_layers_7_mlp_dense_4h_to_h_weight5, gpt_neox_layers_7_mlp_dense_4h_to_h_bias5, alloc169) + R.vm.kill_object(alloc168) + 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((2048, 2048), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[92] + gv271: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc170: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv271, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape287, gpt_neox_layers_7_attention_dense_weight5, gpt_neox_layers_7_attention_dense_bias5, alloc169, alloc163, alloc170) + R.vm.kill_object(alloc163) + R.vm.kill_object(reshape287) + R.vm.kill_object(alloc169) + 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((2048,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[98] + gv272: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc171: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv272, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc170, gpt_neox_layers_8_input_layernorm_weight5, gpt_neox_layers_8_input_layernorm_bias5, alloc171) + 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((6144, 2048), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[102] + gv273: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc172: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv273, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc171, gpt_neox_layers_8_attention_query_key_value_weight5, gpt_neox_layers_8_attention_query_key_value_bias5, alloc172) + R.vm.kill_object(alloc171) + 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) + gv274: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape288: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc172, gv274, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc172) + 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape289: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape288, gv275, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape288) + gv276: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc173: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv276, R.dtype("float16")) + _173: 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)), reshape289, alloc173) + R.vm.kill_object(reshape289) + gv277: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape290: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc173, gv277, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc173) + gv278: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape291: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape290, gv278, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape290) + gpt_neox_layers_8_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[100] + gv279: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc174: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv279, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc170, gpt_neox_layers_8_post_attention_layernorm_weight5, gpt_neox_layers_8_post_attention_layernorm_bias5, alloc174) + 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((8192, 2048), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[106] + gv280: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc175: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv280, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc174, gpt_neox_layers_8_mlp_dense_h_to_4h_weight5, gpt_neox_layers_8_mlp_dense_h_to_4h_bias5, alloc175) + R.vm.kill_object(alloc174) + 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((2048, 8192), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[108] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc176: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv281, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc175, gpt_neox_layers_8_mlp_dense_4h_to_h_weight5, gpt_neox_layers_8_mlp_dense_4h_to_h_bias5, alloc176) + R.vm.kill_object(alloc175) + 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((2048, 2048), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[104] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc177: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv282, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape291, gpt_neox_layers_8_attention_dense_weight5, gpt_neox_layers_8_attention_dense_bias5, alloc176, alloc170, alloc177) + R.vm.kill_object(alloc170) + R.vm.kill_object(reshape291) + R.vm.kill_object(alloc176) + 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((2048,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[110] + gv283: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc178: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv283, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc177, gpt_neox_layers_9_input_layernorm_weight5, gpt_neox_layers_9_input_layernorm_bias5, alloc178) + 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((6144, 2048), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[114] + gv284: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc179: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv284, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc178, gpt_neox_layers_9_attention_query_key_value_weight5, gpt_neox_layers_9_attention_query_key_value_bias5, alloc179) + R.vm.kill_object(alloc178) + 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) + gv285: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape292: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc179, gv285, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc179) + gv286: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape293: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape292, gv286, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape292) + gv287: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc180: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv287, R.dtype("float16")) + _180: 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)), reshape293, alloc180) + R.vm.kill_object(reshape293) + gv288: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape294: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc180, gv288, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc180) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape295: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape294, gv289, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape294) + gpt_neox_layers_9_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[112] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc181: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv290, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc177, gpt_neox_layers_9_post_attention_layernorm_weight5, gpt_neox_layers_9_post_attention_layernorm_bias5, alloc181) + 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((8192, 2048), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[118] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc182: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv291, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc181, gpt_neox_layers_9_mlp_dense_h_to_4h_weight5, gpt_neox_layers_9_mlp_dense_h_to_4h_bias5, alloc182) + R.vm.kill_object(alloc181) + 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((2048, 8192), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[120] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc183: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv292, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc182, gpt_neox_layers_9_mlp_dense_4h_to_h_weight5, gpt_neox_layers_9_mlp_dense_4h_to_h_bias5, alloc183) + R.vm.kill_object(alloc182) + 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((2048, 2048), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[116] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc184: 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_matmul6_add6_add9_add9(reshape295, gpt_neox_layers_9_attention_dense_weight5, gpt_neox_layers_9_attention_dense_bias5, alloc183, alloc177, alloc184) + R.vm.kill_object(alloc177) + R.vm.kill_object(reshape295) + R.vm.kill_object(alloc183) + 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((2048,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[122] + gv294: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc185: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv294, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc184, gpt_neox_layers_10_input_layernorm_weight5, gpt_neox_layers_10_input_layernorm_bias5, alloc185) + 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((6144, 2048), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[126] + gv295: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc186: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv295, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc185, gpt_neox_layers_10_attention_query_key_value_weight5, gpt_neox_layers_10_attention_query_key_value_bias5, alloc186) + R.vm.kill_object(alloc185) + 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) + gv296: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape296: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc186, gv296, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc186) + gv297: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape297: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape296, gv297, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape296) + gv298: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc187: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv298, R.dtype("float16")) + _187: 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)), reshape297, alloc187) + R.vm.kill_object(reshape297) + gv299: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape298: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc187, gv299, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc187) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape299: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape298, gv300, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape298) + gpt_neox_layers_10_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[124] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc188: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv301, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc184, gpt_neox_layers_10_post_attention_layernorm_weight5, gpt_neox_layers_10_post_attention_layernorm_bias5, alloc188) + 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((8192, 2048), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[130] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc189: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv302, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc188, gpt_neox_layers_10_mlp_dense_h_to_4h_weight5, gpt_neox_layers_10_mlp_dense_h_to_4h_bias5, alloc189) + R.vm.kill_object(alloc188) + 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((2048, 8192), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[132] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc190: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv303, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc189, gpt_neox_layers_10_mlp_dense_4h_to_h_weight5, gpt_neox_layers_10_mlp_dense_4h_to_h_bias5, alloc190) + R.vm.kill_object(alloc189) + 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((2048, 2048), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[128] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc191: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv304, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape299, gpt_neox_layers_10_attention_dense_weight5, gpt_neox_layers_10_attention_dense_bias5, alloc190, alloc184, alloc191) + R.vm.kill_object(alloc184) + R.vm.kill_object(reshape299) + R.vm.kill_object(alloc190) + 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((2048,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[134] + gv305: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc192: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv305, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc191, gpt_neox_layers_11_input_layernorm_weight5, gpt_neox_layers_11_input_layernorm_bias5, alloc192) + 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((6144, 2048), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[138] + gv306: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc193: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv306, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc192, gpt_neox_layers_11_attention_query_key_value_weight5, gpt_neox_layers_11_attention_query_key_value_bias5, alloc193) + R.vm.kill_object(alloc192) + 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) + gv307: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape300: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc193, gv307, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc193) + gv308: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape301: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape300, gv308, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape300) + gv309: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc194: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv309, R.dtype("float16")) + _194: 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)), reshape301, alloc194) + R.vm.kill_object(reshape301) + gv310: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape302: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc194, gv310, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc194) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape303: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape302, gv311, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape302) + gpt_neox_layers_11_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[136] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc195: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv312, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc191, gpt_neox_layers_11_post_attention_layernorm_weight5, gpt_neox_layers_11_post_attention_layernorm_bias5, alloc195) + 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((8192, 2048), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[142] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc196: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv313, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc195, gpt_neox_layers_11_mlp_dense_h_to_4h_weight5, gpt_neox_layers_11_mlp_dense_h_to_4h_bias5, alloc196) + R.vm.kill_object(alloc195) + 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((2048, 8192), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[144] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc197: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv314, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc196, gpt_neox_layers_11_mlp_dense_4h_to_h_weight5, gpt_neox_layers_11_mlp_dense_4h_to_h_bias5, alloc197) + R.vm.kill_object(alloc196) + 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((2048, 2048), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[140] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc198: 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_matmul6_add6_add9_add9(reshape303, gpt_neox_layers_11_attention_dense_weight5, gpt_neox_layers_11_attention_dense_bias5, alloc197, alloc191, alloc198) + R.vm.kill_object(alloc191) + R.vm.kill_object(reshape303) + R.vm.kill_object(alloc197) + 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((2048,), dtype="float16") = packed_params[145] + gpt_neox_layers_12_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[146] + gv316: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc199: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv316, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc198, gpt_neox_layers_12_input_layernorm_weight5, gpt_neox_layers_12_input_layernorm_bias5, alloc199) + 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((6144, 2048), dtype="float16") = packed_params[149] + gpt_neox_layers_12_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[150] + gv317: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc200: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv317, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc199, gpt_neox_layers_12_attention_query_key_value_weight5, gpt_neox_layers_12_attention_query_key_value_bias5, alloc200) + R.vm.kill_object(alloc199) + 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) + gv318: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape304: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc200, gv318, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc200) + gv319: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape305: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape304, gv319, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape304) + gv320: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc201: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv320, R.dtype("float16")) + _201: 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)), reshape305, alloc201) + R.vm.kill_object(reshape305) + gv321: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape306: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc201, gv321, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc201) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape307: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape306, gv322, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape306) + gpt_neox_layers_12_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[147] + gpt_neox_layers_12_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[148] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc202: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv323, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc198, gpt_neox_layers_12_post_attention_layernorm_weight5, gpt_neox_layers_12_post_attention_layernorm_bias5, alloc202) + 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((8192, 2048), dtype="float16") = packed_params[153] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[154] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc203: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv324, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc202, gpt_neox_layers_12_mlp_dense_h_to_4h_weight5, gpt_neox_layers_12_mlp_dense_h_to_4h_bias5, alloc203) + R.vm.kill_object(alloc202) + 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((2048, 8192), dtype="float16") = packed_params[155] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[156] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc204: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv325, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc203, gpt_neox_layers_12_mlp_dense_4h_to_h_weight5, gpt_neox_layers_12_mlp_dense_4h_to_h_bias5, alloc204) + R.vm.kill_object(alloc203) + 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((2048, 2048), dtype="float16") = packed_params[151] + gpt_neox_layers_12_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[152] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc205: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv326, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape307, gpt_neox_layers_12_attention_dense_weight5, gpt_neox_layers_12_attention_dense_bias5, alloc204, alloc198, alloc205) + R.vm.kill_object(alloc198) + R.vm.kill_object(reshape307) + R.vm.kill_object(alloc204) + 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((2048,), dtype="float16") = packed_params[157] + gpt_neox_layers_13_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[158] + gv327: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc206: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv327, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc205, gpt_neox_layers_13_input_layernorm_weight5, gpt_neox_layers_13_input_layernorm_bias5, alloc206) + 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((6144, 2048), dtype="float16") = packed_params[161] + gpt_neox_layers_13_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[162] + gv328: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc207: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv328, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc206, gpt_neox_layers_13_attention_query_key_value_weight5, gpt_neox_layers_13_attention_query_key_value_bias5, alloc207) + R.vm.kill_object(alloc206) + 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) + gv329: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape308: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc207, gv329, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc207) + gv330: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape309: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape308, gv330, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape308) + gv331: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc208: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv331, R.dtype("float16")) + _208: 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)), reshape309, alloc208) + R.vm.kill_object(reshape309) + gv332: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape310: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc208, gv332, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc208) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape311: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape310, gv333, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape310) + gpt_neox_layers_13_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[159] + gpt_neox_layers_13_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[160] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc209: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv334, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc205, gpt_neox_layers_13_post_attention_layernorm_weight5, gpt_neox_layers_13_post_attention_layernorm_bias5, alloc209) + 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((8192, 2048), dtype="float16") = packed_params[165] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[166] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc210: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv335, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc209, gpt_neox_layers_13_mlp_dense_h_to_4h_weight5, gpt_neox_layers_13_mlp_dense_h_to_4h_bias5, alloc210) + R.vm.kill_object(alloc209) + 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((2048, 8192), dtype="float16") = packed_params[167] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[168] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc211: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv336, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc210, gpt_neox_layers_13_mlp_dense_4h_to_h_weight5, gpt_neox_layers_13_mlp_dense_4h_to_h_bias5, alloc211) + R.vm.kill_object(alloc210) + 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((2048, 2048), dtype="float16") = packed_params[163] + gpt_neox_layers_13_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[164] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc212: 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_matmul6_add6_add9_add9(reshape311, gpt_neox_layers_13_attention_dense_weight5, gpt_neox_layers_13_attention_dense_bias5, alloc211, alloc205, alloc212) + R.vm.kill_object(alloc205) + R.vm.kill_object(reshape311) + R.vm.kill_object(alloc211) + 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((2048,), dtype="float16") = packed_params[169] + gpt_neox_layers_14_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[170] + gv338: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc213: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv338, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc212, gpt_neox_layers_14_input_layernorm_weight5, gpt_neox_layers_14_input_layernorm_bias5, alloc213) + 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((6144, 2048), dtype="float16") = packed_params[173] + gpt_neox_layers_14_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[174] + gv339: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc214: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv339, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc213, gpt_neox_layers_14_attention_query_key_value_weight5, gpt_neox_layers_14_attention_query_key_value_bias5, alloc214) + R.vm.kill_object(alloc213) + 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) + gv340: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape312: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc214, gv340, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc214) + gv341: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape313: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape312, gv341, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape312) + gv342: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc215: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv342, R.dtype("float16")) + _215: 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)), reshape313, alloc215) + R.vm.kill_object(reshape313) + gv343: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape314: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc215, gv343, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc215) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape315: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape314, gv344, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape314) + gpt_neox_layers_14_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[171] + gpt_neox_layers_14_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[172] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc216: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv345, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc212, gpt_neox_layers_14_post_attention_layernorm_weight5, gpt_neox_layers_14_post_attention_layernorm_bias5, alloc216) + 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((8192, 2048), dtype="float16") = packed_params[177] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[178] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc217: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv346, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc216, gpt_neox_layers_14_mlp_dense_h_to_4h_weight5, gpt_neox_layers_14_mlp_dense_h_to_4h_bias5, alloc217) + R.vm.kill_object(alloc216) + 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((2048, 8192), dtype="float16") = packed_params[179] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[180] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc218: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv347, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc217, gpt_neox_layers_14_mlp_dense_4h_to_h_weight5, gpt_neox_layers_14_mlp_dense_4h_to_h_bias5, alloc218) + R.vm.kill_object(alloc217) + 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((2048, 2048), dtype="float16") = packed_params[175] + gpt_neox_layers_14_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[176] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc219: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv348, R.dtype("float16")) + R.vm.kill_object(storage8) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape315, gpt_neox_layers_14_attention_dense_weight5, gpt_neox_layers_14_attention_dense_bias5, alloc218, alloc212, alloc219) + R.vm.kill_object(alloc212) + R.vm.kill_object(reshape315) + R.vm.kill_object(alloc218) + 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((2048,), dtype="float16") = packed_params[181] + gpt_neox_layers_15_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[182] + gv349: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc220: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv349, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc219, gpt_neox_layers_15_input_layernorm_weight5, gpt_neox_layers_15_input_layernorm_bias5, alloc220) + 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((6144, 2048), dtype="float16") = packed_params[185] + gpt_neox_layers_15_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[186] + gv350: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc221: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv350, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc220, gpt_neox_layers_15_attention_query_key_value_weight5, gpt_neox_layers_15_attention_query_key_value_bias5, alloc221) + R.vm.kill_object(alloc220) + 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) + gv351: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape316: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc221, gv351, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc221) + gv352: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape317: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape316, gv352, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape316) + gv353: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc222: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv353, R.dtype("float16")) + _222: 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)), reshape317, alloc222) + R.vm.kill_object(reshape317) + gv354: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape318: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc222, gv354, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc222) + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape319: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape318, gv355, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape318) + gpt_neox_layers_15_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[183] + gpt_neox_layers_15_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[184] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc223: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv356, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc219, gpt_neox_layers_15_post_attention_layernorm_weight5, gpt_neox_layers_15_post_attention_layernorm_bias5, alloc223) + 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((8192, 2048), dtype="float16") = packed_params[189] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[190] + 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(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc224: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv357, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc223, gpt_neox_layers_15_mlp_dense_h_to_4h_weight5, gpt_neox_layers_15_mlp_dense_h_to_4h_bias5, alloc224) + R.vm.kill_object(alloc223) + 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((2048, 8192), dtype="float16") = packed_params[191] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[192] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc225: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv358, R.dtype("float16")) + R.vm.kill_object(storage7) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc224, gpt_neox_layers_15_mlp_dense_4h_to_h_weight5, gpt_neox_layers_15_mlp_dense_4h_to_h_bias5, alloc225) + R.vm.kill_object(alloc224) + 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((2048, 2048), dtype="float16") = packed_params[187] + gpt_neox_layers_15_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[188] + 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(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc226: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv359, R.dtype("float16")) + R.vm.kill_object(storage9) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape319, gpt_neox_layers_15_attention_dense_weight5, gpt_neox_layers_15_attention_dense_bias5, alloc225, alloc219, alloc226) + R.vm.kill_object(alloc219) + R.vm.kill_object(reshape319) + R.vm.kill_object(alloc225) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_bias5) + gpt_neox_final_layer_norm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[193] + gpt_neox_final_layer_norm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[194] + gv360: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc227: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv360, R.dtype("float16")) + R.vm.kill_object(storage6) + cls.gpt_neox_q0f16_layer_norm1(alloc226, gpt_neox_final_layer_norm_weight5, gpt_neox_final_layer_norm_bias5, alloc227) + R.vm.kill_object(alloc226) + 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, 2048), dtype="float16") = packed_params[195] + gv361: 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),)) + storage10: R.Object = R.vm.alloc_storage(gv361, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv362: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_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),)) + alloc228: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv362, R.dtype("float32")) + R.vm.kill_object(storage10) + cls.gpt_neox_q0f16_fused_NT_matmul9_cast5(alloc227, embed_out_weight5, alloc228) + R.vm.kill_object(alloc227) + R.vm.kill_object(embed_out_weight5) + gv5: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc228, paged_kv_cache + R.vm.kill_object(alloc228) + gv363: R.Tensor(dtype="float32", ndim=3) = gv5[0] + R.call_packed("vm.builtin.match_shape", gv363, 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,)) + gv364: 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", gv364, R.shape([0, 16]), R.prim_value(8), R.prim_value(8), R.prim_value(256), 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, 2048), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 2048), 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, 2048), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(196), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv551: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv551, 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, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv552: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[195] + R.call_packed("vm.builtin.check_tensor_info", gv552, 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, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), 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(2048), R.str("ErrorContext(fn=decode, loc=param[0], param=input_embed, annotation=R.Tensor((1, 1, 2048), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv551, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv552, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(0), R.prim_value(0), R.prim_value(2048), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func3(shape_heap) + gpt_neox_layers_0_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[2] + storage18: R.Object = R.vm.alloc_storage(R.shape([4096]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc345: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), 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, alloc345) + 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((6144, 2048), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[6] + storage19: R.Object = R.vm.alloc_storage(R.shape([12288]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc346: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc345, gpt_neox_layers_0_attention_query_key_value_weight2, gpt_neox_layers_0_attention_query_key_value_bias2, alloc346) + R.vm.kill_object(alloc345) + 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) + lv196: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc346, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc346) + alloc347: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _347: 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)), lv196, alloc347) + R.vm.kill_object(lv196) + lv197: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc347, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc347) + gpt_neox_layers_0_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[4] + alloc348: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), 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, alloc348) + 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((8192, 2048), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[10] + storage20: R.Object = R.vm.alloc_storage(R.shape([16384]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc349: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc348, gpt_neox_layers_0_mlp_dense_h_to_4h_weight2, gpt_neox_layers_0_mlp_dense_h_to_4h_bias2, alloc349) + R.vm.kill_object(alloc348) + 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((2048, 8192), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[12] + alloc350: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc349, gpt_neox_layers_0_mlp_dense_4h_to_h_weight2, gpt_neox_layers_0_mlp_dense_4h_to_h_bias2, alloc350) + R.vm.kill_object(alloc349) + 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((2048, 2048), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[8] + alloc351: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv197, gpt_neox_layers_0_attention_dense_weight2, gpt_neox_layers_0_attention_dense_bias2, alloc350, input_embed, alloc351) + R.vm.kill_object(lv197) + R.vm.kill_object(alloc350) + 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((2048,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[14] + alloc352: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc351, gpt_neox_layers_1_input_layernorm_weight2, gpt_neox_layers_1_input_layernorm_bias2, alloc352) + 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((6144, 2048), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[18] + alloc353: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc352, gpt_neox_layers_1_attention_query_key_value_weight2, gpt_neox_layers_1_attention_query_key_value_bias2, alloc353) + R.vm.kill_object(alloc352) + 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) + lv202: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc353, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc353) + alloc354: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _354: 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)), lv202, alloc354) + R.vm.kill_object(lv202) + lv203: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc354, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc354) + gpt_neox_layers_1_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[16] + alloc355: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc351, gpt_neox_layers_1_post_attention_layernorm_weight2, gpt_neox_layers_1_post_attention_layernorm_bias2, alloc355) + 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((8192, 2048), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[22] + storage21: R.Object = R.vm.alloc_storage(R.shape([16384]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc356: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc355, gpt_neox_layers_1_mlp_dense_h_to_4h_weight2, gpt_neox_layers_1_mlp_dense_h_to_4h_bias2, alloc356) + R.vm.kill_object(alloc355) + 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((2048, 8192), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[24] + alloc357: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc356, gpt_neox_layers_1_mlp_dense_4h_to_h_weight2, gpt_neox_layers_1_mlp_dense_4h_to_h_bias2, alloc357) + R.vm.kill_object(alloc356) + 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((2048, 2048), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[20] + alloc358: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv203, gpt_neox_layers_1_attention_dense_weight2, gpt_neox_layers_1_attention_dense_bias2, alloc357, alloc351, alloc358) + R.vm.kill_object(alloc351) + R.vm.kill_object(lv203) + R.vm.kill_object(alloc357) + 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((2048,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[26] + alloc359: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc358, gpt_neox_layers_2_input_layernorm_weight2, gpt_neox_layers_2_input_layernorm_bias2, alloc359) + 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((6144, 2048), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[30] + alloc360: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc359, gpt_neox_layers_2_attention_query_key_value_weight2, gpt_neox_layers_2_attention_query_key_value_bias2, alloc360) + R.vm.kill_object(alloc359) + 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) + lv208: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc360, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc360) + alloc361: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _361: 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)), lv208, alloc361) + R.vm.kill_object(lv208) + lv209: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc361, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc361) + gpt_neox_layers_2_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[28] + alloc362: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc358, gpt_neox_layers_2_post_attention_layernorm_weight2, gpt_neox_layers_2_post_attention_layernorm_bias2, alloc362) + 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((8192, 2048), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[34] + alloc363: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc362, gpt_neox_layers_2_mlp_dense_h_to_4h_weight2, gpt_neox_layers_2_mlp_dense_h_to_4h_bias2, alloc363) + R.vm.kill_object(alloc362) + 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((2048, 8192), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[36] + alloc364: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc363, gpt_neox_layers_2_mlp_dense_4h_to_h_weight2, gpt_neox_layers_2_mlp_dense_4h_to_h_bias2, alloc364) + R.vm.kill_object(alloc363) + 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((2048, 2048), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[32] + alloc365: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv209, gpt_neox_layers_2_attention_dense_weight2, gpt_neox_layers_2_attention_dense_bias2, alloc364, alloc358, alloc365) + R.vm.kill_object(alloc358) + R.vm.kill_object(lv209) + R.vm.kill_object(alloc364) + 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((2048,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[38] + alloc366: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc365, gpt_neox_layers_3_input_layernorm_weight2, gpt_neox_layers_3_input_layernorm_bias2, alloc366) + 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((6144, 2048), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[42] + alloc367: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc366, gpt_neox_layers_3_attention_query_key_value_weight2, gpt_neox_layers_3_attention_query_key_value_bias2, alloc367) + R.vm.kill_object(alloc366) + 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) + lv214: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc367, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc367) + alloc368: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _368: 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)), lv214, alloc368) + R.vm.kill_object(lv214) + lv215: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc368, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc368) + gpt_neox_layers_3_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[40] + alloc369: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc365, gpt_neox_layers_3_post_attention_layernorm_weight2, gpt_neox_layers_3_post_attention_layernorm_bias2, alloc369) + 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((8192, 2048), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[46] + alloc370: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc369, gpt_neox_layers_3_mlp_dense_h_to_4h_weight2, gpt_neox_layers_3_mlp_dense_h_to_4h_bias2, alloc370) + R.vm.kill_object(alloc369) + 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((2048, 8192), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[48] + alloc371: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc370, gpt_neox_layers_3_mlp_dense_4h_to_h_weight2, gpt_neox_layers_3_mlp_dense_4h_to_h_bias2, alloc371) + R.vm.kill_object(alloc370) + 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((2048, 2048), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[44] + alloc372: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv215, gpt_neox_layers_3_attention_dense_weight2, gpt_neox_layers_3_attention_dense_bias2, alloc371, alloc365, alloc372) + R.vm.kill_object(alloc365) + R.vm.kill_object(lv215) + R.vm.kill_object(alloc371) + 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((2048,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[50] + alloc373: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc372, gpt_neox_layers_4_input_layernorm_weight2, gpt_neox_layers_4_input_layernorm_bias2, alloc373) + 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((6144, 2048), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[54] + alloc374: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc373, gpt_neox_layers_4_attention_query_key_value_weight2, gpt_neox_layers_4_attention_query_key_value_bias2, alloc374) + R.vm.kill_object(alloc373) + 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) + lv220: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc374, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc374) + alloc375: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _375: 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)), lv220, alloc375) + R.vm.kill_object(lv220) + lv221: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc375, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc375) + gpt_neox_layers_4_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[52] + alloc376: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc372, gpt_neox_layers_4_post_attention_layernorm_weight2, gpt_neox_layers_4_post_attention_layernorm_bias2, alloc376) + 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((8192, 2048), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[58] + alloc377: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc376, gpt_neox_layers_4_mlp_dense_h_to_4h_weight2, gpt_neox_layers_4_mlp_dense_h_to_4h_bias2, alloc377) + R.vm.kill_object(alloc376) + 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((2048, 8192), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[60] + alloc378: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc377, gpt_neox_layers_4_mlp_dense_4h_to_h_weight2, gpt_neox_layers_4_mlp_dense_4h_to_h_bias2, alloc378) + R.vm.kill_object(alloc377) + 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((2048, 2048), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[56] + alloc379: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv221, gpt_neox_layers_4_attention_dense_weight2, gpt_neox_layers_4_attention_dense_bias2, alloc378, alloc372, alloc379) + R.vm.kill_object(alloc372) + R.vm.kill_object(lv221) + R.vm.kill_object(alloc378) + 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((2048,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[62] + alloc380: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc379, gpt_neox_layers_5_input_layernorm_weight2, gpt_neox_layers_5_input_layernorm_bias2, alloc380) + 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((6144, 2048), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[66] + alloc381: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc380, gpt_neox_layers_5_attention_query_key_value_weight2, gpt_neox_layers_5_attention_query_key_value_bias2, alloc381) + R.vm.kill_object(alloc380) + 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) + lv226: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc381, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc381) + alloc382: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _382: 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)), lv226, alloc382) + R.vm.kill_object(lv226) + lv227: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc382, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc382) + gpt_neox_layers_5_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[64] + alloc383: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc379, gpt_neox_layers_5_post_attention_layernorm_weight2, gpt_neox_layers_5_post_attention_layernorm_bias2, alloc383) + 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((8192, 2048), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[70] + alloc384: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc383, gpt_neox_layers_5_mlp_dense_h_to_4h_weight2, gpt_neox_layers_5_mlp_dense_h_to_4h_bias2, alloc384) + R.vm.kill_object(alloc383) + 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((2048, 8192), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[72] + alloc385: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc384, gpt_neox_layers_5_mlp_dense_4h_to_h_weight2, gpt_neox_layers_5_mlp_dense_4h_to_h_bias2, alloc385) + R.vm.kill_object(alloc384) + 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((2048, 2048), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[68] + alloc386: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv227, gpt_neox_layers_5_attention_dense_weight2, gpt_neox_layers_5_attention_dense_bias2, alloc385, alloc379, alloc386) + R.vm.kill_object(alloc379) + R.vm.kill_object(lv227) + R.vm.kill_object(alloc385) + 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((2048,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[74] + alloc387: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc386, gpt_neox_layers_6_input_layernorm_weight2, gpt_neox_layers_6_input_layernorm_bias2, alloc387) + 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((6144, 2048), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[78] + alloc388: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc387, gpt_neox_layers_6_attention_query_key_value_weight2, gpt_neox_layers_6_attention_query_key_value_bias2, alloc388) + R.vm.kill_object(alloc387) + 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) + lv232: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc388, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc388) + alloc389: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _389: 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)), lv232, alloc389) + R.vm.kill_object(lv232) + lv233: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc389, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc389) + gpt_neox_layers_6_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[76] + alloc390: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc386, gpt_neox_layers_6_post_attention_layernorm_weight2, gpt_neox_layers_6_post_attention_layernorm_bias2, alloc390) + 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((8192, 2048), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[82] + alloc391: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc390, gpt_neox_layers_6_mlp_dense_h_to_4h_weight2, gpt_neox_layers_6_mlp_dense_h_to_4h_bias2, alloc391) + R.vm.kill_object(alloc390) + 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((2048, 8192), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[84] + alloc392: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc391, gpt_neox_layers_6_mlp_dense_4h_to_h_weight2, gpt_neox_layers_6_mlp_dense_4h_to_h_bias2, alloc392) + R.vm.kill_object(alloc391) + 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((2048, 2048), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[80] + alloc393: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv233, gpt_neox_layers_6_attention_dense_weight2, gpt_neox_layers_6_attention_dense_bias2, alloc392, alloc386, alloc393) + R.vm.kill_object(alloc386) + R.vm.kill_object(lv233) + R.vm.kill_object(alloc392) + 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((2048,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[86] + alloc394: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc393, gpt_neox_layers_7_input_layernorm_weight2, gpt_neox_layers_7_input_layernorm_bias2, alloc394) + 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((6144, 2048), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[90] + alloc395: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc394, gpt_neox_layers_7_attention_query_key_value_weight2, gpt_neox_layers_7_attention_query_key_value_bias2, alloc395) + R.vm.kill_object(alloc394) + 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) + lv238: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc395, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc395) + alloc396: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _396: 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)), lv238, alloc396) + R.vm.kill_object(lv238) + lv239: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc396, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc396) + gpt_neox_layers_7_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[88] + alloc397: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc393, gpt_neox_layers_7_post_attention_layernorm_weight2, gpt_neox_layers_7_post_attention_layernorm_bias2, alloc397) + 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((8192, 2048), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[94] + alloc398: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc397, gpt_neox_layers_7_mlp_dense_h_to_4h_weight2, gpt_neox_layers_7_mlp_dense_h_to_4h_bias2, alloc398) + R.vm.kill_object(alloc397) + 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((2048, 8192), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[96] + alloc399: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc398, gpt_neox_layers_7_mlp_dense_4h_to_h_weight2, gpt_neox_layers_7_mlp_dense_4h_to_h_bias2, alloc399) + R.vm.kill_object(alloc398) + 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((2048, 2048), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[92] + alloc400: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv239, gpt_neox_layers_7_attention_dense_weight2, gpt_neox_layers_7_attention_dense_bias2, alloc399, alloc393, alloc400) + R.vm.kill_object(alloc393) + R.vm.kill_object(lv239) + R.vm.kill_object(alloc399) + 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((2048,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[98] + alloc401: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc400, gpt_neox_layers_8_input_layernorm_weight2, gpt_neox_layers_8_input_layernorm_bias2, alloc401) + 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((6144, 2048), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[102] + alloc402: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc401, gpt_neox_layers_8_attention_query_key_value_weight2, gpt_neox_layers_8_attention_query_key_value_bias2, alloc402) + R.vm.kill_object(alloc401) + 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) + lv244: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc402, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc402) + alloc403: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _403: 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)), lv244, alloc403) + R.vm.kill_object(lv244) + lv245: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc403, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc403) + gpt_neox_layers_8_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[100] + alloc404: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc400, gpt_neox_layers_8_post_attention_layernorm_weight2, gpt_neox_layers_8_post_attention_layernorm_bias2, alloc404) + 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((8192, 2048), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[106] + alloc405: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc404, gpt_neox_layers_8_mlp_dense_h_to_4h_weight2, gpt_neox_layers_8_mlp_dense_h_to_4h_bias2, alloc405) + R.vm.kill_object(alloc404) + 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((2048, 8192), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[108] + alloc406: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc405, gpt_neox_layers_8_mlp_dense_4h_to_h_weight2, gpt_neox_layers_8_mlp_dense_4h_to_h_bias2, alloc406) + R.vm.kill_object(alloc405) + 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((2048, 2048), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[104] + alloc407: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv245, gpt_neox_layers_8_attention_dense_weight2, gpt_neox_layers_8_attention_dense_bias2, alloc406, alloc400, alloc407) + R.vm.kill_object(alloc400) + R.vm.kill_object(lv245) + R.vm.kill_object(alloc406) + 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((2048,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[110] + alloc408: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc407, gpt_neox_layers_9_input_layernorm_weight2, gpt_neox_layers_9_input_layernorm_bias2, alloc408) + 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((6144, 2048), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[114] + alloc409: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc408, gpt_neox_layers_9_attention_query_key_value_weight2, gpt_neox_layers_9_attention_query_key_value_bias2, alloc409) + R.vm.kill_object(alloc408) + 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) + lv250: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc409, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc409) + alloc410: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _410: 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)), lv250, alloc410) + R.vm.kill_object(lv250) + lv251: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc410, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc410) + gpt_neox_layers_9_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[112] + alloc411: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc407, gpt_neox_layers_9_post_attention_layernorm_weight2, gpt_neox_layers_9_post_attention_layernorm_bias2, alloc411) + 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((8192, 2048), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[118] + alloc412: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc411, gpt_neox_layers_9_mlp_dense_h_to_4h_weight2, gpt_neox_layers_9_mlp_dense_h_to_4h_bias2, alloc412) + R.vm.kill_object(alloc411) + 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((2048, 8192), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[120] + alloc413: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc412, gpt_neox_layers_9_mlp_dense_4h_to_h_weight2, gpt_neox_layers_9_mlp_dense_4h_to_h_bias2, alloc413) + R.vm.kill_object(alloc412) + 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((2048, 2048), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[116] + alloc414: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv251, gpt_neox_layers_9_attention_dense_weight2, gpt_neox_layers_9_attention_dense_bias2, alloc413, alloc407, alloc414) + R.vm.kill_object(alloc407) + R.vm.kill_object(lv251) + R.vm.kill_object(alloc413) + 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((2048,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[122] + alloc415: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc414, gpt_neox_layers_10_input_layernorm_weight2, gpt_neox_layers_10_input_layernorm_bias2, alloc415) + 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((6144, 2048), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[126] + alloc416: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc415, gpt_neox_layers_10_attention_query_key_value_weight2, gpt_neox_layers_10_attention_query_key_value_bias2, alloc416) + R.vm.kill_object(alloc415) + 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) + lv256: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc416, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc416) + alloc417: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _417: 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)), lv256, alloc417) + R.vm.kill_object(lv256) + lv257: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc417, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc417) + gpt_neox_layers_10_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[124] + alloc418: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc414, gpt_neox_layers_10_post_attention_layernorm_weight2, gpt_neox_layers_10_post_attention_layernorm_bias2, alloc418) + 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((8192, 2048), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[130] + alloc419: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc418, gpt_neox_layers_10_mlp_dense_h_to_4h_weight2, gpt_neox_layers_10_mlp_dense_h_to_4h_bias2, alloc419) + R.vm.kill_object(alloc418) + 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((2048, 8192), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[132] + alloc420: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc419, gpt_neox_layers_10_mlp_dense_4h_to_h_weight2, gpt_neox_layers_10_mlp_dense_4h_to_h_bias2, alloc420) + R.vm.kill_object(alloc419) + 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((2048, 2048), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[128] + alloc421: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv257, gpt_neox_layers_10_attention_dense_weight2, gpt_neox_layers_10_attention_dense_bias2, alloc420, alloc414, alloc421) + R.vm.kill_object(alloc414) + R.vm.kill_object(lv257) + R.vm.kill_object(alloc420) + 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((2048,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[134] + alloc422: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc421, gpt_neox_layers_11_input_layernorm_weight2, gpt_neox_layers_11_input_layernorm_bias2, alloc422) + 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((6144, 2048), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[138] + alloc423: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc422, gpt_neox_layers_11_attention_query_key_value_weight2, gpt_neox_layers_11_attention_query_key_value_bias2, alloc423) + R.vm.kill_object(alloc422) + 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) + lv262: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc423, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc423) + alloc424: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _424: 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)), lv262, alloc424) + R.vm.kill_object(lv262) + lv263: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc424, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc424) + gpt_neox_layers_11_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[136] + alloc425: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc421, gpt_neox_layers_11_post_attention_layernorm_weight2, gpt_neox_layers_11_post_attention_layernorm_bias2, alloc425) + 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((8192, 2048), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[142] + alloc426: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc425, gpt_neox_layers_11_mlp_dense_h_to_4h_weight2, gpt_neox_layers_11_mlp_dense_h_to_4h_bias2, alloc426) + R.vm.kill_object(alloc425) + 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((2048, 8192), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[144] + alloc427: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc426, gpt_neox_layers_11_mlp_dense_4h_to_h_weight2, gpt_neox_layers_11_mlp_dense_4h_to_h_bias2, alloc427) + R.vm.kill_object(alloc426) + 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((2048, 2048), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[140] + alloc428: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv263, gpt_neox_layers_11_attention_dense_weight2, gpt_neox_layers_11_attention_dense_bias2, alloc427, alloc421, alloc428) + R.vm.kill_object(alloc421) + R.vm.kill_object(lv263) + R.vm.kill_object(alloc427) + 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((2048,), dtype="float16") = packed_params[145] + gpt_neox_layers_12_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[146] + alloc429: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc428, gpt_neox_layers_12_input_layernorm_weight2, gpt_neox_layers_12_input_layernorm_bias2, alloc429) + 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((6144, 2048), dtype="float16") = packed_params[149] + gpt_neox_layers_12_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[150] + alloc430: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc429, gpt_neox_layers_12_attention_query_key_value_weight2, gpt_neox_layers_12_attention_query_key_value_bias2, alloc430) + R.vm.kill_object(alloc429) + 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) + lv268: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc430, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc430) + alloc431: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _431: 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)), lv268, alloc431) + R.vm.kill_object(lv268) + lv269: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc431, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc431) + gpt_neox_layers_12_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[147] + gpt_neox_layers_12_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[148] + alloc432: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc428, gpt_neox_layers_12_post_attention_layernorm_weight2, gpt_neox_layers_12_post_attention_layernorm_bias2, alloc432) + 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((8192, 2048), dtype="float16") = packed_params[153] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[154] + alloc433: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc432, gpt_neox_layers_12_mlp_dense_h_to_4h_weight2, gpt_neox_layers_12_mlp_dense_h_to_4h_bias2, alloc433) + R.vm.kill_object(alloc432) + 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((2048, 8192), dtype="float16") = packed_params[155] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[156] + alloc434: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc433, gpt_neox_layers_12_mlp_dense_4h_to_h_weight2, gpt_neox_layers_12_mlp_dense_4h_to_h_bias2, alloc434) + R.vm.kill_object(alloc433) + 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((2048, 2048), dtype="float16") = packed_params[151] + gpt_neox_layers_12_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[152] + alloc435: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv269, gpt_neox_layers_12_attention_dense_weight2, gpt_neox_layers_12_attention_dense_bias2, alloc434, alloc428, alloc435) + R.vm.kill_object(alloc428) + R.vm.kill_object(lv269) + R.vm.kill_object(alloc434) + 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((2048,), dtype="float16") = packed_params[157] + gpt_neox_layers_13_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[158] + alloc436: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc435, gpt_neox_layers_13_input_layernorm_weight2, gpt_neox_layers_13_input_layernorm_bias2, alloc436) + 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((6144, 2048), dtype="float16") = packed_params[161] + gpt_neox_layers_13_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[162] + alloc437: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc436, gpt_neox_layers_13_attention_query_key_value_weight2, gpt_neox_layers_13_attention_query_key_value_bias2, alloc437) + R.vm.kill_object(alloc436) + 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) + lv274: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc437, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc437) + alloc438: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _438: 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)), lv274, alloc438) + R.vm.kill_object(lv274) + lv275: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc438, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc438) + gpt_neox_layers_13_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[159] + gpt_neox_layers_13_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[160] + alloc439: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc435, gpt_neox_layers_13_post_attention_layernorm_weight2, gpt_neox_layers_13_post_attention_layernorm_bias2, alloc439) + 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((8192, 2048), dtype="float16") = packed_params[165] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[166] + alloc440: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc439, gpt_neox_layers_13_mlp_dense_h_to_4h_weight2, gpt_neox_layers_13_mlp_dense_h_to_4h_bias2, alloc440) + R.vm.kill_object(alloc439) + 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((2048, 8192), dtype="float16") = packed_params[167] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[168] + alloc441: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc440, gpt_neox_layers_13_mlp_dense_4h_to_h_weight2, gpt_neox_layers_13_mlp_dense_4h_to_h_bias2, alloc441) + R.vm.kill_object(alloc440) + 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((2048, 2048), dtype="float16") = packed_params[163] + gpt_neox_layers_13_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[164] + alloc442: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv275, gpt_neox_layers_13_attention_dense_weight2, gpt_neox_layers_13_attention_dense_bias2, alloc441, alloc435, alloc442) + R.vm.kill_object(alloc435) + R.vm.kill_object(lv275) + R.vm.kill_object(alloc441) + 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((2048,), dtype="float16") = packed_params[169] + gpt_neox_layers_14_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[170] + alloc443: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc442, gpt_neox_layers_14_input_layernorm_weight2, gpt_neox_layers_14_input_layernorm_bias2, alloc443) + 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((6144, 2048), dtype="float16") = packed_params[173] + gpt_neox_layers_14_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[174] + alloc444: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc443, gpt_neox_layers_14_attention_query_key_value_weight2, gpt_neox_layers_14_attention_query_key_value_bias2, alloc444) + R.vm.kill_object(alloc443) + 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) + lv280: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc444, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc444) + alloc445: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _445: 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)), lv280, alloc445) + R.vm.kill_object(lv280) + lv281: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc445, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc445) + gpt_neox_layers_14_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[171] + gpt_neox_layers_14_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[172] + alloc446: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc442, gpt_neox_layers_14_post_attention_layernorm_weight2, gpt_neox_layers_14_post_attention_layernorm_bias2, alloc446) + 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((8192, 2048), dtype="float16") = packed_params[177] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[178] + alloc447: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc446, gpt_neox_layers_14_mlp_dense_h_to_4h_weight2, gpt_neox_layers_14_mlp_dense_h_to_4h_bias2, alloc447) + R.vm.kill_object(alloc446) + 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((2048, 8192), dtype="float16") = packed_params[179] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[180] + alloc448: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc447, gpt_neox_layers_14_mlp_dense_4h_to_h_weight2, gpt_neox_layers_14_mlp_dense_4h_to_h_bias2, alloc448) + R.vm.kill_object(alloc447) + 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((2048, 2048), dtype="float16") = packed_params[175] + gpt_neox_layers_14_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[176] + alloc449: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage20, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + R.vm.kill_object(storage20) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv281, gpt_neox_layers_14_attention_dense_weight2, gpt_neox_layers_14_attention_dense_bias2, alloc448, alloc442, alloc449) + R.vm.kill_object(alloc442) + R.vm.kill_object(lv281) + R.vm.kill_object(alloc448) + 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((2048,), dtype="float16") = packed_params[181] + gpt_neox_layers_15_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[182] + alloc450: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc449, gpt_neox_layers_15_input_layernorm_weight2, gpt_neox_layers_15_input_layernorm_bias2, alloc450) + 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((6144, 2048), dtype="float16") = packed_params[185] + gpt_neox_layers_15_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[186] + alloc451: R.Tensor((1, 1, 6144), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 6144]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc450, gpt_neox_layers_15_attention_query_key_value_weight2, gpt_neox_layers_15_attention_query_key_value_bias2, alloc451) + R.vm.kill_object(alloc450) + 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) + lv286: R.Tensor((1, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc451, R.shape([1, 24, 256]), sinfo_args=(R.Tensor((1, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc451) + alloc452: R.Tensor((1, 8, 256), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 8, 256]), R.dtype("float16")) + _452: 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)), lv286, alloc452) + R.vm.kill_object(lv286) + lv287: R.Tensor((1, 1, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc452, R.shape([1, 1, 2048]), sinfo_args=(R.Tensor((1, 1, 2048), dtype="float16"),)) + R.vm.kill_object(alloc452) + gpt_neox_layers_15_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[183] + gpt_neox_layers_15_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[184] + alloc453: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc449, gpt_neox_layers_15_post_attention_layernorm_weight2, gpt_neox_layers_15_post_attention_layernorm_bias2, alloc453) + 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((8192, 2048), dtype="float16") = packed_params[189] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[190] + alloc454: R.Tensor((1, 1, 8192), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 8192]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc453, gpt_neox_layers_15_mlp_dense_h_to_4h_weight2, gpt_neox_layers_15_mlp_dense_h_to_4h_bias2, alloc454) + R.vm.kill_object(alloc453) + 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((2048, 8192), dtype="float16") = packed_params[191] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[192] + alloc455: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage19, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + R.vm.kill_object(storage19) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc454, gpt_neox_layers_15_mlp_dense_4h_to_h_weight2, gpt_neox_layers_15_mlp_dense_4h_to_h_bias2, alloc455) + R.vm.kill_object(alloc454) + 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((2048, 2048), dtype="float16") = packed_params[187] + gpt_neox_layers_15_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[188] + alloc456: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage21, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + R.vm.kill_object(storage21) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv287, gpt_neox_layers_15_attention_dense_weight2, gpt_neox_layers_15_attention_dense_bias2, alloc455, alloc449, alloc456) + R.vm.kill_object(alloc449) + R.vm.kill_object(lv287) + R.vm.kill_object(alloc455) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_bias2) + gpt_neox_final_layer_norm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[193] + gpt_neox_final_layer_norm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[194] + alloc457: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + R.vm.kill_object(storage18) + cls.gpt_neox_q0f16_layer_norm2(alloc456, gpt_neox_final_layer_norm_weight2, gpt_neox_final_layer_norm_bias2, alloc457) + R.vm.kill_object(alloc456) + 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, 2048), dtype="float16") = packed_params[195] + gv553: 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),)) + storage22: R.Object = R.vm.alloc_storage(gv553, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv554: R.Shape(ndim=3) = 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),)) + alloc458: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv554, R.dtype("float32")) + R.vm.kill_object(storage22) + cls.gpt_neox_q0f16_fused_NT_matmul14_cast8(alloc457, embed_out_weight2, alloc458) + R.vm.kill_object(alloc457) + R.vm.kill_object(embed_out_weight2) + gv2: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc458, paged_kv_cache + R.vm.kill_object(alloc458) + gv555: R.Tensor(dtype="float32", ndim=3) = gv2[0] + R.call_packed("vm.builtin.match_shape", gv555, 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", 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 2048), dtype="float16"))) -> R.Tensor(("seq_len", 2048), 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(196), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv365: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv365, 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, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv366: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[195] + R.call_packed("vm.builtin.check_tensor_info", gv366, 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, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), 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", gv365, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(2048), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv366, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(2048), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gpt_neox_embed_in_weight: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[0] + storage11: R.Object = R.vm.alloc_storage(R.shape([8388608]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv367: 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(2048), sinfo_args=(R.Shape(ndim=2),)) + alloc229: R.Tensor(dtype="float16", ndim=2) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv367, R.dtype("float16")) + R.vm.kill_object(storage11) + cls.gpt_neox_q0f16_take1(gpt_neox_embed_in_weight, input_ids, alloc229) + R.vm.kill_object(gpt_neox_embed_in_weight) + R.call_packed("vm.builtin.match_shape", alloc229, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(0), R.prim_value(0), R.prim_value(2048), R.str("ErrorContext(fn=embed, loc=return, annotation=R.Tensor((seq_len, 2048), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + return alloc229 + + @R.function + def prefill(input_embed: R.Tensor((1, "seq_len", 2048), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 2048), 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, 2048), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(196), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv, 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, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv1: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[195] + R.call_packed("vm.builtin.check_tensor_info", gv1, 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, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), 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(2048), R.str("ErrorContext(fn=prefill, loc=param[0], param=input_embed, annotation=R.Tensor((1, seq_len, 2048), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(2048), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv1, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(2048), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((6144, 2048), dtype=\"float16\"), R.Tensor((6144,), dtype=\"float16\"), R.Tensor((2048, 2048), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((8192, 2048), dtype=\"float16\"), R.Tensor((8192,), dtype=\"float32\"), R.Tensor((2048, 8192), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float32\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((2048,), dtype=\"float16\"), R.Tensor((vocab_size, 2048), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func(shape_heap) + gpt_neox_layers_0_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[2] + storage: R.Object = R.vm.alloc_storage(R.shape([8388608]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv2: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv2, 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, alloc) + 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((6144, 2048), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[6] + storage1: R.Object = R.vm.alloc_storage(R.shape([25165824]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv3: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc1: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv3, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc, gpt_neox_layers_0_attention_query_key_value_weight1, gpt_neox_layers_0_attention_query_key_value_bias1, alloc1) + R.vm.kill_object(alloc) + 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) + gv4: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc1, gv4, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc1) + gv5: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape1: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape, gv5, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape) + gv6: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc2: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv6, R.dtype("float16")) + _2: 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, alloc2) + R.vm.kill_object(reshape1) + gv7: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape2: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc2, gv7, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc2) + gv8: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape3: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape2, gv8, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape2) + gpt_neox_layers_0_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[4] + gv9: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc3: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv9, 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, alloc3) + 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((8192, 2048), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[10] + storage2: R.Object = R.vm.alloc_storage(R.shape([33554432]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv10: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc4: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv10, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc3, gpt_neox_layers_0_mlp_dense_h_to_4h_weight1, gpt_neox_layers_0_mlp_dense_h_to_4h_bias1, alloc4) + R.vm.kill_object(alloc3) + 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((2048, 8192), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[12] + gv11: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc5: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv11, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc4, gpt_neox_layers_0_mlp_dense_4h_to_h_weight1, gpt_neox_layers_0_mlp_dense_4h_to_h_bias1, alloc5) + R.vm.kill_object(alloc4) + 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((2048, 2048), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[8] + gv12: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc6: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv12, 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, alloc5, input_embed, alloc6) + R.vm.kill_object(reshape3) + R.vm.kill_object(alloc5) + 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((2048,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[14] + gv13: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc7: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv13, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc6, gpt_neox_layers_1_input_layernorm_weight1, gpt_neox_layers_1_input_layernorm_bias1, alloc7) + 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((6144, 2048), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[18] + gv14: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc8: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv14, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc7, gpt_neox_layers_1_attention_query_key_value_weight1, gpt_neox_layers_1_attention_query_key_value_bias1, alloc8) + R.vm.kill_object(alloc7) + 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) + gv15: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape4: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc8, gv15, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc8) + gv16: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape5: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape4, gv16, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape4) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc9: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv17, R.dtype("float16")) + _9: 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, alloc9) + R.vm.kill_object(reshape5) + gv18: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape6: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc9, gv18, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc9) + gv19: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape7: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape6, gv19, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape6) + gpt_neox_layers_1_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[16] + gv20: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc10: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv20, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc6, gpt_neox_layers_1_post_attention_layernorm_weight1, gpt_neox_layers_1_post_attention_layernorm_bias1, alloc10) + 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((8192, 2048), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[22] + storage3: R.Object = R.vm.alloc_storage(R.shape([33554432]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv21: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc11: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv21, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc10, gpt_neox_layers_1_mlp_dense_h_to_4h_weight1, gpt_neox_layers_1_mlp_dense_h_to_4h_bias1, alloc11) + R.vm.kill_object(alloc10) + 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((2048, 8192), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[24] + gv22: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc12: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv22, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc11, gpt_neox_layers_1_mlp_dense_4h_to_h_weight1, gpt_neox_layers_1_mlp_dense_4h_to_h_bias1, alloc12) + R.vm.kill_object(alloc11) + 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((2048, 2048), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[20] + gv23: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc13: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv23, 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, alloc12, alloc6, alloc13) + R.vm.kill_object(alloc6) + R.vm.kill_object(reshape7) + R.vm.kill_object(alloc12) + 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((2048,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[26] + gv24: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc14: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv24, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc13, gpt_neox_layers_2_input_layernorm_weight1, gpt_neox_layers_2_input_layernorm_bias1, alloc14) + 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((6144, 2048), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[30] + gv25: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc15: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv25, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc14, gpt_neox_layers_2_attention_query_key_value_weight1, gpt_neox_layers_2_attention_query_key_value_bias1, alloc15) + R.vm.kill_object(alloc14) + 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) + gv26: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape8: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc15, gv26, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc15) + gv27: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape9: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape8, gv27, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape8) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc16: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv28, R.dtype("float16")) + _16: 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, alloc16) + R.vm.kill_object(reshape9) + gv29: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape10: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc16, gv29, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc16) + gv30: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape11: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape10, gv30, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape10) + gpt_neox_layers_2_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[28] + gv31: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc17: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv31, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc13, gpt_neox_layers_2_post_attention_layernorm_weight1, gpt_neox_layers_2_post_attention_layernorm_bias1, alloc17) + 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((8192, 2048), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[34] + gv32: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc18: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv32, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc17, gpt_neox_layers_2_mlp_dense_h_to_4h_weight1, gpt_neox_layers_2_mlp_dense_h_to_4h_bias1, alloc18) + R.vm.kill_object(alloc17) + 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((2048, 8192), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[36] + gv33: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc19: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv33, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc18, gpt_neox_layers_2_mlp_dense_4h_to_h_weight1, gpt_neox_layers_2_mlp_dense_4h_to_h_bias1, alloc19) + R.vm.kill_object(alloc18) + 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((2048, 2048), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[32] + gv34: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc20: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv34, 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, alloc19, alloc13, alloc20) + R.vm.kill_object(alloc13) + R.vm.kill_object(reshape11) + R.vm.kill_object(alloc19) + 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((2048,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[38] + gv35: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc21: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv35, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc20, gpt_neox_layers_3_input_layernorm_weight1, gpt_neox_layers_3_input_layernorm_bias1, alloc21) + 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((6144, 2048), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[42] + gv36: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc22: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv36, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc21, gpt_neox_layers_3_attention_query_key_value_weight1, gpt_neox_layers_3_attention_query_key_value_bias1, alloc22) + R.vm.kill_object(alloc21) + 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) + gv37: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape12: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc22, gv37, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc22) + gv38: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape13: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape12, gv38, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape12) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc23: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv39, R.dtype("float16")) + _23: 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, alloc23) + R.vm.kill_object(reshape13) + gv40: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape14: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc23, gv40, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc23) + gv41: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape15: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape14, gv41, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape14) + gpt_neox_layers_3_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[40] + gv42: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc24: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv42, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc20, gpt_neox_layers_3_post_attention_layernorm_weight1, gpt_neox_layers_3_post_attention_layernorm_bias1, alloc24) + 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((8192, 2048), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[46] + gv43: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc25: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv43, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc24, gpt_neox_layers_3_mlp_dense_h_to_4h_weight1, gpt_neox_layers_3_mlp_dense_h_to_4h_bias1, alloc25) + R.vm.kill_object(alloc24) + 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((2048, 8192), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[48] + gv44: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc26: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv44, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc25, gpt_neox_layers_3_mlp_dense_4h_to_h_weight1, gpt_neox_layers_3_mlp_dense_4h_to_h_bias1, alloc26) + R.vm.kill_object(alloc25) + 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((2048, 2048), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[44] + gv45: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc27: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv45, 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, alloc26, alloc20, alloc27) + R.vm.kill_object(alloc20) + R.vm.kill_object(reshape15) + R.vm.kill_object(alloc26) + 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((2048,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[50] + gv46: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc28: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv46, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc27, gpt_neox_layers_4_input_layernorm_weight1, gpt_neox_layers_4_input_layernorm_bias1, alloc28) + 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((6144, 2048), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[54] + gv47: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc29: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv47, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc28, gpt_neox_layers_4_attention_query_key_value_weight1, gpt_neox_layers_4_attention_query_key_value_bias1, alloc29) + R.vm.kill_object(alloc28) + 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) + gv48: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape16: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc29, gv48, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc29) + gv49: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape17: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape16, gv49, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape16) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc30: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv50, R.dtype("float16")) + _30: 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, alloc30) + R.vm.kill_object(reshape17) + gv51: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape18: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc30, gv51, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc30) + gv52: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape19: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape18, gv52, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape18) + gpt_neox_layers_4_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[52] + gv53: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc31: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv53, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc27, gpt_neox_layers_4_post_attention_layernorm_weight1, gpt_neox_layers_4_post_attention_layernorm_bias1, alloc31) + 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((8192, 2048), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[58] + gv54: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc32: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv54, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc31, gpt_neox_layers_4_mlp_dense_h_to_4h_weight1, gpt_neox_layers_4_mlp_dense_h_to_4h_bias1, alloc32) + R.vm.kill_object(alloc31) + 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((2048, 8192), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[60] + gv55: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc33: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv55, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc32, gpt_neox_layers_4_mlp_dense_4h_to_h_weight1, gpt_neox_layers_4_mlp_dense_4h_to_h_bias1, alloc33) + R.vm.kill_object(alloc32) + 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((2048, 2048), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[56] + gv56: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc34: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv56, 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, alloc33, alloc27, alloc34) + R.vm.kill_object(alloc27) + R.vm.kill_object(reshape19) + R.vm.kill_object(alloc33) + 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((2048,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[62] + gv57: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc35: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv57, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc34, gpt_neox_layers_5_input_layernorm_weight1, gpt_neox_layers_5_input_layernorm_bias1, alloc35) + 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((6144, 2048), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[66] + gv58: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc36: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv58, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc35, gpt_neox_layers_5_attention_query_key_value_weight1, gpt_neox_layers_5_attention_query_key_value_bias1, alloc36) + R.vm.kill_object(alloc35) + 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) + gv59: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape20: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc36, gv59, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc36) + gv60: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape21: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape20, gv60, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape20) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc37: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv61, R.dtype("float16")) + _37: 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, alloc37) + R.vm.kill_object(reshape21) + gv62: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape22: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc37, gv62, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc37) + gv63: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape23: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape22, gv63, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape22) + gpt_neox_layers_5_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[64] + gv64: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc38: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv64, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc34, gpt_neox_layers_5_post_attention_layernorm_weight1, gpt_neox_layers_5_post_attention_layernorm_bias1, alloc38) + 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((8192, 2048), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[70] + gv65: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc39: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv65, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc38, gpt_neox_layers_5_mlp_dense_h_to_4h_weight1, gpt_neox_layers_5_mlp_dense_h_to_4h_bias1, alloc39) + R.vm.kill_object(alloc38) + 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((2048, 8192), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[72] + gv66: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc40: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv66, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc39, gpt_neox_layers_5_mlp_dense_4h_to_h_weight1, gpt_neox_layers_5_mlp_dense_4h_to_h_bias1, alloc40) + R.vm.kill_object(alloc39) + 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((2048, 2048), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[68] + gv67: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc41: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv67, 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, alloc40, alloc34, alloc41) + R.vm.kill_object(alloc34) + R.vm.kill_object(reshape23) + R.vm.kill_object(alloc40) + 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((2048,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[74] + gv68: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc42: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv68, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc41, gpt_neox_layers_6_input_layernorm_weight1, gpt_neox_layers_6_input_layernorm_bias1, alloc42) + 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((6144, 2048), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[78] + gv69: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc43: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv69, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc42, gpt_neox_layers_6_attention_query_key_value_weight1, gpt_neox_layers_6_attention_query_key_value_bias1, alloc43) + R.vm.kill_object(alloc42) + 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) + gv70: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape24: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc43, gv70, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc43) + gv71: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape25: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape24, gv71, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape24) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc44: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv72, R.dtype("float16")) + _44: 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, alloc44) + R.vm.kill_object(reshape25) + gv73: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape26: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc44, gv73, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc44) + gv74: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape27: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape26, gv74, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape26) + gpt_neox_layers_6_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[76] + gv75: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc45: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv75, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc41, gpt_neox_layers_6_post_attention_layernorm_weight1, gpt_neox_layers_6_post_attention_layernorm_bias1, alloc45) + 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((8192, 2048), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[82] + gv76: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc46: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv76, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc45, gpt_neox_layers_6_mlp_dense_h_to_4h_weight1, gpt_neox_layers_6_mlp_dense_h_to_4h_bias1, alloc46) + R.vm.kill_object(alloc45) + 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((2048, 8192), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[84] + gv77: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc47: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv77, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc46, gpt_neox_layers_6_mlp_dense_4h_to_h_weight1, gpt_neox_layers_6_mlp_dense_4h_to_h_bias1, alloc47) + R.vm.kill_object(alloc46) + 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((2048, 2048), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[80] + gv78: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc48: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv78, 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, alloc47, alloc41, alloc48) + R.vm.kill_object(alloc41) + R.vm.kill_object(reshape27) + R.vm.kill_object(alloc47) + 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((2048,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[86] + gv79: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc49: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv79, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc48, gpt_neox_layers_7_input_layernorm_weight1, gpt_neox_layers_7_input_layernorm_bias1, alloc49) + 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((6144, 2048), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[90] + gv80: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc50: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv80, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc49, gpt_neox_layers_7_attention_query_key_value_weight1, gpt_neox_layers_7_attention_query_key_value_bias1, alloc50) + R.vm.kill_object(alloc49) + 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) + gv81: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape28: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc50, gv81, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc50) + gv82: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape29: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape28, gv82, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape28) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc51: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv83, R.dtype("float16")) + _51: 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, alloc51) + R.vm.kill_object(reshape29) + gv84: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape30: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc51, gv84, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc51) + gv85: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape31: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape30, gv85, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape30) + gpt_neox_layers_7_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[88] + gv86: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc52: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv86, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc48, gpt_neox_layers_7_post_attention_layernorm_weight1, gpt_neox_layers_7_post_attention_layernorm_bias1, alloc52) + 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((8192, 2048), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[94] + gv87: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc53: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv87, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc52, gpt_neox_layers_7_mlp_dense_h_to_4h_weight1, gpt_neox_layers_7_mlp_dense_h_to_4h_bias1, alloc53) + R.vm.kill_object(alloc52) + 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((2048, 8192), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[96] + gv88: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc54: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv88, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc53, gpt_neox_layers_7_mlp_dense_4h_to_h_weight1, gpt_neox_layers_7_mlp_dense_4h_to_h_bias1, alloc54) + R.vm.kill_object(alloc53) + 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((2048, 2048), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[92] + gv89: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc55: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv89, 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, alloc54, alloc48, alloc55) + R.vm.kill_object(alloc48) + R.vm.kill_object(reshape31) + R.vm.kill_object(alloc54) + 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((2048,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[98] + gv90: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc56: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv90, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc55, gpt_neox_layers_8_input_layernorm_weight1, gpt_neox_layers_8_input_layernorm_bias1, alloc56) + 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((6144, 2048), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[102] + gv91: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc57: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv91, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc56, gpt_neox_layers_8_attention_query_key_value_weight1, gpt_neox_layers_8_attention_query_key_value_bias1, alloc57) + R.vm.kill_object(alloc56) + 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) + gv92: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape32: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc57, gv92, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc57) + gv93: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape33: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape32, gv93, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape32) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc58: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv94, R.dtype("float16")) + _58: 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, alloc58) + R.vm.kill_object(reshape33) + gv95: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape34: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc58, gv95, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc58) + gv96: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape35: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape34, gv96, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape34) + gpt_neox_layers_8_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[100] + gv97: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc59: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv97, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc55, gpt_neox_layers_8_post_attention_layernorm_weight1, gpt_neox_layers_8_post_attention_layernorm_bias1, alloc59) + 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((8192, 2048), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[106] + gv98: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc60: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv98, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc59, gpt_neox_layers_8_mlp_dense_h_to_4h_weight1, gpt_neox_layers_8_mlp_dense_h_to_4h_bias1, alloc60) + R.vm.kill_object(alloc59) + 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((2048, 8192), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[108] + gv99: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc61: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv99, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc60, gpt_neox_layers_8_mlp_dense_4h_to_h_weight1, gpt_neox_layers_8_mlp_dense_4h_to_h_bias1, alloc61) + R.vm.kill_object(alloc60) + 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((2048, 2048), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[104] + gv100: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc62: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv100, 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, alloc61, alloc55, alloc62) + R.vm.kill_object(alloc55) + R.vm.kill_object(reshape35) + R.vm.kill_object(alloc61) + 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((2048,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[110] + gv101: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc63: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv101, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc62, gpt_neox_layers_9_input_layernorm_weight1, gpt_neox_layers_9_input_layernorm_bias1, alloc63) + 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((6144, 2048), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[114] + gv102: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc64: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv102, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc63, gpt_neox_layers_9_attention_query_key_value_weight1, gpt_neox_layers_9_attention_query_key_value_bias1, alloc64) + R.vm.kill_object(alloc63) + 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) + gv103: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape36: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc64, gv103, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc64) + gv104: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape37: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape36, gv104, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape36) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc65: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv105, R.dtype("float16")) + _65: 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, alloc65) + R.vm.kill_object(reshape37) + gv106: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape38: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc65, gv106, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc65) + gv107: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape39: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape38, gv107, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape38) + gpt_neox_layers_9_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[112] + gv108: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc66: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv108, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc62, gpt_neox_layers_9_post_attention_layernorm_weight1, gpt_neox_layers_9_post_attention_layernorm_bias1, alloc66) + 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((8192, 2048), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[118] + gv109: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc67: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv109, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc66, gpt_neox_layers_9_mlp_dense_h_to_4h_weight1, gpt_neox_layers_9_mlp_dense_h_to_4h_bias1, alloc67) + R.vm.kill_object(alloc66) + 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((2048, 8192), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[120] + gv110: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc68: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv110, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc67, gpt_neox_layers_9_mlp_dense_4h_to_h_weight1, gpt_neox_layers_9_mlp_dense_4h_to_h_bias1, alloc68) + R.vm.kill_object(alloc67) + 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((2048, 2048), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[116] + gv111: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc69: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv111, 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, alloc68, alloc62, alloc69) + R.vm.kill_object(alloc62) + R.vm.kill_object(reshape39) + R.vm.kill_object(alloc68) + 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((2048,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[122] + gv112: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc70: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv112, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc69, gpt_neox_layers_10_input_layernorm_weight1, gpt_neox_layers_10_input_layernorm_bias1, alloc70) + 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((6144, 2048), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[126] + gv113: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc71: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv113, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc70, gpt_neox_layers_10_attention_query_key_value_weight1, gpt_neox_layers_10_attention_query_key_value_bias1, alloc71) + R.vm.kill_object(alloc70) + 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) + gv114: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape40: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc71, gv114, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc71) + gv115: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape41: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape40, gv115, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape40) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc72: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv116, R.dtype("float16")) + _72: 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, alloc72) + R.vm.kill_object(reshape41) + gv117: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape42: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc72, gv117, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc72) + gv118: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape43: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape42, gv118, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape42) + gpt_neox_layers_10_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[124] + gv119: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc73: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv119, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc69, gpt_neox_layers_10_post_attention_layernorm_weight1, gpt_neox_layers_10_post_attention_layernorm_bias1, alloc73) + 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((8192, 2048), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[130] + gv120: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc74: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv120, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc73, gpt_neox_layers_10_mlp_dense_h_to_4h_weight1, gpt_neox_layers_10_mlp_dense_h_to_4h_bias1, alloc74) + R.vm.kill_object(alloc73) + 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((2048, 8192), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[132] + gv121: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc75: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv121, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc74, gpt_neox_layers_10_mlp_dense_4h_to_h_weight1, gpt_neox_layers_10_mlp_dense_4h_to_h_bias1, alloc75) + R.vm.kill_object(alloc74) + 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((2048, 2048), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[128] + gv122: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc76: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv122, 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, alloc75, alloc69, alloc76) + R.vm.kill_object(alloc69) + R.vm.kill_object(reshape43) + R.vm.kill_object(alloc75) + 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((2048,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[134] + gv123: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc77: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv123, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc76, gpt_neox_layers_11_input_layernorm_weight1, gpt_neox_layers_11_input_layernorm_bias1, alloc77) + 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((6144, 2048), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[138] + gv124: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc78: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv124, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc77, gpt_neox_layers_11_attention_query_key_value_weight1, gpt_neox_layers_11_attention_query_key_value_bias1, alloc78) + R.vm.kill_object(alloc77) + 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) + gv125: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape44: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc78, gv125, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc78) + gv126: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape45: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape44, gv126, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape44) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc79: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv127, R.dtype("float16")) + _79: 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, alloc79) + R.vm.kill_object(reshape45) + gv128: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape46: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc79, gv128, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc79) + gv129: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape47: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape46, gv129, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape46) + gpt_neox_layers_11_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[136] + gv130: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc80: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv130, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc76, gpt_neox_layers_11_post_attention_layernorm_weight1, gpt_neox_layers_11_post_attention_layernorm_bias1, alloc80) + 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((8192, 2048), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[142] + gv131: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc81: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv131, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc80, gpt_neox_layers_11_mlp_dense_h_to_4h_weight1, gpt_neox_layers_11_mlp_dense_h_to_4h_bias1, alloc81) + R.vm.kill_object(alloc80) + 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((2048, 8192), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[144] + gv132: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc82: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv132, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc81, gpt_neox_layers_11_mlp_dense_4h_to_h_weight1, gpt_neox_layers_11_mlp_dense_4h_to_h_bias1, alloc82) + R.vm.kill_object(alloc81) + 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((2048, 2048), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[140] + gv133: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc83: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv133, 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, alloc82, alloc76, alloc83) + R.vm.kill_object(alloc76) + R.vm.kill_object(reshape47) + R.vm.kill_object(alloc82) + 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((2048,), dtype="float16") = packed_params[145] + gpt_neox_layers_12_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[146] + gv134: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc84: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv134, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc83, gpt_neox_layers_12_input_layernorm_weight1, gpt_neox_layers_12_input_layernorm_bias1, alloc84) + 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((6144, 2048), dtype="float16") = packed_params[149] + gpt_neox_layers_12_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[150] + gv135: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc85: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv135, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc84, gpt_neox_layers_12_attention_query_key_value_weight1, gpt_neox_layers_12_attention_query_key_value_bias1, alloc85) + R.vm.kill_object(alloc84) + 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) + gv136: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape48: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc85, gv136, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc85) + gv137: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape49: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape48, gv137, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape48) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc86: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv138, R.dtype("float16")) + _86: 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, alloc86) + R.vm.kill_object(reshape49) + gv139: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape50: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc86, gv139, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc86) + gv140: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape51: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape50, gv140, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape50) + gpt_neox_layers_12_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[147] + gpt_neox_layers_12_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[148] + gv141: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc87: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv141, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc83, gpt_neox_layers_12_post_attention_layernorm_weight1, gpt_neox_layers_12_post_attention_layernorm_bias1, alloc87) + 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((8192, 2048), dtype="float16") = packed_params[153] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[154] + gv142: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc88: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv142, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc87, gpt_neox_layers_12_mlp_dense_h_to_4h_weight1, gpt_neox_layers_12_mlp_dense_h_to_4h_bias1, alloc88) + R.vm.kill_object(alloc87) + 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((2048, 8192), dtype="float16") = packed_params[155] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[156] + gv143: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc89: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv143, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc88, gpt_neox_layers_12_mlp_dense_4h_to_h_weight1, gpt_neox_layers_12_mlp_dense_4h_to_h_bias1, alloc89) + R.vm.kill_object(alloc88) + 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((2048, 2048), dtype="float16") = packed_params[151] + gpt_neox_layers_12_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[152] + gv144: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc90: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv144, 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, alloc89, alloc83, alloc90) + R.vm.kill_object(alloc83) + R.vm.kill_object(reshape51) + R.vm.kill_object(alloc89) + 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((2048,), dtype="float16") = packed_params[157] + gpt_neox_layers_13_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[158] + gv145: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc91: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv145, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc90, gpt_neox_layers_13_input_layernorm_weight1, gpt_neox_layers_13_input_layernorm_bias1, alloc91) + 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((6144, 2048), dtype="float16") = packed_params[161] + gpt_neox_layers_13_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[162] + gv146: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc92: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv146, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc91, gpt_neox_layers_13_attention_query_key_value_weight1, gpt_neox_layers_13_attention_query_key_value_bias1, alloc92) + R.vm.kill_object(alloc91) + 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) + gv147: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape52: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc92, gv147, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc92) + gv148: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape53: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape52, gv148, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape52) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc93: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv149, R.dtype("float16")) + _93: 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, alloc93) + R.vm.kill_object(reshape53) + gv150: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape54: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc93, gv150, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc93) + gv151: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape55: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape54, gv151, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape54) + gpt_neox_layers_13_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[159] + gpt_neox_layers_13_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[160] + gv152: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc94: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv152, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc90, gpt_neox_layers_13_post_attention_layernorm_weight1, gpt_neox_layers_13_post_attention_layernorm_bias1, alloc94) + 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((8192, 2048), dtype="float16") = packed_params[165] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[166] + gv153: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc95: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv153, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc94, gpt_neox_layers_13_mlp_dense_h_to_4h_weight1, gpt_neox_layers_13_mlp_dense_h_to_4h_bias1, alloc95) + R.vm.kill_object(alloc94) + 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((2048, 8192), dtype="float16") = packed_params[167] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[168] + gv154: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc96: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv154, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc95, gpt_neox_layers_13_mlp_dense_4h_to_h_weight1, gpt_neox_layers_13_mlp_dense_4h_to_h_bias1, alloc96) + R.vm.kill_object(alloc95) + 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((2048, 2048), dtype="float16") = packed_params[163] + gpt_neox_layers_13_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[164] + gv155: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc97: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv155, 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, alloc96, alloc90, alloc97) + R.vm.kill_object(alloc90) + R.vm.kill_object(reshape55) + R.vm.kill_object(alloc96) + 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((2048,), dtype="float16") = packed_params[169] + gpt_neox_layers_14_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[170] + gv156: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc98: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv156, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc97, gpt_neox_layers_14_input_layernorm_weight1, gpt_neox_layers_14_input_layernorm_bias1, alloc98) + 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((6144, 2048), dtype="float16") = packed_params[173] + gpt_neox_layers_14_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[174] + gv157: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc99: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv157, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc98, gpt_neox_layers_14_attention_query_key_value_weight1, gpt_neox_layers_14_attention_query_key_value_bias1, alloc99) + R.vm.kill_object(alloc98) + 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) + gv158: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape56: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc99, gv158, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc99) + gv159: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape57: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape56, gv159, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape56) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc100: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv160, R.dtype("float16")) + _100: 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, alloc100) + R.vm.kill_object(reshape57) + gv161: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape58: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc100, gv161, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc100) + gv162: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape59: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape58, gv162, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape58) + gpt_neox_layers_14_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[171] + gpt_neox_layers_14_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[172] + gv163: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc101: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv163, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc97, gpt_neox_layers_14_post_attention_layernorm_weight1, gpt_neox_layers_14_post_attention_layernorm_bias1, alloc101) + 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((8192, 2048), dtype="float16") = packed_params[177] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[178] + gv164: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc102: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv164, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc101, gpt_neox_layers_14_mlp_dense_h_to_4h_weight1, gpt_neox_layers_14_mlp_dense_h_to_4h_bias1, alloc102) + R.vm.kill_object(alloc101) + 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((2048, 8192), dtype="float16") = packed_params[179] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[180] + gv165: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc103: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv165, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc102, gpt_neox_layers_14_mlp_dense_4h_to_h_weight1, gpt_neox_layers_14_mlp_dense_4h_to_h_bias1, alloc103) + R.vm.kill_object(alloc102) + 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((2048, 2048), dtype="float16") = packed_params[175] + gpt_neox_layers_14_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[176] + gv166: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc104: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv166, R.dtype("float16")) + R.vm.kill_object(storage2) + 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, alloc103, alloc97, alloc104) + R.vm.kill_object(alloc97) + R.vm.kill_object(reshape59) + R.vm.kill_object(alloc103) + 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((2048,), dtype="float16") = packed_params[181] + gpt_neox_layers_15_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[182] + gv167: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc105: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv167, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc104, gpt_neox_layers_15_input_layernorm_weight1, gpt_neox_layers_15_input_layernorm_bias1, alloc105) + 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((6144, 2048), dtype="float16") = packed_params[185] + gpt_neox_layers_15_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[186] + gv168: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(6144), sinfo_args=(R.Shape(ndim=3),)) + alloc106: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv168, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc105, gpt_neox_layers_15_attention_query_key_value_weight1, gpt_neox_layers_15_attention_query_key_value_bias1, alloc106) + R.vm.kill_object(alloc105) + 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) + gv169: 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(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape60: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc106, gv169, sinfo_args=(R.Tensor((1, seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(alloc106) + gv170: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + reshape61: R.Tensor((seq_len, 24, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape60, gv170, sinfo_args=(R.Tensor((seq_len, 24, 256), dtype="float16"),)) + R.vm.kill_object(reshape60) + 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=3),)) + alloc107: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv171, R.dtype("float16")) + _107: 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, alloc107) + R.vm.kill_object(reshape61) + gv172: 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(8), R.prim_value(0), R.prim_value(256), sinfo_args=(R.Shape(ndim=4),)) + reshape62: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc107, gv172, sinfo_args=(R.Tensor((1, seq_len, 8, 256), dtype="float16"),)) + R.vm.kill_object(alloc107) + gv173: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + reshape63: R.Tensor((1, seq_len, 2048), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape62, gv173, sinfo_args=(R.Tensor((1, seq_len, 2048), dtype="float16"),)) + R.vm.kill_object(reshape62) + gpt_neox_layers_15_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[183] + gpt_neox_layers_15_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[184] + gv174: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc108: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv174, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc104, gpt_neox_layers_15_post_attention_layernorm_weight1, gpt_neox_layers_15_post_attention_layernorm_bias1, alloc108) + 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((8192, 2048), dtype="float16") = packed_params[189] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[190] + gv175: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(8192), sinfo_args=(R.Shape(ndim=3),)) + alloc109: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv175, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc108, gpt_neox_layers_15_mlp_dense_h_to_4h_weight1, gpt_neox_layers_15_mlp_dense_h_to_4h_bias1, alloc109) + R.vm.kill_object(alloc108) + 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((2048, 8192), dtype="float16") = packed_params[191] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[192] + gv176: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc110: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv176, R.dtype("float16")) + R.vm.kill_object(storage1) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc109, gpt_neox_layers_15_mlp_dense_4h_to_h_weight1, gpt_neox_layers_15_mlp_dense_4h_to_h_bias1, alloc110) + R.vm.kill_object(alloc109) + 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((2048, 2048), dtype="float16") = packed_params[187] + gpt_neox_layers_15_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[188] + gv177: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc111: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv177, R.dtype("float16")) + R.vm.kill_object(storage3) + 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, alloc110, alloc104, alloc111) + R.vm.kill_object(alloc104) + R.vm.kill_object(reshape63) + R.vm.kill_object(alloc110) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_15_attention_dense_bias1) + gpt_neox_final_layer_norm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[193] + gpt_neox_final_layer_norm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[194] + gv178: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2048), sinfo_args=(R.Shape(ndim=3),)) + alloc112: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage, R.prim_value(0), gv178, R.dtype("float16")) + R.vm.kill_object(storage) + cls.gpt_neox_q0f16_layer_norm1(alloc111, gpt_neox_final_layer_norm_weight1, gpt_neox_final_layer_norm_bias1, alloc112) + R.vm.kill_object(alloc111) + R.vm.kill_object(gpt_neox_final_layer_norm_weight1) + R.vm.kill_object(gpt_neox_final_layer_norm_bias1) + storage4: R.Object = R.vm.alloc_storage(R.shape([4096]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc113: R.Tensor((1, 1, 2048), dtype="float16") = R.vm.alloc_tensor(storage4, R.prim_value(0), R.shape([1, 1, 2048]), R.dtype("float16")) + R.vm.kill_object(storage4) + cls.gpt_neox_q0f16_index(alloc112, alloc113) + R.vm.kill_object(alloc112) + embed_out_weight1: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[195] + gv179: 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),)) + storage5: R.Object = R.vm.alloc_storage(gv179, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv180: R.Shape(ndim=3) = 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),)) + alloc114: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv180, R.dtype("float32")) + R.vm.kill_object(storage5) + cls.gpt_neox_q0f16_fused_NT_matmul14_cast8(alloc113, embed_out_weight1, alloc114) + R.vm.kill_object(alloc113) + R.vm.kill_object(embed_out_weight1) + gv1_1: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc114, paged_kv_cache + R.vm.kill_object(alloc114) + gv181: R.Tensor(dtype="float32", ndim=3) = gv1_1[0] + R.call_packed("vm.builtin.match_shape", gv181, 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_1 + + @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_func4(shape_heap) + gv556: 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, gv556, sinfo_args=(R.Tensor((batch_size, vocab_size), dtype="float32"),)) + gv557: 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),)) + storage23: R.Object = R.vm.alloc_storage(gv557, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv558: 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),)) + alloc459: R.Tensor(dtype="float32", ndim=2) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv558, R.dtype("float32")) + R.vm.kill_object(storage23) + gv559: 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),)) + storage24: R.Object = R.vm.alloc_storage(gv559, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv560: 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),)) + alloc460: R.Tensor(dtype="float32", ndim=2) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv560, R.dtype("float32")) + R.vm.kill_object(storage24) + cls.gpt_neox_q0f16_chunk_lse(lv, temperature, alloc459, alloc460) + lv1: R.Tuple(R.Tensor(dtype="float32", ndim=2), R.Tensor(dtype="float32", ndim=2)) = alloc459, alloc460 + gv561: 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),)) + storage25: R.Object = R.vm.alloc_storage(gv561, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv562: 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),)) + alloc461: R.Tensor(dtype="float32", ndim=2) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv562, R.dtype("float32")) + R.vm.kill_object(storage25) + cls.gpt_neox_q0f16_softmax_with_chunked_sum(lv, temperature, alloc459, alloc460, alloc461) + R.vm.kill_object(lv) + R.vm.kill_object(alloc459) + R.vm.kill_object(alloc460) + gv563: R.Shape(ndim=3) = R.call_packed("vm.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: R.Tensor((batch_size, 1, vocab_size), dtype="float32") = R.call_packed("vm.builtin.reshape", alloc461, gv563, sinfo_args=(R.Tensor((batch_size, 1, vocab_size), dtype="float32"),)) + R.vm.kill_object(alloc461) + return gv \ No newline at end of file