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 +1,8272 @@ +# 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, 12, 64), "float16") + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(pages_handle, (max_num_pages, 2, 12, 16, 64), "float16", offset_factor=1) + page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (B,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) + output = T.match_buffer(output_handle, (B, 12, 64), "float16") + lse = T.match_buffer(lse_handle, (B, 12)) + # with T.block("root"): + sm_scale: T.float32 = T.float32(0.18033688011112042) + for bx in T.thread_binding(B, thread="blockIdx.x"): + for fused_by_bz in T.thread_binding(12, thread="blockIdx.y"): + for ty in T.thread_binding(1, thread="threadIdx.y"): + for tx in T.thread_binding(16, thread="threadIdx.x"): + for tz in T.thread_binding(16, thread="threadIdx.z"): + with T.block("attn"): + T.reads(page_table_indptr[bx:bx + 2], length_info[bx], q_rope_position[bx], Q[bx, fused_by_bz // 12 + ty + fused_by_bz % 12, tx * 4 - 32:tx * 4 - 32 + 68]) + T.writes(output[bx, fused_by_bz % 12 + fused_by_bz // 12 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 12 + fused_by_bz // 12 + ty]) + Q_local = T.alloc_buffer((4,), "float16", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + O_allreduce = T.alloc_buffer((16, 1, 64), scope="shared") + md_allreduce = T.alloc_buffer((16, 1, 2), scope="shared") + S_reduce_local = T.alloc_buffer((1,), scope="local") + t0 = T.alloc_buffer((1,), scope="local") + S_local = T.alloc_buffer((1,), scope="local") + QK_local = T.alloc_buffer((4,), scope="local") + V_local = T.alloc_buffer((4,), "float16", scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_prev = T.alloc_buffer((1,), scope="local") + other_m = T.alloc_buffer((1,), scope="local") + other_d = T.alloc_buffer((1,), scope="local") + exp_mprev = T.alloc_buffer((1,), scope="local") + exp_otherm = T.alloc_buffer((1,), scope="local") + other_o = T.alloc_buffer((4,), scope="local") + st_m = T.alloc_buffer((1,), scope="local") + st_d = T.alloc_buffer((1,), scope="local") + O_local = T.alloc_buffer((4,), scope="local") + by: T.int32 = fused_by_bz % 12 + bz: T.int32 = fused_by_bz // 12 + batch_idx: T.int32 = bx + cur_page_indptr_begin: T.int32 = page_table_indptr[batch_idx] + cur_page_indptr_end: T.int32 = page_table_indptr[batch_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[batch_idx], 0) + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for vec in T.vectorized(4): + freq = T.float32() + Q_local[vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", Q[bx, by + bz + ty, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, Q[bx, by + bz + ty, tx * 4 + vec + 32] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) + for iterator in range((kv_chunk_len[0] + 15) // 16): + tile_start_s: T.int32 = tz + ty + tile_start_g: T.int32 = iterator * 16 + tz + ty + for j in range(1): + with T.block("KV_load"): + T.reads() + T.writes() + row_g: T.int32 = tile_start_g + j + if row_g < kv_chunk_len[0]: + seq_offset: T.int32 = row_g + page_no: T.int32 = page_table_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + for vec in T.vectorized(4): + freq = T.float32() + K_smem[tile_start_s + j, tx * 4 + vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, pages[page_no, 0, by, page_offset, tx * 4 + vec + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[batch_idx] + row_g) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, tx * 4 + vec]) + V_smem[tile_start_s + j, tx * 4 + vec] = pages[page_no, 1, by, page_offset, tx * 4 + vec] + else: + for vec in T.vectorized(4): + K_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + V_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + T.tvm_storage_sync("shared") + m_prev[0] = st_m[0] + for j in range(1): + for vec in T.vectorized(4): + QK_local[vec] = T.Cast("float32", Q_local[vec]) * T.Cast("float32", K_smem[tz + j, tx * 4 + vec]) * attn_score_scaling_factor * sm_scale + S_reduce_local[0] = T.float32(0.0) + for vec in T.unroll(4): + S_reduce_local[0] = S_reduce_local[0] + QK_local[vec] + with T.block("block_cross_thread"): + T.reads(S_reduce_local[0]) + T.writes(t0[0]) + T.attr(T.comm_reducer(lambda x0, y0: x0 + y0, [T.float32(0.0)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))) + T.tvm_thread_allreduce(T.uint32(1), S_reduce_local[0], T.bool(True), t0[0], tx) + S_local[j] = T.float32(-50000.0) + if iterator * 16 + tz + j < kv_chunk_len[0]: + S_local[j] = t0[0] + st_m[0] = T.max(st_m[0], S_local[j]) + o_scale: T.float32 = T.exp2(m_prev[0] - st_m[0]) + st_d[0] = st_d[0] * o_scale + for j in range(1): + S_local[j] = T.exp2(S_local[j] - st_m[0]) + st_d[0] = st_d[0] + S_local[j] + for j in T.vectorized(4): + O_local[j] = O_local[j] * o_scale + for j in range(1): + for vec in T.vectorized(4): + V_local[vec] = V_smem[tz + j, tx * 4 + vec] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] + T.Cast("float32", V_local[vec]) * S_local[j] + for vec in T.vectorized(4): + O_allreduce[tz, ty, tx * 4 + vec] = O_local[vec] + md_allreduce[tz, ty, 0] = st_m[0] + md_allreduce[tz, ty, 1] = st_d[0] + T.tvm_storage_sync("shared") + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for j in range(16): + m_prev[0] = st_m[0] + d_prev[0] = st_d[0] + other_m[0] = md_allreduce[j, ty, 0] + other_d[0] = md_allreduce[j, ty, 1] + for vec in T.vectorized(4): + other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] + st_m[0] = T.max(st_m[0], other_m[0]) + st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) + exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) + exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] / st_d[0] + for vec in T.vectorized(4): + output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) + lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) + + @T.prim_func + def gpt_neox_q0f16_batch_decode_paged_kv_sliding_window(_0: T.int32, Q_handle: T.handle, pages_handle: T.handle, page_table_indptr_handle: T.handle, page_table_values_handle: T.handle, var_length_info: T.handle, k_rope_pos_offset_handle: T.handle, q_rope_position_handle: T.handle, output_handle: T.handle, lse_handle: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + B = T.int32(is_size_var=True) + Q = T.match_buffer(Q_handle, (B, 12, 64), "float16") + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(pages_handle, (max_num_pages, 2, 12, 16, 64), "float16", offset_factor=1) + page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (3, B), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) + output = T.match_buffer(output_handle, (B, 12, 64), "float16") + lse = T.match_buffer(lse_handle, (B, 12)) + # with T.block("root"): + sm_scale: T.float32 = T.float32(0.18033688011112042) + for bx in T.thread_binding(B, thread="blockIdx.x"): + for fused_by_bz in T.thread_binding(12, thread="blockIdx.y"): + for ty in T.thread_binding(1, thread="threadIdx.y"): + for tx in T.thread_binding(16, thread="threadIdx.x"): + for tz in T.thread_binding(16, thread="threadIdx.z"): + with T.block("attn"): + T.reads(page_table_indptr[bx:bx + 2], length_info[0:3, bx], q_rope_position[bx], Q[bx, fused_by_bz // 12 + ty + fused_by_bz % 12, tx * 4 - 32:tx * 4 - 32 + 68]) + T.writes(output[bx, fused_by_bz % 12 + fused_by_bz // 12 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 12 + fused_by_bz // 12 + ty]) + Q_local = T.alloc_buffer((4,), "float16", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + O_allreduce = T.alloc_buffer((16, 1, 64), scope="shared") + md_allreduce = T.alloc_buffer((16, 1, 2), scope="shared") + S_reduce_local = T.alloc_buffer((1,), scope="local") + t0 = T.alloc_buffer((1,), scope="local") + S_local = T.alloc_buffer((1,), scope="local") + QK_local = T.alloc_buffer((4,), scope="local") + V_local = T.alloc_buffer((4,), "float16", scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_prev = T.alloc_buffer((1,), scope="local") + other_m = T.alloc_buffer((1,), scope="local") + other_d = T.alloc_buffer((1,), scope="local") + exp_mprev = T.alloc_buffer((1,), scope="local") + exp_otherm = T.alloc_buffer((1,), scope="local") + other_o = T.alloc_buffer((4,), scope="local") + st_m = T.alloc_buffer((1,), scope="local") + st_d = T.alloc_buffer((1,), scope="local") + O_local = T.alloc_buffer((4,), scope="local") + by: T.int32 = fused_by_bz % 12 + bz: T.int32 = fused_by_bz // 12 + batch_idx: T.int32 = bx + cur_page_indptr_begin: T.int32 = page_table_indptr[batch_idx] + cur_page_indptr_end: T.int32 = page_table_indptr[batch_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[0, batch_idx] - length_info[1, batch_idx] + length_info[2, batch_idx], 0) + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for vec in T.vectorized(4): + freq = T.float32() + Q_local[vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", Q[bx, by + bz + ty, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, Q[bx, by + bz + ty, tx * 4 + vec + 32] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) + for iterator in range((kv_chunk_len[0] + 15) // 16): + tile_start_s: T.int32 = tz + ty + tile_start_g: T.int32 = iterator * 16 + tz + ty + for j in range(1): + with T.block("KV_load"): + T.reads() + T.writes() + row_g: T.int32 = tile_start_g + j + if row_g < kv_chunk_len[0]: + seq_offset: T.int32 = T.if_then_else(row_g < length_info[2, batch_idx], row_g, row_g - length_info[2, batch_idx] + length_info[1, batch_idx]) + page_no: T.int32 = page_table_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + for vec in T.vectorized(4): + freq = T.float32() + K_smem[tile_start_s + j, tx * 4 + vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, pages[page_no, 0, by, page_offset, tx * 4 + vec + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[batch_idx] + row_g) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, tx * 4 + vec]) + V_smem[tile_start_s + j, tx * 4 + vec] = pages[page_no, 1, by, page_offset, tx * 4 + vec] + else: + for vec in T.vectorized(4): + K_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + V_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + T.tvm_storage_sync("shared") + m_prev[0] = st_m[0] + for j in range(1): + for vec in T.vectorized(4): + QK_local[vec] = T.Cast("float32", Q_local[vec]) * T.Cast("float32", K_smem[tz + j, tx * 4 + vec]) * attn_score_scaling_factor * sm_scale + S_reduce_local[0] = T.float32(0.0) + for vec in T.unroll(4): + S_reduce_local[0] = S_reduce_local[0] + QK_local[vec] + with T.block("block_cross_thread"): + T.reads(S_reduce_local[0]) + T.writes(t0[0]) + T.attr(T.comm_reducer(lambda x0, y0: x0 + y0, [T.float32(0.0)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))) + T.tvm_thread_allreduce(T.uint32(1), S_reduce_local[0], T.bool(True), t0[0], tx) + S_local[j] = T.float32(-50000.0) + if iterator * 16 + tz + j < kv_chunk_len[0]: + S_local[j] = t0[0] + st_m[0] = T.max(st_m[0], S_local[j]) + o_scale: T.float32 = T.exp2(m_prev[0] - st_m[0]) + st_d[0] = st_d[0] * o_scale + for j in range(1): + S_local[j] = T.exp2(S_local[j] - st_m[0]) + st_d[0] = st_d[0] + S_local[j] + for j in T.vectorized(4): + O_local[j] = O_local[j] * o_scale + for j in range(1): + for vec in T.vectorized(4): + V_local[vec] = V_smem[tz + j, tx * 4 + vec] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] + T.Cast("float32", V_local[vec]) * S_local[j] + for vec in T.vectorized(4): + O_allreduce[tz, ty, tx * 4 + vec] = O_local[vec] + md_allreduce[tz, ty, 0] = st_m[0] + md_allreduce[tz, ty, 1] = st_d[0] + T.tvm_storage_sync("shared") + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for j in range(16): + m_prev[0] = st_m[0] + d_prev[0] = st_d[0] + other_m[0] = md_allreduce[j, ty, 0] + other_d[0] = md_allreduce[j, ty, 1] + for vec in T.vectorized(4): + other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] + st_m[0] = T.max(st_m[0], other_m[0]) + st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) + exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) + exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] / st_d[0] + for vec in T.vectorized(4): + output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) + lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) + + @T.prim_func + def gpt_neox_q0f16_batch_prefill_paged_kv(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + total_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (total_len, 12, 64), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 12, 16, 64), "float16", offset_factor=1) + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 12, 64), "float16") + lse = T.match_buffer(var_lse, (total_len, 12)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(12, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 16), scope="shared") + S_local = T.alloc_buffer((32, 16), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(4): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) + j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 15) // 16): + L_kv_start: T.int32 = iterator_1 * 16 + for lz_ly_fused_0 in range(2): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + freq = T.float32() + K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, pages[page_no, 0, by, page_offset, j + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, j]) + else: + K_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_ly_fused_0 in range(2): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:16, 0:64]) + T.writes(S_local[0:32, 0:16]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + k = T.axis.reduce(64, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 2): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(16): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) + T.writes(S_smem[row, 0:16]) + for j in range(16): + if row < 32: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(16): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:16], V_smem[0:16, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(2, 8, 4, 4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + k = T.axis.reduce(16, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def gpt_neox_q0f16_batch_prefill_paged_kv_sliding_window(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + total_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (total_len, 12, 64), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 12, 16, 64), "float16", offset_factor=1) + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (3, batch_size), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 12, 64), "float16") + lse = T.match_buffer(var_lse, (total_len, 12)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(12, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 16), scope="shared") + S_local = T.alloc_buffer((32, 16), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[0, b_idx] - length_info[1, b_idx] + length_info[2, b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(4): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) + j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 15) // 16): + L_kv_start: T.int32 = iterator_1 * 16 + for lz_ly_fused_0 in range(2): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = T.if_then_else(cur_L < length_info[2, b_idx], cur_L, cur_L - length_info[2, b_idx] + length_info[1, b_idx]) + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + freq = T.float32() + K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, pages[page_no, 0, by, page_offset, j + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, j]) + else: + K_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_ly_fused_0 in range(2): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = T.if_then_else(cur_L < length_info[2, b_idx], cur_L, cur_L - length_info[2, b_idx] + length_info[1, b_idx]) + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:16, 0:64]) + T.writes(S_local[0:32, 0:16]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + k = T.axis.reduce(64, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 2): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(16): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) + T.writes(S_smem[row, 0:16]) + for j in range(16): + if row < 32: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(16): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:16], V_smem[0:16, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(2, 8, 4, 4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + k = T.axis.reduce(16, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def gpt_neox_q0f16_batch_prefill_ragged_kv(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_k_rope_pos_offset: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + qo_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (qo_len, 12, 64), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + kv_len = T.int32(is_size_var=True) + k = T.match_buffer(var_k, (kv_len, 12, 64), "float16") + v = T.match_buffer(var_v, (kv_len, 12, 64), "float16") + kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (qo_len, 12, 64), "float16") + lse = T.match_buffer(var_lse, (qo_len, 12)) + # with T.block("root"): + for lbx in T.thread_binding(8, thread="blockIdx.x"): + for lby in T.thread_binding(12, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((64, 32), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 32), scope="shared") + S_local = T.alloc_buffer((32, 32), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + q_indptr_val: T.int32 = q_indptr[b_idx] + LH_start: T.int32 = tile_id[0] * 32 + kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1_0 * 4 + lj_1_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_1_lj_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_1_lj_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_0, lj_0_0 in T.grid(1, 2): + for lj_1 in T.vectorized(8): + with T.block("Q_load"): + i = T.axis.spatial(32, li_0 * 32 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) // 4) + j = T.axis.spatial(64, lj_0_0 * 32 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) % 4 * 8 + lj_1) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 31) // 32): + L_kv_start: T.int32 = iterator_1 * 32 + L_kv_base: T.int32 = kv_indptr[b_idx] + for lz_1_ly_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for lz_1_ly_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lz_0, ly_0_0 in T.grid(1, 2): + for ly_1 in T.vectorized(8): + with T.block("K_load"): + i = T.axis.spatial(32, lz_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 4) + j = T.axis.spatial(64, ly_0_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 4 * 8 + ly_1) + T.reads(kv_chunk_len[0], k_rope_pos_offset[b_idx], k[L_kv_base + L_kv_start + i, by, j - 32:j - 32 + 65]) + T.writes(K_smem[j, i]) + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + freq = T.float32() + K_smem[j, i] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", k[L_kv_base + cur_L, by, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, k[L_kv_base + cur_L, by, j + 32] * T.float16(-1.0), k[L_kv_base + cur_L, by, j - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), k[L_kv_base + cur_L, by, j]) + else: + K_smem[j, i] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_1_ly_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for lz_1_ly_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lz_0, ly_0_0 in T.grid(1, 2): + for ly_1 in T.vectorized(8): + with T.block("V_load"): + i = T.axis.spatial(32, lz_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 4) + j = T.axis.spatial(64, ly_0_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 4 * 8 + ly_1) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + V_smem[i, j] = v[L_kv_base + cur_L, by, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:64, 0:32]) + T.writes(S_local[0:32, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init in T.unroll(2): + for lj_1_0_init in T.unroll(1): + for lj_1_1_init in T.vectorized(4): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_0_init * 4 + lj_1_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0 in range(4): + for li_1 in T.unroll(2): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + for lk_1 in range(16): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1_0 * 4 + lj_1_1) + k_1 = T.axis.reduce(64, lk_0 * 16 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k_1], K_smem[k_1, j]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k_1]) * T.Cast("float32", K_smem[k_1, j]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(2): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1_0 * 4 + lj_1_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 32: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init in T.unroll(4): + for lj_1_0_init in T.unroll(1): + for lj_1_1_init in T.vectorized(4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_0_init * 4 + lj_1_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1 in T.grid(2, 16): + for li_1 in T.unroll(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1_0 * 4 + lj_1_1) + k_1 = T.axis.reduce(32, lk_0 * 16 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k_1], V_smem[k_1, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k_1] * T.Cast("float32", V_smem[k_1, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1_0 * 4 + lj_1_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 8 + + @T.prim_func + def gpt_neox_q0f16_batch_tree_attn(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_mn_indptr: T.handle, var_mask: T.handle, var_output: T.handle, var_lse: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32, batch_size: T.int32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + qo_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (qo_len, 12, 64), "float16") + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + kv_len = T.int32(is_size_var=True) + k = T.match_buffer(var_k, (kv_len, 12, 64), "float16") + v = T.match_buffer(var_v, (kv_len, 12, 64), "float16") + kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) + mn_indptr = T.match_buffer(var_mn_indptr, (batch_size + 1,), "int32", offset_factor=1) + tree_size = T.int32(is_size_var=True) + mask = T.match_buffer(var_mask, (tree_size, 2), "int32", offset_factor=1) + output = T.match_buffer(var_output, (qo_len, 12, 64), "float16") + lse = T.match_buffer(var_lse, (qo_len, 12)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(12, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 32), scope="shared") + S_local = T.alloc_buffer((32, 32), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(4): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) + j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 31) // 32): + L_kv_start: T.int32 = iterator_1 * 32 + L_kv_base: T.int32 = kv_indptr[b_idx] + for lz_ly_fused_0 in range(4): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("KV_load"): + i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_base + L_kv_start + i + if L_kv_start + i < kv_chunk_len[0]: + freq = T.float32() + K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", k[cur_L, by, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, k[cur_L, by, j + 32] * T.float16(-1.0), k[cur_L, by, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), k[cur_L, by, j]) + V_smem[i, j] = v[cur_L, by, j] + else: + K_smem[i, j] = T.float16(0.0) + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:32, 0:64]) + T.writes(S_local[0:32, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 4): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 4, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + k_1 = T.axis.reduce(64, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k_1], K_smem[j, k_1]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k_1]) * T.Cast("float32", K_smem[j, k_1]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 4): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 32: + row_: T.int32 = LH_start + row + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(4, 8, 4, 4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + k_1 = T.axis.reduce(32, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k_1], V_smem[k_1, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k_1] * T.Cast("float32", V_smem[k_1, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def gpt_neox_q0f16_chunk_lse(var_A: T.handle, var_temperature: T.handle, var_chunked_sum: T.handle, var_chunked_max: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size, vocab_size = T.int32(is_size_var=True), T.int32(is_size_var=True) + A = T.match_buffer(var_A, (batch_size, vocab_size)) + temperature = T.match_buffer(var_temperature, (batch_size,)) + num_chunks = T.int32(is_size_var=True) + chunked_sum = T.match_buffer(var_chunked_sum, (batch_size, num_chunks)) + chunked_max = T.match_buffer(var_chunked_max, (batch_size, num_chunks)) + # with T.block("root"): + temp_max_shared = T.alloc_buffer((batch_size, num_chunks), scope="shared") + temp_sum_shared = T.alloc_buffer((batch_size, num_chunks), scope="shared") + for ax0_ax1_fused in T.thread_binding(batch_size * num_chunks, thread="blockIdx.x"): + for ax0, ax1 in T.grid(1, 1): + for ax2_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + for ax2_fused_0 in T.serial(16, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("max"): + v0 = T.axis.spatial(batch_size, ax0_ax1_fused % (num_chunks * batch_size) // num_chunks + ax0) + v1 = T.axis.spatial(num_chunks, ax0_ax1_fused % num_chunks + ax1) + v2 = T.axis.reduce(4096, ax2_fused_0 * 256 + ax2_fused_1) + T.reads(temperature[v0], A[v0, v1 * 4096 + v2]) + T.writes(temp_max_shared[v0, v1]) + with T.init(): + temp_max_shared[v0, v1] = T.float32(-340282346638528859811704183484516925440.0) + temp_max_shared[v0, v1] = T.max(temp_max_shared[v0, v1], T.if_then_else(v1 * 4096 + v2 < vocab_size, T.if_then_else(temperature[v0] > T.float32(1.0000000000000001e-05), A[v0, v1 * 4096 + v2] / temperature[v0], A[v0, v1 * 4096 + v2]), T.float32(-340282346638528859811704183484516925440.0))) + for ax0, ax1 in T.grid(1, 1): + for ax2_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + for ax2_fused_0 in T.serial(16, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("sum_exp"): + v0 = T.axis.spatial(batch_size, ax0_ax1_fused % (num_chunks * batch_size) // num_chunks + ax0) + v1 = T.axis.spatial(num_chunks, ax0_ax1_fused % num_chunks + ax1) + v2 = T.axis.reduce(4096, ax2_fused_0 * 256 + ax2_fused_1) + T.reads(temperature[v0], A[v0, v1 * 4096 + v2], temp_max_shared[v0, v1]) + T.writes(temp_sum_shared[v0, v1]) + with T.init(): + temp_sum_shared[v0, v1] = T.float32(0.0) + temp_sum_shared[v0, v1] = temp_sum_shared[v0, v1] + T.if_then_else(v1 * 4096 + v2 < vocab_size, T.Select(temperature[v0] > T.float32(1.0000000000000001e-05), T.exp(T.if_then_else(v1 * 4096 + v2 < vocab_size, T.if_then_else(temperature[v0] > T.float32(1.0000000000000001e-05), A[v0, v1 * 4096 + v2] / temperature[v0], A[v0, v1 * 4096 + v2]), T.float32(-340282346638528859811704183484516925440.0)) - temp_max_shared[v0, v1]), T.Cast("float32", T.if_then_else(v1 * 4096 + v2 < vocab_size, T.if_then_else(temperature[v0] > T.float32(1.0000000000000001e-05), A[v0, v1 * 4096 + v2] / temperature[v0], A[v0, v1 * 4096 + v2]), T.float32(-340282346638528859811704183484516925440.0)) == temp_max_shared[v0, v1])), T.float32(0.0)) + for ax2_1 in T.thread_binding(256, thread="threadIdx.x"): + for ax2_0 in T.serial(1, annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("log"): + v0 = T.axis.spatial(batch_size, ax0_ax1_fused % (num_chunks * batch_size) // num_chunks) + v1 = T.axis.spatial(num_chunks, ax0_ax1_fused % num_chunks) + v2 = T.axis.spatial(1, ax2_0 * 256 + ax2_1) + T.where(ax2_0 * 256 + ax2_1 < 1) + T.reads(temperature[v0], temp_sum_shared[v0, v1], temp_max_shared[v0, v1]) + T.writes(chunked_sum[v0, v1], chunked_max[v0, v1]) + chunked_sum[v0, v1] = T.Select(temperature[v0] > T.float32(1.0000000000000001e-05), T.log(temp_sum_shared[v0, v1]), temp_sum_shared[v0, v1]) + chunked_max[v0, v1] = temp_max_shared[v0, v1] + + @T.prim_func + def gpt_neox_q0f16_compact_kv_copy(var_pages: T.handle, var_copy_length_indptr: T.handle, var_copy_src_dst_pos: T.handle, batch_size: T.int32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + num_pages = T.int32() + pages = T.match_buffer(var_pages, (num_pages, 2, 12, 16, 64), "float16", offset_factor=1) + copy_length_indptr = T.match_buffer(var_copy_length_indptr, (batch_size + 1,), "int32", offset_factor=1) + total_copy_length = T.int32() + copy_src_dst_pos = T.match_buffer(var_copy_src_dst_pos, (2, total_copy_length), "int32", offset_factor=1) + with T.block("root"): + T.reads() + T.writes() + for bhd_o in T.thread_binding(batch_size * 3, thread="blockIdx.x"): + for bhd_i in T.thread_binding(256, thread="threadIdx.x"): + b: T.int32 = (bhd_o * 256 + bhd_i) // 768 + h: T.int32 = (bhd_o * 256 + bhd_i) // 64 % 12 + d: T.int32 = (bhd_o * 256 + bhd_i) % 64 + if bhd_o * 256 + bhd_i < batch_size * 12 * 64: + for i in range(copy_length_indptr[b + 1] - copy_length_indptr[b]): + src_pos: T.int32 = copy_src_dst_pos[0, copy_length_indptr[b] + i] + dst_pos: T.int32 = copy_src_dst_pos[1, copy_length_indptr[b] + i] + pages[dst_pos // 16, 0, h, dst_pos % 16, d] = pages[src_pos // 16, 0, h, src_pos % 16, d] + pages[dst_pos // 16, 1, h, dst_pos % 16, d] = pages[src_pos // 16, 1, h, src_pos % 16, d] + + @T.prim_func + def gpt_neox_q0f16_copy_single_page(var_pages: T.handle, src_page_id: T.int32, tgt_page_id: T.int32, copy_length: T.int32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + num_pages, page_size = T.int32(), T.int32() + pages = T.match_buffer(var_pages, (num_pages, 2, 12, page_size, 64), "float16", offset_factor=1) + # with T.block("root"): + for b in T.thread_binding(copy_length * 3, thread="blockIdx.x"): + for t in T.thread_binding(256, thread="threadIdx.x"): + with T.block("copy"): + vh = T.axis.spatial(12, (b * 256 + t) // (copy_length * 64)) + vp = T.axis.spatial(copy_length, (b * 256 + t) % (copy_length * 64) // 64) + vd = T.axis.spatial(64, (b * 256 + t) % 64) + T.where(b * 256 + t < copy_length * 12 * 64) + T.reads(pages[src_page_id, 0:2, vh, vp, vd]) + T.writes(pages[tgt_page_id, 0:2, vh, vp, vd]) + pages[tgt_page_id, 0, vh, vp, vd] = pages[src_page_id, 0, vh, vp, vd] + pages[tgt_page_id, 1, vh, vp, vd] = pages[src_page_id, 1, vh, vp, vd] + + @T.prim_func + def gpt_neox_q0f16_fused_NT_matmul10_add10(layer_norm25: T.Buffer((1, 1, 768), "float16"), gpt_neox_layers_0_attention_query_key_value_weight2: T.Buffer((2304, 768), "float16"), gpt_neox_layers_0_attention_query_key_value_bias2: T.Buffer((2304,), "float16"), T_add_intermediate: T.Buffer((1, 1, 2304), "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, 2304), "float16", scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((256, 1, 1, 2304), "float16", scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((32, 1, 1, 2304), "float16", scope="local") + gpt_neox_layers_0_attention_query_key_value_weight2_local = T.alloc_buffer((2304, 768), "float16", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(1152, 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(2304, 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(3, 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(2304, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(768, 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(2304, 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_norm25[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_norm25[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(2304, 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(2304, 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(2304, 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(2304, 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(reshape51: T.Buffer((1, 1, 768), "float16"), gpt_neox_layers_0_attention_dense_weight2: T.Buffer((768, 768), "float16"), gpt_neox_layers_0_attention_dense_bias2: T.Buffer((768,), "float16"), astype26: T.Buffer((1, 1, 768), "float16"), input_embed: T.Buffer((1, 1, 768), "float16"), T_add_intermediate_1_2: T.Buffer((1, 1, 768), "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, 768), "float16", scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((256, 1, 1, 768), "float16", scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((32, 1, 1, 768), "float16", scope="local") + gpt_neox_layers_0_attention_dense_weight2_local = T.alloc_buffer((768, 768), "float16", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(384, 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(768, 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(3, 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(768, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(768, 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(768, 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], reshape51[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] + reshape51[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(768, 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(768, 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(768, 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(768, u_fused_ax0_fused_fused_0 * 2 + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) + T.reads(astype26[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] = astype26[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_norm26: T.Buffer((1, 1, 768), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_weight2: T.Buffer((3072, 768), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: T.Buffer((3072,), "float32"), compute_intermediate: T.Buffer((1, 1, 3072), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + NT_matmul_intermediate_local = T.alloc_buffer((1, 1, 3072), scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((256, 1, 1, 3072), scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((32, 1, 1, 3072), scope="local") + gpt_neox_layers_0_mlp_dense_h_to_4h_weight2_local = T.alloc_buffer((3072, 768), "float16", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(1536, thread="blockIdx.x"): + for u_fused_ax0_fused_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for u_fused_ax0_fused_fused_2_init in range(1): + for ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused = T.axis.spatial(256, ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + ax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1_init) + v0 = T.axis.spatial(3072, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0]) + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0] = T.float32(0.0) + for ax1_fused_u_fused_0 in T.serial(3, 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(3072, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(768, 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(3072, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2) + vax1_fused_u_fused_2, vax1_fused_u_fused_0 = T.axis.remap("RR", [ax1_fused_u_fused_2, ax1_fused_u_fused_0]) + T.reads(NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused, 0, 0, v0], layer_norm26[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_norm26[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(3072, u_fused_ax0_fused_fused_0 * 2 + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] = T.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(3072, u_fused_ax0_fused_fused_0 * 2 + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) + T.reads(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0], NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1, 0, 0, v0]) + T.writes(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] = NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0] + NT_matmul_intermediate_rf_local[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 * 8 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_1, 0, 0, v0] + for ax1_fused_2 in range(1): + for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial(3072, u_fused_ax0_fused_fused_0 * 2 + ax1_fused_0_ax1_fused_1_fused + ax1_fused_2) + T.reads(NT_matmul_intermediate_rf_local_1[vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused_0, 0, 0, v0]) + T.writes(NT_matmul_intermediate_local[0, 0, v0]) + with T.init(): + NT_matmul_intermediate_local[0, 0, v0] = T.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(3072, u_fused_ax0_fused_fused_0 * 2 + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) + T.reads(NT_matmul_intermediate_local[0, 0, v0], gpt_neox_layers_0_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(astype25: T.Buffer((1, 1, 3072), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_weight2: T.Buffer((768, 3072), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: T.Buffer((768,), "float32"), compute_intermediate: T.Buffer((1, 1, 768), "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, 768), scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((256, 1, 1, 768), scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((32, 1, 1, 768), scope="local") + gpt_neox_layers_0_mlp_dense_4h_to_h_weight2_local = T.alloc_buffer((768, 3072), "float16", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(384, 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(768, 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(12, 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(768, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(3072, 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(768, 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], astype25[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", astype25[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(768, 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(768, 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(768, 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(768, 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_norm49: T.Buffer((1, 1, 768), "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, 768), "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, 768), "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(3, 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(768, 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_norm49[0, 0, vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused], 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_norm49[0, 0, vax1_fused_u_fused_2 * 8 + vax1_fused_u_fused_0 * 256 + vax1_fused_u_fused_1_ax1_fused_u_fused_3_fused] * 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_reshape147: T.handle, gpt_neox_layers_0_attention_dense_weight4: T.Buffer((768, 768), "float16"), gpt_neox_layers_0_attention_dense_bias4: T.Buffer((768,), "float16"), p_astype76: 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() + reshape147 = T.match_buffer(p_reshape147, (batch_size, 1, 768), "float16") + astype76 = T.match_buffer(p_astype76, (batch_size, 1, 768), "float16") + input_embeds = T.match_buffer(p_input_embeds, (batch_size, 1, 768), "float16") + T_add_intermediate_1_2 = T.match_buffer(p_output0, (batch_size, 1, 768), "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, 768), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 1) // 2 * 2, 1, 768), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 1) // 2 * 2, 1, 768), "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(192, 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(768, 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(3, 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(768, 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], reshape147[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, reshape147[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(768, 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(768, 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(768, 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(768, 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(astype76[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] = astype76[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, 768), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 3) // 4 * 4, 1, 768), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 3) // 4 * 4, 1, 768), "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(192, 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(768, 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(3, 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(768, 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], reshape147[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, reshape147[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(768, 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(768, 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(768, 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(768, 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(astype76[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] = astype76[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, 768), "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(3, 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(768, 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(48, 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(768, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(768, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], reshape147[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, reshape147[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(768, 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(astype76[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] = astype76[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_norm76: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_weight4: T.Buffer((3072, 768), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: T.Buffer((3072,), "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_norm76 = T.match_buffer(p_layer_norm76, (batch_size, 1, 768), "float16") + compute_intermediate = T.match_buffer(p_output0, (batch_size, 1, 3072), "float16") + # with T.block("root"): + if T.tvm_thread_invariant(batch_size <= 2): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + 1) // 2 * 2, 1, 3072), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 1) // 2 * 2, 1, 3072), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 1) // 2 * 2, 1, 3072), scope="local") + for ax0_0 in T.thread_binding((batch_size + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(768, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(2, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax0_1_init) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = T.float32(0.0) + for ax2_fused_0 in T.serial(3, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(2, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax0_1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1], layer_norm76[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_norm76[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(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = T.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(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1] + for ax2_fused_2, ax1 in T.grid(2, 2): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((batch_size + 1) // 2 * 2, ax0_0 * 2 + ax1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_local[v0, 0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[v0, 0, v1] = T.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(3072, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (batch_size + 1) // 2 < 0 or ax0_0 == 0) and ax0_0 * 2 + ax0 < batch_size) + T.reads(NT_matmul_intermediate_pad_local[v0, 0, v1], gpt_neox_layers_0_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, 3072), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 3) // 4 * 4, 1, 3072), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 3) // 4 * 4, 1, 3072), scope="local") + for ax0_0 in T.thread_binding((batch_size + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(768, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(4, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax0_1_init) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1] = T.float32(0.0) + for ax2_fused_0 in T.serial(3, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(4, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax0_1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, 0, v1], layer_norm76[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_norm76[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(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = T.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(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, v0, 0, v1] + for ax2_fused_2, ax1 in T.grid(2, 4): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((batch_size + 3) // 4 * 4, ax0_0 * 4 + ax1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, 0, v1]) + T.writes(NT_matmul_intermediate_pad_local[v0, 0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[v0, 0, v1] = T.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(3072, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (batch_size + 3) // 4 < 0 or ax0_0 == 0) and ax0_0 * 4 + ax0 < batch_size) + T.reads(NT_matmul_intermediate_pad_local[v0, 0, v1], gpt_neox_layers_0_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, 3072), scope="local") + for ax0_ax1_0_fused in T.thread_binding((batch_size + 7) // 8, thread="blockIdx.y"): + for ax2_0 in T.thread_binding(12, thread="blockIdx.x"): + for ax1_1 in T.thread_binding(1, thread="vthread.y"): + for ax2_1 in T.thread_binding(1, thread="vthread.x"): + for ax1_2 in T.thread_binding(4, thread="threadIdx.y"): + for ax2_2 in T.thread_binding(32, thread="threadIdx.x", annotations={"pragma_auto_unroll_max_step": 16, "pragma_unroll_explicit": 1}): + for ax1_3_init, ax2_3_0_init in T.grid(2, 1): + for ax2_3_1_init in T.vectorized(8): + with T.block("NT_matmul_init"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3_init) + v2 = T.axis.spatial(3072, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0_init * 8 + ax2_3_1_init) + T.reads() + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = T.float32(0.0) + for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(48, 16, 2, 1): + for ax2_3_1 in T.vectorized(8): + with T.block("NT_matmul_update"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((batch_size + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3) + v2 = T.axis.spatial(3072, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(768, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm76[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_norm76[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(3072, ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1) + T.where(ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1 < batch_size) + T.reads(NT_matmul_intermediate_reindex_pad_local[v0, v1, v2], gpt_neox_layers_0_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_astype75: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_weight4: T.Buffer((768, 3072), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: T.Buffer((768,), "float32"), p_output0: T.handle): + T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + astype75 = T.match_buffer(p_astype75, (batch_size, 1, 3072), "float16") + compute_intermediate = T.match_buffer(p_output0, (batch_size, 1, 768), "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, 768), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 1) // 2 * 2, 1, 768), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 1) // 2 * 2, 1, 768), scope="local") + for ax0_0 in T.thread_binding((batch_size + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(192, 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(768, 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(12, 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(768, 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], astype75[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, astype75[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(768, 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(768, 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(768, 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(768, 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, 768), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 3) // 4 * 4, 1, 768), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 3) // 4 * 4, 1, 768), scope="local") + for ax0_0 in T.thread_binding((batch_size + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(192, 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(768, 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(12, 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(768, 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], astype75[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, astype75[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(768, 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(768, 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(768, 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(768, 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, 768), 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(3, 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(768, 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(192, 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(768, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(3072, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], astype75[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, astype75[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(768, 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_norm99: 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_norm99 = T.match_buffer(p_layer_norm99, (batch_size, 1, 768), "float16") + vocab_size = T.int32() + embed_out_weight4 = T.match_buffer(p_embed_out_weight4, (vocab_size, 768), "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(48, 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(768, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm99[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_norm99[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_norm50: T.handle, gpt_neox_layers_0_attention_query_key_value_weight3: T.Buffer((2304, 768), "float16"), gpt_neox_layers_0_attention_query_key_value_bias3: T.Buffer((2304,), "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_norm50 = T.match_buffer(p_layer_norm50, (1, seq_len, 768), "float16") + T_add_intermediate = T.match_buffer(p_output0, (1, seq_len, 2304), "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, 2304), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 1) // 2 * 2, 2304), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 1) // 2 * 2, 2304), "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(576, 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(2304, 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(3, 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(2304, 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_norm50[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_norm50[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(2304, 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(2304, 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(2304, 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(2304, 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, 2304), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 3) // 4 * 4, 2304), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 3) // 4 * 4, 2304), "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(576, 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(2304, 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(3, 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(2304, 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_norm50[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_norm50[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(2304, 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(2304, 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(2304, 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(2304, 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, 2304), "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(9, 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(2304, 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(48, 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(2304, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(768, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm50[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_norm50[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(2304, 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_reshape99: T.handle, gpt_neox_layers_0_attention_dense_weight3: T.Buffer((768, 768), "float16"), gpt_neox_layers_0_attention_dense_bias3: T.Buffer((768,), "float16"), p_astype51: 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() + reshape99 = T.match_buffer(p_reshape99, (1, seq_len, 768), "float16") + astype51 = T.match_buffer(p_astype51, (1, seq_len, 768), "float16") + input_embeds = T.match_buffer(p_input_embeds, (1, seq_len, 768), "float16") + T_add_intermediate_1_2 = T.match_buffer(p_output0, (1, seq_len, 768), "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, 768), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 1) // 2 * 2, 768), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 1) // 2 * 2, 768), "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(192, 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(768, 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(3, 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(768, 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], reshape99[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, reshape99[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(768, 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(768, 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(768, 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(768, 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(astype51[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] = astype51[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, 768), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 3) // 4 * 4, 768), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 3) // 4 * 4, 768), "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(192, 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(768, 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(3, 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(768, 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], reshape99[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, reshape99[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(768, 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(768, 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(768, 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(768, 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(astype51[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] = astype51[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, 768), "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(3, 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(768, 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(48, 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(768, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(768, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], reshape99[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, reshape99[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(768, 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(astype51[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] = astype51[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_norm51: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_weight3: T.Buffer((3072, 768), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: T.Buffer((3072,), "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_norm51 = T.match_buffer(p_layer_norm51, (1, seq_len, 768), "float16") + compute_intermediate = T.match_buffer(p_output0, (1, seq_len, 3072), "float16") + # with T.block("root"): + if T.tvm_thread_invariant(seq_len <= 2): + with T.block("root"): + T.reads() + T.writes() + NT_matmul_intermediate_pad_local = T.alloc_buffer((1, (seq_len + 1) // 2 * 2, 3072), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 1) // 2 * 2, 3072), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 1) // 2 * 2, 3072), scope="local") + for ax0_0 in T.thread_binding((seq_len + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(768, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(2, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax0_1_init) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = T.float32(0.0) + for ax2_fused_0 in T.serial(3, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(2, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax0_1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1], layer_norm51[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_norm51[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(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = T.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(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1] + for ax2_fused_2, ax1 in T.grid(2, 2): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((seq_len + 1) // 2 * 2, ax0_0 * 2 + ax1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_local[0, v0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[0, v0, v1] = T.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(3072, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (seq_len + 1) // 2 < 0 or ax0_0 == 0) and ax0_0 * 2 + ax0 < seq_len) + T.reads(NT_matmul_intermediate_pad_local[0, v0, v1], gpt_neox_layers_0_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, 3072), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 3) // 4 * 4, 3072), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 3) // 4 * 4, 3072), scope="local") + for ax0_0 in T.thread_binding((seq_len + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(768, thread="blockIdx.x"): + for ax1_fused_1 in T.thread_binding(2, thread="threadIdx.x"): + for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(32, thread="threadIdx.y"): + for ax0_1_init, ax1_fused_2_init in T.grid(4, 2): + for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(8): + with T.block("NT_matmul_rf_init"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1_init) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax0_1_init) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2_init) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1] = T.float32(0.0) + for ax2_fused_0 in T.serial(3, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): + for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(4, 2, 1): + for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(8): + with T.block("NT_matmul_rf_update"): + vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(256, ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax2_fused_1_ax2_fused_3_fused_1) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax0_1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax1_fused_2) + vax2_fused_2, vax2_fused_0 = T.axis.remap("RR", [ax2_fused_2, ax2_fused_0]) + T.reads(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, 0, v0, v1], layer_norm51[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_norm51[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(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads() + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = T.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(3072, ax1_fused_0 * 4 + ax3_fused_0_ax3_fused_1_fused * 2 + ax3_fused_2_0 * 2 + ax3_fused_2_1) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * 8 + vax2_fused_1_ax2_fused_3_fused_1, 0, v0, v1] + for ax2_fused_2, ax1 in T.grid(2, 4): + for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(2, thread="threadIdx.x"): + for ax0 in T.thread_binding(32, thread="threadIdx.y"): + with T.block("NT_matmul"): + vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(32, ax0) + v0 = T.axis.spatial((seq_len + 3) // 4 * 4, ax0_0 * 4 + ax1) + v1 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax2_fused_0_ax2_fused_1_fused * 2 + ax2_fused_2) + T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, 0, v0, v1]) + T.writes(NT_matmul_intermediate_pad_local[0, v0, v1]) + with T.init(): + NT_matmul_intermediate_pad_local[0, v0, v1] = T.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(3072, ax1_fused_0 * 4 + ax1_fused_0_ax1_fused_1_fused * 2 + ax1_fused_2) + T.where((ax0_0 - (seq_len + 3) // 4 < 0 or ax0_0 == 0) and ax0_0 * 4 + ax0 < seq_len) + T.reads(NT_matmul_intermediate_pad_local[0, v0, v1], gpt_neox_layers_0_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, 3072), scope="local") + for ax0_ax1_0_fused in T.thread_binding((seq_len + 7) // 8, thread="blockIdx.y"): + for ax2_0 in T.thread_binding(12, thread="blockIdx.x"): + for ax1_1 in T.thread_binding(1, thread="vthread.y"): + for ax2_1 in T.thread_binding(1, thread="vthread.x"): + for ax1_2 in T.thread_binding(4, thread="threadIdx.y"): + for ax2_2 in T.thread_binding(32, thread="threadIdx.x", annotations={"pragma_auto_unroll_max_step": 16, "pragma_unroll_explicit": 1}): + for ax1_3_init, ax2_3_0_init in T.grid(2, 1): + for ax2_3_1_init in T.vectorized(8): + with T.block("NT_matmul_init"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3_init) + v2 = T.axis.spatial(3072, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0_init * 8 + ax2_3_1_init) + T.reads() + T.writes(NT_matmul_intermediate_reindex_pad_local[0, v1, v2]) + NT_matmul_intermediate_reindex_pad_local[0, v1, v2] = T.float32(0.0) + for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(48, 16, 2, 1): + for ax2_3_1 in T.vectorized(8): + with T.block("NT_matmul_update"): + v0 = T.axis.spatial(1, 0) + v1 = T.axis.spatial((seq_len + 7) // 8 * 8, ax0_ax1_0_fused * 8 + ax1_1 * 8 + ax1_2 * 2 + ax1_3) + v2 = T.axis.spatial(3072, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(768, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm51[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_norm51[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(3072, ax2_0 * 256 + ax2_2 * 8 + ax2_0_1 * 8 + ax2_1_1) + T.where(ax0_ax1_0_fused * 8 + ax1_2 * 2 + ax1 < seq_len) + T.reads(NT_matmul_intermediate_reindex_pad_local[v0, v1, v2], gpt_neox_layers_0_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_astype50: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_weight3: T.Buffer((768, 3072), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: T.Buffer((768,), "float32"), p_output0: T.handle): + T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + astype50 = T.match_buffer(p_astype50, (1, seq_len, 3072), "float16") + compute_intermediate = T.match_buffer(p_output0, (1, seq_len, 768), "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, 768), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 1) // 2 * 2, 768), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 1) // 2 * 2, 768), scope="local") + for ax0_0 in T.thread_binding((seq_len + 1) // 2, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(192, 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(768, 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(12, 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(768, 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], astype50[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, astype50[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(768, 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(768, 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(768, 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(768, 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, 768), scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, 1, (seq_len + 3) // 4 * 4, 768), scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, 1, (seq_len + 3) // 4 * 4, 768), scope="local") + for ax0_0 in T.thread_binding((seq_len + 3) // 4, thread="blockIdx.y"): + for ax1_fused_0 in T.thread_binding(192, 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(768, 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(12, 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(768, 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], astype50[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, astype50[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(768, 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(768, 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(768, 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(768, 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, 768), 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(3, 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(768, 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(192, 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(768, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(3072, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], astype50[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, astype50[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(768, 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, 768), "float16") + vocab_size = T.int32() + embed_out_weight3 = T.match_buffer(p_embed_out_weight3, (vocab_size, 768), "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(48, 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(768, 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_norm75: T.handle, gpt_neox_layers_0_attention_query_key_value_weight4: T.Buffer((2304, 768), "float16"), gpt_neox_layers_0_attention_query_key_value_bias4: T.Buffer((2304,), "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_norm75 = T.match_buffer(p_layer_norm75, (batch_size, 1, 768), "float16") + T_add_intermediate = T.match_buffer(p_output0, (batch_size, 1, 2304), "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, 2304), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 1) // 2 * 2, 1, 2304), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 1) // 2 * 2, 1, 2304), "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(576, 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(2304, 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(3, 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(2304, 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_norm75[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_norm75[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(2304, 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(2304, 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(2304, 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(2304, 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, 2304), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((256, (batch_size + 3) // 4 * 4, 1, 2304), "float16", scope="local") + NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((32, (batch_size + 3) // 4 * 4, 1, 2304), "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(576, 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(2304, 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(3, 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(2304, 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_norm75[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_norm75[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(2304, 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(2304, 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(2304, 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(2304, 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, 2304), "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(9, 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(2304, 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(48, 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(2304, ax2_0 * 256 + ax2_1 * 256 + ax2_2 * 8 + ax2_3_0 * 8 + ax2_3_1) + v3 = T.axis.reduce(768, ax3_0 * 16 + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[0, v1, v2], layer_norm75[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_norm75[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(2304, 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(lv13: T.Buffer((1, 12, 64), "float16"), T_reshape_intermediate_1: T.Buffer((1, 1, 768), "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(3, 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(768, ax0_fused_0 * 256 + ax0_fused_1) + T.reads(lv13[0, v0 // 64, v0 % 64]) + T.writes(T_reshape_intermediate_1[0, 0, v0]) + T_reshape_intermediate_1[0, 0, v0] = lv13[0, v0 // 64, v0 % 64] + + @T.prim_func + def gpt_neox_q0f16_fused_reshape8_reshape9(add72: T.Buffer((1, 1, 2304), "float16"), T_reshape_intermediate_1: T.Buffer((1, 36, 64), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(9, 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(36, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) // 64) + v1 = T.axis.spatial(64, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 64) + T.reads(add72[0, 0, v0 * 64 + v1]) + T.writes(T_reshape_intermediate_1[0, v0, v1]) + T_reshape_intermediate_1[0, v0, v1] = add72[0, 0, v0 * 64 + v1] + + @T.prim_func + def gpt_neox_q0f16_fused_rope(var_qkv: T.handle, var_position_map: T.handle, var_q: T.handle, var_k: T.handle, var_v: T.handle, apply_rope: T.int32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + qkv = T.match_buffer(var_qkv, (seq_len, 36, 64), "float16") + position_map = T.match_buffer(var_position_map, (seq_len,), "int32", offset_factor=1) + q = T.match_buffer(var_q, (seq_len, 12, 64), "float16") + k = T.match_buffer(var_k, (seq_len, 12, 64), "float16") + v = T.match_buffer(var_v, (seq_len, 12, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 9, 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) // 2304) + v1 = T.axis.spatial(36, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 2304 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(position_map[v0], qkv[v0, v1, v2 + -8:v2 + -8 + 17]) + T.writes(q[v0, v1, v2], k[v0, v1 + -12, v2], v[v0, v1 + -24, v2]) + if v1 < 12: + freq = T.float32() + q[v0, v1, v2] = T.if_then_else(apply_rope > 0 and v2 < 16, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", qkv[v0, v1, v2]) + T.sin(freq) * T.Cast("float32", T.if_then_else(v2 < 8, qkv[v0, v1, v2 + 8] * T.float16(-1.0), qkv[v0, v1, v2 + -8]))), where={freq: T.Cast("float32", position_map[v0]) / T.pow(T.float32(10000.0), T.Cast("float32", v2 * 2 % 16) / T.float32(16.0))}), qkv[v0, v1, v2]) + else: + if v1 < 24: + freq = T.float32() + k[v0, v1 + -12, v2] = T.if_then_else(apply_rope > 0 and v2 < 16, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", qkv[v0, v1, v2]) + T.sin(freq) * T.Cast("float32", T.if_then_else(v2 < 8, qkv[v0, v1, v2 + 8] * T.float16(-1.0), qkv[v0, v1, v2 + -8]))), where={freq: T.Cast("float32", position_map[v0]) / T.pow(T.float32(10000.0), T.Cast("float32", v2 * 2 % 16) / T.float32(16.0))}), qkv[v0, v1, v2]) + else: + v[v0, v1 + -24, 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_norm24: T.handle, index: T.Buffer((1, 1, 768), "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_norm24 = T.match_buffer(var_layer_norm24, (1, seq_len, 768), "float16") + # with T.block("root"): + for ax0_fused_0 in T.thread_binding(3, thread="blockIdx.x"): + for ax0_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("index"): + v0 = T.axis.spatial(768, ax0_fused_0 * 256 + ax0_fused_1) + T.reads(layer_norm24[0, seq_len - 1, v0]) + T.writes(index[0, 0, v0]) + index[0, 0, v0] = layer_norm24[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((768,), "float16"), gpt_neox_layers_0_input_layernorm_bias4: T.Buffer((768,), "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, 768), "float16") + T_layer_norm = T.match_buffer(var_T_layer_norm, (batch_size, 1, 768), "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(3, 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(768, 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(3, 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(768, 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(0.0013020833333333333)) * T.rsqrt(input_embeds_red_temp_v1_shared[v0, 0] * T.float32(0.0013020833333333333) - input_embeds_red_temp_v0_shared[v0, 0] * T.float32(0.0013020833333333333) * (input_embeds_red_temp_v0_shared[v0, 0] * T.float32(0.0013020833333333333)) + 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((768,), "float16"), gpt_neox_layers_0_input_layernorm_bias3: T.Buffer((768,), "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, 768), "float16") + T_layer_norm = T.match_buffer(var_T_layer_norm, (1, seq_len, 768), "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(3, 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(768, 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(3, 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(768, 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(0.0013020833333333333)) * T.rsqrt(input_embeds_red_temp_v1_shared[0, v0] * T.float32(0.0013020833333333333) - input_embeds_red_temp_v0_shared[0, v0] * T.float32(0.0013020833333333333) * (input_embeds_red_temp_v0_shared[0, v0] * T.float32(0.0013020833333333333)) + 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, 768), "float16"), gpt_neox_layers_0_input_layernorm_weight2: T.Buffer((768,), "float16"), gpt_neox_layers_0_input_layernorm_bias2: T.Buffer((768,), "float16"), T_layer_norm: T.Buffer((1, 1, 768), "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(3, 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(768, 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(3, 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(768, 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(0.0013020833333333333)) * T.rsqrt(input_embed_red_temp_v1_shared[0, 0] * T.float32(0.0013020833333333333) - input_embed_red_temp_v0_shared[0, 0] * T.float32(0.0013020833333333333) * (input_embed_red_temp_v0_shared[0, 0] * T.float32(0.0013020833333333333)) + T.float32(1.0000000000000001e-05))) * gpt_neox_layers_0_input_layernorm_weight2[v1] + gpt_neox_layers_0_input_layernorm_bias2[v1] + + @T.prim_func + def gpt_neox_q0f16_merge_state_inplace(v: T.handle, s: T.handle, v_other: T.handle, s_other: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + N, H, D = T.int32(is_size_var=True), T.int32(is_size_var=True), T.int32(is_size_var=True) + V = T.match_buffer(v, (N, H, D), "float16") + S = T.match_buffer(s, (N, H)) + V_other = T.match_buffer(v_other, (N, H, D), "float16") + S_other = T.match_buffer(s_other, (N, H)) + # with T.block("root"): + for bx in T.thread_binding(N, thread="blockIdx.x"): + for by in T.thread_binding(1, thread="blockIdx.y"): + for ty in T.thread_binding(12, thread="threadIdx.y"): + for tx in T.thread_binding(16, thread="threadIdx.x"): + with T.block("merge"): + T.reads(S[bx, ty + by * 12], S_other[bx, ty + by * 12], V[bx, ty + by * 12, tx * 4:tx * 4 + 4], V_other[bx, ty + by * 12, tx * 4:tx * 4 + 4]) + T.writes(V[bx, ty + by * 12, tx * 4:tx * 4 + 4], S[bx, ty + by * 12]) + 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 * 12] + s_other_val[0] = S_other[bx, ty + by * 12] + 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 * 12, tx * 4 + vec] + for vec in T.vectorized(4): + v_other_vec[vec] = V_other[bx, ty + by * 12, 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 * 12, tx * 4 + vec] = v_vec[vec] + S[bx, ty + by * 12] = T.log2(s_val[0] + s_other_val[0]) + s_max[0] + + @T.prim_func + def gpt_neox_q0f16_reshape(var_add216: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + add216 = T.match_buffer(var_add216, (batch_size, 1, 2304), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, 1, 36, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * 9, 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) // 2304) + v1 = T.axis.spatial(36, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 2304 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(add216[v0, 0, v1 * 64 + v2]) + T.writes(T_reshape[v0, 0, v1, v2]) + T_reshape[v0, 0, v1, v2] = add216[v0, 0, v1 * 64 + v2] + + @T.prim_func + def gpt_neox_q0f16_reshape1(var_reshape144: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + reshape144 = T.match_buffer(var_reshape144, (batch_size, 1, 36, 64), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, 36, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * 9, 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) // 2304) + v1 = T.axis.spatial(36, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 2304 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(reshape144[v0, 0, v1, v2]) + T.writes(T_reshape[v0, v1, v2]) + T_reshape[v0, v1, v2] = reshape144[v0, 0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_reshape2(var_lv37: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + lv37 = T.match_buffer(var_lv37, (batch_size, 12, 64), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, 1, 12, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * 3, 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) // 768) + v1 = T.axis.spatial(12, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 768 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(lv37[v0, v1, v2]) + T.writes(T_reshape[v0, 0, v1, v2]) + T_reshape[v0, 0, v1, v2] = lv37[v0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_reshape3(var_reshape146: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int32() + reshape146 = T.match_buffer(var_reshape146, (batch_size, 1, 12, 64), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, 1, 768), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(batch_size * 3, 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) // 768) + v1 = T.axis.spatial(768, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 768) + T.reads(reshape146[v0, 0, v1 // 64, v1 % 64]) + T.writes(T_reshape[v0, 0, v1]) + T_reshape[v0, 0, v1] = reshape146[v0, 0, v1 // 64, v1 % 64] + + @T.prim_func + def gpt_neox_q0f16_reshape4(var_add144: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + add144 = T.match_buffer(var_add144, (1, seq_len, 2304), "float16") + T_reshape = T.match_buffer(var_T_reshape, (1, seq_len, 36, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 9, 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) // 2304) + v1 = T.axis.spatial(36, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 2304 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(add144[0, v0, v1 * 64 + v2]) + T.writes(T_reshape[0, v0, v1, v2]) + T_reshape[0, v0, v1, v2] = add144[0, v0, v1 * 64 + v2] + + @T.prim_func + def gpt_neox_q0f16_reshape5(var_reshape96: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + reshape96 = T.match_buffer(var_reshape96, (1, seq_len, 36, 64), "float16") + T_reshape = T.match_buffer(var_T_reshape, (seq_len, 36, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 9, 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) // 2304) + v1 = T.axis.spatial(36, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 2304 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(reshape96[0, v0, v1, v2]) + T.writes(T_reshape[v0, v1, v2]) + T_reshape[v0, v1, v2] = reshape96[0, v0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_reshape6(var_lv25: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + lv25 = T.match_buffer(var_lv25, (seq_len, 12, 64), "float16") + T_reshape = T.match_buffer(var_T_reshape, (1, seq_len, 12, 64), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 3, 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) // 768) + v1 = T.axis.spatial(12, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 768 // 64) + v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) + T.reads(lv25[v0, v1, v2]) + T.writes(T_reshape[0, v0, v1, v2]) + T_reshape[0, v0, v1, v2] = lv25[v0, v1, v2] + + @T.prim_func + def gpt_neox_q0f16_reshape7(var_reshape98: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int32() + reshape98 = T.match_buffer(var_reshape98, (1, seq_len, 12, 64), "float16") + T_reshape = T.match_buffer(var_T_reshape, (1, seq_len, 768), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(seq_len * 3, 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) // 768) + v1 = T.axis.spatial(768, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 768) + T.reads(reshape98[0, v0, v1 // 64, v1 % 64]) + T.writes(T_reshape[0, v0, v1]) + T_reshape[0, v0, v1] = reshape98[0, v0, v1 // 64, v1 % 64] + + @T.prim_func + def gpt_neox_q0f16_scatter_probs(var_src: T.handle, var_indices: T.handle, var_dst: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size, n = T.int32(is_size_var=True), T.int32(is_size_var=True) + src = T.match_buffer(var_src, (batch_size, n)) + indices = T.match_buffer(var_indices, (batch_size,), "int32") + m = T.int32(is_size_var=True) + dst = T.match_buffer(var_dst, (m, n)) + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding((batch_size * n + 255) // 256, thread="blockIdx.x"): + for ax0_ax1_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("scatter_2d"): + v0 = T.axis.spatial(batch_size, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % (n * batch_size) // n) + v1 = T.axis.spatial(n, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % n) + T.where(ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1 < batch_size * n) + T.reads(src[v0, v1], indices[v0]) + T.writes(dst[indices[v0], v1]) + dst[indices[v0], v1] = src[v0, v1] + + @T.prim_func + def gpt_neox_q0f16_shape_func(H: T.Buffer((T.int64(5),), "int64")): + T.func_attr({"tir.is_host_func": 1}) + H[T.int64(4)] = T.int64(128) * H[T.int64(2)] * 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(3)] = (H[T.int64(1)] + T.int64(4096) - T.int64(1)) // T.int64(4096) + H[T.int64(2)] = T.int64(512) * ((H[T.int64(1)] + T.int64(4096) - T.int64(1)) // T.int64(4096)) + H[T.int64(4)] = T.int64(512) * H[T.int64(1)] + + @T.prim_func + def gpt_neox_q0f16_shape_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(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_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_norm74: 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_norm74 = T.match_buffer(var_layer_norm74, (1, seq_len, 768), "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, 768), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(batch_size * 3, 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) // 768) + v1 = T.axis.spatial(768, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 768) + T.reads(layer_norm74[0, logit_positions[v0], v1], logit_positions[v0]) + T.writes(T_take[0, v0, v1]) + T_take[0, v0, v1] = layer_norm74[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, 768), "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, 768), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(seq_len * 3, 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) // 768) + v1 = T.axis.spatial(768, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 768) + 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, 12, page_size, 64), "float16", offset_factor=1) + seqlen = T.int32(is_size_var=True) + position_map = T.match_buffer(var_position_map, (seqlen,), "int32", offset_factor=1) + k_data = T.match_buffer(var_k_data, (12, seqlen, 12, 64), "float16") + v_data = T.match_buffer(var_v_data, (12, seqlen, 12, 64), "float16") + # with T.block("root"): + for p_h_d_fused_0 in T.thread_binding(seqlen * 3, 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) // 768) + vh = T.axis.spatial(12, (p_h_d_fused_0 * 256 + p_h_d_fused_1) % 768 // 64) + vd = T.axis.spatial(64, (p_h_d_fused_0 * 256 + p_h_d_fused_1) % 64) + T.reads(position_map[vp], pages[position_map[vp] // page_size, 0:2, vh, position_map[vp] % page_size, vd]) + T.writes(k_data[layer_id, vp, vh, vd], v_data[layer_id, vp, vh, vd]) + position: T.int32 = position_map[vp] + k_data[layer_id, vp, vh, vd] = pages[position // page_size, 0, vh, position % page_size, vd] + v_data[layer_id, vp, vh, vd] = pages[position // page_size, 1, vh, position % page_size, vd] + + @T.prim_func + def gpt_neox_q0f16_tir_kv_cache_transpose_append(var_pages: T.handle, var_k_data: T.handle, var_v_data: T.handle, var_position_map: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + num_pages = T.int32() + pages = T.match_buffer(var_pages, (num_pages, 2, 12, 16, 64), "float16", offset_factor=1) + ntoken = T.int32(is_size_var=True) + k_data = T.match_buffer(var_k_data, (ntoken, 12, 64), "float16") + v_data = T.match_buffer(var_v_data, (ntoken, 12, 64), "float16") + position_map = T.match_buffer(var_position_map, (ntoken,), "int32", offset_factor=1) + # with T.block("root"): + for global_pos_h_f_fused_0 in T.thread_binding(ntoken * 3, 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) // 768] != -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) // 768) + vh = T.axis.spatial(12, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) % 768 // 64) + vf = T.axis.spatial(64, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) % 64) + T.reads(position_map[vgpos], k_data[vgpos, vh, vf]) + T.writes(pages[position_map[vgpos] // 16, 0, vh, position_map[vgpos] % 16, vf]) + position: T.int32 = position_map[vgpos] + pages[position // 16, 0, vh, position % 16, vf] = k_data[vgpos, vh, vf] + with T.block("v_transpose_append"): + vgpos = T.axis.spatial(ntoken, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) // 768) + vh = T.axis.spatial(12, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) % 768 // 64) + vf = T.axis.spatial(64, (global_pos_h_f_fused_0 * 256 + global_pos_h_f_fused_1) % 64) + T.reads(position_map[vgpos], v_data[vgpos, vh, vf]) + T.writes(pages[position_map[vgpos] // 16, 1, vh, position_map[vgpos] % 16, vf]) + position: T.int32 = position_map[vgpos] + pages[position // 16, 1, vh, position % 16, vf] = v_data[vgpos, vh, vf] + + @T.prim_func + def gpt_neox_q0f16_tree_attn_paged_kv(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32, tree_order_indptr_handle: T.handle, tree_order_handle: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + total_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (total_len, 12, 64), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 12, 32, 64), "float16") + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 12, 64), "float16") + lse = T.match_buffer(var_lse, (total_len, 12)) + 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(12, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 32), scope="shared") + S_local = T.alloc_buffer((32, 32), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 32 + length_info[b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(4): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) + j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 31) // 32): + L_kv_start: T.int32 = iterator_1 * 32 + for lz_ly_fused_0 in range(4): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 32] + page_offset: T.int32 = seq_offset % 32 + K_smem[i, j] = pages[page_no, 0, by, page_offset, j] + else: + K_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_ly_fused_0 in range(4): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 32] + page_offset: T.int32 = seq_offset % 32 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:32, 0:64]) + T.writes(S_local[0:32, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 4): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 4, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + k = T.axis.reduce(64, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 4): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 32: + row_: T.int32 = LH_start + row + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(4, 8, 4, 4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + k = T.axis.reduce(32, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @R.function + def _metadata() -> R.Object: + shape_heap: R.Object = R.null_value() + return R.str("{\"model_type\": \"gpt_neox\", \"quantization\": \"q0f16\", \"context_window_size\": 2048, \"sliding_window_size\": -1, \"attention_sink_size\": -1, \"prefill_chunk_size\": 2048, \"tensor_parallel_shards\": 1, \"pipeline_parallel_stages\": 1, \"disaggregation\": false, \"kv_state_kind\": \"kv_cache\", \"max_batch_size\": 128, \"params\": [{\"name\": \"gpt_neox.embed_in.weight\", \"shape\": [\"vocab_size\", 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.input_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.input_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.post_attention_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.post_attention_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.query_key_value.weight\", \"shape\": [2304, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.query_key_value.bias\", \"shape\": [2304], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.dense.weight\", \"shape\": [768, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.dense.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_h_to_4h.weight\", \"shape\": [3072, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_h_to_4h.bias\", \"shape\": [3072], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_4h_to_h.weight\", \"shape\": [768, 3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_4h_to_h.bias\", \"shape\": [768], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.input_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.input_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.post_attention_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.post_attention_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.query_key_value.weight\", \"shape\": [2304, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.query_key_value.bias\", \"shape\": [2304], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.dense.weight\", \"shape\": [768, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.dense.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_h_to_4h.weight\", \"shape\": [3072, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_h_to_4h.bias\", \"shape\": [3072], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_4h_to_h.weight\", \"shape\": [768, 3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_4h_to_h.bias\", \"shape\": [768], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.input_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.input_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.post_attention_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.post_attention_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.query_key_value.weight\", \"shape\": [2304, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.query_key_value.bias\", \"shape\": [2304], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.dense.weight\", \"shape\": [768, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.dense.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_h_to_4h.weight\", \"shape\": [3072, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_h_to_4h.bias\", \"shape\": [3072], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_4h_to_h.weight\", \"shape\": [768, 3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_4h_to_h.bias\", \"shape\": [768], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.input_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.input_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.post_attention_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.post_attention_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.query_key_value.weight\", \"shape\": [2304, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.query_key_value.bias\", \"shape\": [2304], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.dense.weight\", \"shape\": [768, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.dense.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_h_to_4h.weight\", \"shape\": [3072, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_h_to_4h.bias\", \"shape\": [3072], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_4h_to_h.weight\", \"shape\": [768, 3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_4h_to_h.bias\", \"shape\": [768], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.input_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.input_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.post_attention_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.post_attention_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.query_key_value.weight\", \"shape\": [2304, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.query_key_value.bias\", \"shape\": [2304], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.dense.weight\", \"shape\": [768, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.dense.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_h_to_4h.weight\", \"shape\": [3072, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_h_to_4h.bias\", \"shape\": [3072], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_4h_to_h.weight\", \"shape\": [768, 3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_4h_to_h.bias\", \"shape\": [768], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.input_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.input_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.post_attention_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.post_attention_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.query_key_value.weight\", \"shape\": [2304, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.query_key_value.bias\", \"shape\": [2304], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.dense.weight\", \"shape\": [768, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.dense.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_h_to_4h.weight\", \"shape\": [3072, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_h_to_4h.bias\", \"shape\": [3072], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_4h_to_h.weight\", \"shape\": [768, 3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_4h_to_h.bias\", \"shape\": [768], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.input_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.input_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.post_attention_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.post_attention_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.query_key_value.weight\", \"shape\": [2304, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.query_key_value.bias\", \"shape\": [2304], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.dense.weight\", \"shape\": [768, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.dense.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_h_to_4h.weight\", \"shape\": [3072, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_h_to_4h.bias\", \"shape\": [3072], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_4h_to_h.weight\", \"shape\": [768, 3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_4h_to_h.bias\", \"shape\": [768], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.input_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.input_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.post_attention_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.post_attention_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.query_key_value.weight\", \"shape\": [2304, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.query_key_value.bias\", \"shape\": [2304], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.dense.weight\", \"shape\": [768, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.dense.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_h_to_4h.weight\", \"shape\": [3072, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_h_to_4h.bias\", \"shape\": [3072], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_4h_to_h.weight\", \"shape\": [768, 3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_4h_to_h.bias\", \"shape\": [768], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.input_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.input_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.post_attention_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.post_attention_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.query_key_value.weight\", \"shape\": [2304, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.query_key_value.bias\", \"shape\": [2304], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.dense.weight\", \"shape\": [768, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.dense.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_h_to_4h.weight\", \"shape\": [3072, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_h_to_4h.bias\", \"shape\": [3072], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_4h_to_h.weight\", \"shape\": [768, 3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_4h_to_h.bias\", \"shape\": [768], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.input_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.input_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.post_attention_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.post_attention_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.query_key_value.weight\", \"shape\": [2304, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.query_key_value.bias\", \"shape\": [2304], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.dense.weight\", \"shape\": [768, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.dense.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_h_to_4h.weight\", \"shape\": [3072, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_h_to_4h.bias\", \"shape\": [3072], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_4h_to_h.weight\", \"shape\": [768, 3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_4h_to_h.bias\", \"shape\": [768], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.input_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.input_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.post_attention_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.post_attention_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.query_key_value.weight\", \"shape\": [2304, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.query_key_value.bias\", \"shape\": [2304], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.dense.weight\", \"shape\": [768, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.dense.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_h_to_4h.weight\", \"shape\": [3072, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_h_to_4h.bias\", \"shape\": [3072], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_4h_to_h.weight\", \"shape\": [768, 3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_4h_to_h.bias\", \"shape\": [768], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.input_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.input_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.post_attention_layernorm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.post_attention_layernorm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.query_key_value.weight\", \"shape\": [2304, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.query_key_value.bias\", \"shape\": [2304], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.dense.weight\", \"shape\": [768, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.dense.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_h_to_4h.weight\", \"shape\": [3072, 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_h_to_4h.bias\", \"shape\": [3072], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_4h_to_h.weight\", \"shape\": [768, 3072], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_4h_to_h.bias\", \"shape\": [768], \"dtype\": \"float32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.final_layer_norm.weight\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.final_layer_norm.bias\", \"shape\": [768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"embed_out.weight\", \"shape\": [\"vocab_size\", 768], \"dtype\": \"float16\", \"preprocs\": [], \"pipeline_stages\": [0]}], \"kv_cache\": {\"num_hidden_layers\": 12, \"num_attention_heads\": 12, \"num_key_value_heads\": 12, \"head_dim\": 64}, \"memory_usage\": {\"alloc_embedding_tensor\": 3145728, \"batch_decode\": 2359296, \"batch_prefill\": 37945344, \"batch_verify\": 37748736, \"create_tir_paged_kv_cache\": 0, \"decode\": 18432, \"embed\": 3145728, \"prefill\": 37750272, \"softmax_with_temperature\": 0}}") + + @R.function + def alloc_embedding_tensor() -> R.Tensor((2048, 768), 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([3145728]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv: R.Tensor((2048, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([2048, 768]), R.dtype("float16")) + R.vm.kill_object(storage) + return gv + + @R.function + def batch_decode(input_embeds: R.Tensor(("batch_size", 1, 768), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor(("vocab_size", 768), 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, 768), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(148), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv431: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv431, 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, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv432: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[147] + R.call_packed("vm.builtin.check_tensor_info", gv432, 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, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), 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(768), R.str("ErrorContext(fn=batch_decode, loc=param[0], param=input_embeds, annotation=R.Tensor((batch_size, 1, 768), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv431, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(768), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv432, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(768), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func5(shape_heap) + gpt_neox_layers_0_input_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[2] + storage26: R.Object = R.vm.alloc_storage(R.shape([196608]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv433: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc350: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv433, 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, alloc350) + 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((2304, 768), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[6] + storage27: R.Object = R.vm.alloc_storage(R.shape([589824]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv434: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc351: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv434, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc350, gpt_neox_layers_0_attention_query_key_value_weight4, gpt_neox_layers_0_attention_query_key_value_bias4, alloc351) + R.vm.kill_object(alloc350) + 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) + gv435: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape144: R.Tensor((batch_size, 1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc351, gv435, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc351) + gv436: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape145: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape144, gv436, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape144) + gv437: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc352: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv437, R.dtype("float16")) + _351: 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)), reshape145, alloc352) + R.vm.kill_object(reshape145) + gv438: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape146: R.Tensor((batch_size, 1, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc352, gv438, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc352) + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape147: R.Tensor((batch_size, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape146, gv439, sinfo_args=(R.Tensor((batch_size, 1, 768), dtype="float16"),)) + R.vm.kill_object(reshape146) + gpt_neox_layers_0_post_attention_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[4] + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc353: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv440, 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, alloc353) + 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((3072, 768), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[10] + storage28: R.Object = R.vm.alloc_storage(R.shape([786432]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv441: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc354: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv441, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc353, gpt_neox_layers_0_mlp_dense_h_to_4h_weight4, gpt_neox_layers_0_mlp_dense_h_to_4h_bias4, alloc354) + R.vm.kill_object(alloc353) + 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((768, 3072), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[12] + gv442: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc355: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv442, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc354, gpt_neox_layers_0_mlp_dense_4h_to_h_weight4, gpt_neox_layers_0_mlp_dense_4h_to_h_bias4, alloc355) + R.vm.kill_object(alloc354) + 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((768, 768), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[8] + gv443: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc356: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv443, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape147, gpt_neox_layers_0_attention_dense_weight4, gpt_neox_layers_0_attention_dense_bias4, alloc355, input_embeds, alloc356) + R.vm.kill_object(reshape147) + R.vm.kill_object(alloc355) + 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((768,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[14] + gv444: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc357: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv444, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc356, gpt_neox_layers_1_input_layernorm_weight4, gpt_neox_layers_1_input_layernorm_bias4, alloc357) + 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((2304, 768), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[18] + gv445: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc358: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv445, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc357, gpt_neox_layers_1_attention_query_key_value_weight4, gpt_neox_layers_1_attention_query_key_value_bias4, alloc358) + R.vm.kill_object(alloc357) + 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) + gv446: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape148: R.Tensor((batch_size, 1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc358, gv446, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc358) + gv447: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape149: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape148, gv447, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape148) + gv448: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc359: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv448, R.dtype("float16")) + _358: 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)), reshape149, alloc359) + R.vm.kill_object(reshape149) + gv449: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape150: R.Tensor((batch_size, 1, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc359, gv449, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc359) + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape151: R.Tensor((batch_size, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape150, gv450, sinfo_args=(R.Tensor((batch_size, 1, 768), dtype="float16"),)) + R.vm.kill_object(reshape150) + gpt_neox_layers_1_post_attention_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[16] + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc360: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv451, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc356, gpt_neox_layers_1_post_attention_layernorm_weight4, gpt_neox_layers_1_post_attention_layernorm_bias4, alloc360) + 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((3072, 768), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[22] + storage29: R.Object = R.vm.alloc_storage(R.shape([786432]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv452: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc361: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv452, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc360, gpt_neox_layers_1_mlp_dense_h_to_4h_weight4, gpt_neox_layers_1_mlp_dense_h_to_4h_bias4, alloc361) + R.vm.kill_object(alloc360) + 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((768, 3072), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[24] + gv453: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc362: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv453, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc361, gpt_neox_layers_1_mlp_dense_4h_to_h_weight4, gpt_neox_layers_1_mlp_dense_4h_to_h_bias4, alloc362) + R.vm.kill_object(alloc361) + 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((768, 768), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[20] + gv454: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc363: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv454, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape151, gpt_neox_layers_1_attention_dense_weight4, gpt_neox_layers_1_attention_dense_bias4, alloc362, alloc356, alloc363) + R.vm.kill_object(alloc356) + R.vm.kill_object(reshape151) + R.vm.kill_object(alloc362) + 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((768,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[26] + gv455: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc364: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv455, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc363, gpt_neox_layers_2_input_layernorm_weight4, gpt_neox_layers_2_input_layernorm_bias4, alloc364) + 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((2304, 768), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[30] + gv456: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc365: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv456, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc364, gpt_neox_layers_2_attention_query_key_value_weight4, gpt_neox_layers_2_attention_query_key_value_bias4, alloc365) + R.vm.kill_object(alloc364) + 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) + gv457: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape152: R.Tensor((batch_size, 1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc365, gv457, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc365) + gv458: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape153: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape152, gv458, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape152) + gv459: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc366: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv459, R.dtype("float16")) + _365: 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)), reshape153, alloc366) + R.vm.kill_object(reshape153) + gv460: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape154: R.Tensor((batch_size, 1, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc366, gv460, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc366) + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape155: R.Tensor((batch_size, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape154, gv461, sinfo_args=(R.Tensor((batch_size, 1, 768), dtype="float16"),)) + R.vm.kill_object(reshape154) + gpt_neox_layers_2_post_attention_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[28] + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc367: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv462, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc363, gpt_neox_layers_2_post_attention_layernorm_weight4, gpt_neox_layers_2_post_attention_layernorm_bias4, alloc367) + 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((3072, 768), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[34] + gv463: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc368: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv463, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc367, gpt_neox_layers_2_mlp_dense_h_to_4h_weight4, gpt_neox_layers_2_mlp_dense_h_to_4h_bias4, alloc368) + R.vm.kill_object(alloc367) + 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((768, 3072), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[36] + gv464: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc369: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv464, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc368, gpt_neox_layers_2_mlp_dense_4h_to_h_weight4, gpt_neox_layers_2_mlp_dense_4h_to_h_bias4, alloc369) + R.vm.kill_object(alloc368) + 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((768, 768), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[32] + gv465: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc370: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv465, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape155, gpt_neox_layers_2_attention_dense_weight4, gpt_neox_layers_2_attention_dense_bias4, alloc369, alloc363, alloc370) + R.vm.kill_object(alloc363) + R.vm.kill_object(reshape155) + R.vm.kill_object(alloc369) + 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((768,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[38] + gv466: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc371: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv466, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc370, gpt_neox_layers_3_input_layernorm_weight4, gpt_neox_layers_3_input_layernorm_bias4, alloc371) + 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((2304, 768), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[42] + gv467: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc372: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv467, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc371, gpt_neox_layers_3_attention_query_key_value_weight4, gpt_neox_layers_3_attention_query_key_value_bias4, alloc372) + R.vm.kill_object(alloc371) + 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) + gv468: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape156: R.Tensor((batch_size, 1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc372, gv468, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc372) + gv469: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape157: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape156, gv469, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape156) + gv470: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc373: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv470, R.dtype("float16")) + _372: 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)), reshape157, alloc373) + R.vm.kill_object(reshape157) + gv471: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape158: R.Tensor((batch_size, 1, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc373, gv471, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc373) + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape159: R.Tensor((batch_size, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape158, gv472, sinfo_args=(R.Tensor((batch_size, 1, 768), dtype="float16"),)) + R.vm.kill_object(reshape158) + gpt_neox_layers_3_post_attention_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[40] + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc374: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv473, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc370, gpt_neox_layers_3_post_attention_layernorm_weight4, gpt_neox_layers_3_post_attention_layernorm_bias4, alloc374) + 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((3072, 768), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[46] + gv474: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc375: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv474, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc374, gpt_neox_layers_3_mlp_dense_h_to_4h_weight4, gpt_neox_layers_3_mlp_dense_h_to_4h_bias4, alloc375) + R.vm.kill_object(alloc374) + 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((768, 3072), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[48] + gv475: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc376: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv475, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc375, gpt_neox_layers_3_mlp_dense_4h_to_h_weight4, gpt_neox_layers_3_mlp_dense_4h_to_h_bias4, alloc376) + R.vm.kill_object(alloc375) + 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((768, 768), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[44] + gv476: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc377: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv476, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape159, gpt_neox_layers_3_attention_dense_weight4, gpt_neox_layers_3_attention_dense_bias4, alloc376, alloc370, alloc377) + R.vm.kill_object(alloc370) + R.vm.kill_object(reshape159) + R.vm.kill_object(alloc376) + 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((768,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[50] + gv477: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc378: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv477, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc377, gpt_neox_layers_4_input_layernorm_weight4, gpt_neox_layers_4_input_layernorm_bias4, alloc378) + 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((2304, 768), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[54] + gv478: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc379: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv478, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc378, gpt_neox_layers_4_attention_query_key_value_weight4, gpt_neox_layers_4_attention_query_key_value_bias4, alloc379) + R.vm.kill_object(alloc378) + 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) + gv479: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape160: R.Tensor((batch_size, 1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc379, gv479, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc379) + gv480: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape161: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape160, gv480, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape160) + gv481: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc380: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv481, R.dtype("float16")) + _379: 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)), reshape161, alloc380) + R.vm.kill_object(reshape161) + gv482: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape162: R.Tensor((batch_size, 1, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc380, gv482, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc380) + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape163: R.Tensor((batch_size, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape162, gv483, sinfo_args=(R.Tensor((batch_size, 1, 768), dtype="float16"),)) + R.vm.kill_object(reshape162) + gpt_neox_layers_4_post_attention_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[52] + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc381: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv484, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc377, gpt_neox_layers_4_post_attention_layernorm_weight4, gpt_neox_layers_4_post_attention_layernorm_bias4, alloc381) + 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((3072, 768), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[58] + gv485: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc382: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv485, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc381, gpt_neox_layers_4_mlp_dense_h_to_4h_weight4, gpt_neox_layers_4_mlp_dense_h_to_4h_bias4, alloc382) + R.vm.kill_object(alloc381) + 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((768, 3072), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[60] + gv486: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc383: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv486, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc382, gpt_neox_layers_4_mlp_dense_4h_to_h_weight4, gpt_neox_layers_4_mlp_dense_4h_to_h_bias4, alloc383) + R.vm.kill_object(alloc382) + 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((768, 768), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[56] + gv487: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc384: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv487, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape163, gpt_neox_layers_4_attention_dense_weight4, gpt_neox_layers_4_attention_dense_bias4, alloc383, alloc377, alloc384) + R.vm.kill_object(alloc377) + R.vm.kill_object(reshape163) + R.vm.kill_object(alloc383) + 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((768,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[62] + gv488: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc385: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv488, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc384, gpt_neox_layers_5_input_layernorm_weight4, gpt_neox_layers_5_input_layernorm_bias4, alloc385) + 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((2304, 768), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[66] + gv489: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc386: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv489, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc385, gpt_neox_layers_5_attention_query_key_value_weight4, gpt_neox_layers_5_attention_query_key_value_bias4, alloc386) + R.vm.kill_object(alloc385) + 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) + gv490: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape164: R.Tensor((batch_size, 1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc386, gv490, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc386) + gv491: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape165: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape164, gv491, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape164) + gv492: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc387: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv492, R.dtype("float16")) + _386: 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)), reshape165, alloc387) + R.vm.kill_object(reshape165) + gv493: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape166: R.Tensor((batch_size, 1, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc387, gv493, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc387) + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape167: R.Tensor((batch_size, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape166, gv494, sinfo_args=(R.Tensor((batch_size, 1, 768), dtype="float16"),)) + R.vm.kill_object(reshape166) + gpt_neox_layers_5_post_attention_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[64] + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc388: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv495, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc384, gpt_neox_layers_5_post_attention_layernorm_weight4, gpt_neox_layers_5_post_attention_layernorm_bias4, alloc388) + 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((3072, 768), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[70] + gv496: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc389: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv496, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc388, gpt_neox_layers_5_mlp_dense_h_to_4h_weight4, gpt_neox_layers_5_mlp_dense_h_to_4h_bias4, alloc389) + R.vm.kill_object(alloc388) + 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((768, 3072), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[72] + gv497: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc390: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv497, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc389, gpt_neox_layers_5_mlp_dense_4h_to_h_weight4, gpt_neox_layers_5_mlp_dense_4h_to_h_bias4, alloc390) + R.vm.kill_object(alloc389) + 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((768, 768), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[68] + gv498: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc391: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv498, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape167, gpt_neox_layers_5_attention_dense_weight4, gpt_neox_layers_5_attention_dense_bias4, alloc390, alloc384, alloc391) + R.vm.kill_object(alloc384) + R.vm.kill_object(reshape167) + R.vm.kill_object(alloc390) + 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((768,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[74] + gv499: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc392: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv499, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc391, gpt_neox_layers_6_input_layernorm_weight4, gpt_neox_layers_6_input_layernorm_bias4, alloc392) + 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((2304, 768), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[78] + gv500: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc393: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv500, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc392, gpt_neox_layers_6_attention_query_key_value_weight4, gpt_neox_layers_6_attention_query_key_value_bias4, alloc393) + R.vm.kill_object(alloc392) + 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) + gv501: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape168: R.Tensor((batch_size, 1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc393, gv501, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc393) + gv502: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape169: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape168, gv502, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape168) + gv503: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc394: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv503, R.dtype("float16")) + _393: 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)), reshape169, alloc394) + R.vm.kill_object(reshape169) + gv504: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape170: R.Tensor((batch_size, 1, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc394, gv504, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc394) + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape171: R.Tensor((batch_size, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape170, gv505, sinfo_args=(R.Tensor((batch_size, 1, 768), dtype="float16"),)) + R.vm.kill_object(reshape170) + gpt_neox_layers_6_post_attention_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[76] + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc395: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv506, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc391, gpt_neox_layers_6_post_attention_layernorm_weight4, gpt_neox_layers_6_post_attention_layernorm_bias4, alloc395) + 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((3072, 768), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[82] + gv507: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc396: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv507, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc395, gpt_neox_layers_6_mlp_dense_h_to_4h_weight4, gpt_neox_layers_6_mlp_dense_h_to_4h_bias4, alloc396) + R.vm.kill_object(alloc395) + 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((768, 3072), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[84] + gv508: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc397: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv508, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc396, gpt_neox_layers_6_mlp_dense_4h_to_h_weight4, gpt_neox_layers_6_mlp_dense_4h_to_h_bias4, alloc397) + R.vm.kill_object(alloc396) + 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((768, 768), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[80] + gv509: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc398: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv509, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape171, gpt_neox_layers_6_attention_dense_weight4, gpt_neox_layers_6_attention_dense_bias4, alloc397, alloc391, alloc398) + R.vm.kill_object(alloc391) + R.vm.kill_object(reshape171) + R.vm.kill_object(alloc397) + 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((768,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[86] + gv510: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc399: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv510, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc398, gpt_neox_layers_7_input_layernorm_weight4, gpt_neox_layers_7_input_layernorm_bias4, alloc399) + 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((2304, 768), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[90] + gv511: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc400: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv511, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc399, gpt_neox_layers_7_attention_query_key_value_weight4, gpt_neox_layers_7_attention_query_key_value_bias4, alloc400) + R.vm.kill_object(alloc399) + 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) + gv512: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape172: R.Tensor((batch_size, 1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc400, gv512, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc400) + gv513: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape173: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape172, gv513, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape172) + gv514: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc401: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv514, R.dtype("float16")) + _400: 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)), reshape173, alloc401) + R.vm.kill_object(reshape173) + gv515: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape174: R.Tensor((batch_size, 1, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc401, gv515, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc401) + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape175: R.Tensor((batch_size, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape174, gv516, sinfo_args=(R.Tensor((batch_size, 1, 768), dtype="float16"),)) + R.vm.kill_object(reshape174) + gpt_neox_layers_7_post_attention_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[88] + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc402: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv517, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc398, gpt_neox_layers_7_post_attention_layernorm_weight4, gpt_neox_layers_7_post_attention_layernorm_bias4, alloc402) + 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((3072, 768), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[94] + gv518: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc403: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv518, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc402, gpt_neox_layers_7_mlp_dense_h_to_4h_weight4, gpt_neox_layers_7_mlp_dense_h_to_4h_bias4, alloc403) + R.vm.kill_object(alloc402) + 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((768, 3072), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[96] + gv519: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc404: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv519, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc403, gpt_neox_layers_7_mlp_dense_4h_to_h_weight4, gpt_neox_layers_7_mlp_dense_4h_to_h_bias4, alloc404) + R.vm.kill_object(alloc403) + 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((768, 768), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[92] + gv520: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc405: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv520, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape175, gpt_neox_layers_7_attention_dense_weight4, gpt_neox_layers_7_attention_dense_bias4, alloc404, alloc398, alloc405) + R.vm.kill_object(alloc398) + R.vm.kill_object(reshape175) + R.vm.kill_object(alloc404) + 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((768,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[98] + gv521: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc406: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv521, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc405, gpt_neox_layers_8_input_layernorm_weight4, gpt_neox_layers_8_input_layernorm_bias4, alloc406) + 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((2304, 768), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[102] + gv522: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc407: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv522, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc406, gpt_neox_layers_8_attention_query_key_value_weight4, gpt_neox_layers_8_attention_query_key_value_bias4, alloc407) + R.vm.kill_object(alloc406) + 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) + gv523: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape176: R.Tensor((batch_size, 1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc407, gv523, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc407) + gv524: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape177: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape176, gv524, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape176) + gv525: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc408: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv525, R.dtype("float16")) + _407: 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)), reshape177, alloc408) + R.vm.kill_object(reshape177) + gv526: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape178: R.Tensor((batch_size, 1, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc408, gv526, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc408) + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape179: R.Tensor((batch_size, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape178, gv527, sinfo_args=(R.Tensor((batch_size, 1, 768), dtype="float16"),)) + R.vm.kill_object(reshape178) + gpt_neox_layers_8_post_attention_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[100] + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc409: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv528, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc405, gpt_neox_layers_8_post_attention_layernorm_weight4, gpt_neox_layers_8_post_attention_layernorm_bias4, alloc409) + 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((3072, 768), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[106] + gv529: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc410: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv529, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc409, gpt_neox_layers_8_mlp_dense_h_to_4h_weight4, gpt_neox_layers_8_mlp_dense_h_to_4h_bias4, alloc410) + R.vm.kill_object(alloc409) + 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((768, 3072), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[108] + gv530: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc411: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv530, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc410, gpt_neox_layers_8_mlp_dense_4h_to_h_weight4, gpt_neox_layers_8_mlp_dense_4h_to_h_bias4, alloc411) + R.vm.kill_object(alloc410) + 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((768, 768), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[104] + gv531: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc412: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv531, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape179, gpt_neox_layers_8_attention_dense_weight4, gpt_neox_layers_8_attention_dense_bias4, alloc411, alloc405, alloc412) + R.vm.kill_object(alloc405) + R.vm.kill_object(reshape179) + R.vm.kill_object(alloc411) + 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((768,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[110] + gv532: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc413: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv532, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc412, gpt_neox_layers_9_input_layernorm_weight4, gpt_neox_layers_9_input_layernorm_bias4, alloc413) + 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((2304, 768), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[114] + gv533: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc414: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv533, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc413, gpt_neox_layers_9_attention_query_key_value_weight4, gpt_neox_layers_9_attention_query_key_value_bias4, alloc414) + R.vm.kill_object(alloc413) + 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) + gv534: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape180: R.Tensor((batch_size, 1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc414, gv534, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc414) + gv535: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape181: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape180, gv535, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape180) + gv536: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc415: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv536, R.dtype("float16")) + _414: 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)), reshape181, alloc415) + R.vm.kill_object(reshape181) + gv537: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape182: R.Tensor((batch_size, 1, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc415, gv537, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc415) + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape183: R.Tensor((batch_size, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape182, gv538, sinfo_args=(R.Tensor((batch_size, 1, 768), dtype="float16"),)) + R.vm.kill_object(reshape182) + gpt_neox_layers_9_post_attention_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[112] + 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(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc416: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv539, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc412, gpt_neox_layers_9_post_attention_layernorm_weight4, gpt_neox_layers_9_post_attention_layernorm_bias4, alloc416) + 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((3072, 768), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[118] + gv540: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc417: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv540, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc416, gpt_neox_layers_9_mlp_dense_h_to_4h_weight4, gpt_neox_layers_9_mlp_dense_h_to_4h_bias4, alloc417) + R.vm.kill_object(alloc416) + 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((768, 3072), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[120] + gv541: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc418: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv541, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc417, gpt_neox_layers_9_mlp_dense_4h_to_h_weight4, gpt_neox_layers_9_mlp_dense_4h_to_h_bias4, alloc418) + R.vm.kill_object(alloc417) + 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((768, 768), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[116] + gv542: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc419: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv542, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape183, gpt_neox_layers_9_attention_dense_weight4, gpt_neox_layers_9_attention_dense_bias4, alloc418, alloc412, alloc419) + R.vm.kill_object(alloc412) + R.vm.kill_object(reshape183) + R.vm.kill_object(alloc418) + 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((768,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[122] + gv543: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc420: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv543, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc419, gpt_neox_layers_10_input_layernorm_weight4, gpt_neox_layers_10_input_layernorm_bias4, alloc420) + 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((2304, 768), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[126] + gv544: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc421: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv544, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc420, gpt_neox_layers_10_attention_query_key_value_weight4, gpt_neox_layers_10_attention_query_key_value_bias4, alloc421) + R.vm.kill_object(alloc420) + 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) + gv545: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape184: R.Tensor((batch_size, 1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc421, gv545, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc421) + gv546: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape185: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape184, gv546, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape184) + gv547: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc422: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv547, R.dtype("float16")) + _421: 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)), reshape185, alloc422) + R.vm.kill_object(reshape185) + gv548: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape186: R.Tensor((batch_size, 1, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc422, gv548, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc422) + gv549: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape187: R.Tensor((batch_size, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape186, gv549, sinfo_args=(R.Tensor((batch_size, 1, 768), dtype="float16"),)) + R.vm.kill_object(reshape186) + gpt_neox_layers_10_post_attention_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[124] + gv550: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc423: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv550, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc419, gpt_neox_layers_10_post_attention_layernorm_weight4, gpt_neox_layers_10_post_attention_layernorm_bias4, alloc423) + 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((3072, 768), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[130] + gv551: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc424: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv551, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc423, gpt_neox_layers_10_mlp_dense_h_to_4h_weight4, gpt_neox_layers_10_mlp_dense_h_to_4h_bias4, alloc424) + R.vm.kill_object(alloc423) + 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((768, 3072), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[132] + gv552: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc425: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv552, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc424, gpt_neox_layers_10_mlp_dense_4h_to_h_weight4, gpt_neox_layers_10_mlp_dense_4h_to_h_bias4, alloc425) + R.vm.kill_object(alloc424) + 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((768, 768), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[128] + gv553: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc426: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv553, R.dtype("float16")) + R.vm.kill_object(storage28) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape187, gpt_neox_layers_10_attention_dense_weight4, gpt_neox_layers_10_attention_dense_bias4, alloc425, alloc419, alloc426) + R.vm.kill_object(alloc419) + R.vm.kill_object(reshape187) + R.vm.kill_object(alloc425) + 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((768,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[134] + gv554: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc427: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv554, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc426, gpt_neox_layers_11_input_layernorm_weight4, gpt_neox_layers_11_input_layernorm_bias4, alloc427) + 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((2304, 768), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[138] + gv555: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc428: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv555, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul_add(alloc427, gpt_neox_layers_11_attention_query_key_value_weight4, gpt_neox_layers_11_attention_query_key_value_bias4, alloc428) + R.vm.kill_object(alloc427) + 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) + gv556: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape188: R.Tensor((batch_size, 1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc428, gv556, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc428) + gv557: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape189: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape188, gv557, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape188) + gv558: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc429: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv558, R.dtype("float16")) + _428: 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)), reshape189, alloc429) + R.vm.kill_object(reshape189) + gv559: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape190: R.Tensor((batch_size, 1, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc429, gv559, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc429) + gv560: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape191: R.Tensor((batch_size, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape190, gv560, sinfo_args=(R.Tensor((batch_size, 1, 768), dtype="float16"),)) + R.vm.kill_object(reshape190) + gpt_neox_layers_11_post_attention_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[136] + gv561: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc430: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv561, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm(alloc426, gpt_neox_layers_11_post_attention_layernorm_weight4, gpt_neox_layers_11_post_attention_layernorm_bias4, alloc430) + 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((3072, 768), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[142] + gv562: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc431: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv562, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul2_add2_gelu_cast(alloc430, gpt_neox_layers_11_mlp_dense_h_to_4h_weight4, gpt_neox_layers_11_mlp_dense_h_to_4h_bias4, alloc431) + R.vm.kill_object(alloc430) + 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((768, 3072), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[144] + 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(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc432: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv563, R.dtype("float16")) + R.vm.kill_object(storage27) + cls.gpt_neox_q0f16_fused_NT_matmul3_add3_cast1(alloc431, gpt_neox_layers_11_mlp_dense_4h_to_h_weight4, gpt_neox_layers_11_mlp_dense_4h_to_h_bias4, alloc432) + R.vm.kill_object(alloc431) + 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((768, 768), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[140] + gv564: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc433: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage29, R.prim_value(0), gv564, R.dtype("float16")) + R.vm.kill_object(storage29) + cls.gpt_neox_q0f16_fused_NT_matmul1_add1_add4_add4(reshape191, gpt_neox_layers_11_attention_dense_weight4, gpt_neox_layers_11_attention_dense_bias4, alloc432, alloc426, alloc433) + R.vm.kill_object(alloc426) + R.vm.kill_object(reshape191) + R.vm.kill_object(alloc432) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_weight4) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias4) + gpt_neox_final_layer_norm_weight4: R.Tensor((768,), dtype="float16") = packed_params[145] + gpt_neox_final_layer_norm_bias4: R.Tensor((768,), dtype="float16") = packed_params[146] + gv565: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc434: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv565, R.dtype("float16")) + R.vm.kill_object(storage26) + cls.gpt_neox_q0f16_layer_norm(alloc433, gpt_neox_final_layer_norm_weight4, gpt_neox_final_layer_norm_bias4, alloc434) + R.vm.kill_object(alloc433) + 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, 768), dtype="float16") = packed_params[147] + gv566: 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(gv566, 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(1), R.prim_value(1), sinfo_args=(R.Shape(ndim=3),)) + alloc435: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage30, R.prim_value(0), gv567, R.dtype("float32")) + R.vm.kill_object(storage30) + cls.gpt_neox_q0f16_fused_NT_matmul4_cast2(alloc434, embed_out_weight4, alloc435) + R.vm.kill_object(alloc434) + R.vm.kill_object(embed_out_weight4) + gv4: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc435, paged_kv_cache + R.vm.kill_object(alloc435) + gv568: R.Tensor(dtype="float32", ndim=3) = gv4[0] + R.call_packed("vm.builtin.match_shape", gv568, 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", 768), dtype="float16"), logit_positions: R.Tensor(("batch_size",), dtype="int32"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor(("vocab_size", 768), 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, 768), 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(148), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv3: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv3, 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, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv4: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[147] + R.call_packed("vm.builtin.check_tensor_info", gv4, 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, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), 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(768), R.str("ErrorContext(fn=batch_prefill, loc=param[0], param=input_embeds, annotation=R.Tensor((1, seq_len, 768), 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", gv3, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(2), R.prim_value(0), R.prim_value(768), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv4, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(2), R.prim_value(0), R.prim_value(768), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func(shape_heap) + gpt_neox_layers_0_input_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[2] + storage1: R.Object = R.vm.alloc_storage(R.shape([3145728]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv5: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc1: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv5, 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, alloc1) + 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((2304, 768), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[6] + storage2: R.Object = R.vm.alloc_storage(R.shape([9437184]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv6: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc2: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv6, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc1, gpt_neox_layers_0_attention_query_key_value_weight3, gpt_neox_layers_0_attention_query_key_value_bias3, alloc2) + R.vm.kill_object(alloc1) + 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) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape96: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc2, gv7, sinfo_args=(R.Tensor((1, seq_len, 36, 64), 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(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape97: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape96, gv8, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape96) + gv9: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), 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")) + _3: 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)), reshape97, alloc3) + R.vm.kill_object(reshape97) + gv10: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape98: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc3, gv10, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc3) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape99: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape98, gv11, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape98) + gpt_neox_layers_0_post_attention_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[4] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc4: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv12, 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, alloc4) + 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((3072, 768), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[10] + storage3: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc5: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv13, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc4, gpt_neox_layers_0_mlp_dense_h_to_4h_weight3, gpt_neox_layers_0_mlp_dense_h_to_4h_bias3, alloc5) + R.vm.kill_object(alloc4) + 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((768, 3072), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[12] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc6: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv14, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc5, gpt_neox_layers_0_mlp_dense_4h_to_h_weight3, gpt_neox_layers_0_mlp_dense_4h_to_h_bias3, alloc6) + R.vm.kill_object(alloc5) + 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((768, 768), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[8] + gv15: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc7: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv15, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape99, gpt_neox_layers_0_attention_dense_weight3, gpt_neox_layers_0_attention_dense_bias3, alloc6, input_embeds, alloc7) + R.vm.kill_object(reshape99) + R.vm.kill_object(alloc6) + 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((768,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[14] + gv16: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc8: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv16, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc7, gpt_neox_layers_1_input_layernorm_weight3, gpt_neox_layers_1_input_layernorm_bias3, alloc8) + 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((2304, 768), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[18] + gv17: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc9: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv17, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc8, gpt_neox_layers_1_attention_query_key_value_weight3, gpt_neox_layers_1_attention_query_key_value_bias3, alloc9) + R.vm.kill_object(alloc8) + 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) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape100: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc9, gv18, sinfo_args=(R.Tensor((1, seq_len, 36, 64), 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(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape101: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape100, gv19, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape100) + gv20: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), 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")) + _10: 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)), reshape101, alloc10) + R.vm.kill_object(reshape101) + gv21: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape102: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc10, gv21, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc10) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape103: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape102, gv22, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape102) + gpt_neox_layers_1_post_attention_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[16] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc11: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv23, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc7, gpt_neox_layers_1_post_attention_layernorm_weight3, gpt_neox_layers_1_post_attention_layernorm_bias3, alloc11) + 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((3072, 768), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[22] + storage4: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc12: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv24, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc11, gpt_neox_layers_1_mlp_dense_h_to_4h_weight3, gpt_neox_layers_1_mlp_dense_h_to_4h_bias3, alloc12) + R.vm.kill_object(alloc11) + 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((768, 3072), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[24] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc13: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv25, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc12, gpt_neox_layers_1_mlp_dense_4h_to_h_weight3, gpt_neox_layers_1_mlp_dense_4h_to_h_bias3, alloc13) + R.vm.kill_object(alloc12) + 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((768, 768), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[20] + gv26: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc14: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv26, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape103, gpt_neox_layers_1_attention_dense_weight3, gpt_neox_layers_1_attention_dense_bias3, alloc13, alloc7, alloc14) + R.vm.kill_object(alloc7) + R.vm.kill_object(reshape103) + R.vm.kill_object(alloc13) + 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((768,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[26] + gv27: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc15: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv27, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc14, gpt_neox_layers_2_input_layernorm_weight3, gpt_neox_layers_2_input_layernorm_bias3, alloc15) + 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((2304, 768), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[30] + gv28: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc16: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv28, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc15, gpt_neox_layers_2_attention_query_key_value_weight3, gpt_neox_layers_2_attention_query_key_value_bias3, alloc16) + R.vm.kill_object(alloc15) + 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) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape104: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc16, gv29, sinfo_args=(R.Tensor((1, seq_len, 36, 64), 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(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape105: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape104, gv30, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape104) + gv31: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), 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")) + _17: 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)), reshape105, alloc17) + R.vm.kill_object(reshape105) + gv32: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape106: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc17, gv32, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc17) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape107: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape106, gv33, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape106) + gpt_neox_layers_2_post_attention_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[28] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc18: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv34, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc14, gpt_neox_layers_2_post_attention_layernorm_weight3, gpt_neox_layers_2_post_attention_layernorm_bias3, alloc18) + 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((3072, 768), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[34] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc19: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv35, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc18, gpt_neox_layers_2_mlp_dense_h_to_4h_weight3, gpt_neox_layers_2_mlp_dense_h_to_4h_bias3, alloc19) + R.vm.kill_object(alloc18) + 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((768, 3072), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[36] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc20: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv36, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc19, gpt_neox_layers_2_mlp_dense_4h_to_h_weight3, gpt_neox_layers_2_mlp_dense_4h_to_h_bias3, alloc20) + R.vm.kill_object(alloc19) + 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((768, 768), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[32] + gv37: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc21: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv37, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape107, gpt_neox_layers_2_attention_dense_weight3, gpt_neox_layers_2_attention_dense_bias3, alloc20, alloc14, alloc21) + R.vm.kill_object(alloc14) + R.vm.kill_object(reshape107) + R.vm.kill_object(alloc20) + 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((768,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[38] + gv38: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc22: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv38, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc21, gpt_neox_layers_3_input_layernorm_weight3, gpt_neox_layers_3_input_layernorm_bias3, alloc22) + 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((2304, 768), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[42] + gv39: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc23: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv39, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc22, gpt_neox_layers_3_attention_query_key_value_weight3, gpt_neox_layers_3_attention_query_key_value_bias3, alloc23) + R.vm.kill_object(alloc22) + 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) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape108: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc23, gv40, sinfo_args=(R.Tensor((1, seq_len, 36, 64), 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(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape109: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape108, gv41, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape108) + gv42: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), 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")) + _24: 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)), reshape109, alloc24) + R.vm.kill_object(reshape109) + gv43: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape110: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc24, gv43, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc24) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape111: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape110, gv44, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape110) + gpt_neox_layers_3_post_attention_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[40] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc25: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv45, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc21, gpt_neox_layers_3_post_attention_layernorm_weight3, gpt_neox_layers_3_post_attention_layernorm_bias3, alloc25) + 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((3072, 768), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[46] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc26: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv46, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc25, gpt_neox_layers_3_mlp_dense_h_to_4h_weight3, gpt_neox_layers_3_mlp_dense_h_to_4h_bias3, alloc26) + R.vm.kill_object(alloc25) + 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((768, 3072), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[48] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc27: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv47, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc26, gpt_neox_layers_3_mlp_dense_4h_to_h_weight3, gpt_neox_layers_3_mlp_dense_4h_to_h_bias3, alloc27) + R.vm.kill_object(alloc26) + 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((768, 768), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[44] + gv48: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc28: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv48, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape111, gpt_neox_layers_3_attention_dense_weight3, gpt_neox_layers_3_attention_dense_bias3, alloc27, alloc21, alloc28) + R.vm.kill_object(alloc21) + R.vm.kill_object(reshape111) + R.vm.kill_object(alloc27) + 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((768,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[50] + gv49: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc29: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv49, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc28, gpt_neox_layers_4_input_layernorm_weight3, gpt_neox_layers_4_input_layernorm_bias3, alloc29) + 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((2304, 768), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[54] + gv50: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc30: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv50, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc29, gpt_neox_layers_4_attention_query_key_value_weight3, gpt_neox_layers_4_attention_query_key_value_bias3, alloc30) + R.vm.kill_object(alloc29) + 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) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape112: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc30, gv51, sinfo_args=(R.Tensor((1, seq_len, 36, 64), 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(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape113: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape112, gv52, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape112) + gv53: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), 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")) + _31: 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)), reshape113, alloc31) + R.vm.kill_object(reshape113) + gv54: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape114: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc31, gv54, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc31) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape115: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape114, gv55, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape114) + gpt_neox_layers_4_post_attention_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[52] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc32: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv56, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc28, gpt_neox_layers_4_post_attention_layernorm_weight3, gpt_neox_layers_4_post_attention_layernorm_bias3, alloc32) + 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((3072, 768), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[58] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc33: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv57, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc32, gpt_neox_layers_4_mlp_dense_h_to_4h_weight3, gpt_neox_layers_4_mlp_dense_h_to_4h_bias3, alloc33) + R.vm.kill_object(alloc32) + 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((768, 3072), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[60] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc34: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv58, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc33, gpt_neox_layers_4_mlp_dense_4h_to_h_weight3, gpt_neox_layers_4_mlp_dense_4h_to_h_bias3, alloc34) + R.vm.kill_object(alloc33) + 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((768, 768), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[56] + gv59: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc35: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv59, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape115, gpt_neox_layers_4_attention_dense_weight3, gpt_neox_layers_4_attention_dense_bias3, alloc34, alloc28, alloc35) + R.vm.kill_object(alloc28) + R.vm.kill_object(reshape115) + R.vm.kill_object(alloc34) + 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((768,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[62] + gv60: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc36: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv60, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc35, gpt_neox_layers_5_input_layernorm_weight3, gpt_neox_layers_5_input_layernorm_bias3, alloc36) + 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((2304, 768), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[66] + gv61: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc37: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv61, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc36, gpt_neox_layers_5_attention_query_key_value_weight3, gpt_neox_layers_5_attention_query_key_value_bias3, alloc37) + R.vm.kill_object(alloc36) + 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) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape116: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc37, gv62, sinfo_args=(R.Tensor((1, seq_len, 36, 64), 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(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape117: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape116, gv63, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape116) + gv64: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), 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")) + _38: 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)), reshape117, alloc38) + R.vm.kill_object(reshape117) + gv65: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape118: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc38, gv65, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc38) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape119: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape118, gv66, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape118) + gpt_neox_layers_5_post_attention_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[64] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc39: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv67, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc35, gpt_neox_layers_5_post_attention_layernorm_weight3, gpt_neox_layers_5_post_attention_layernorm_bias3, alloc39) + 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((3072, 768), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[70] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc40: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv68, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc39, gpt_neox_layers_5_mlp_dense_h_to_4h_weight3, gpt_neox_layers_5_mlp_dense_h_to_4h_bias3, alloc40) + R.vm.kill_object(alloc39) + 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((768, 3072), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[72] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc41: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv69, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc40, gpt_neox_layers_5_mlp_dense_4h_to_h_weight3, gpt_neox_layers_5_mlp_dense_4h_to_h_bias3, alloc41) + R.vm.kill_object(alloc40) + 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((768, 768), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[68] + gv70: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc42: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv70, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape119, gpt_neox_layers_5_attention_dense_weight3, gpt_neox_layers_5_attention_dense_bias3, alloc41, alloc35, alloc42) + R.vm.kill_object(alloc35) + R.vm.kill_object(reshape119) + R.vm.kill_object(alloc41) + 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((768,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[74] + gv71: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc43: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv71, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc42, gpt_neox_layers_6_input_layernorm_weight3, gpt_neox_layers_6_input_layernorm_bias3, alloc43) + 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((2304, 768), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[78] + gv72: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc44: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv72, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc43, gpt_neox_layers_6_attention_query_key_value_weight3, gpt_neox_layers_6_attention_query_key_value_bias3, alloc44) + R.vm.kill_object(alloc43) + 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) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape120: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc44, gv73, sinfo_args=(R.Tensor((1, seq_len, 36, 64), 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(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape121: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape120, gv74, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape120) + gv75: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), 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")) + _45: 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)), reshape121, alloc45) + R.vm.kill_object(reshape121) + gv76: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape122: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc45, gv76, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc45) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape123: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape122, gv77, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape122) + gpt_neox_layers_6_post_attention_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[76] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc46: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv78, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc42, gpt_neox_layers_6_post_attention_layernorm_weight3, gpt_neox_layers_6_post_attention_layernorm_bias3, alloc46) + 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((3072, 768), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[82] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc47: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv79, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc46, gpt_neox_layers_6_mlp_dense_h_to_4h_weight3, gpt_neox_layers_6_mlp_dense_h_to_4h_bias3, alloc47) + R.vm.kill_object(alloc46) + 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((768, 3072), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[84] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc48: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv80, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc47, gpt_neox_layers_6_mlp_dense_4h_to_h_weight3, gpt_neox_layers_6_mlp_dense_4h_to_h_bias3, alloc48) + R.vm.kill_object(alloc47) + 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((768, 768), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[80] + gv81: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc49: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv81, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape123, gpt_neox_layers_6_attention_dense_weight3, gpt_neox_layers_6_attention_dense_bias3, alloc48, alloc42, alloc49) + R.vm.kill_object(alloc42) + R.vm.kill_object(reshape123) + R.vm.kill_object(alloc48) + 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((768,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[86] + gv82: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc50: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv82, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc49, gpt_neox_layers_7_input_layernorm_weight3, gpt_neox_layers_7_input_layernorm_bias3, alloc50) + 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((2304, 768), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[90] + gv83: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc51: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv83, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc50, gpt_neox_layers_7_attention_query_key_value_weight3, gpt_neox_layers_7_attention_query_key_value_bias3, alloc51) + R.vm.kill_object(alloc50) + 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) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape124: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc51, gv84, sinfo_args=(R.Tensor((1, seq_len, 36, 64), 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(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape125: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape124, gv85, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape124) + gv86: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), 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")) + _52: 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)), reshape125, alloc52) + R.vm.kill_object(reshape125) + gv87: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape126: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc52, gv87, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc52) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape127: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape126, gv88, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape126) + gpt_neox_layers_7_post_attention_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[88] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc53: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv89, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc49, gpt_neox_layers_7_post_attention_layernorm_weight3, gpt_neox_layers_7_post_attention_layernorm_bias3, alloc53) + 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((3072, 768), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[94] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc54: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv90, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc53, gpt_neox_layers_7_mlp_dense_h_to_4h_weight3, gpt_neox_layers_7_mlp_dense_h_to_4h_bias3, alloc54) + R.vm.kill_object(alloc53) + 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((768, 3072), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[96] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc55: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv91, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc54, gpt_neox_layers_7_mlp_dense_4h_to_h_weight3, gpt_neox_layers_7_mlp_dense_4h_to_h_bias3, alloc55) + R.vm.kill_object(alloc54) + 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((768, 768), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[92] + gv92: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc56: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv92, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape127, gpt_neox_layers_7_attention_dense_weight3, gpt_neox_layers_7_attention_dense_bias3, alloc55, alloc49, alloc56) + R.vm.kill_object(alloc49) + R.vm.kill_object(reshape127) + R.vm.kill_object(alloc55) + 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((768,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[98] + gv93: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc57: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv93, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc56, gpt_neox_layers_8_input_layernorm_weight3, gpt_neox_layers_8_input_layernorm_bias3, alloc57) + 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((2304, 768), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[102] + gv94: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc58: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv94, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc57, gpt_neox_layers_8_attention_query_key_value_weight3, gpt_neox_layers_8_attention_query_key_value_bias3, alloc58) + R.vm.kill_object(alloc57) + 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) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape128: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc58, gv95, sinfo_args=(R.Tensor((1, seq_len, 36, 64), 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(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape129: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape128, gv96, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape128) + gv97: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), 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")) + _59: 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)), reshape129, alloc59) + R.vm.kill_object(reshape129) + gv98: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape130: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc59, gv98, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc59) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape131: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape130, gv99, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape130) + gpt_neox_layers_8_post_attention_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[100] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc60: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv100, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc56, gpt_neox_layers_8_post_attention_layernorm_weight3, gpt_neox_layers_8_post_attention_layernorm_bias3, alloc60) + 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((3072, 768), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[106] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc61: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv101, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc60, gpt_neox_layers_8_mlp_dense_h_to_4h_weight3, gpt_neox_layers_8_mlp_dense_h_to_4h_bias3, alloc61) + R.vm.kill_object(alloc60) + 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((768, 3072), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[108] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc62: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv102, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc61, gpt_neox_layers_8_mlp_dense_4h_to_h_weight3, gpt_neox_layers_8_mlp_dense_4h_to_h_bias3, alloc62) + R.vm.kill_object(alloc61) + 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((768, 768), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[104] + gv103: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc63: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv103, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape131, gpt_neox_layers_8_attention_dense_weight3, gpt_neox_layers_8_attention_dense_bias3, alloc62, alloc56, alloc63) + R.vm.kill_object(alloc56) + R.vm.kill_object(reshape131) + R.vm.kill_object(alloc62) + 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((768,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[110] + gv104: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc64: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv104, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc63, gpt_neox_layers_9_input_layernorm_weight3, gpt_neox_layers_9_input_layernorm_bias3, alloc64) + 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((2304, 768), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[114] + gv105: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc65: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv105, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc64, gpt_neox_layers_9_attention_query_key_value_weight3, gpt_neox_layers_9_attention_query_key_value_bias3, alloc65) + R.vm.kill_object(alloc64) + 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) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape132: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc65, gv106, sinfo_args=(R.Tensor((1, seq_len, 36, 64), 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(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape133: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape132, gv107, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape132) + gv108: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), 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")) + _66: 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)), reshape133, alloc66) + R.vm.kill_object(reshape133) + gv109: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape134: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc66, gv109, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc66) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape135: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape134, gv110, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape134) + gpt_neox_layers_9_post_attention_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[112] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc67: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv111, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc63, gpt_neox_layers_9_post_attention_layernorm_weight3, gpt_neox_layers_9_post_attention_layernorm_bias3, alloc67) + 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((3072, 768), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[118] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc68: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv112, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc67, gpt_neox_layers_9_mlp_dense_h_to_4h_weight3, gpt_neox_layers_9_mlp_dense_h_to_4h_bias3, alloc68) + R.vm.kill_object(alloc67) + 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((768, 3072), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[120] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc69: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv113, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc68, gpt_neox_layers_9_mlp_dense_4h_to_h_weight3, gpt_neox_layers_9_mlp_dense_4h_to_h_bias3, alloc69) + R.vm.kill_object(alloc68) + 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((768, 768), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[116] + gv114: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc70: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv114, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape135, gpt_neox_layers_9_attention_dense_weight3, gpt_neox_layers_9_attention_dense_bias3, alloc69, alloc63, alloc70) + R.vm.kill_object(alloc63) + R.vm.kill_object(reshape135) + R.vm.kill_object(alloc69) + 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((768,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[122] + gv115: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc71: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv115, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc70, gpt_neox_layers_10_input_layernorm_weight3, gpt_neox_layers_10_input_layernorm_bias3, alloc71) + 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((2304, 768), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[126] + gv116: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc72: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv116, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc71, gpt_neox_layers_10_attention_query_key_value_weight3, gpt_neox_layers_10_attention_query_key_value_bias3, alloc72) + R.vm.kill_object(alloc71) + 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) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape136: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc72, gv117, sinfo_args=(R.Tensor((1, seq_len, 36, 64), 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(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape137: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape136, gv118, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape136) + gv119: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), 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")) + _73: 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)), reshape137, alloc73) + R.vm.kill_object(reshape137) + gv120: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape138: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc73, gv120, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc73) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape139: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape138, gv121, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape138) + gpt_neox_layers_10_post_attention_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[124] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc74: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv122, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc70, gpt_neox_layers_10_post_attention_layernorm_weight3, gpt_neox_layers_10_post_attention_layernorm_bias3, alloc74) + 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((3072, 768), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[130] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc75: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv123, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc74, gpt_neox_layers_10_mlp_dense_h_to_4h_weight3, gpt_neox_layers_10_mlp_dense_h_to_4h_bias3, alloc75) + R.vm.kill_object(alloc74) + 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((768, 3072), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[132] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc76: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv124, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc75, gpt_neox_layers_10_mlp_dense_4h_to_h_weight3, gpt_neox_layers_10_mlp_dense_4h_to_h_bias3, alloc76) + R.vm.kill_object(alloc75) + 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((768, 768), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[128] + gv125: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc77: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage3, R.prim_value(0), gv125, R.dtype("float16")) + R.vm.kill_object(storage3) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape139, gpt_neox_layers_10_attention_dense_weight3, gpt_neox_layers_10_attention_dense_bias3, alloc76, alloc70, alloc77) + R.vm.kill_object(alloc70) + R.vm.kill_object(reshape139) + R.vm.kill_object(alloc76) + 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((768,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[134] + gv126: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc78: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv126, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc77, gpt_neox_layers_11_input_layernorm_weight3, gpt_neox_layers_11_input_layernorm_bias3, alloc78) + 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((2304, 768), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[138] + gv127: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc79: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv127, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc78, gpt_neox_layers_11_attention_query_key_value_weight3, gpt_neox_layers_11_attention_query_key_value_bias3, alloc79) + R.vm.kill_object(alloc78) + 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) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape140: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc79, gv128, sinfo_args=(R.Tensor((1, seq_len, 36, 64), 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(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape141: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape140, gv129, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape140) + gv130: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), 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")) + _80: 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)), reshape141, alloc80) + R.vm.kill_object(reshape141) + gv131: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape142: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc80, gv131, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc80) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape143: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape142, gv132, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape142) + gpt_neox_layers_11_post_attention_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[136] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc81: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv133, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc77, gpt_neox_layers_11_post_attention_layernorm_weight3, gpt_neox_layers_11_post_attention_layernorm_bias3, alloc81) + 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((3072, 768), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[142] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc82: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv134, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc81, gpt_neox_layers_11_mlp_dense_h_to_4h_weight3, gpt_neox_layers_11_mlp_dense_h_to_4h_bias3, alloc82) + R.vm.kill_object(alloc81) + 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((768, 3072), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[144] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc83: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage2, R.prim_value(0), gv135, R.dtype("float16")) + R.vm.kill_object(storage2) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc82, gpt_neox_layers_11_mlp_dense_4h_to_h_weight3, gpt_neox_layers_11_mlp_dense_4h_to_h_bias3, alloc83) + R.vm.kill_object(alloc82) + 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((768, 768), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[140] + gv136: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc84: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv136, R.dtype("float16")) + R.vm.kill_object(storage4) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape143, gpt_neox_layers_11_attention_dense_weight3, gpt_neox_layers_11_attention_dense_bias3, alloc83, alloc77, alloc84) + R.vm.kill_object(alloc77) + R.vm.kill_object(reshape143) + R.vm.kill_object(alloc83) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_weight3) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias3) + gpt_neox_final_layer_norm_weight3: R.Tensor((768,), dtype="float16") = packed_params[145] + gpt_neox_final_layer_norm_bias3: R.Tensor((768,), dtype="float16") = packed_params[146] + gv137: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc85: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage1, R.prim_value(0), gv137, R.dtype("float16")) + R.vm.kill_object(storage1) + cls.gpt_neox_q0f16_layer_norm1(alloc84, gpt_neox_final_layer_norm_weight3, gpt_neox_final_layer_norm_bias3, alloc85) + R.vm.kill_object(alloc84) + R.vm.kill_object(gpt_neox_final_layer_norm_weight3) + R.vm.kill_object(gpt_neox_final_layer_norm_bias3) + storage5: R.Object = R.vm.alloc_storage(R.shape([196608]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv138: R.Shape(ndim=3) = R.call_packed("vm.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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc86: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv138, R.dtype("float16")) + R.vm.kill_object(storage5) + cls.gpt_neox_q0f16_take(alloc85, logit_positions, alloc86) + R.vm.kill_object(alloc85) + embed_out_weight3: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[147] + gv139: 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),)) + storage6: R.Object = R.vm.alloc_storage(gv139, R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(1), R.prim_value(1), R.prim_value(2), sinfo_args=(R.Shape(ndim=3),)) + alloc87: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv140, R.dtype("float32")) + R.vm.kill_object(storage6) + cls.gpt_neox_q0f16_fused_NT_matmul9_cast5(alloc86, embed_out_weight3, alloc87) + R.vm.kill_object(alloc86) + R.vm.kill_object(embed_out_weight3) + gv3_1: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc87, paged_kv_cache + R.vm.kill_object(alloc87) + gv141: R.Tensor(dtype="float32", ndim=3) = gv3_1[0] + R.call_packed("vm.builtin.match_shape", gv141, 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_1 + + @R.function + def batch_verify(input_embeds: R.Tensor((1, "seq_len", 768), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor(("vocab_size", 768), 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, 768), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(148), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv142: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv142, 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, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv143: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[147] + R.call_packed("vm.builtin.check_tensor_info", gv143, 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, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), 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(768), R.str("ErrorContext(fn=batch_verify, loc=param[0], param=input_embeds, annotation=R.Tensor((1, seq_len, 768), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv142, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(768), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv143, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(768), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func1(shape_heap) + gpt_neox_layers_0_input_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[2] + storage7: R.Object = R.vm.alloc_storage(R.shape([3145728]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc88: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv144, 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, alloc88) + 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((2304, 768), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[6] + storage8: R.Object = R.vm.alloc_storage(R.shape([9437184]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc89: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv145, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc88, gpt_neox_layers_0_attention_query_key_value_weight5, gpt_neox_layers_0_attention_query_key_value_bias5, alloc89) + R.vm.kill_object(alloc88) + 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) + gv146: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape192: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc89, gv146, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc89) + gv147: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape193: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape192, gv147, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape192) + 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc90: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv148, R.dtype("float16")) + _90: 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, alloc90) + R.vm.kill_object(reshape193) + gv149: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape194: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc90, gv149, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc90) + gv150: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape195: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape194, gv150, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape194) + gpt_neox_layers_0_post_attention_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[4] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc91: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv151, 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, alloc91) + 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((3072, 768), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[10] + storage9: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc92: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv152, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc91, gpt_neox_layers_0_mlp_dense_h_to_4h_weight5, gpt_neox_layers_0_mlp_dense_h_to_4h_bias5, alloc92) + R.vm.kill_object(alloc91) + 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((768, 3072), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[12] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc93: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv153, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc92, gpt_neox_layers_0_mlp_dense_4h_to_h_weight5, gpt_neox_layers_0_mlp_dense_4h_to_h_bias5, alloc93) + R.vm.kill_object(alloc92) + 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((768, 768), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[8] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc94: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv154, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape195, gpt_neox_layers_0_attention_dense_weight5, gpt_neox_layers_0_attention_dense_bias5, alloc93, input_embeds, alloc94) + R.vm.kill_object(reshape195) + R.vm.kill_object(alloc93) + 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((768,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[14] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc95: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv155, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc94, gpt_neox_layers_1_input_layernorm_weight5, gpt_neox_layers_1_input_layernorm_bias5, alloc95) + 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((2304, 768), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[18] + 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(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc96: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv156, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc95, gpt_neox_layers_1_attention_query_key_value_weight5, gpt_neox_layers_1_attention_query_key_value_bias5, alloc96) + R.vm.kill_object(alloc95) + 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) + gv157: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape196: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc96, gv157, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc96) + gv158: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape197: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape196, gv158, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape196) + 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc97: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv159, R.dtype("float16")) + _97: 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, alloc97) + R.vm.kill_object(reshape197) + gv160: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape198: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc97, gv160, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc97) + gv161: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape199: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape198, gv161, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape198) + gpt_neox_layers_1_post_attention_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[16] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc98: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv162, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc94, gpt_neox_layers_1_post_attention_layernorm_weight5, gpt_neox_layers_1_post_attention_layernorm_bias5, alloc98) + 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((3072, 768), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[22] + storage10: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc99: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv163, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc98, gpt_neox_layers_1_mlp_dense_h_to_4h_weight5, gpt_neox_layers_1_mlp_dense_h_to_4h_bias5, alloc99) + R.vm.kill_object(alloc98) + 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((768, 3072), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[24] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc100: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv164, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc99, gpt_neox_layers_1_mlp_dense_4h_to_h_weight5, gpt_neox_layers_1_mlp_dense_4h_to_h_bias5, alloc100) + R.vm.kill_object(alloc99) + 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((768, 768), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[20] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc101: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv165, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape199, gpt_neox_layers_1_attention_dense_weight5, gpt_neox_layers_1_attention_dense_bias5, alloc100, alloc94, alloc101) + R.vm.kill_object(alloc94) + R.vm.kill_object(reshape199) + R.vm.kill_object(alloc100) + 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((768,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[26] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc102: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv166, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc101, gpt_neox_layers_2_input_layernorm_weight5, gpt_neox_layers_2_input_layernorm_bias5, alloc102) + 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((2304, 768), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[30] + 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(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc103: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv167, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc102, gpt_neox_layers_2_attention_query_key_value_weight5, gpt_neox_layers_2_attention_query_key_value_bias5, alloc103) + R.vm.kill_object(alloc102) + 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) + gv168: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape200: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc103, gv168, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc103) + gv169: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape201: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape200, gv169, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape200) + 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc104: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv170, R.dtype("float16")) + _104: 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, alloc104) + R.vm.kill_object(reshape201) + gv171: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape202: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc104, gv171, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc104) + gv172: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape203: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape202, gv172, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape202) + gpt_neox_layers_2_post_attention_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[28] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc105: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv173, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc101, gpt_neox_layers_2_post_attention_layernorm_weight5, gpt_neox_layers_2_post_attention_layernorm_bias5, alloc105) + 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((3072, 768), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[34] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc106: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv174, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc105, gpt_neox_layers_2_mlp_dense_h_to_4h_weight5, gpt_neox_layers_2_mlp_dense_h_to_4h_bias5, alloc106) + R.vm.kill_object(alloc105) + 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((768, 3072), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[36] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc107: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv175, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc106, gpt_neox_layers_2_mlp_dense_4h_to_h_weight5, gpt_neox_layers_2_mlp_dense_4h_to_h_bias5, alloc107) + R.vm.kill_object(alloc106) + 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((768, 768), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[32] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc108: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv176, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape203, gpt_neox_layers_2_attention_dense_weight5, gpt_neox_layers_2_attention_dense_bias5, alloc107, alloc101, alloc108) + R.vm.kill_object(alloc101) + R.vm.kill_object(reshape203) + R.vm.kill_object(alloc107) + 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((768,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[38] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc109: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv177, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc108, gpt_neox_layers_3_input_layernorm_weight5, gpt_neox_layers_3_input_layernorm_bias5, alloc109) + 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((2304, 768), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[42] + 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(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc110: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv178, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc109, gpt_neox_layers_3_attention_query_key_value_weight5, gpt_neox_layers_3_attention_query_key_value_bias5, alloc110) + R.vm.kill_object(alloc109) + 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) + gv179: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape204: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc110, gv179, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc110) + gv180: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape205: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape204, gv180, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape204) + gv181: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc111: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv181, R.dtype("float16")) + _111: 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, alloc111) + R.vm.kill_object(reshape205) + gv182: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape206: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc111, gv182, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc111) + gv183: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape207: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape206, gv183, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape206) + gpt_neox_layers_3_post_attention_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[40] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc112: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv184, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc108, gpt_neox_layers_3_post_attention_layernorm_weight5, gpt_neox_layers_3_post_attention_layernorm_bias5, alloc112) + 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((3072, 768), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[46] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc113: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv185, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc112, gpt_neox_layers_3_mlp_dense_h_to_4h_weight5, gpt_neox_layers_3_mlp_dense_h_to_4h_bias5, alloc113) + R.vm.kill_object(alloc112) + 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((768, 3072), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[48] + gv186: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc114: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv186, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc113, gpt_neox_layers_3_mlp_dense_4h_to_h_weight5, gpt_neox_layers_3_mlp_dense_4h_to_h_bias5, alloc114) + R.vm.kill_object(alloc113) + 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((768, 768), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[44] + gv187: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc115: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv187, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape207, gpt_neox_layers_3_attention_dense_weight5, gpt_neox_layers_3_attention_dense_bias5, alloc114, alloc108, alloc115) + R.vm.kill_object(alloc108) + R.vm.kill_object(reshape207) + R.vm.kill_object(alloc114) + 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((768,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[50] + gv188: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc116: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv188, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc115, gpt_neox_layers_4_input_layernorm_weight5, gpt_neox_layers_4_input_layernorm_bias5, alloc116) + 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((2304, 768), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[54] + gv189: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc117: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv189, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc116, gpt_neox_layers_4_attention_query_key_value_weight5, gpt_neox_layers_4_attention_query_key_value_bias5, alloc117) + R.vm.kill_object(alloc116) + 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) + gv190: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape208: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc117, gv190, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc117) + gv191: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape209: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape208, gv191, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape208) + gv192: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc118: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv192, R.dtype("float16")) + _118: 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, alloc118) + R.vm.kill_object(reshape209) + gv193: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape210: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc118, gv193, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc118) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape211: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape210, gv194, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape210) + gpt_neox_layers_4_post_attention_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[52] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc119: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv195, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc115, gpt_neox_layers_4_post_attention_layernorm_weight5, gpt_neox_layers_4_post_attention_layernorm_bias5, alloc119) + 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((3072, 768), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[58] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc120: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv196, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc119, gpt_neox_layers_4_mlp_dense_h_to_4h_weight5, gpt_neox_layers_4_mlp_dense_h_to_4h_bias5, alloc120) + R.vm.kill_object(alloc119) + 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((768, 3072), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[60] + gv197: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc121: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv197, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc120, gpt_neox_layers_4_mlp_dense_4h_to_h_weight5, gpt_neox_layers_4_mlp_dense_4h_to_h_bias5, alloc121) + R.vm.kill_object(alloc120) + 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((768, 768), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[56] + gv198: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc122: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv198, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape211, gpt_neox_layers_4_attention_dense_weight5, gpt_neox_layers_4_attention_dense_bias5, alloc121, alloc115, alloc122) + R.vm.kill_object(alloc115) + R.vm.kill_object(reshape211) + R.vm.kill_object(alloc121) + 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((768,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[62] + gv199: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc123: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv199, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc122, gpt_neox_layers_5_input_layernorm_weight5, gpt_neox_layers_5_input_layernorm_bias5, alloc123) + 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((2304, 768), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[66] + gv200: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc124: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv200, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc123, gpt_neox_layers_5_attention_query_key_value_weight5, gpt_neox_layers_5_attention_query_key_value_bias5, alloc124) + R.vm.kill_object(alloc123) + 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) + gv201: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape212: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc124, gv201, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc124) + gv202: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape213: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape212, gv202, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape212) + gv203: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc125: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv203, R.dtype("float16")) + _125: 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, alloc125) + R.vm.kill_object(reshape213) + gv204: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape214: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc125, gv204, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc125) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape215: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape214, gv205, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape214) + gpt_neox_layers_5_post_attention_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[64] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc126: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv206, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc122, gpt_neox_layers_5_post_attention_layernorm_weight5, gpt_neox_layers_5_post_attention_layernorm_bias5, alloc126) + 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((3072, 768), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[70] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc127: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv207, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc126, gpt_neox_layers_5_mlp_dense_h_to_4h_weight5, gpt_neox_layers_5_mlp_dense_h_to_4h_bias5, alloc127) + R.vm.kill_object(alloc126) + 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((768, 3072), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[72] + gv208: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc128: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv208, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc127, gpt_neox_layers_5_mlp_dense_4h_to_h_weight5, gpt_neox_layers_5_mlp_dense_4h_to_h_bias5, alloc128) + R.vm.kill_object(alloc127) + 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((768, 768), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[68] + gv209: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc129: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv209, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape215, gpt_neox_layers_5_attention_dense_weight5, gpt_neox_layers_5_attention_dense_bias5, alloc128, alloc122, alloc129) + R.vm.kill_object(alloc122) + R.vm.kill_object(reshape215) + R.vm.kill_object(alloc128) + 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((768,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[74] + gv210: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc130: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv210, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc129, gpt_neox_layers_6_input_layernorm_weight5, gpt_neox_layers_6_input_layernorm_bias5, alloc130) + 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((2304, 768), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[78] + gv211: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc131: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv211, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc130, gpt_neox_layers_6_attention_query_key_value_weight5, gpt_neox_layers_6_attention_query_key_value_bias5, alloc131) + R.vm.kill_object(alloc130) + 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) + gv212: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape216: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc131, gv212, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc131) + gv213: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape217: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape216, gv213, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape216) + gv214: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc132: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv214, R.dtype("float16")) + _132: 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, alloc132) + R.vm.kill_object(reshape217) + gv215: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape218: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc132, gv215, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc132) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape219: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape218, gv216, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape218) + gpt_neox_layers_6_post_attention_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[76] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc133: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv217, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc129, gpt_neox_layers_6_post_attention_layernorm_weight5, gpt_neox_layers_6_post_attention_layernorm_bias5, alloc133) + 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((3072, 768), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[82] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc134: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv218, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc133, gpt_neox_layers_6_mlp_dense_h_to_4h_weight5, gpt_neox_layers_6_mlp_dense_h_to_4h_bias5, alloc134) + R.vm.kill_object(alloc133) + 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((768, 3072), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[84] + gv219: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc135: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv219, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc134, gpt_neox_layers_6_mlp_dense_4h_to_h_weight5, gpt_neox_layers_6_mlp_dense_4h_to_h_bias5, alloc135) + R.vm.kill_object(alloc134) + 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((768, 768), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[80] + gv220: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc136: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv220, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape219, gpt_neox_layers_6_attention_dense_weight5, gpt_neox_layers_6_attention_dense_bias5, alloc135, alloc129, alloc136) + R.vm.kill_object(alloc129) + R.vm.kill_object(reshape219) + R.vm.kill_object(alloc135) + 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((768,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[86] + gv221: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc137: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv221, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc136, gpt_neox_layers_7_input_layernorm_weight5, gpt_neox_layers_7_input_layernorm_bias5, alloc137) + 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((2304, 768), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[90] + gv222: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc138: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv222, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc137, gpt_neox_layers_7_attention_query_key_value_weight5, gpt_neox_layers_7_attention_query_key_value_bias5, alloc138) + R.vm.kill_object(alloc137) + 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) + gv223: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape220: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc138, gv223, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc138) + gv224: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape221: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape220, gv224, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape220) + gv225: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc139: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv225, R.dtype("float16")) + _139: 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, alloc139) + R.vm.kill_object(reshape221) + gv226: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape222: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc139, gv226, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc139) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape223: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape222, gv227, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape222) + gpt_neox_layers_7_post_attention_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[88] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc140: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv228, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc136, gpt_neox_layers_7_post_attention_layernorm_weight5, gpt_neox_layers_7_post_attention_layernorm_bias5, alloc140) + 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((3072, 768), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[94] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc141: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv229, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc140, gpt_neox_layers_7_mlp_dense_h_to_4h_weight5, gpt_neox_layers_7_mlp_dense_h_to_4h_bias5, alloc141) + R.vm.kill_object(alloc140) + 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((768, 3072), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[96] + gv230: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc142: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv230, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc141, gpt_neox_layers_7_mlp_dense_4h_to_h_weight5, gpt_neox_layers_7_mlp_dense_4h_to_h_bias5, alloc142) + R.vm.kill_object(alloc141) + 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((768, 768), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[92] + gv231: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc143: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv231, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape223, gpt_neox_layers_7_attention_dense_weight5, gpt_neox_layers_7_attention_dense_bias5, alloc142, alloc136, alloc143) + R.vm.kill_object(alloc136) + R.vm.kill_object(reshape223) + R.vm.kill_object(alloc142) + 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((768,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[98] + gv232: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc144: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv232, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc143, gpt_neox_layers_8_input_layernorm_weight5, gpt_neox_layers_8_input_layernorm_bias5, alloc144) + 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((2304, 768), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[102] + gv233: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc145: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv233, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc144, gpt_neox_layers_8_attention_query_key_value_weight5, gpt_neox_layers_8_attention_query_key_value_bias5, alloc145) + R.vm.kill_object(alloc144) + 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) + gv234: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape224: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc145, gv234, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc145) + gv235: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape225: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape224, gv235, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape224) + gv236: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc146: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv236, R.dtype("float16")) + _146: 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, alloc146) + R.vm.kill_object(reshape225) + gv237: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape226: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc146, gv237, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc146) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape227: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape226, gv238, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape226) + gpt_neox_layers_8_post_attention_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[100] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc147: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv239, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc143, gpt_neox_layers_8_post_attention_layernorm_weight5, gpt_neox_layers_8_post_attention_layernorm_bias5, alloc147) + 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((3072, 768), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[106] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc148: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv240, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc147, gpt_neox_layers_8_mlp_dense_h_to_4h_weight5, gpt_neox_layers_8_mlp_dense_h_to_4h_bias5, alloc148) + R.vm.kill_object(alloc147) + 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((768, 3072), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[108] + gv241: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc149: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv241, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc148, gpt_neox_layers_8_mlp_dense_4h_to_h_weight5, gpt_neox_layers_8_mlp_dense_4h_to_h_bias5, alloc149) + R.vm.kill_object(alloc148) + 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((768, 768), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[104] + gv242: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc150: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv242, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape227, gpt_neox_layers_8_attention_dense_weight5, gpt_neox_layers_8_attention_dense_bias5, alloc149, alloc143, alloc150) + R.vm.kill_object(alloc143) + R.vm.kill_object(reshape227) + R.vm.kill_object(alloc149) + 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((768,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[110] + gv243: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc151: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv243, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc150, gpt_neox_layers_9_input_layernorm_weight5, gpt_neox_layers_9_input_layernorm_bias5, alloc151) + 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((2304, 768), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[114] + gv244: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc152: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv244, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc151, gpt_neox_layers_9_attention_query_key_value_weight5, gpt_neox_layers_9_attention_query_key_value_bias5, alloc152) + R.vm.kill_object(alloc151) + 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) + gv245: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape228: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc152, gv245, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc152) + gv246: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape229: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape228, gv246, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape228) + gv247: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc153: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv247, R.dtype("float16")) + _153: 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, alloc153) + R.vm.kill_object(reshape229) + gv248: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape230: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc153, gv248, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc153) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape231: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape230, gv249, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape230) + gpt_neox_layers_9_post_attention_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[112] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc154: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv250, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc150, gpt_neox_layers_9_post_attention_layernorm_weight5, gpt_neox_layers_9_post_attention_layernorm_bias5, alloc154) + 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((3072, 768), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[118] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc155: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv251, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc154, gpt_neox_layers_9_mlp_dense_h_to_4h_weight5, gpt_neox_layers_9_mlp_dense_h_to_4h_bias5, alloc155) + R.vm.kill_object(alloc154) + 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((768, 3072), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[120] + gv252: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc156: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv252, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc155, gpt_neox_layers_9_mlp_dense_4h_to_h_weight5, gpt_neox_layers_9_mlp_dense_4h_to_h_bias5, alloc156) + R.vm.kill_object(alloc155) + 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((768, 768), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[116] + gv253: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc157: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv253, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape231, gpt_neox_layers_9_attention_dense_weight5, gpt_neox_layers_9_attention_dense_bias5, alloc156, alloc150, alloc157) + R.vm.kill_object(alloc150) + R.vm.kill_object(reshape231) + R.vm.kill_object(alloc156) + 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((768,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[122] + gv254: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc158: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv254, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc157, gpt_neox_layers_10_input_layernorm_weight5, gpt_neox_layers_10_input_layernorm_bias5, alloc158) + 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((2304, 768), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[126] + gv255: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc159: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv255, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc158, gpt_neox_layers_10_attention_query_key_value_weight5, gpt_neox_layers_10_attention_query_key_value_bias5, alloc159) + R.vm.kill_object(alloc158) + 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) + gv256: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape232: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc159, gv256, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc159) + gv257: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape233: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape232, gv257, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape232) + gv258: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc160: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv258, R.dtype("float16")) + _160: 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, alloc160) + R.vm.kill_object(reshape233) + gv259: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape234: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc160, gv259, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc160) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape235: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape234, gv260, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape234) + gpt_neox_layers_10_post_attention_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[124] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc161: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv261, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc157, gpt_neox_layers_10_post_attention_layernorm_weight5, gpt_neox_layers_10_post_attention_layernorm_bias5, alloc161) + 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((3072, 768), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[130] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc162: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv262, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc161, gpt_neox_layers_10_mlp_dense_h_to_4h_weight5, gpt_neox_layers_10_mlp_dense_h_to_4h_bias5, alloc162) + R.vm.kill_object(alloc161) + 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((768, 3072), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[132] + gv263: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc163: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv263, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc162, gpt_neox_layers_10_mlp_dense_4h_to_h_weight5, gpt_neox_layers_10_mlp_dense_4h_to_h_bias5, alloc163) + R.vm.kill_object(alloc162) + 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((768, 768), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[128] + gv264: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc164: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv264, R.dtype("float16")) + R.vm.kill_object(storage9) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape235, gpt_neox_layers_10_attention_dense_weight5, gpt_neox_layers_10_attention_dense_bias5, alloc163, alloc157, alloc164) + R.vm.kill_object(alloc157) + R.vm.kill_object(reshape235) + R.vm.kill_object(alloc163) + 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((768,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[134] + gv265: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc165: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv265, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc164, gpt_neox_layers_11_input_layernorm_weight5, gpt_neox_layers_11_input_layernorm_bias5, alloc165) + 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((2304, 768), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[138] + gv266: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc166: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv266, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc165, gpt_neox_layers_11_attention_query_key_value_weight5, gpt_neox_layers_11_attention_query_key_value_bias5, alloc166) + R.vm.kill_object(alloc165) + 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) + gv267: 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape236: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc166, gv267, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc166) + gv268: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape237: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape236, gv268, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape236) + gv269: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc167: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv269, R.dtype("float16")) + _167: 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, alloc167) + R.vm.kill_object(reshape237) + gv270: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape238: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc167, gv270, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc167) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape239: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape238, gv271, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape238) + gpt_neox_layers_11_post_attention_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[136] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc168: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv272, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc164, gpt_neox_layers_11_post_attention_layernorm_weight5, gpt_neox_layers_11_post_attention_layernorm_bias5, alloc168) + 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((3072, 768), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[142] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc169: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv273, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc168, gpt_neox_layers_11_mlp_dense_h_to_4h_weight5, gpt_neox_layers_11_mlp_dense_h_to_4h_bias5, alloc169) + R.vm.kill_object(alloc168) + 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((768, 3072), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[144] + gv274: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc170: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv274, R.dtype("float16")) + R.vm.kill_object(storage8) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc169, gpt_neox_layers_11_mlp_dense_4h_to_h_weight5, gpt_neox_layers_11_mlp_dense_4h_to_h_bias5, alloc170) + R.vm.kill_object(alloc169) + 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((768, 768), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[140] + gv275: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc171: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv275, R.dtype("float16")) + R.vm.kill_object(storage10) + cls.gpt_neox_q0f16_fused_NT_matmul6_add6_add9_add9(reshape239, gpt_neox_layers_11_attention_dense_weight5, gpt_neox_layers_11_attention_dense_bias5, alloc170, alloc164, alloc171) + R.vm.kill_object(alloc164) + R.vm.kill_object(reshape239) + R.vm.kill_object(alloc170) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_weight5) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias5) + gpt_neox_final_layer_norm_weight5: R.Tensor((768,), dtype="float16") = packed_params[145] + gpt_neox_final_layer_norm_bias5: R.Tensor((768,), dtype="float16") = packed_params[146] + gv276: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc172: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv276, R.dtype("float16")) + R.vm.kill_object(storage7) + cls.gpt_neox_q0f16_layer_norm1(alloc171, gpt_neox_final_layer_norm_weight5, gpt_neox_final_layer_norm_bias5, alloc172) + R.vm.kill_object(alloc171) + 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, 768), dtype="float16") = packed_params[147] + gv277: 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),)) + storage11: R.Object = R.vm.alloc_storage(gv277, R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(1), R.prim_value(1), sinfo_args=(R.Shape(ndim=3),)) + alloc173: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv278, R.dtype("float32")) + R.vm.kill_object(storage11) + cls.gpt_neox_q0f16_fused_NT_matmul9_cast5(alloc172, embed_out_weight5, alloc173) + R.vm.kill_object(alloc172) + R.vm.kill_object(embed_out_weight5) + gv5: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc173, paged_kv_cache + R.vm.kill_object(alloc173) + gv279: R.Tensor(dtype="float32", ndim=3) = gv5[0] + R.call_packed("vm.builtin.match_shape", gv279, 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,)) + gv569: 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", gv569, R.shape([0, 12]), R.prim_value(12), R.prim_value(12), R.prim_value(64), R.prim_value(1), R.prim_value(1), R.prim_value(10000), R.const(0.0, "float16"), cls.gpt_neox_q0f16_tir_kv_cache_transpose_append, cls.gpt_neox_q0f16_batch_prefill_paged_kv, cls.gpt_neox_q0f16_batch_decode_paged_kv, cls.gpt_neox_q0f16_batch_prefill_paged_kv_sliding_window, cls.gpt_neox_q0f16_batch_decode_paged_kv_sliding_window, cls.gpt_neox_q0f16_batch_prefill_ragged_kv, cls.gpt_neox_q0f16_merge_state_inplace, cls.gpt_neox_q0f16_fused_rope, cls.gpt_neox_q0f16_copy_single_page, cls.gpt_neox_q0f16_tir_kv_cache_debug_get_kv, cls.gpt_neox_q0f16_compact_kv_copy, cls.gpt_neox_q0f16_batch_tree_attn, cls.gpt_neox_q0f16_tree_attn_paged_kv, R.prim_value(0), R.prim_value(0), sinfo_args=(R.Object,)) + return paged_kv_cache + + @R.function + def decode(input_embed: R.Tensor((1, 1, 768), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor(("vocab_size", 768), 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, 768), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(148), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv288: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv288, 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, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv289: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[147] + R.call_packed("vm.builtin.check_tensor_info", gv289, 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, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), 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(768), R.str("ErrorContext(fn=decode, loc=param[0], param=input_embed, annotation=R.Tensor((1, 1, 768), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv288, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv289, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(0), R.prim_value(0), R.prim_value(768), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func3(shape_heap) + gpt_neox_layers_0_input_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[2] + storage15: R.Object = R.vm.alloc_storage(R.shape([1536]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc177: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 1, 768]), 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, alloc177) + 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((2304, 768), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[6] + storage16: R.Object = R.vm.alloc_storage(R.shape([4608]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc178: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc177, gpt_neox_layers_0_attention_query_key_value_weight2, gpt_neox_layers_0_attention_query_key_value_bias2, alloc178) + R.vm.kill_object(alloc177) + 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) + lv148: R.Tensor((1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc178, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc178) + alloc179: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) + _178: 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)), lv148, alloc179) + R.vm.kill_object(lv148) + lv149: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc179, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) + R.vm.kill_object(alloc179) + gpt_neox_layers_0_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[4] + alloc180: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), 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, alloc180) + 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((3072, 768), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[10] + storage17: R.Object = R.vm.alloc_storage(R.shape([6144]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc181: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage17, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc180, gpt_neox_layers_0_mlp_dense_h_to_4h_weight2, gpt_neox_layers_0_mlp_dense_h_to_4h_bias2, alloc181) + R.vm.kill_object(alloc180) + 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((768, 3072), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[12] + alloc182: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc181, gpt_neox_layers_0_mlp_dense_4h_to_h_weight2, gpt_neox_layers_0_mlp_dense_4h_to_h_bias2, alloc182) + R.vm.kill_object(alloc181) + 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((768, 768), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[8] + alloc183: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage17, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv149, gpt_neox_layers_0_attention_dense_weight2, gpt_neox_layers_0_attention_dense_bias2, alloc182, input_embed, alloc183) + R.vm.kill_object(lv149) + R.vm.kill_object(alloc182) + 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((768,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[14] + alloc184: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc183, gpt_neox_layers_1_input_layernorm_weight2, gpt_neox_layers_1_input_layernorm_bias2, alloc184) + 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((2304, 768), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[18] + alloc185: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc184, gpt_neox_layers_1_attention_query_key_value_weight2, gpt_neox_layers_1_attention_query_key_value_bias2, alloc185) + R.vm.kill_object(alloc184) + 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) + lv154: R.Tensor((1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc185, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc185) + alloc186: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) + _185: 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)), lv154, alloc186) + R.vm.kill_object(lv154) + lv155: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc186, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) + R.vm.kill_object(alloc186) + gpt_neox_layers_1_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[16] + alloc187: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc183, gpt_neox_layers_1_post_attention_layernorm_weight2, gpt_neox_layers_1_post_attention_layernorm_bias2, alloc187) + 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((3072, 768), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[22] + storage18: R.Object = R.vm.alloc_storage(R.shape([6144]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc188: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc187, gpt_neox_layers_1_mlp_dense_h_to_4h_weight2, gpt_neox_layers_1_mlp_dense_h_to_4h_bias2, alloc188) + R.vm.kill_object(alloc187) + 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((768, 3072), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[24] + alloc189: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc188, gpt_neox_layers_1_mlp_dense_4h_to_h_weight2, gpt_neox_layers_1_mlp_dense_4h_to_h_bias2, alloc189) + R.vm.kill_object(alloc188) + 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((768, 768), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[20] + alloc190: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv155, gpt_neox_layers_1_attention_dense_weight2, gpt_neox_layers_1_attention_dense_bias2, alloc189, alloc183, alloc190) + R.vm.kill_object(alloc183) + R.vm.kill_object(lv155) + R.vm.kill_object(alloc189) + 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((768,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[26] + alloc191: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc190, gpt_neox_layers_2_input_layernorm_weight2, gpt_neox_layers_2_input_layernorm_bias2, alloc191) + 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((2304, 768), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[30] + alloc192: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc191, gpt_neox_layers_2_attention_query_key_value_weight2, gpt_neox_layers_2_attention_query_key_value_bias2, alloc192) + R.vm.kill_object(alloc191) + 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) + lv160: R.Tensor((1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc192, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc192) + alloc193: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) + _192: 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)), lv160, alloc193) + R.vm.kill_object(lv160) + lv161: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc193, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) + R.vm.kill_object(alloc193) + gpt_neox_layers_2_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[28] + alloc194: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc190, gpt_neox_layers_2_post_attention_layernorm_weight2, gpt_neox_layers_2_post_attention_layernorm_bias2, alloc194) + 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((3072, 768), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[34] + alloc195: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage17, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc194, gpt_neox_layers_2_mlp_dense_h_to_4h_weight2, gpt_neox_layers_2_mlp_dense_h_to_4h_bias2, alloc195) + R.vm.kill_object(alloc194) + 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((768, 3072), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[36] + alloc196: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc195, gpt_neox_layers_2_mlp_dense_4h_to_h_weight2, gpt_neox_layers_2_mlp_dense_4h_to_h_bias2, alloc196) + R.vm.kill_object(alloc195) + 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((768, 768), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[32] + alloc197: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage17, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv161, gpt_neox_layers_2_attention_dense_weight2, gpt_neox_layers_2_attention_dense_bias2, alloc196, alloc190, alloc197) + R.vm.kill_object(alloc190) + R.vm.kill_object(lv161) + R.vm.kill_object(alloc196) + 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((768,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[38] + alloc198: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc197, gpt_neox_layers_3_input_layernorm_weight2, gpt_neox_layers_3_input_layernorm_bias2, alloc198) + 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((2304, 768), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[42] + alloc199: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc198, gpt_neox_layers_3_attention_query_key_value_weight2, gpt_neox_layers_3_attention_query_key_value_bias2, alloc199) + R.vm.kill_object(alloc198) + 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) + lv166: R.Tensor((1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc199, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc199) + alloc200: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) + _199: 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)), lv166, alloc200) + R.vm.kill_object(lv166) + lv167: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc200, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) + R.vm.kill_object(alloc200) + gpt_neox_layers_3_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[40] + alloc201: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc197, gpt_neox_layers_3_post_attention_layernorm_weight2, gpt_neox_layers_3_post_attention_layernorm_bias2, alloc201) + 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((3072, 768), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[46] + alloc202: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc201, gpt_neox_layers_3_mlp_dense_h_to_4h_weight2, gpt_neox_layers_3_mlp_dense_h_to_4h_bias2, alloc202) + R.vm.kill_object(alloc201) + 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((768, 3072), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[48] + alloc203: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc202, gpt_neox_layers_3_mlp_dense_4h_to_h_weight2, gpt_neox_layers_3_mlp_dense_4h_to_h_bias2, alloc203) + R.vm.kill_object(alloc202) + 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((768, 768), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[44] + alloc204: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv167, gpt_neox_layers_3_attention_dense_weight2, gpt_neox_layers_3_attention_dense_bias2, alloc203, alloc197, alloc204) + R.vm.kill_object(alloc197) + R.vm.kill_object(lv167) + R.vm.kill_object(alloc203) + 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((768,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[50] + alloc205: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc204, gpt_neox_layers_4_input_layernorm_weight2, gpt_neox_layers_4_input_layernorm_bias2, alloc205) + 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((2304, 768), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[54] + alloc206: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc205, gpt_neox_layers_4_attention_query_key_value_weight2, gpt_neox_layers_4_attention_query_key_value_bias2, alloc206) + R.vm.kill_object(alloc205) + 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) + lv172: R.Tensor((1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc206, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc206) + alloc207: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) + _206: 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)), lv172, alloc207) + R.vm.kill_object(lv172) + lv173: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc207, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) + R.vm.kill_object(alloc207) + gpt_neox_layers_4_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[52] + alloc208: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc204, gpt_neox_layers_4_post_attention_layernorm_weight2, gpt_neox_layers_4_post_attention_layernorm_bias2, alloc208) + 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((3072, 768), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[58] + alloc209: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage17, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc208, gpt_neox_layers_4_mlp_dense_h_to_4h_weight2, gpt_neox_layers_4_mlp_dense_h_to_4h_bias2, alloc209) + R.vm.kill_object(alloc208) + 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((768, 3072), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[60] + alloc210: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc209, gpt_neox_layers_4_mlp_dense_4h_to_h_weight2, gpt_neox_layers_4_mlp_dense_4h_to_h_bias2, alloc210) + R.vm.kill_object(alloc209) + 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((768, 768), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[56] + alloc211: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage17, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv173, gpt_neox_layers_4_attention_dense_weight2, gpt_neox_layers_4_attention_dense_bias2, alloc210, alloc204, alloc211) + R.vm.kill_object(alloc204) + R.vm.kill_object(lv173) + R.vm.kill_object(alloc210) + 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((768,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[62] + alloc212: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc211, gpt_neox_layers_5_input_layernorm_weight2, gpt_neox_layers_5_input_layernorm_bias2, alloc212) + 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((2304, 768), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[66] + alloc213: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc212, gpt_neox_layers_5_attention_query_key_value_weight2, gpt_neox_layers_5_attention_query_key_value_bias2, alloc213) + R.vm.kill_object(alloc212) + 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) + lv178: R.Tensor((1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc213, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc213) + alloc214: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) + _213: 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)), lv178, alloc214) + R.vm.kill_object(lv178) + lv179: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc214, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) + R.vm.kill_object(alloc214) + gpt_neox_layers_5_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[64] + alloc215: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc211, gpt_neox_layers_5_post_attention_layernorm_weight2, gpt_neox_layers_5_post_attention_layernorm_bias2, alloc215) + 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((3072, 768), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[70] + alloc216: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc215, gpt_neox_layers_5_mlp_dense_h_to_4h_weight2, gpt_neox_layers_5_mlp_dense_h_to_4h_bias2, alloc216) + R.vm.kill_object(alloc215) + 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((768, 3072), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[72] + alloc217: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc216, gpt_neox_layers_5_mlp_dense_4h_to_h_weight2, gpt_neox_layers_5_mlp_dense_4h_to_h_bias2, alloc217) + R.vm.kill_object(alloc216) + 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((768, 768), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[68] + alloc218: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv179, gpt_neox_layers_5_attention_dense_weight2, gpt_neox_layers_5_attention_dense_bias2, alloc217, alloc211, alloc218) + R.vm.kill_object(alloc211) + R.vm.kill_object(lv179) + R.vm.kill_object(alloc217) + 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((768,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[74] + alloc219: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc218, gpt_neox_layers_6_input_layernorm_weight2, gpt_neox_layers_6_input_layernorm_bias2, alloc219) + 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((2304, 768), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[78] + alloc220: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc219, gpt_neox_layers_6_attention_query_key_value_weight2, gpt_neox_layers_6_attention_query_key_value_bias2, alloc220) + R.vm.kill_object(alloc219) + 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) + lv184: R.Tensor((1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc220, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc220) + alloc221: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) + _220: 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)), lv184, alloc221) + R.vm.kill_object(lv184) + lv185: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc221, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) + R.vm.kill_object(alloc221) + gpt_neox_layers_6_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[76] + alloc222: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc218, gpt_neox_layers_6_post_attention_layernorm_weight2, gpt_neox_layers_6_post_attention_layernorm_bias2, alloc222) + 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((3072, 768), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[82] + alloc223: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage17, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc222, gpt_neox_layers_6_mlp_dense_h_to_4h_weight2, gpt_neox_layers_6_mlp_dense_h_to_4h_bias2, alloc223) + R.vm.kill_object(alloc222) + 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((768, 3072), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[84] + alloc224: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc223, gpt_neox_layers_6_mlp_dense_4h_to_h_weight2, gpt_neox_layers_6_mlp_dense_4h_to_h_bias2, alloc224) + R.vm.kill_object(alloc223) + 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((768, 768), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[80] + alloc225: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage17, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv185, gpt_neox_layers_6_attention_dense_weight2, gpt_neox_layers_6_attention_dense_bias2, alloc224, alloc218, alloc225) + R.vm.kill_object(alloc218) + R.vm.kill_object(lv185) + R.vm.kill_object(alloc224) + 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((768,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[86] + alloc226: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc225, gpt_neox_layers_7_input_layernorm_weight2, gpt_neox_layers_7_input_layernorm_bias2, alloc226) + 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((2304, 768), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[90] + alloc227: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc226, gpt_neox_layers_7_attention_query_key_value_weight2, gpt_neox_layers_7_attention_query_key_value_bias2, alloc227) + R.vm.kill_object(alloc226) + 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) + lv190: R.Tensor((1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc227, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc227) + alloc228: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) + _227: 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)), lv190, alloc228) + R.vm.kill_object(lv190) + lv191: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc228, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) + R.vm.kill_object(alloc228) + gpt_neox_layers_7_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[88] + alloc229: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc225, gpt_neox_layers_7_post_attention_layernorm_weight2, gpt_neox_layers_7_post_attention_layernorm_bias2, alloc229) + 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((3072, 768), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[94] + alloc230: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc229, gpt_neox_layers_7_mlp_dense_h_to_4h_weight2, gpt_neox_layers_7_mlp_dense_h_to_4h_bias2, alloc230) + R.vm.kill_object(alloc229) + 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((768, 3072), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[96] + alloc231: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc230, gpt_neox_layers_7_mlp_dense_4h_to_h_weight2, gpt_neox_layers_7_mlp_dense_4h_to_h_bias2, alloc231) + R.vm.kill_object(alloc230) + 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((768, 768), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[92] + alloc232: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv191, gpt_neox_layers_7_attention_dense_weight2, gpt_neox_layers_7_attention_dense_bias2, alloc231, alloc225, alloc232) + R.vm.kill_object(alloc225) + R.vm.kill_object(lv191) + R.vm.kill_object(alloc231) + 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((768,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[98] + alloc233: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc232, gpt_neox_layers_8_input_layernorm_weight2, gpt_neox_layers_8_input_layernorm_bias2, alloc233) + 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((2304, 768), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[102] + alloc234: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc233, gpt_neox_layers_8_attention_query_key_value_weight2, gpt_neox_layers_8_attention_query_key_value_bias2, alloc234) + R.vm.kill_object(alloc233) + 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) + lv196: R.Tensor((1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc234, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc234) + alloc235: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) + _234: 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)), lv196, alloc235) + R.vm.kill_object(lv196) + lv197: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc235, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) + R.vm.kill_object(alloc235) + gpt_neox_layers_8_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[100] + alloc236: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc232, gpt_neox_layers_8_post_attention_layernorm_weight2, gpt_neox_layers_8_post_attention_layernorm_bias2, alloc236) + 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((3072, 768), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[106] + alloc237: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage17, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc236, gpt_neox_layers_8_mlp_dense_h_to_4h_weight2, gpt_neox_layers_8_mlp_dense_h_to_4h_bias2, alloc237) + R.vm.kill_object(alloc236) + 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((768, 3072), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[108] + alloc238: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc237, gpt_neox_layers_8_mlp_dense_4h_to_h_weight2, gpt_neox_layers_8_mlp_dense_4h_to_h_bias2, alloc238) + R.vm.kill_object(alloc237) + 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((768, 768), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[104] + alloc239: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage17, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv197, gpt_neox_layers_8_attention_dense_weight2, gpt_neox_layers_8_attention_dense_bias2, alloc238, alloc232, alloc239) + R.vm.kill_object(alloc232) + R.vm.kill_object(lv197) + R.vm.kill_object(alloc238) + 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((768,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[110] + alloc240: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc239, gpt_neox_layers_9_input_layernorm_weight2, gpt_neox_layers_9_input_layernorm_bias2, alloc240) + 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((2304, 768), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[114] + alloc241: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc240, gpt_neox_layers_9_attention_query_key_value_weight2, gpt_neox_layers_9_attention_query_key_value_bias2, alloc241) + R.vm.kill_object(alloc240) + 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) + lv202: R.Tensor((1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc241, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc241) + alloc242: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) + _241: 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)), lv202, alloc242) + R.vm.kill_object(lv202) + lv203: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc242, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) + R.vm.kill_object(alloc242) + gpt_neox_layers_9_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[112] + alloc243: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc239, gpt_neox_layers_9_post_attention_layernorm_weight2, gpt_neox_layers_9_post_attention_layernorm_bias2, alloc243) + 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((3072, 768), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[118] + alloc244: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc243, gpt_neox_layers_9_mlp_dense_h_to_4h_weight2, gpt_neox_layers_9_mlp_dense_h_to_4h_bias2, alloc244) + R.vm.kill_object(alloc243) + 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((768, 3072), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[120] + alloc245: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc244, gpt_neox_layers_9_mlp_dense_4h_to_h_weight2, gpt_neox_layers_9_mlp_dense_4h_to_h_bias2, alloc245) + R.vm.kill_object(alloc244) + 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((768, 768), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[116] + alloc246: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv203, gpt_neox_layers_9_attention_dense_weight2, gpt_neox_layers_9_attention_dense_bias2, alloc245, alloc239, alloc246) + R.vm.kill_object(alloc239) + R.vm.kill_object(lv203) + R.vm.kill_object(alloc245) + 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((768,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[122] + alloc247: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc246, gpt_neox_layers_10_input_layernorm_weight2, gpt_neox_layers_10_input_layernorm_bias2, alloc247) + 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((2304, 768), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[126] + alloc248: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc247, gpt_neox_layers_10_attention_query_key_value_weight2, gpt_neox_layers_10_attention_query_key_value_bias2, alloc248) + R.vm.kill_object(alloc247) + 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) + lv208: R.Tensor((1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc248, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc248) + alloc249: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) + _248: 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)), lv208, alloc249) + R.vm.kill_object(lv208) + lv209: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc249, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) + R.vm.kill_object(alloc249) + gpt_neox_layers_10_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[124] + alloc250: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc246, gpt_neox_layers_10_post_attention_layernorm_weight2, gpt_neox_layers_10_post_attention_layernorm_bias2, alloc250) + 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((3072, 768), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[130] + alloc251: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage17, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc250, gpt_neox_layers_10_mlp_dense_h_to_4h_weight2, gpt_neox_layers_10_mlp_dense_h_to_4h_bias2, alloc251) + R.vm.kill_object(alloc250) + 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((768, 3072), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[132] + alloc252: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc251, gpt_neox_layers_10_mlp_dense_4h_to_h_weight2, gpt_neox_layers_10_mlp_dense_4h_to_h_bias2, alloc252) + R.vm.kill_object(alloc251) + 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((768, 768), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[128] + alloc253: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage17, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + R.vm.kill_object(storage17) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv209, gpt_neox_layers_10_attention_dense_weight2, gpt_neox_layers_10_attention_dense_bias2, alloc252, alloc246, alloc253) + R.vm.kill_object(alloc246) + R.vm.kill_object(lv209) + R.vm.kill_object(alloc252) + 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((768,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[134] + alloc254: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc253, gpt_neox_layers_11_input_layernorm_weight2, gpt_neox_layers_11_input_layernorm_bias2, alloc254) + 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((2304, 768), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[138] + alloc255: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul10_add10(alloc254, gpt_neox_layers_11_attention_query_key_value_weight2, gpt_neox_layers_11_attention_query_key_value_bias2, alloc255) + R.vm.kill_object(alloc254) + 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) + lv214: R.Tensor((1, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc255, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc255) + alloc256: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) + _255: 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)), lv214, alloc256) + R.vm.kill_object(lv214) + lv215: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc256, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) + R.vm.kill_object(alloc256) + gpt_neox_layers_11_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[136] + alloc257: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm2(alloc253, gpt_neox_layers_11_post_attention_layernorm_weight2, gpt_neox_layers_11_post_attention_layernorm_bias2, alloc257) + 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((3072, 768), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[142] + alloc258: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul12_add12_gelu2_cast6(alloc257, gpt_neox_layers_11_mlp_dense_h_to_4h_weight2, gpt_neox_layers_11_mlp_dense_h_to_4h_bias2, alloc258) + R.vm.kill_object(alloc257) + 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((768, 3072), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[144] + alloc259: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage16, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + R.vm.kill_object(storage16) + cls.gpt_neox_q0f16_fused_NT_matmul13_add13_cast7(alloc258, gpt_neox_layers_11_mlp_dense_4h_to_h_weight2, gpt_neox_layers_11_mlp_dense_4h_to_h_bias2, alloc259) + R.vm.kill_object(alloc258) + 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((768, 768), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[140] + alloc260: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage18, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + R.vm.kill_object(storage18) + cls.gpt_neox_q0f16_fused_NT_matmul11_add11_add14_add14(lv215, gpt_neox_layers_11_attention_dense_weight2, gpt_neox_layers_11_attention_dense_bias2, alloc259, alloc253, alloc260) + R.vm.kill_object(alloc253) + R.vm.kill_object(lv215) + R.vm.kill_object(alloc259) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_weight2) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias2) + gpt_neox_final_layer_norm_weight2: R.Tensor((768,), dtype="float16") = packed_params[145] + gpt_neox_final_layer_norm_bias2: R.Tensor((768,), dtype="float16") = packed_params[146] + alloc261: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage15, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + R.vm.kill_object(storage15) + cls.gpt_neox_q0f16_layer_norm2(alloc260, gpt_neox_final_layer_norm_weight2, gpt_neox_final_layer_norm_bias2, alloc261) + R.vm.kill_object(alloc260) + 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, 768), dtype="float16") = packed_params[147] + gv290: 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),)) + storage19: R.Object = R.vm.alloc_storage(gv290, R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), sinfo_args=(R.Shape(ndim=3),)) + alloc262: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage19, R.prim_value(0), gv291, R.dtype("float32")) + R.vm.kill_object(storage19) + cls.gpt_neox_q0f16_fused_NT_matmul14_cast8(alloc261, embed_out_weight2, alloc262) + R.vm.kill_object(alloc261) + R.vm.kill_object(embed_out_weight2) + gv2: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc262, paged_kv_cache + R.vm.kill_object(alloc262) + gv292: R.Tensor(dtype="float32", ndim=3) = gv2[0] + R.call_packed("vm.builtin.match_shape", gv292, 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", 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor(("vocab_size", 768), dtype="float16"))) -> R.Tensor(("seq_len", 768), 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(148), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv: R.Tensor((vocab_size, 768), 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=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv1: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[147] + R.call_packed("vm.builtin.check_tensor_info", gv1, 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, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), 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", gv, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(768), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), 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(768), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gpt_neox_embed_in_weight: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[0] + storage: R.Object = R.vm.alloc_storage(R.shape([3145728]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv2: R.Shape(ndim=2) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=2),)) + alloc: R.Tensor(dtype="float16", ndim=2) = R.vm.alloc_tensor(storage, R.prim_value(0), gv2, R.dtype("float16")) + R.vm.kill_object(storage) + cls.gpt_neox_q0f16_take1(gpt_neox_embed_in_weight, input_ids, alloc) + R.vm.kill_object(gpt_neox_embed_in_weight) + R.call_packed("vm.builtin.match_shape", alloc, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(0), R.prim_value(0), R.prim_value(768), R.str("ErrorContext(fn=embed, loc=return, annotation=R.Tensor((seq_len, 768), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + return alloc + + @R.function + def prefill(input_embed: R.Tensor((1, "seq_len", 768), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 768), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 768), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 768), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 3072), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor(("vocab_size", 768), 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, 768), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tuple_info", packed_params, R.prim_value(148), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv293: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[0] + R.call_packed("vm.builtin.check_tensor_info", gv293, 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, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + gv294: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[147] + R.call_packed("vm.builtin.check_tensor_info", gv294, 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, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), 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(768), R.str("ErrorContext(fn=prefill, loc=param[0], param=input_embed, annotation=R.Tensor((1, seq_len, 768), dtype=\"float16\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv293, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(768), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", gv294, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(768), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 768), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 768), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 768), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 3072), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 768), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func4(shape_heap) + gpt_neox_layers_0_input_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[2] + storage20: R.Object = R.vm.alloc_storage(R.shape([3145728]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc263: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv295, 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, alloc263) + 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((2304, 768), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[6] + storage21: R.Object = R.vm.alloc_storage(R.shape([9437184]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv296: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc264: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv296, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc263, gpt_neox_layers_0_attention_query_key_value_weight1, gpt_neox_layers_0_attention_query_key_value_bias1, alloc264) + R.vm.kill_object(alloc263) + 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) + gv297: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc264, gv297, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc264) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape1: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape, gv298, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape) + gv299: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc265: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv299, R.dtype("float16")) + _264: 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, alloc265) + R.vm.kill_object(reshape1) + gv300: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape2: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc265, gv300, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc265) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape3: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape2, gv301, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape2) + gpt_neox_layers_0_post_attention_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[4] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc266: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv302, 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, alloc266) + 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((3072, 768), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[10] + storage22: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc267: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv303, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc266, gpt_neox_layers_0_mlp_dense_h_to_4h_weight1, gpt_neox_layers_0_mlp_dense_h_to_4h_bias1, alloc267) + R.vm.kill_object(alloc266) + 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((768, 3072), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[12] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc268: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv304, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc267, gpt_neox_layers_0_mlp_dense_4h_to_h_weight1, gpt_neox_layers_0_mlp_dense_4h_to_h_bias1, alloc268) + R.vm.kill_object(alloc267) + 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((768, 768), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[8] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc269: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv305, 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, alloc268, input_embed, alloc269) + R.vm.kill_object(reshape3) + R.vm.kill_object(alloc268) + 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((768,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[14] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc270: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv306, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc269, gpt_neox_layers_1_input_layernorm_weight1, gpt_neox_layers_1_input_layernorm_bias1, alloc270) + 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((2304, 768), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[18] + gv307: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc271: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv307, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc270, gpt_neox_layers_1_attention_query_key_value_weight1, gpt_neox_layers_1_attention_query_key_value_bias1, alloc271) + R.vm.kill_object(alloc270) + 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) + gv308: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape4: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc271, gv308, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc271) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape5: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape4, gv309, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape4) + gv310: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc272: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv310, R.dtype("float16")) + _271: 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, alloc272) + R.vm.kill_object(reshape5) + gv311: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape6: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc272, gv311, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc272) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape7: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape6, gv312, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape6) + gpt_neox_layers_1_post_attention_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[16] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc273: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv313, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc269, gpt_neox_layers_1_post_attention_layernorm_weight1, gpt_neox_layers_1_post_attention_layernorm_bias1, alloc273) + 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((3072, 768), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[22] + storage23: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc274: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv314, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc273, gpt_neox_layers_1_mlp_dense_h_to_4h_weight1, gpt_neox_layers_1_mlp_dense_h_to_4h_bias1, alloc274) + R.vm.kill_object(alloc273) + 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((768, 3072), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[24] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc275: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv315, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc274, gpt_neox_layers_1_mlp_dense_4h_to_h_weight1, gpt_neox_layers_1_mlp_dense_4h_to_h_bias1, alloc275) + R.vm.kill_object(alloc274) + 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((768, 768), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[20] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc276: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv316, 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, alloc275, alloc269, alloc276) + R.vm.kill_object(alloc269) + R.vm.kill_object(reshape7) + R.vm.kill_object(alloc275) + 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((768,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[26] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc277: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv317, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc276, gpt_neox_layers_2_input_layernorm_weight1, gpt_neox_layers_2_input_layernorm_bias1, alloc277) + 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((2304, 768), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[30] + gv318: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc278: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv318, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc277, gpt_neox_layers_2_attention_query_key_value_weight1, gpt_neox_layers_2_attention_query_key_value_bias1, alloc278) + R.vm.kill_object(alloc277) + 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) + gv319: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape8: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc278, gv319, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc278) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape9: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape8, gv320, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape8) + gv321: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc279: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv321, R.dtype("float16")) + _278: 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, alloc279) + R.vm.kill_object(reshape9) + gv322: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape10: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc279, gv322, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc279) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape11: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape10, gv323, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape10) + gpt_neox_layers_2_post_attention_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[28] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc280: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv324, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc276, gpt_neox_layers_2_post_attention_layernorm_weight1, gpt_neox_layers_2_post_attention_layernorm_bias1, alloc280) + 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((3072, 768), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[34] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc281: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv325, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc280, gpt_neox_layers_2_mlp_dense_h_to_4h_weight1, gpt_neox_layers_2_mlp_dense_h_to_4h_bias1, alloc281) + R.vm.kill_object(alloc280) + 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((768, 3072), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[36] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc282: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv326, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc281, gpt_neox_layers_2_mlp_dense_4h_to_h_weight1, gpt_neox_layers_2_mlp_dense_4h_to_h_bias1, alloc282) + R.vm.kill_object(alloc281) + 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((768, 768), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[32] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc283: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv327, 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, alloc282, alloc276, alloc283) + R.vm.kill_object(alloc276) + R.vm.kill_object(reshape11) + R.vm.kill_object(alloc282) + 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((768,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[38] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc284: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv328, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc283, gpt_neox_layers_3_input_layernorm_weight1, gpt_neox_layers_3_input_layernorm_bias1, alloc284) + 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((2304, 768), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[42] + gv329: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc285: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv329, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc284, gpt_neox_layers_3_attention_query_key_value_weight1, gpt_neox_layers_3_attention_query_key_value_bias1, alloc285) + R.vm.kill_object(alloc284) + 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) + gv330: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape12: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc285, gv330, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc285) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape13: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape12, gv331, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape12) + gv332: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc286: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv332, R.dtype("float16")) + _285: 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, alloc286) + R.vm.kill_object(reshape13) + gv333: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape14: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc286, gv333, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc286) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape15: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape14, gv334, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape14) + gpt_neox_layers_3_post_attention_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[40] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc287: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv335, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc283, gpt_neox_layers_3_post_attention_layernorm_weight1, gpt_neox_layers_3_post_attention_layernorm_bias1, alloc287) + 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((3072, 768), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[46] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc288: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv336, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc287, gpt_neox_layers_3_mlp_dense_h_to_4h_weight1, gpt_neox_layers_3_mlp_dense_h_to_4h_bias1, alloc288) + R.vm.kill_object(alloc287) + 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((768, 3072), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[48] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc289: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv337, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc288, gpt_neox_layers_3_mlp_dense_4h_to_h_weight1, gpt_neox_layers_3_mlp_dense_4h_to_h_bias1, alloc289) + R.vm.kill_object(alloc288) + 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((768, 768), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[44] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc290: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv338, 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, alloc289, alloc283, alloc290) + R.vm.kill_object(alloc283) + R.vm.kill_object(reshape15) + R.vm.kill_object(alloc289) + 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((768,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[50] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc291: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv339, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc290, gpt_neox_layers_4_input_layernorm_weight1, gpt_neox_layers_4_input_layernorm_bias1, alloc291) + 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((2304, 768), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[54] + gv340: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc292: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv340, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc291, gpt_neox_layers_4_attention_query_key_value_weight1, gpt_neox_layers_4_attention_query_key_value_bias1, alloc292) + R.vm.kill_object(alloc291) + 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) + gv341: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape16: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc292, gv341, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc292) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape17: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape16, gv342, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape16) + gv343: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc293: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv343, R.dtype("float16")) + _292: 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, alloc293) + R.vm.kill_object(reshape17) + gv344: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape18: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc293, gv344, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc293) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape19: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape18, gv345, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape18) + gpt_neox_layers_4_post_attention_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[52] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc294: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv346, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc290, gpt_neox_layers_4_post_attention_layernorm_weight1, gpt_neox_layers_4_post_attention_layernorm_bias1, alloc294) + 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((3072, 768), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[58] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc295: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv347, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc294, gpt_neox_layers_4_mlp_dense_h_to_4h_weight1, gpt_neox_layers_4_mlp_dense_h_to_4h_bias1, alloc295) + R.vm.kill_object(alloc294) + 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((768, 3072), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[60] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc296: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv348, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc295, gpt_neox_layers_4_mlp_dense_4h_to_h_weight1, gpt_neox_layers_4_mlp_dense_4h_to_h_bias1, alloc296) + R.vm.kill_object(alloc295) + 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((768, 768), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[56] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc297: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv349, 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, alloc296, alloc290, alloc297) + R.vm.kill_object(alloc290) + R.vm.kill_object(reshape19) + R.vm.kill_object(alloc296) + 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((768,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[62] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc298: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv350, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc297, gpt_neox_layers_5_input_layernorm_weight1, gpt_neox_layers_5_input_layernorm_bias1, alloc298) + 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((2304, 768), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[66] + gv351: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc299: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv351, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc298, gpt_neox_layers_5_attention_query_key_value_weight1, gpt_neox_layers_5_attention_query_key_value_bias1, alloc299) + R.vm.kill_object(alloc298) + 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) + gv352: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape20: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc299, gv352, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc299) + 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(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape21: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape20, gv353, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape20) + gv354: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc300: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv354, R.dtype("float16")) + _299: 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, alloc300) + R.vm.kill_object(reshape21) + gv355: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape22: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc300, gv355, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc300) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape23: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape22, gv356, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape22) + gpt_neox_layers_5_post_attention_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[64] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc301: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv357, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc297, gpt_neox_layers_5_post_attention_layernorm_weight1, gpt_neox_layers_5_post_attention_layernorm_bias1, alloc301) + 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((3072, 768), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[70] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc302: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv358, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc301, gpt_neox_layers_5_mlp_dense_h_to_4h_weight1, gpt_neox_layers_5_mlp_dense_h_to_4h_bias1, alloc302) + R.vm.kill_object(alloc301) + 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((768, 3072), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[72] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc303: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv359, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc302, gpt_neox_layers_5_mlp_dense_4h_to_h_weight1, gpt_neox_layers_5_mlp_dense_4h_to_h_bias1, alloc303) + R.vm.kill_object(alloc302) + 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((768, 768), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[68] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc304: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv360, 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, alloc303, alloc297, alloc304) + R.vm.kill_object(alloc297) + R.vm.kill_object(reshape23) + R.vm.kill_object(alloc303) + 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((768,), dtype="float16") = packed_params[73] + gpt_neox_layers_6_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[74] + gv361: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc305: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv361, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc304, gpt_neox_layers_6_input_layernorm_weight1, gpt_neox_layers_6_input_layernorm_bias1, alloc305) + 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((2304, 768), dtype="float16") = packed_params[77] + gpt_neox_layers_6_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[78] + 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(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc306: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv362, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc305, gpt_neox_layers_6_attention_query_key_value_weight1, gpt_neox_layers_6_attention_query_key_value_bias1, alloc306) + R.vm.kill_object(alloc305) + 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) + gv363: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape24: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc306, gv363, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc306) + gv364: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape25: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape24, gv364, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape24) + gv365: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc307: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv365, R.dtype("float16")) + _306: 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, alloc307) + R.vm.kill_object(reshape25) + gv366: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape26: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc307, gv366, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc307) + gv367: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + reshape27: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape26, gv367, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape26) + gpt_neox_layers_6_post_attention_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[75] + gpt_neox_layers_6_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[76] + gv368: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc308: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv368, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc304, gpt_neox_layers_6_post_attention_layernorm_weight1, gpt_neox_layers_6_post_attention_layernorm_bias1, alloc308) + 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((3072, 768), dtype="float16") = packed_params[81] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[82] + gv369: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc309: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv369, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc308, gpt_neox_layers_6_mlp_dense_h_to_4h_weight1, gpt_neox_layers_6_mlp_dense_h_to_4h_bias1, alloc309) + R.vm.kill_object(alloc308) + 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((768, 3072), dtype="float16") = packed_params[83] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[84] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc310: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv370, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc309, gpt_neox_layers_6_mlp_dense_4h_to_h_weight1, gpt_neox_layers_6_mlp_dense_4h_to_h_bias1, alloc310) + R.vm.kill_object(alloc309) + 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((768, 768), dtype="float16") = packed_params[79] + gpt_neox_layers_6_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[80] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc311: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv371, 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, alloc310, alloc304, alloc311) + R.vm.kill_object(alloc304) + R.vm.kill_object(reshape27) + R.vm.kill_object(alloc310) + 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((768,), dtype="float16") = packed_params[85] + gpt_neox_layers_7_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[86] + gv372: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc312: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv372, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc311, gpt_neox_layers_7_input_layernorm_weight1, gpt_neox_layers_7_input_layernorm_bias1, alloc312) + 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((2304, 768), dtype="float16") = packed_params[89] + gpt_neox_layers_7_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[90] + gv373: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc313: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv373, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc312, gpt_neox_layers_7_attention_query_key_value_weight1, gpt_neox_layers_7_attention_query_key_value_bias1, alloc313) + R.vm.kill_object(alloc312) + 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) + gv374: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape28: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc313, gv374, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc313) + gv375: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape29: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape28, gv375, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape28) + gv376: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc314: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv376, R.dtype("float16")) + _313: 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, alloc314) + R.vm.kill_object(reshape29) + gv377: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape30: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc314, gv377, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc314) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape31: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape30, gv378, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape30) + gpt_neox_layers_7_post_attention_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[87] + gpt_neox_layers_7_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[88] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc315: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv379, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc311, gpt_neox_layers_7_post_attention_layernorm_weight1, gpt_neox_layers_7_post_attention_layernorm_bias1, alloc315) + 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((3072, 768), dtype="float16") = packed_params[93] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[94] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc316: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv380, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc315, gpt_neox_layers_7_mlp_dense_h_to_4h_weight1, gpt_neox_layers_7_mlp_dense_h_to_4h_bias1, alloc316) + R.vm.kill_object(alloc315) + 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((768, 3072), dtype="float16") = packed_params[95] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[96] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc317: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv381, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc316, gpt_neox_layers_7_mlp_dense_4h_to_h_weight1, gpt_neox_layers_7_mlp_dense_4h_to_h_bias1, alloc317) + R.vm.kill_object(alloc316) + 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((768, 768), dtype="float16") = packed_params[91] + gpt_neox_layers_7_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[92] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc318: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv382, 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, alloc317, alloc311, alloc318) + R.vm.kill_object(alloc311) + R.vm.kill_object(reshape31) + R.vm.kill_object(alloc317) + 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((768,), dtype="float16") = packed_params[97] + gpt_neox_layers_8_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[98] + gv383: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc319: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv383, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc318, gpt_neox_layers_8_input_layernorm_weight1, gpt_neox_layers_8_input_layernorm_bias1, alloc319) + 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((2304, 768), dtype="float16") = packed_params[101] + gpt_neox_layers_8_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[102] + gv384: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc320: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv384, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc319, gpt_neox_layers_8_attention_query_key_value_weight1, gpt_neox_layers_8_attention_query_key_value_bias1, alloc320) + R.vm.kill_object(alloc319) + 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) + gv385: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape32: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc320, gv385, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc320) + gv386: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape33: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape32, gv386, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape32) + gv387: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc321: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv387, R.dtype("float16")) + _320: 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, alloc321) + R.vm.kill_object(reshape33) + gv388: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape34: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc321, gv388, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc321) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape35: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape34, gv389, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape34) + gpt_neox_layers_8_post_attention_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[99] + gpt_neox_layers_8_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[100] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc322: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv390, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc318, gpt_neox_layers_8_post_attention_layernorm_weight1, gpt_neox_layers_8_post_attention_layernorm_bias1, alloc322) + 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((3072, 768), dtype="float16") = packed_params[105] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[106] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc323: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv391, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc322, gpt_neox_layers_8_mlp_dense_h_to_4h_weight1, gpt_neox_layers_8_mlp_dense_h_to_4h_bias1, alloc323) + R.vm.kill_object(alloc322) + 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((768, 3072), dtype="float16") = packed_params[107] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[108] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc324: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv392, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc323, gpt_neox_layers_8_mlp_dense_4h_to_h_weight1, gpt_neox_layers_8_mlp_dense_4h_to_h_bias1, alloc324) + R.vm.kill_object(alloc323) + 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((768, 768), dtype="float16") = packed_params[103] + gpt_neox_layers_8_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[104] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc325: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv393, 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, alloc324, alloc318, alloc325) + R.vm.kill_object(alloc318) + R.vm.kill_object(reshape35) + R.vm.kill_object(alloc324) + 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((768,), dtype="float16") = packed_params[109] + gpt_neox_layers_9_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[110] + gv394: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc326: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv394, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc325, gpt_neox_layers_9_input_layernorm_weight1, gpt_neox_layers_9_input_layernorm_bias1, alloc326) + 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((2304, 768), dtype="float16") = packed_params[113] + gpt_neox_layers_9_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[114] + gv395: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc327: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv395, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc326, gpt_neox_layers_9_attention_query_key_value_weight1, gpt_neox_layers_9_attention_query_key_value_bias1, alloc327) + R.vm.kill_object(alloc326) + 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) + gv396: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape36: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc327, gv396, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc327) + gv397: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape37: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape36, gv397, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape36) + gv398: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc328: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv398, R.dtype("float16")) + _327: 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, alloc328) + R.vm.kill_object(reshape37) + gv399: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape38: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc328, gv399, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc328) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape39: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape38, gv400, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape38) + gpt_neox_layers_9_post_attention_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[111] + gpt_neox_layers_9_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[112] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc329: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv401, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc325, gpt_neox_layers_9_post_attention_layernorm_weight1, gpt_neox_layers_9_post_attention_layernorm_bias1, alloc329) + 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((3072, 768), dtype="float16") = packed_params[117] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[118] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc330: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv402, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc329, gpt_neox_layers_9_mlp_dense_h_to_4h_weight1, gpt_neox_layers_9_mlp_dense_h_to_4h_bias1, alloc330) + R.vm.kill_object(alloc329) + 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((768, 3072), dtype="float16") = packed_params[119] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[120] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc331: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv403, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc330, gpt_neox_layers_9_mlp_dense_4h_to_h_weight1, gpt_neox_layers_9_mlp_dense_4h_to_h_bias1, alloc331) + R.vm.kill_object(alloc330) + 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((768, 768), dtype="float16") = packed_params[115] + gpt_neox_layers_9_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[116] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc332: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv404, 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, alloc331, alloc325, alloc332) + R.vm.kill_object(alloc325) + R.vm.kill_object(reshape39) + R.vm.kill_object(alloc331) + 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((768,), dtype="float16") = packed_params[121] + gpt_neox_layers_10_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[122] + gv405: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc333: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv405, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc332, gpt_neox_layers_10_input_layernorm_weight1, gpt_neox_layers_10_input_layernorm_bias1, alloc333) + 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((2304, 768), dtype="float16") = packed_params[125] + gpt_neox_layers_10_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[126] + gv406: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc334: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv406, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc333, gpt_neox_layers_10_attention_query_key_value_weight1, gpt_neox_layers_10_attention_query_key_value_bias1, alloc334) + R.vm.kill_object(alloc333) + 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) + gv407: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape40: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc334, gv407, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc334) + gv408: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape41: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape40, gv408, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape40) + gv409: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc335: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv409, R.dtype("float16")) + _334: 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, alloc335) + R.vm.kill_object(reshape41) + gv410: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape42: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc335, gv410, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc335) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape43: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape42, gv411, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape42) + gpt_neox_layers_10_post_attention_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[123] + gpt_neox_layers_10_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[124] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc336: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv412, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc332, gpt_neox_layers_10_post_attention_layernorm_weight1, gpt_neox_layers_10_post_attention_layernorm_bias1, alloc336) + 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((3072, 768), dtype="float16") = packed_params[129] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[130] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc337: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv413, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc336, gpt_neox_layers_10_mlp_dense_h_to_4h_weight1, gpt_neox_layers_10_mlp_dense_h_to_4h_bias1, alloc337) + R.vm.kill_object(alloc336) + 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((768, 3072), dtype="float16") = packed_params[131] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[132] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc338: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv414, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc337, gpt_neox_layers_10_mlp_dense_4h_to_h_weight1, gpt_neox_layers_10_mlp_dense_4h_to_h_bias1, alloc338) + R.vm.kill_object(alloc337) + 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((768, 768), dtype="float16") = packed_params[127] + gpt_neox_layers_10_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[128] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc339: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv415, R.dtype("float16")) + R.vm.kill_object(storage22) + 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, alloc338, alloc332, alloc339) + R.vm.kill_object(alloc332) + R.vm.kill_object(reshape43) + R.vm.kill_object(alloc338) + 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((768,), dtype="float16") = packed_params[133] + gpt_neox_layers_11_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[134] + gv416: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc340: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv416, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc339, gpt_neox_layers_11_input_layernorm_weight1, gpt_neox_layers_11_input_layernorm_bias1, alloc340) + 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((2304, 768), dtype="float16") = packed_params[137] + gpt_neox_layers_11_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[138] + gv417: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) + alloc341: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv417, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul5_add5(alloc340, gpt_neox_layers_11_attention_query_key_value_weight1, gpt_neox_layers_11_attention_query_key_value_bias1, alloc341) + R.vm.kill_object(alloc340) + 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) + gv418: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape44: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc341, gv418, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(alloc341) + gv419: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(36), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + reshape45: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape44, gv419, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) + R.vm.kill_object(reshape44) + gv420: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) + alloc342: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv420, R.dtype("float16")) + _341: 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, alloc342) + R.vm.kill_object(reshape45) + gv421: 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=4),)) + reshape46: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc342, gv421, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) + R.vm.kill_object(alloc342) + 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(768), sinfo_args=(R.Shape(ndim=3),)) + reshape47: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape46, gv422, sinfo_args=(R.Tensor((1, seq_len, 768), dtype="float16"),)) + R.vm.kill_object(reshape46) + gpt_neox_layers_11_post_attention_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[135] + gpt_neox_layers_11_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[136] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc343: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv423, R.dtype("float16")) + cls.gpt_neox_q0f16_layer_norm1(alloc339, gpt_neox_layers_11_post_attention_layernorm_weight1, gpt_neox_layers_11_post_attention_layernorm_bias1, alloc343) + 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((3072, 768), dtype="float16") = packed_params[141] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[142] + 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(3072), sinfo_args=(R.Shape(ndim=3),)) + alloc344: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv424, R.dtype("float16")) + cls.gpt_neox_q0f16_fused_NT_matmul7_add7_gelu1_cast3(alloc343, gpt_neox_layers_11_mlp_dense_h_to_4h_weight1, gpt_neox_layers_11_mlp_dense_h_to_4h_bias1, alloc344) + R.vm.kill_object(alloc343) + 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((768, 3072), dtype="float16") = packed_params[143] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[144] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc345: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv425, R.dtype("float16")) + R.vm.kill_object(storage21) + cls.gpt_neox_q0f16_fused_NT_matmul8_add8_cast4(alloc344, gpt_neox_layers_11_mlp_dense_4h_to_h_weight1, gpt_neox_layers_11_mlp_dense_4h_to_h_bias1, alloc345) + R.vm.kill_object(alloc344) + 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((768, 768), dtype="float16") = packed_params[139] + gpt_neox_layers_11_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[140] + 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(768), sinfo_args=(R.Shape(ndim=3),)) + alloc346: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv426, R.dtype("float16")) + R.vm.kill_object(storage23) + 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, alloc345, alloc339, alloc346) + R.vm.kill_object(alloc339) + R.vm.kill_object(reshape47) + R.vm.kill_object(alloc345) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_weight1) + R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias1) + gpt_neox_final_layer_norm_weight1: R.Tensor((768,), dtype="float16") = packed_params[145] + gpt_neox_final_layer_norm_bias1: R.Tensor((768,), dtype="float16") = packed_params[146] + gv427: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) + alloc347: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv427, R.dtype("float16")) + R.vm.kill_object(storage20) + cls.gpt_neox_q0f16_layer_norm1(alloc346, gpt_neox_final_layer_norm_weight1, gpt_neox_final_layer_norm_bias1, alloc347) + R.vm.kill_object(alloc346) + R.vm.kill_object(gpt_neox_final_layer_norm_weight1) + R.vm.kill_object(gpt_neox_final_layer_norm_bias1) + storage24: R.Object = R.vm.alloc_storage(R.shape([1536]), R.prim_value(0), R.dtype("uint8"), R.str("global")) + alloc348: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage24, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) + R.vm.kill_object(storage24) + cls.gpt_neox_q0f16_index(alloc347, alloc348) + R.vm.kill_object(alloc347) + embed_out_weight1: R.Tensor((vocab_size, 768), dtype="float16") = packed_params[147] + gv428: R.Shape(ndim=1) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(1), R.prim_value(1), R.prim_value(3), sinfo_args=(R.Shape(ndim=1),)) + storage25: R.Object = R.vm.alloc_storage(gv428, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv429: R.Shape(ndim=3) = 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),)) + alloc349: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv429, R.dtype("float32")) + R.vm.kill_object(storage25) + cls.gpt_neox_q0f16_fused_NT_matmul14_cast8(alloc348, embed_out_weight1, alloc349) + R.vm.kill_object(alloc348) + R.vm.kill_object(embed_out_weight1) + gv1: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc349, paged_kv_cache + R.vm.kill_object(alloc349) + gv430: R.Tensor(dtype="float32", ndim=3) = gv1[0] + R.call_packed("vm.builtin.match_shape", gv430, shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(0), R.prim_value(1), R.prim_value(3), R.prim_value(1), R.str("ErrorContext(fn=prefill, loc=return, annotation=R.Tuple(R.Tensor((1, 1, vocab_size), dtype=\"float32\"), R.Object)) "), sinfo_args=(R.Tuple,)) + return gv1 + + @R.function + def softmax_with_temperature(logits: R.Tensor(("batch_size", 1, "vocab_size"), dtype="float32"), temperature: R.Tensor(("batch_size",), dtype="float32")) -> R.Tensor(("batch_size", 1, "vocab_size"), dtype="float32"): + batch_size = T.int64(is_size_var=True) + vocab_size = T.int64(is_size_var=True) + R.func_attr({"relax.force_pure": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) + cls = Module + shape_heap: R.Tensor(dtype="int64", ndim=1) = R.call_builtin_with_ctx("vm.builtin.alloc_shape_heap", (R.prim_value(5),), sinfo_args=(R.Tensor(dtype="int64", ndim=1),)) + R.call_packed("vm.builtin.check_tensor_info", logits, R.prim_value(3), R.dtype("float32"), R.str("ErrorContext(fn=softmax_with_temperature, loc=param[0], param=logits, annotation=R.Tensor((batch_size, 1, vocab_size), dtype=\"float32\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.check_tensor_info", temperature, R.prim_value(1), R.dtype("float32"), R.str("ErrorContext(fn=softmax_with_temperature, loc=param[1], param=temperature, annotation=R.Tensor((batch_size,), dtype=\"float32\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", logits, shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(1), R.str("ErrorContext(fn=softmax_with_temperature, loc=param[0], param=logits, annotation=R.Tensor((batch_size, 1, vocab_size), dtype=\"float32\")) "), sinfo_args=(R.Tuple,)) + R.call_packed("vm.builtin.match_shape", temperature, shape_heap, R.prim_value(1), R.prim_value(3), R.prim_value(0), R.str("ErrorContext(fn=softmax_with_temperature, loc=param[1], param=temperature, annotation=R.Tensor((batch_size,), dtype=\"float32\")) "), sinfo_args=(R.Tuple,)) + cls.gpt_neox_q0f16_shape_func2(shape_heap) + gv280: 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, gv280, sinfo_args=(R.Tensor((batch_size, vocab_size), dtype="float32"),)) + gv281: 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),)) + storage12: R.Object = R.vm.alloc_storage(gv281, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv282: 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),)) + alloc174: R.Tensor(dtype="float32", ndim=2) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv282, R.dtype("float32")) + R.vm.kill_object(storage12) + gv283: 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),)) + storage13: R.Object = R.vm.alloc_storage(gv283, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv284: 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),)) + alloc175: R.Tensor(dtype="float32", ndim=2) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv284, R.dtype("float32")) + R.vm.kill_object(storage13) + cls.gpt_neox_q0f16_chunk_lse(lv, temperature, alloc174, alloc175) + lv1: R.Tuple(R.Tensor(dtype="float32", ndim=2), R.Tensor(dtype="float32", ndim=2)) = alloc174, alloc175 + gv285: 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),)) + storage14: R.Object = R.vm.alloc_storage(gv285, R.prim_value(0), R.dtype("uint8"), R.str("global")) + gv286: 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),)) + alloc176: R.Tensor(dtype="float32", ndim=2) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv286, R.dtype("float32")) + R.vm.kill_object(storage14) + cls.gpt_neox_q0f16_softmax_with_chunked_sum(lv, temperature, alloc174, alloc175, alloc176) + R.vm.kill_object(lv) + R.vm.kill_object(alloc174) + R.vm.kill_object(alloc175) + gv287: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(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", alloc176, gv287, sinfo_args=(R.Tensor((batch_size, 1, vocab_size), dtype="float32"),)) + R.vm.kill_object(alloc176) + return gv \ No newline at end of file