# 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_q4f16_1_"}) @T.prim_func def gpt_neox_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_0_attention_query_key_value_q_weight2: T.Buffer((2304, 96), "uint32"), gpt_neox_layers_0_attention_query_key_value_q_scale2: T.Buffer((2304, 24), "float16"), layer_norm25: T.Buffer((1, 1, 768), "float16"), gpt_neox_layers_0_attention_query_key_value_bias2: T.Buffer((2304,), "float16"), T_add_intermediate_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_q_weight2_local = T.alloc_buffer((2304, 96), "uint32", 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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init in T.vectorized(8): with T.block("NT_matmul_rf_init"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused = T.axis.spatial(256, ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] = T.float16(0.0) for ax1_0_fused_ax1_1_fused_0 in T.serial(3, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_ax1_fused_0 in range(1): for ax0_ax1_fused_1 in T.vectorized(1): with T.block("gpt_neox_layers_0_attention_query_key_value_q_weight2_local"): v0 = T.axis.spatial(2304, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) v1 = T.axis.spatial(96, ax1_0_fused_ax1_1_fused_0 * 32 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0) T.reads(gpt_neox_layers_0_attention_query_key_value_q_weight2[v0, v1]) T.writes(gpt_neox_layers_0_attention_query_key_value_q_weight2_local[v0, v1]) gpt_neox_layers_0_attention_query_key_value_q_weight2_local[v0, v1] = gpt_neox_layers_0_attention_query_key_value_q_weight2[v0, v1] for u_fused_ax0_fused_fused_2, ax1_0_fused_ax1_1_fused_2 in T.grid(1, 1): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1 in T.vectorized(8): with T.block("NT_matmul_rf_update"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused = T.axis.spatial(256, ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_2, vax1_0_fused_ax1_1_fused_0 = T.axis.remap("RR", [ax1_0_fused_ax1_1_fused_2, ax1_0_fused_ax1_1_fused_0]) T.reads(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0], layer_norm25[0, 0, vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused], gpt_neox_layers_0_attention_query_key_value_q_weight2_local[v0, vax1_0_fused_ax1_1_fused_0 * 32 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // 8 + vax1_0_fused_ax1_1_fused_2], gpt_neox_layers_0_attention_query_key_value_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // 32]) T.writes(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] = NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] + layer_norm25[0, 0, vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused] * ((T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_query_key_value_q_weight2_local[v0, vax1_0_fused_ax1_1_fused_0 * 32 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // 8 + vax1_0_fused_ax1_1_fused_2], T.Cast("uint32", (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_query_key_value_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // 32]) 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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] = T.float16(0.0) for ax1 in range(8): with T.block("NT_matmul_rf_update"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0], NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, 0, 0, v0]) T.writes(NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] = NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] + NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_intermediate[0, 0, v0]) T_add_intermediate_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_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_0_attention_query_key_value_q_weight3: T.Buffer((2304, 96), "uint32"), gpt_neox_layers_0_attention_query_key_value_q_scale3: T.Buffer((2304, 24), "float16"), p_layer_norm50: T.handle, 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_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() dequantize_intermediate_local = T.alloc_buffer((2304, 768), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(2304, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(768, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_attention_query_key_value_q_weight3[v0, v1 // 8], gpt_neox_layers_0_attention_query_key_value_q_scale3[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_query_key_value_q_weight3[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_query_key_value_q_scale3[v0, v1 // 32] 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], dequantize_intermediate_local[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)) * dequantize_intermediate_local[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_intermediate[0, v0, v1]) T_add_intermediate_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() dequantize_intermediate_local = T.alloc_buffer((2304, 768), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(2304, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(768, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_attention_query_key_value_q_weight3[v0, v1 // 8], gpt_neox_layers_0_attention_query_key_value_q_scale3[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_query_key_value_q_weight3[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_query_key_value_q_scale3[v0, v1 // 32] 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], dequantize_intermediate_local[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)) * dequantize_intermediate_local[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_intermediate[0, v0, v1]) T_add_intermediate_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_q_weight3[v2, v3 // 8], gpt_neox_layers_0_attention_query_key_value_q_scale3[v2, v3 // 32]) 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)) * ((T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_query_key_value_q_weight3[v2, v3 // 8], T.Cast("uint32", v3 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_query_key_value_q_scale3[v2, v3 // 32]) 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_intermediate[0, v1, v2]) T_add_intermediate_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_q4f16_1_fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_0_attention_query_key_value_q_weight4: T.Buffer((2304, 96), "uint32"), gpt_neox_layers_0_attention_query_key_value_q_scale4: T.Buffer((2304, 24), "float16"), p_layer_norm75: T.handle, 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_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() dequantize_intermediate_local = T.alloc_buffer((2304, 768), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(2304, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(768, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_attention_query_key_value_q_weight4[v0, v1 // 8], gpt_neox_layers_0_attention_query_key_value_q_scale4[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_query_key_value_q_weight4[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_query_key_value_q_scale4[v0, v1 // 32] 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], dequantize_intermediate_local[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)) * dequantize_intermediate_local[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_intermediate[v0, 0, v1]) T_add_intermediate_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() dequantize_intermediate_local = T.alloc_buffer((2304, 768), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(2304, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(768, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_attention_query_key_value_q_weight4[v0, v1 // 8], gpt_neox_layers_0_attention_query_key_value_q_scale4[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_query_key_value_q_weight4[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_query_key_value_q_scale4[v0, v1 // 32] 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], dequantize_intermediate_local[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)) * dequantize_intermediate_local[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_intermediate[v0, 0, v1]) T_add_intermediate_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_q_weight4[v2, v3 // 8], gpt_neox_layers_0_attention_query_key_value_q_scale4[v2, v3 // 32]) 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)) * ((T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_query_key_value_q_weight4[v2, v3 // 8], T.Cast("uint32", v3 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_query_key_value_q_scale4[v2, v3 // 32]) 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_intermediate[v1, 0, v2]) T_add_intermediate_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_q4f16_1_fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_0_attention_dense_q_weight2: T.Buffer((768, 96), "uint32"), gpt_neox_layers_0_attention_dense_q_scale2: T.Buffer((768, 24), "float16"), lv149: T.Buffer((1, 1, 768), "float16"), gpt_neox_layers_0_attention_dense_bias2: T.Buffer((768,), "float16"), lv110: T.Buffer((1, 1, 768), "float16"), input_embed: T.Buffer((1, 1, 768), "float16"), T_add_intermediate_1_2_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), "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_q_weight2_local = T.alloc_buffer((768, 96), "uint32", 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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init in T.vectorized(8): with T.block("NT_matmul_rf_init"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused = T.axis.spatial(256, ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] = T.float16(0.0) for ax1_0_fused_ax1_1_fused_0 in T.serial(3, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_ax1_fused_0 in range(1): for ax0_ax1_fused_1 in T.vectorized(1): with T.block("gpt_neox_layers_0_attention_dense_q_weight2_local"): v0 = T.axis.spatial(768, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) v1 = T.axis.spatial(96, ax1_0_fused_ax1_1_fused_0 * 32 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0) T.reads(gpt_neox_layers_0_attention_dense_q_weight2[v0, v1]) T.writes(gpt_neox_layers_0_attention_dense_q_weight2_local[v0, v1]) gpt_neox_layers_0_attention_dense_q_weight2_local[v0, v1] = gpt_neox_layers_0_attention_dense_q_weight2[v0, v1] for u_fused_ax0_fused_fused_2, ax1_0_fused_ax1_1_fused_2 in T.grid(1, 1): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1 in T.vectorized(8): with T.block("NT_matmul_rf_update"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused = T.axis.spatial(256, ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_2, vax1_0_fused_ax1_1_fused_0 = T.axis.remap("RR", [ax1_0_fused_ax1_1_fused_2, ax1_0_fused_ax1_1_fused_0]) T.reads(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0], lv149[0, 0, vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused], gpt_neox_layers_0_attention_dense_q_weight2_local[v0, vax1_0_fused_ax1_1_fused_0 * 32 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // 8 + vax1_0_fused_ax1_1_fused_2], gpt_neox_layers_0_attention_dense_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // 32]) T.writes(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] = NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] + lv149[0, 0, vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused] * ((T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_dense_q_weight2_local[v0, vax1_0_fused_ax1_1_fused_0 * 32 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // 8 + vax1_0_fused_ax1_1_fused_2], T.Cast("uint32", (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // 32]) 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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] = T.float16(0.0) for ax1 in range(8): with T.block("NT_matmul_rf_update"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0], NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, 0, 0, v0]) T.writes(NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] = NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] + NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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(lv110[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_intermediate[0, 0, v0]) T_add_intermediate_1_2_intermediate[0, 0, v0] = lv110[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_q4f16_1_fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_0_attention_dense_q_weight4: T.Buffer((768, 96), "uint32"), gpt_neox_layers_0_attention_dense_q_scale4: T.Buffer((768, 24), "float16"), p_reshape147: T.handle, gpt_neox_layers_0_attention_dense_bias4: T.Buffer((768,), "float16"), p_lv2: 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") lv2 = T.match_buffer(p_lv2, (batch_size, 1, 768), "float16") input_embeds = T.match_buffer(p_input_embeds, (batch_size, 1, 768), "float16") T_add_intermediate_1_2_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() dequantize_intermediate_local = T.alloc_buffer((768, 768), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(768, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(768, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_attention_dense_q_weight4[v0, v1 // 8], gpt_neox_layers_0_attention_dense_q_scale4[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_dense_q_weight4[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale4[v0, v1 // 32] 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], dequantize_intermediate_local[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)) * dequantize_intermediate_local[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(lv2[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_intermediate[v0, 0, v1]) T_add_intermediate_1_2_intermediate[v0, 0, v1] = lv2[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() dequantize_intermediate_local = T.alloc_buffer((768, 768), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(768, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(768, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_attention_dense_q_weight4[v0, v1 // 8], gpt_neox_layers_0_attention_dense_q_scale4[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_dense_q_weight4[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale4[v0, v1 // 32] 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], dequantize_intermediate_local[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)) * dequantize_intermediate_local[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(lv2[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_intermediate[v0, 0, v1]) T_add_intermediate_1_2_intermediate[v0, 0, v1] = lv2[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_q_weight4[v2, v3 // 8], gpt_neox_layers_0_attention_dense_q_scale4[v2, v3 // 32]) 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)) * ((T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_dense_q_weight4[v2, v3 // 8], T.Cast("uint32", v3 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale4[v2, v3 // 32]) 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(lv2[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_intermediate[v1, 0, v2]) T_add_intermediate_1_2_intermediate[v1, 0, v2] = lv2[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_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_0_attention_dense_q_weight3: T.Buffer((768, 96), "uint32"), gpt_neox_layers_0_attention_dense_q_scale3: T.Buffer((768, 24), "float16"), p_reshape99: T.handle, gpt_neox_layers_0_attention_dense_bias3: T.Buffer((768,), "float16"), p_lv38: 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") lv38 = T.match_buffer(p_lv38, (1, seq_len, 768), "float16") input_embeds = T.match_buffer(p_input_embeds, (1, seq_len, 768), "float16") T_add_intermediate_1_2_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() dequantize_intermediate_local = T.alloc_buffer((768, 768), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(768, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(768, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_attention_dense_q_weight3[v0, v1 // 8], gpt_neox_layers_0_attention_dense_q_scale3[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_dense_q_weight3[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale3[v0, v1 // 32] 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], dequantize_intermediate_local[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)) * dequantize_intermediate_local[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(lv38[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_intermediate[0, v0, v1]) T_add_intermediate_1_2_intermediate[0, v0, v1] = lv38[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() dequantize_intermediate_local = T.alloc_buffer((768, 768), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(768, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(768, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_attention_dense_q_weight3[v0, v1 // 8], gpt_neox_layers_0_attention_dense_q_scale3[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_dense_q_weight3[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale3[v0, v1 // 32] 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], dequantize_intermediate_local[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)) * dequantize_intermediate_local[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(lv38[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_intermediate[0, v0, v1]) T_add_intermediate_1_2_intermediate[0, v0, v1] = lv38[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_q_weight3[v2, v3 // 8], gpt_neox_layers_0_attention_dense_q_scale3[v2, v3 // 32]) 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)) * ((T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_attention_dense_q_weight3[v2, v3 // 8], T.Cast("uint32", v3 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale3[v2, v3 // 32]) 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(lv38[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_intermediate[0, v1, v2]) T_add_intermediate_1_2_intermediate[0, v1, v2] = lv38[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_q4f16_1_fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2: T.Buffer((3072, 96), "uint32"), gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale2: T.Buffer((3072, 24), "float16"), layer_norm26: T.Buffer((1, 1, 768), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: T.Buffer((3072,), "float32"), compute_intermediate_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_q_weight2_local = T.alloc_buffer((3072, 96), "uint32", 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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init in T.vectorized(8): with T.block("NT_matmul_rf_init"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused = T.axis.spatial(256, ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] = T.float32(0.0) for ax1_0_fused_ax1_1_fused_0 in T.serial(3, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_ax1_fused_0 in range(1): for ax0_ax1_fused_1 in T.vectorized(1): with T.block("gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2_local"): v0 = T.axis.spatial(3072, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) v1 = T.axis.spatial(96, ax1_0_fused_ax1_1_fused_0 * 32 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0) T.reads(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2[v0, v1]) T.writes(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2_local[v0, v1]) gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2_local[v0, v1] = gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2[v0, v1] for u_fused_ax0_fused_fused_2, ax1_0_fused_ax1_1_fused_2 in T.grid(1, 1): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1 in T.vectorized(8): with T.block("NT_matmul_rf_update"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused = T.axis.spatial(256, ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_2, vax1_0_fused_ax1_1_fused_0 = T.axis.remap("RR", [ax1_0_fused_ax1_1_fused_2, ax1_0_fused_ax1_1_fused_0]) T.reads(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0], layer_norm26[0, 0, vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused], gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2_local[v0, vax1_0_fused_ax1_1_fused_0 * 32 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // 8 + vax1_0_fused_ax1_1_fused_2], gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // 32]) T.writes(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] = NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] + T.Cast("float32", layer_norm26[0, 0, vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused]) * T.Cast("float32", (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2_local[v0, vax1_0_fused_ax1_1_fused_0 * 32 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // 8 + vax1_0_fused_ax1_1_fused_2], T.Cast("uint32", (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // 32]) 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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] = T.float32(0.0) for ax1 in range(8): with T.block("NT_matmul_rf_update"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0], NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, 0, 0, v0]) T.writes(NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] = NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] + NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_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_intermediate[0, 0, v0]) compute_intermediate_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_q4f16_1_fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4: T.Buffer((3072, 96), "uint32"), gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4: T.Buffer((3072, 24), "float16"), p_layer_norm76: T.handle, 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_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() dequantize_intermediate_local = T.alloc_buffer((3072, 768), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(768, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4[v0, v1 // 8], gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4[v0, v1 // 32] 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], dequantize_intermediate_local[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", dequantize_intermediate_local[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_intermediate[v0, 0, v1]) compute_intermediate_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() dequantize_intermediate_local = T.alloc_buffer((3072, 768), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(768, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4[v0, v1 // 8], gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4[v0, v1 // 32] 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], dequantize_intermediate_local[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", dequantize_intermediate_local[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_intermediate[v0, 0, v1]) compute_intermediate_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_q_weight4[v2, v3 // 8], gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4[v2, v3 // 32]) 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", (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4[v2, v3 // 8], T.Cast("uint32", v3 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4[v2, v3 // 32]) 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_intermediate[v1, 0, v2]) compute_intermediate_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_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3: T.Buffer((3072, 96), "uint32"), gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3: T.Buffer((3072, 24), "float16"), p_layer_norm51: T.handle, 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_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() dequantize_intermediate_local = T.alloc_buffer((3072, 768), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(768, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3[v0, v1 // 8], gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3[v0, v1 // 32] 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], dequantize_intermediate_local[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", dequantize_intermediate_local[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_intermediate[0, v0, v1]) compute_intermediate_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() dequantize_intermediate_local = T.alloc_buffer((3072, 768), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(3072, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(768, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3[v0, v1 // 8], gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3[v0, v1 // 32] 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], dequantize_intermediate_local[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", dequantize_intermediate_local[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_intermediate[0, v0, v1]) compute_intermediate_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_q_weight3[v2, v3 // 8], gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3[v2, v3 // 32]) 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", (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3[v2, v3 // 8], T.Cast("uint32", v3 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3[v2, v3 // 32]) 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_intermediate[0, v1, v2]) compute_intermediate_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_q4f16_1_fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2: T.Buffer((768, 384), "uint32"), gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale2: T.Buffer((768, 96), "float16"), lv150: T.Buffer((1, 1, 3072), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: T.Buffer((768,), "float32"), compute_intermediate_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_q_weight2_local = T.alloc_buffer((768, 384), "uint32", 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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init in T.vectorized(8): with T.block("NT_matmul_rf_init"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused = T.axis.spatial(256, ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] = T.float32(0.0) for ax1_0_fused_ax1_1_fused_0 in T.serial(12, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_ax1_fused_0 in range(1): for ax0_ax1_fused_1 in T.vectorized(1): with T.block("gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2_local"): v0 = T.axis.spatial(768, u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1) v1 = T.axis.spatial(384, ax1_0_fused_ax1_1_fused_0 * 32 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0) T.reads(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2[v0, v1]) T.writes(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2_local[v0, v1]) gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2_local[v0, v1] = gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2[v0, v1] for u_fused_ax0_fused_fused_2, ax1_0_fused_ax1_1_fused_2 in T.grid(1, 1): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1 in T.vectorized(8): with T.block("NT_matmul_rf_update"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused = T.axis.spatial(256, ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_2, vax1_0_fused_ax1_1_fused_0 = T.axis.remap("RR", [ax1_0_fused_ax1_1_fused_2, ax1_0_fused_ax1_1_fused_0]) T.reads(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0], lv150[0, 0, vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused], gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2_local[v0, vax1_0_fused_ax1_1_fused_0 * 32 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // 8 + vax1_0_fused_ax1_1_fused_2], gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // 32]) T.writes(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] = NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] + T.Cast("float32", lv150[0, 0, vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused]) * T.Cast("float32", (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2_local[v0, vax1_0_fused_ax1_1_fused_0 * 32 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // 8 + vax1_0_fused_ax1_1_fused_2], T.Cast("uint32", (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // 32]) 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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] = T.float32(0.0) for ax1 in range(8): with T.block("NT_matmul_rf_update"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0], NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, 0, 0, v0]) T.writes(NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] = NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] + NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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(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_intermediate[0, 0, v0]) compute_intermediate_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_q4f16_1_fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4: T.Buffer((768, 384), "uint32"), gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4: T.Buffer((768, 96), "float16"), p_lv1: T.handle, 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() lv1 = T.match_buffer(p_lv1, (batch_size, 1, 3072), "float16") compute_intermediate_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() dequantize_intermediate_local = T.alloc_buffer((768, 3072), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(768, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(3072, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4[v0, v1 // 8], gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4[v0, v1 // 32] 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], lv1[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[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, lv1[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", dequantize_intermediate_local[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_intermediate[v0, 0, v1]) compute_intermediate_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() dequantize_intermediate_local = T.alloc_buffer((768, 3072), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(768, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(3072, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4[v0, v1 // 8], gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4[v0, v1 // 32] 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], lv1[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[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, lv1[v0, 0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", dequantize_intermediate_local[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_intermediate[v0, 0, v1]) compute_intermediate_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], lv1[v1, 0, v3], gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4[v2, v3 // 8], gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4[v2, v3 // 32]) 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, lv1[v1, 0, v3], T.float16(0.0))) * T.Cast("float32", (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4[v2, v3 // 8], T.Cast("uint32", v3 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4[v2, v3 // 32]) 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_intermediate[v1, 0, v2]) compute_intermediate_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_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3: T.Buffer((768, 384), "uint32"), gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3: T.Buffer((768, 96), "float16"), p_lv50: T.handle, 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() lv50 = T.match_buffer(p_lv50, (1, seq_len, 3072), "float16") compute_intermediate_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() dequantize_intermediate_local = T.alloc_buffer((768, 3072), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(768, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(3072, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3[v0, v1 // 8], gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3[v0, v1 // 32] 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], lv50[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[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, lv50[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", dequantize_intermediate_local[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_intermediate[0, v0, v1]) compute_intermediate_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() dequantize_intermediate_local = T.alloc_buffer((768, 3072), "float16", scope="local") 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_0_1, ax1 in T.grid(2, 8): for ax0_1 in T.vectorized(1): with T.block("dequantize"): v0 = T.axis.spatial(768, ax1_fused_0 * 4 + ax1_fused_1 * 2 + ax0_0_1 + ax0_1) v1 = T.axis.spatial(3072, ax2_fused_0 * 256 + ax2_fused_1_ax2_fused_3_fused_0 * 8 + ax1) T.reads(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3[v0, v1 // 8], gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3[v0, v1 // 32]) T.writes(dequantize_intermediate_local[v0, v1]) dequantize_intermediate_local[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3[v0, v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3[v0, v1 // 32] 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], lv50[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[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, lv50[0, v0, vax2_fused_2 * 8 + vax2_fused_0 * 256 + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", dequantize_intermediate_local[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_intermediate[0, v0, v1]) compute_intermediate_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], lv50[0, v1, v3], gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3[v2, v3 // 8], gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3[v2, v3 // 32]) 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, lv50[0, v1, v3], T.float16(0.0))) * T.Cast("float32", (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3[v2, v3 // 8], T.Cast("uint32", v3 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3[v2, v3 // 32]) 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_intermediate[0, v1, v2]) compute_intermediate_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_q4f16_1_fused_dequantize_fused_NT_matmul14_cast8(p_embed_out_q_weight2: T.handle, p_embed_out_q_scale2: T.handle, layer_norm49: T.Buffer((1, 1, 768), "float16"), p_output0: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) vocab_size = T.int32() embed_out_q_weight2 = T.match_buffer(p_embed_out_q_weight2, (vocab_size, 96), "uint32") embed_out_q_scale2 = T.match_buffer(p_embed_out_q_scale2, (vocab_size, 24), "float16") compute_intermediate_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_q_weight2_local = T.alloc_buffer((vocab_size, 96), "uint32", 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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init in T.vectorized(8): with T.block("NT_matmul_rf_init"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused = T.axis.spatial(256, ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] = T.float16(0.0) for ax1_0_fused_ax1_1_fused_0 in T.serial(3, annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_ax1_fused_0 in range(1): for ax0_ax1_fused_1 in T.vectorized(1): with T.block("embed_out_q_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(96, ax1_0_fused_ax1_1_fused_0 * 32 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0) T.where(u_fused_ax0_fused_fused_0 * 2 + u_fused_ax0_fused_fused_1 < vocab_size) T.reads(embed_out_q_weight2[v0, v1]) T.writes(embed_out_q_weight2_local[v0, v1]) embed_out_q_weight2_local[v0, v1] = embed_out_q_weight2[v0, v1] for u_fused_ax0_fused_fused_2, ax1_0_fused_ax1_1_fused_2 in T.grid(1, 1): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1 in T.vectorized(8): with T.block("NT_matmul_rf_update"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused = T.axis.spatial(256, ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_2, vax1_0_fused_ax1_1_fused_0 = T.axis.remap("RR", [ax1_0_fused_ax1_1_fused_2, ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0], layer_norm49[0, 0, vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused], embed_out_q_weight2_local[v0, vax1_0_fused_ax1_1_fused_0 * 32 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // 8 + vax1_0_fused_ax1_1_fused_2], embed_out_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // 32]) T.writes(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] = NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, 0, 0, v0] + layer_norm49[0, 0, vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused] * ((T.Cast("float16", T.bitwise_and(T.shift_right(embed_out_q_weight2_local[v0, vax1_0_fused_ax1_1_fused_0 * 32 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // 8 + vax1_0_fused_ax1_1_fused_2], T.Cast("uint32", (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * embed_out_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * 8 + vax1_0_fused_ax1_1_fused_0 * 256 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // 32]) 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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] = T.float16(0.0) for ax1 in range(8): with T.block("NT_matmul_rf_update"): vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0], NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, 0, 0, v0]) T.writes(NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] = NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, 0, 0, v0] + NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * 8 + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_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(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_intermediate[0, 0, v0]) compute_intermediate_intermediate[0, 0, v0] = T.Cast("float32", NT_matmul_intermediate_local[0, 0, v0]) @T.prim_func def gpt_neox_q4f16_1_fused_dequantize_fused_NT_matmul4_cast2(p_embed_out_q_weight4: T.handle, p_embed_out_q_scale4: T.handle, p_layer_norm99: T.handle, p_output0: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) vocab_size = T.int32() embed_out_q_weight4 = T.match_buffer(p_embed_out_q_weight4, (vocab_size, 96), "uint32") embed_out_q_scale4 = T.match_buffer(p_embed_out_q_scale4, (vocab_size, 24), "float16") batch_size = T.int32() layer_norm99 = T.match_buffer(p_layer_norm99, (batch_size, 1, 768), "float16") compute_intermediate_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_q_weight4[v2, v3 // 8], embed_out_q_scale4[v2, v3 // 32]) 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, (T.Cast("float16", T.bitwise_and(T.shift_right(embed_out_q_weight4[v2, v3 // 8], T.Cast("uint32", v3 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * embed_out_q_scale4[v2, v3 // 32], 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_intermediate[v1, 0, v2]) compute_intermediate_intermediate[v1, 0, v2] = T.Cast("float32", NT_matmul_intermediate_reindex_pad_local[v0, v1, v2]) @T.prim_func def gpt_neox_q4f16_1_fused_dequantize_fused_NT_matmul9_cast5(p_embed_out_q_weight3: T.handle, p_embed_out_q_scale3: T.handle, p_take1: T.handle, p_output0: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) vocab_size = T.int32() embed_out_q_weight3 = T.match_buffer(p_embed_out_q_weight3, (vocab_size, 96), "uint32") embed_out_q_scale3 = T.match_buffer(p_embed_out_q_scale3, (vocab_size, 24), "float16") batch_size = T.int32() take1 = T.match_buffer(p_take1, (1, batch_size, 768), "float16") compute_intermediate_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_q_weight3[v2, v3 // 8], embed_out_q_scale3[v2, v3 // 32]) 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, (T.Cast("float16", T.bitwise_and(T.shift_right(embed_out_q_weight3[v2, v3 // 8], T.Cast("uint32", v3 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * embed_out_q_scale3[v2, v3 // 32], 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_intermediate[0, v1, v2]) compute_intermediate_intermediate[0, v1, v2] = T.Cast("float32", NT_matmul_intermediate_reindex_pad_local[v0, v1, v2]) @T.prim_func def gpt_neox_q4f16_1_fused_dequantize_take1(p_gpt_neox_embed_in_q_weight: T.handle, p_gpt_neox_embed_in_q_scale: T.handle, p_input_ids: T.handle, p_output0: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) vocab_size = T.int32() gpt_neox_embed_in_q_weight = T.match_buffer(p_gpt_neox_embed_in_q_weight, (vocab_size, 96), "uint32") gpt_neox_embed_in_q_scale = T.match_buffer(p_gpt_neox_embed_in_q_scale, (vocab_size, 24), "float16") seq_len = T.int32() input_ids = T.match_buffer(p_input_ids, (seq_len,), "int32") T_take_intermediate = T.match_buffer(p_output0, (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(input_ids[v0], gpt_neox_embed_in_q_weight[input_ids[v0], v1 // 8], gpt_neox_embed_in_q_scale[input_ids[v0], v1 // 32]) T.writes(T_take_intermediate[v0, v1]) T_take_intermediate[v0, v1] = (T.Cast("float16", T.bitwise_and(T.shift_right(gpt_neox_embed_in_q_weight[input_ids[v0], v1 // 8], T.Cast("uint32", v1 % 8 * 4)), T.uint32(15))) - T.float16(7.0)) * gpt_neox_embed_in_q_scale[input_ids[v0], v1 // 32] @T.prim_func def gpt_neox_q4f16_1_fused_reshape10_reshape11(lv64: 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(lv64[0, v0 // 64, v0 % 64]) T.writes(T_reshape_intermediate_1[0, 0, v0]) T_reshape_intermediate_1[0, 0, v0] = lv64[0, v0 // 64, v0 % 64] @T.prim_func def gpt_neox_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_reshape2(var_lv186: T.handle, var_T_reshape: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) batch_size = T.int32() lv186 = T.match_buffer(var_lv186, (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(lv186[v0, v1, v2]) T.writes(T_reshape[v0, 0, v1, v2]) T_reshape[v0, 0, v1, v2] = lv186[v0, v1, v2] @T.prim_func def gpt_neox_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_reshape6(var_lv125: T.handle, var_T_reshape: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) seq_len = T.int32() lv125 = T.match_buffer(var_lv125, (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(lv125[v0, v1, v2]) T.writes(T_reshape[0, v0, v1, v2]) T_reshape[0, v0, v1, v2] = lv125[v0, v1, v2] @T.prim_func def gpt_neox_q4f16_1_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_q4f16_1_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_q4f16_1_shape_func(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_q4f16_1_shape_func1(H: T.Buffer((T.int64(5),), "int64")): T.func_attr({"tir.is_host_func": 1}) H[T.int64(4)] = T.int64(512) * H[T.int64(1)] H[T.int64(2)] = T.int64(512) * ((H[T.int64(1)] + T.int64(4096) - T.int64(1)) // T.int64(4096)) H[T.int64(3)] = (H[T.int64(1)] + T.int64(4096) - T.int64(1)) // T.int64(4096) @T.prim_func def gpt_neox_q4f16_1_shape_func2(H: T.Buffer((T.int64(5),), "int64")): T.func_attr({"tir.is_host_func": 1}) H[T.int64(4)] = T.int64(128) * H[T.int64(2)] * T.int64(4) @T.prim_func def gpt_neox_q4f16_1_shape_func3(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_q4f16_1_shape_func4(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_q4f16_1_shape_func5(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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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_q4f16_1_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\": \"q4f16_1\", \"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.q_weight\", \"shape\": [\"vocab_size\", 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.embed_in.q_scale\", \"shape\": [\"vocab_size\", 24], \"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.q_weight\", \"shape\": [2304, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.query_key_value.q_scale\", \"shape\": [2304, 24], \"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.q_weight\", \"shape\": [768, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.attention.dense.q_scale\", \"shape\": [768, 24], \"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.q_weight\", \"shape\": [3072, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_h_to_4h.q_scale\", \"shape\": [3072, 24], \"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.q_weight\", \"shape\": [768, 384], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.0.mlp.dense_4h_to_h.q_scale\", \"shape\": [768, 96], \"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.q_weight\", \"shape\": [2304, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.query_key_value.q_scale\", \"shape\": [2304, 24], \"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.q_weight\", \"shape\": [768, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.attention.dense.q_scale\", \"shape\": [768, 24], \"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.q_weight\", \"shape\": [3072, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_h_to_4h.q_scale\", \"shape\": [3072, 24], \"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.q_weight\", \"shape\": [768, 384], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.1.mlp.dense_4h_to_h.q_scale\", \"shape\": [768, 96], \"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.q_weight\", \"shape\": [2304, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.query_key_value.q_scale\", \"shape\": [2304, 24], \"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.q_weight\", \"shape\": [768, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.attention.dense.q_scale\", \"shape\": [768, 24], \"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.q_weight\", \"shape\": [3072, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_h_to_4h.q_scale\", \"shape\": [3072, 24], \"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.q_weight\", \"shape\": [768, 384], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.2.mlp.dense_4h_to_h.q_scale\", \"shape\": [768, 96], \"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.q_weight\", \"shape\": [2304, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.query_key_value.q_scale\", \"shape\": [2304, 24], \"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.q_weight\", \"shape\": [768, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.attention.dense.q_scale\", \"shape\": [768, 24], \"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.q_weight\", \"shape\": [3072, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_h_to_4h.q_scale\", \"shape\": [3072, 24], \"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.q_weight\", \"shape\": [768, 384], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.3.mlp.dense_4h_to_h.q_scale\", \"shape\": [768, 96], \"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.q_weight\", \"shape\": [2304, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.query_key_value.q_scale\", \"shape\": [2304, 24], \"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.q_weight\", \"shape\": [768, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.attention.dense.q_scale\", \"shape\": [768, 24], \"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.q_weight\", \"shape\": [3072, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_h_to_4h.q_scale\", \"shape\": [3072, 24], \"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.q_weight\", \"shape\": [768, 384], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.4.mlp.dense_4h_to_h.q_scale\", \"shape\": [768, 96], \"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.q_weight\", \"shape\": [2304, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.query_key_value.q_scale\", \"shape\": [2304, 24], \"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.q_weight\", \"shape\": [768, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.attention.dense.q_scale\", \"shape\": [768, 24], \"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.q_weight\", \"shape\": [3072, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_h_to_4h.q_scale\", \"shape\": [3072, 24], \"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.q_weight\", \"shape\": [768, 384], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.5.mlp.dense_4h_to_h.q_scale\", \"shape\": [768, 96], \"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.q_weight\", \"shape\": [2304, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.query_key_value.q_scale\", \"shape\": [2304, 24], \"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.q_weight\", \"shape\": [768, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.attention.dense.q_scale\", \"shape\": [768, 24], \"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.q_weight\", \"shape\": [3072, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_h_to_4h.q_scale\", \"shape\": [3072, 24], \"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.q_weight\", \"shape\": [768, 384], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.6.mlp.dense_4h_to_h.q_scale\", \"shape\": [768, 96], \"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.q_weight\", \"shape\": [2304, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.query_key_value.q_scale\", \"shape\": [2304, 24], \"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.q_weight\", \"shape\": [768, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.attention.dense.q_scale\", \"shape\": [768, 24], \"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.q_weight\", \"shape\": [3072, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_h_to_4h.q_scale\", \"shape\": [3072, 24], \"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.q_weight\", \"shape\": [768, 384], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.7.mlp.dense_4h_to_h.q_scale\", \"shape\": [768, 96], \"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.q_weight\", \"shape\": [2304, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.query_key_value.q_scale\", \"shape\": [2304, 24], \"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.q_weight\", \"shape\": [768, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.attention.dense.q_scale\", \"shape\": [768, 24], \"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.q_weight\", \"shape\": [3072, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_h_to_4h.q_scale\", \"shape\": [3072, 24], \"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.q_weight\", \"shape\": [768, 384], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.8.mlp.dense_4h_to_h.q_scale\", \"shape\": [768, 96], \"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.q_weight\", \"shape\": [2304, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.query_key_value.q_scale\", \"shape\": [2304, 24], \"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.q_weight\", \"shape\": [768, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.attention.dense.q_scale\", \"shape\": [768, 24], \"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.q_weight\", \"shape\": [3072, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_h_to_4h.q_scale\", \"shape\": [3072, 24], \"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.q_weight\", \"shape\": [768, 384], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.9.mlp.dense_4h_to_h.q_scale\", \"shape\": [768, 96], \"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.q_weight\", \"shape\": [2304, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.query_key_value.q_scale\", \"shape\": [2304, 24], \"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.q_weight\", \"shape\": [768, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.attention.dense.q_scale\", \"shape\": [768, 24], \"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.q_weight\", \"shape\": [3072, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_h_to_4h.q_scale\", \"shape\": [3072, 24], \"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.q_weight\", \"shape\": [768, 384], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.10.mlp.dense_4h_to_h.q_scale\", \"shape\": [768, 96], \"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.q_weight\", \"shape\": [2304, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.query_key_value.q_scale\", \"shape\": [2304, 24], \"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.q_weight\", \"shape\": [768, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.attention.dense.q_scale\", \"shape\": [768, 24], \"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.q_weight\", \"shape\": [3072, 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_h_to_4h.q_scale\", \"shape\": [3072, 24], \"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.q_weight\", \"shape\": [768, 384], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"gpt_neox.layers.11.mlp.dense_4h_to_h.q_scale\", \"shape\": [768, 96], \"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.q_weight\", \"shape\": [\"vocab_size\", 96], \"dtype\": \"uint32\", \"preprocs\": [], \"pipeline_stages\": [0]}, {\"name\": \"embed_out.q_scale\", \"shape\": [\"vocab_size\", 24], \"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", 96), dtype="uint32"), R.Tensor(("vocab_size", 24), 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, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor(("vocab_size", 96), dtype="uint32"), R.Tensor(("vocab_size", 24), 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(198), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv302: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[0] R.call_packed("vm.builtin.check_tensor_info", gv302, R.prim_value(2), R.dtype("uint32"), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv303: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[1] R.call_packed("vm.builtin.check_tensor_info", gv303, 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, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv304: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[196] R.call_packed("vm.builtin.check_tensor_info", gv304, R.prim_value(2), R.dtype("uint32"), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv305: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[197] R.call_packed("vm.builtin.check_tensor_info", gv305, 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, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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", gv302, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(96), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv303, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(24), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv304, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(96), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv305, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(24), R.str("ErrorContext(fn=batch_decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) cls.gpt_neox_q4f16_1_shape_func4(shape_heap) gpt_neox_layers_0_input_layernorm_weight4: R.Tensor((768,), dtype="float16") = packed_params[2] gpt_neox_layers_0_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[3] storage20: R.Object = R.vm.alloc_storage(R.shape([196608]), R.prim_value(0), R.dtype("uint8"), R.str("global")) gv306: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv306, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(input_embeds, gpt_neox_layers_0_input_layernorm_weight4, gpt_neox_layers_0_input_layernorm_bias4, alloc263) 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_q_weight4: R.Tensor((2304, 96), dtype="uint32") = packed_params[6] gpt_neox_layers_0_attention_query_key_value_q_scale4: R.Tensor((2304, 24), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[8] storage21: R.Object = R.vm.alloc_storage(R.shape([589824]), R.prim_value(0), R.dtype("uint8"), R.str("global")) gv307: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv307, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_0_attention_query_key_value_q_weight4, gpt_neox_layers_0_attention_query_key_value_q_scale4, alloc263, gpt_neox_layers_0_attention_query_key_value_bias4, alloc264) R.vm.kill_object(alloc263) R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_q_weight4) R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_q_scale4) R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_bias4) gv308: 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", alloc264, gv308, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc264) 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),)) reshape145: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape144, gv309, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape144) 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),)) alloc265: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv310, 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)), reshape145, alloc265) R.vm.kill_object(reshape145) gv311: 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", alloc265, gv311, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc265) gv312: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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, gv312, 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[4] gpt_neox_layers_0_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[5] gv313: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc266: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv313, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(input_embeds, gpt_neox_layers_0_post_attention_layernorm_weight4, gpt_neox_layers_0_post_attention_layernorm_bias4, alloc266) 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_q_weight4: R.Tensor((3072, 96), dtype="uint32") = packed_params[12] gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4: R.Tensor((3072, 24), dtype="float16") = packed_params[13] gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[14] storage22: R.Object = R.vm.alloc_storage(R.shape([786432]), 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(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),)) alloc267: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv314, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4, alloc266, gpt_neox_layers_0_mlp_dense_h_to_4h_bias4, alloc267) R.vm.kill_object(alloc266) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_bias4) gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4: R.Tensor((768, 384), dtype="uint32") = packed_params[15] gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4: R.Tensor((768, 96), dtype="float16") = packed_params[16] gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[17] gv315: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc268: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv315, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4, alloc267, gpt_neox_layers_0_mlp_dense_4h_to_h_bias4, alloc268) R.vm.kill_object(alloc267) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_bias4) gpt_neox_layers_0_attention_dense_q_weight4: R.Tensor((768, 96), dtype="uint32") = packed_params[9] gpt_neox_layers_0_attention_dense_q_scale4: R.Tensor((768, 24), dtype="float16") = packed_params[10] gpt_neox_layers_0_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[11] gv316: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc269: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv316, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_0_attention_dense_q_weight4, gpt_neox_layers_0_attention_dense_q_scale4, reshape147, gpt_neox_layers_0_attention_dense_bias4, alloc268, input_embeds, alloc269) R.vm.kill_object(reshape147) R.vm.kill_object(alloc268) R.vm.kill_object(gpt_neox_layers_0_attention_dense_q_weight4) R.vm.kill_object(gpt_neox_layers_0_attention_dense_q_scale4) 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[18] gpt_neox_layers_1_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[19] gv317: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv317, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc269, gpt_neox_layers_1_input_layernorm_weight4, gpt_neox_layers_1_input_layernorm_bias4, alloc270) 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_q_weight4: R.Tensor((2304, 96), dtype="uint32") = packed_params[22] gpt_neox_layers_1_attention_query_key_value_q_scale4: R.Tensor((2304, 24), dtype="float16") = packed_params[23] gpt_neox_layers_1_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[24] gv318: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv318, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_1_attention_query_key_value_q_weight4, gpt_neox_layers_1_attention_query_key_value_q_scale4, alloc270, gpt_neox_layers_1_attention_query_key_value_bias4, alloc271) R.vm.kill_object(alloc270) R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_q_weight4) R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_q_scale4) R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_bias4) gv319: 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", alloc271, gv319, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc271) 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),)) reshape149: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape148, gv320, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape148) 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),)) alloc272: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv321, 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)), reshape149, alloc272) R.vm.kill_object(reshape149) gv322: 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", alloc272, gv322, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc272) gv323: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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, gv323, 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[20] gpt_neox_layers_1_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[21] gv324: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc273: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv324, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc269, gpt_neox_layers_1_post_attention_layernorm_weight4, gpt_neox_layers_1_post_attention_layernorm_bias4, alloc273) 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_q_weight4: R.Tensor((3072, 96), dtype="uint32") = packed_params[28] gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale4: R.Tensor((3072, 24), dtype="float16") = packed_params[29] gpt_neox_layers_1_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[30] storage23: R.Object = R.vm.alloc_storage(R.shape([786432]), R.prim_value(0), R.dtype("uint8"), R.str("global")) gv325: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc274: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv325, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale4, alloc273, gpt_neox_layers_1_mlp_dense_h_to_4h_bias4, alloc274) R.vm.kill_object(alloc273) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight4) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale4) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_bias4) gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight4: R.Tensor((768, 384), dtype="uint32") = packed_params[31] gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale4: R.Tensor((768, 96), dtype="float16") = packed_params[32] gpt_neox_layers_1_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[33] gv326: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc275: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv326, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale4, alloc274, gpt_neox_layers_1_mlp_dense_4h_to_h_bias4, alloc275) R.vm.kill_object(alloc274) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight4) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale4) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_bias4) gpt_neox_layers_1_attention_dense_q_weight4: R.Tensor((768, 96), dtype="uint32") = packed_params[25] gpt_neox_layers_1_attention_dense_q_scale4: R.Tensor((768, 24), dtype="float16") = packed_params[26] gpt_neox_layers_1_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[27] gv327: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc276: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv327, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_1_attention_dense_q_weight4, gpt_neox_layers_1_attention_dense_q_scale4, reshape151, gpt_neox_layers_1_attention_dense_bias4, alloc275, alloc269, alloc276) R.vm.kill_object(alloc269) R.vm.kill_object(reshape151) R.vm.kill_object(alloc275) R.vm.kill_object(gpt_neox_layers_1_attention_dense_q_weight4) R.vm.kill_object(gpt_neox_layers_1_attention_dense_q_scale4) 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[34] gpt_neox_layers_2_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[35] gv328: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv328, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc276, gpt_neox_layers_2_input_layernorm_weight4, gpt_neox_layers_2_input_layernorm_bias4, alloc277) 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_q_weight4: R.Tensor((2304, 96), dtype="uint32") = packed_params[38] gpt_neox_layers_2_attention_query_key_value_q_scale4: R.Tensor((2304, 24), dtype="float16") = packed_params[39] gpt_neox_layers_2_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[40] gv329: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv329, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_2_attention_query_key_value_q_weight4, gpt_neox_layers_2_attention_query_key_value_q_scale4, alloc277, gpt_neox_layers_2_attention_query_key_value_bias4, alloc278) R.vm.kill_object(alloc277) R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_q_weight4) R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_q_scale4) R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_bias4) gv330: 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", alloc278, gv330, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc278) 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),)) reshape153: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape152, gv331, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape152) 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),)) alloc279: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv332, 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)), reshape153, alloc279) R.vm.kill_object(reshape153) gv333: 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", alloc279, gv333, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc279) gv334: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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, gv334, 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[36] gpt_neox_layers_2_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[37] gv335: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc280: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv335, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc276, gpt_neox_layers_2_post_attention_layernorm_weight4, gpt_neox_layers_2_post_attention_layernorm_bias4, alloc280) 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_q_weight4: R.Tensor((3072, 96), dtype="uint32") = packed_params[44] gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale4: R.Tensor((3072, 24), dtype="float16") = packed_params[45] gpt_neox_layers_2_mlp_dense_h_to_4h_bias4: 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(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),)) alloc281: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv336, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale4, alloc280, gpt_neox_layers_2_mlp_dense_h_to_4h_bias4, alloc281) R.vm.kill_object(alloc280) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight4) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale4) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_bias4) gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight4: R.Tensor((768, 384), dtype="uint32") = packed_params[47] gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale4: R.Tensor((768, 96), dtype="float16") = packed_params[48] gpt_neox_layers_2_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[49] gv337: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc282: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv337, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale4, alloc281, gpt_neox_layers_2_mlp_dense_4h_to_h_bias4, alloc282) R.vm.kill_object(alloc281) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight4) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale4) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_bias4) gpt_neox_layers_2_attention_dense_q_weight4: R.Tensor((768, 96), dtype="uint32") = packed_params[41] gpt_neox_layers_2_attention_dense_q_scale4: R.Tensor((768, 24), dtype="float16") = packed_params[42] gpt_neox_layers_2_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[43] gv338: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc283: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv338, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_2_attention_dense_q_weight4, gpt_neox_layers_2_attention_dense_q_scale4, reshape155, gpt_neox_layers_2_attention_dense_bias4, alloc282, alloc276, alloc283) R.vm.kill_object(alloc276) R.vm.kill_object(reshape155) R.vm.kill_object(alloc282) R.vm.kill_object(gpt_neox_layers_2_attention_dense_q_weight4) R.vm.kill_object(gpt_neox_layers_2_attention_dense_q_scale4) 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[50] gpt_neox_layers_3_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[51] gv339: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv339, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc283, gpt_neox_layers_3_input_layernorm_weight4, gpt_neox_layers_3_input_layernorm_bias4, alloc284) 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_q_weight4: R.Tensor((2304, 96), dtype="uint32") = packed_params[54] gpt_neox_layers_3_attention_query_key_value_q_scale4: R.Tensor((2304, 24), dtype="float16") = packed_params[55] gpt_neox_layers_3_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[56] gv340: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv340, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_3_attention_query_key_value_q_weight4, gpt_neox_layers_3_attention_query_key_value_q_scale4, alloc284, gpt_neox_layers_3_attention_query_key_value_bias4, alloc285) R.vm.kill_object(alloc284) R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_q_weight4) R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_q_scale4) R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_bias4) gv341: 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", alloc285, gv341, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc285) 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),)) reshape157: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape156, gv342, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape156) 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),)) alloc286: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv343, 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)), reshape157, alloc286) R.vm.kill_object(reshape157) gv344: 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", alloc286, gv344, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc286) gv345: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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, gv345, 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[52] gpt_neox_layers_3_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[53] gv346: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc287: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv346, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc283, gpt_neox_layers_3_post_attention_layernorm_weight4, gpt_neox_layers_3_post_attention_layernorm_bias4, alloc287) 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_q_weight4: R.Tensor((3072, 96), dtype="uint32") = packed_params[60] gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale4: R.Tensor((3072, 24), dtype="float16") = packed_params[61] gpt_neox_layers_3_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[62] gv347: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc288: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv347, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale4, alloc287, gpt_neox_layers_3_mlp_dense_h_to_4h_bias4, alloc288) R.vm.kill_object(alloc287) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight4) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale4) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_bias4) gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight4: R.Tensor((768, 384), dtype="uint32") = packed_params[63] gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale4: R.Tensor((768, 96), dtype="float16") = packed_params[64] gpt_neox_layers_3_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[65] gv348: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc289: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv348, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale4, alloc288, gpt_neox_layers_3_mlp_dense_4h_to_h_bias4, alloc289) R.vm.kill_object(alloc288) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight4) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale4) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_bias4) gpt_neox_layers_3_attention_dense_q_weight4: R.Tensor((768, 96), dtype="uint32") = packed_params[57] gpt_neox_layers_3_attention_dense_q_scale4: R.Tensor((768, 24), dtype="float16") = packed_params[58] gpt_neox_layers_3_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[59] gv349: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc290: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv349, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_3_attention_dense_q_weight4, gpt_neox_layers_3_attention_dense_q_scale4, reshape159, gpt_neox_layers_3_attention_dense_bias4, alloc289, alloc283, alloc290) R.vm.kill_object(alloc283) R.vm.kill_object(reshape159) R.vm.kill_object(alloc289) R.vm.kill_object(gpt_neox_layers_3_attention_dense_q_weight4) R.vm.kill_object(gpt_neox_layers_3_attention_dense_q_scale4) 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[66] gpt_neox_layers_4_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[67] gv350: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv350, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc290, gpt_neox_layers_4_input_layernorm_weight4, gpt_neox_layers_4_input_layernorm_bias4, alloc291) 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_q_weight4: R.Tensor((2304, 96), dtype="uint32") = packed_params[70] gpt_neox_layers_4_attention_query_key_value_q_scale4: R.Tensor((2304, 24), dtype="float16") = packed_params[71] gpt_neox_layers_4_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[72] gv351: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv351, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_4_attention_query_key_value_q_weight4, gpt_neox_layers_4_attention_query_key_value_q_scale4, alloc291, gpt_neox_layers_4_attention_query_key_value_bias4, alloc292) R.vm.kill_object(alloc291) R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_q_weight4) R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_q_scale4) R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_bias4) gv352: 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", alloc292, gv352, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc292) 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),)) reshape161: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape160, gv353, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape160) 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),)) alloc293: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv354, 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)), reshape161, alloc293) R.vm.kill_object(reshape161) gv355: 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", alloc293, gv355, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc293) gv356: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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, gv356, 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[68] gpt_neox_layers_4_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[69] gv357: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc294: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv357, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc290, gpt_neox_layers_4_post_attention_layernorm_weight4, gpt_neox_layers_4_post_attention_layernorm_bias4, alloc294) 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_q_weight4: R.Tensor((3072, 96), dtype="uint32") = packed_params[76] gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale4: R.Tensor((3072, 24), dtype="float16") = packed_params[77] gpt_neox_layers_4_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[78] gv358: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc295: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv358, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale4, alloc294, gpt_neox_layers_4_mlp_dense_h_to_4h_bias4, alloc295) R.vm.kill_object(alloc294) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight4) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale4) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_bias4) gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight4: R.Tensor((768, 384), dtype="uint32") = packed_params[79] gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale4: R.Tensor((768, 96), dtype="float16") = packed_params[80] gpt_neox_layers_4_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[81] gv359: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc296: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv359, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale4, alloc295, gpt_neox_layers_4_mlp_dense_4h_to_h_bias4, alloc296) R.vm.kill_object(alloc295) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight4) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale4) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_bias4) gpt_neox_layers_4_attention_dense_q_weight4: R.Tensor((768, 96), dtype="uint32") = packed_params[73] gpt_neox_layers_4_attention_dense_q_scale4: R.Tensor((768, 24), dtype="float16") = packed_params[74] gpt_neox_layers_4_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[75] gv360: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc297: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv360, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_4_attention_dense_q_weight4, gpt_neox_layers_4_attention_dense_q_scale4, reshape163, gpt_neox_layers_4_attention_dense_bias4, alloc296, alloc290, alloc297) R.vm.kill_object(alloc290) R.vm.kill_object(reshape163) R.vm.kill_object(alloc296) R.vm.kill_object(gpt_neox_layers_4_attention_dense_q_weight4) R.vm.kill_object(gpt_neox_layers_4_attention_dense_q_scale4) 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[82] gpt_neox_layers_5_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[83] gv361: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv361, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc297, gpt_neox_layers_5_input_layernorm_weight4, gpt_neox_layers_5_input_layernorm_bias4, alloc298) 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_q_weight4: R.Tensor((2304, 96), dtype="uint32") = packed_params[86] gpt_neox_layers_5_attention_query_key_value_q_scale4: R.Tensor((2304, 24), dtype="float16") = packed_params[87] gpt_neox_layers_5_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[88] gv362: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv362, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_5_attention_query_key_value_q_weight4, gpt_neox_layers_5_attention_query_key_value_q_scale4, alloc298, gpt_neox_layers_5_attention_query_key_value_bias4, alloc299) R.vm.kill_object(alloc298) R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_q_weight4) R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_q_scale4) R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_bias4) gv363: 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", alloc299, gv363, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc299) 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),)) reshape165: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape164, gv364, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape164) 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),)) alloc300: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv365, 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)), reshape165, alloc300) R.vm.kill_object(reshape165) gv366: 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", alloc300, gv366, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc300) gv367: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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, gv367, 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[84] gpt_neox_layers_5_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[85] gv368: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc301: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv368, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc297, gpt_neox_layers_5_post_attention_layernorm_weight4, gpt_neox_layers_5_post_attention_layernorm_bias4, alloc301) 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_q_weight4: R.Tensor((3072, 96), dtype="uint32") = packed_params[92] gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale4: R.Tensor((3072, 24), dtype="float16") = packed_params[93] gpt_neox_layers_5_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[94] gv369: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc302: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv369, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale4, alloc301, gpt_neox_layers_5_mlp_dense_h_to_4h_bias4, alloc302) R.vm.kill_object(alloc301) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight4) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale4) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_bias4) gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight4: R.Tensor((768, 384), dtype="uint32") = packed_params[95] gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale4: R.Tensor((768, 96), dtype="float16") = packed_params[96] gpt_neox_layers_5_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[97] gv370: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc303: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv370, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale4, alloc302, gpt_neox_layers_5_mlp_dense_4h_to_h_bias4, alloc303) R.vm.kill_object(alloc302) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight4) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale4) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_bias4) gpt_neox_layers_5_attention_dense_q_weight4: R.Tensor((768, 96), dtype="uint32") = packed_params[89] gpt_neox_layers_5_attention_dense_q_scale4: R.Tensor((768, 24), dtype="float16") = packed_params[90] gpt_neox_layers_5_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[91] gv371: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc304: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv371, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_5_attention_dense_q_weight4, gpt_neox_layers_5_attention_dense_q_scale4, reshape167, gpt_neox_layers_5_attention_dense_bias4, alloc303, alloc297, alloc304) R.vm.kill_object(alloc297) R.vm.kill_object(reshape167) R.vm.kill_object(alloc303) R.vm.kill_object(gpt_neox_layers_5_attention_dense_q_weight4) R.vm.kill_object(gpt_neox_layers_5_attention_dense_q_scale4) 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[98] gpt_neox_layers_6_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[99] gv372: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv372, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc304, gpt_neox_layers_6_input_layernorm_weight4, gpt_neox_layers_6_input_layernorm_bias4, alloc305) 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_q_weight4: R.Tensor((2304, 96), dtype="uint32") = packed_params[102] gpt_neox_layers_6_attention_query_key_value_q_scale4: R.Tensor((2304, 24), dtype="float16") = packed_params[103] gpt_neox_layers_6_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[104] gv373: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv373, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_6_attention_query_key_value_q_weight4, gpt_neox_layers_6_attention_query_key_value_q_scale4, alloc305, gpt_neox_layers_6_attention_query_key_value_bias4, alloc306) R.vm.kill_object(alloc305) R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_q_weight4) R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_q_scale4) R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_bias4) gv374: 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", alloc306, gv374, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc306) 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),)) reshape169: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape168, gv375, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape168) 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),)) alloc307: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv376, 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)), reshape169, alloc307) R.vm.kill_object(reshape169) gv377: 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", alloc307, gv377, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc307) gv378: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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, gv378, 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[100] gpt_neox_layers_6_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[101] gv379: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc308: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv379, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc304, gpt_neox_layers_6_post_attention_layernorm_weight4, gpt_neox_layers_6_post_attention_layernorm_bias4, alloc308) 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_q_weight4: R.Tensor((3072, 96), dtype="uint32") = packed_params[108] gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale4: R.Tensor((3072, 24), dtype="float16") = packed_params[109] gpt_neox_layers_6_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[110] gv380: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc309: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv380, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale4, alloc308, gpt_neox_layers_6_mlp_dense_h_to_4h_bias4, alloc309) R.vm.kill_object(alloc308) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight4) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale4) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_bias4) gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight4: R.Tensor((768, 384), dtype="uint32") = packed_params[111] gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale4: R.Tensor((768, 96), dtype="float16") = packed_params[112] gpt_neox_layers_6_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[113] gv381: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc310: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv381, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale4, alloc309, gpt_neox_layers_6_mlp_dense_4h_to_h_bias4, alloc310) R.vm.kill_object(alloc309) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight4) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale4) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_bias4) gpt_neox_layers_6_attention_dense_q_weight4: R.Tensor((768, 96), dtype="uint32") = packed_params[105] gpt_neox_layers_6_attention_dense_q_scale4: R.Tensor((768, 24), dtype="float16") = packed_params[106] gpt_neox_layers_6_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[107] gv382: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc311: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv382, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_6_attention_dense_q_weight4, gpt_neox_layers_6_attention_dense_q_scale4, reshape171, gpt_neox_layers_6_attention_dense_bias4, alloc310, alloc304, alloc311) R.vm.kill_object(alloc304) R.vm.kill_object(reshape171) R.vm.kill_object(alloc310) R.vm.kill_object(gpt_neox_layers_6_attention_dense_q_weight4) R.vm.kill_object(gpt_neox_layers_6_attention_dense_q_scale4) 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[114] gpt_neox_layers_7_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[115] gv383: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv383, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc311, gpt_neox_layers_7_input_layernorm_weight4, gpt_neox_layers_7_input_layernorm_bias4, alloc312) 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_q_weight4: R.Tensor((2304, 96), dtype="uint32") = packed_params[118] gpt_neox_layers_7_attention_query_key_value_q_scale4: R.Tensor((2304, 24), dtype="float16") = packed_params[119] gpt_neox_layers_7_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[120] gv384: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv384, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_7_attention_query_key_value_q_weight4, gpt_neox_layers_7_attention_query_key_value_q_scale4, alloc312, gpt_neox_layers_7_attention_query_key_value_bias4, alloc313) R.vm.kill_object(alloc312) R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_q_weight4) R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_q_scale4) R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_bias4) gv385: 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", alloc313, gv385, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc313) 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),)) reshape173: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape172, gv386, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape172) 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),)) alloc314: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv387, 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)), reshape173, alloc314) R.vm.kill_object(reshape173) gv388: 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", alloc314, gv388, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc314) gv389: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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, gv389, 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[116] gpt_neox_layers_7_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[117] gv390: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc315: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv390, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc311, gpt_neox_layers_7_post_attention_layernorm_weight4, gpt_neox_layers_7_post_attention_layernorm_bias4, alloc315) 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_q_weight4: R.Tensor((3072, 96), dtype="uint32") = packed_params[124] gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale4: R.Tensor((3072, 24), dtype="float16") = packed_params[125] gpt_neox_layers_7_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[126] gv391: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc316: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv391, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale4, alloc315, gpt_neox_layers_7_mlp_dense_h_to_4h_bias4, alloc316) R.vm.kill_object(alloc315) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight4) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale4) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_bias4) gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight4: R.Tensor((768, 384), dtype="uint32") = packed_params[127] gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale4: R.Tensor((768, 96), dtype="float16") = packed_params[128] gpt_neox_layers_7_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[129] gv392: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc317: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv392, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale4, alloc316, gpt_neox_layers_7_mlp_dense_4h_to_h_bias4, alloc317) R.vm.kill_object(alloc316) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight4) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale4) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_bias4) gpt_neox_layers_7_attention_dense_q_weight4: R.Tensor((768, 96), dtype="uint32") = packed_params[121] gpt_neox_layers_7_attention_dense_q_scale4: R.Tensor((768, 24), dtype="float16") = packed_params[122] gpt_neox_layers_7_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[123] gv393: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc318: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv393, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_7_attention_dense_q_weight4, gpt_neox_layers_7_attention_dense_q_scale4, reshape175, gpt_neox_layers_7_attention_dense_bias4, alloc317, alloc311, alloc318) R.vm.kill_object(alloc311) R.vm.kill_object(reshape175) R.vm.kill_object(alloc317) R.vm.kill_object(gpt_neox_layers_7_attention_dense_q_weight4) R.vm.kill_object(gpt_neox_layers_7_attention_dense_q_scale4) 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[130] gpt_neox_layers_8_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[131] gv394: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv394, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc318, gpt_neox_layers_8_input_layernorm_weight4, gpt_neox_layers_8_input_layernorm_bias4, alloc319) 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_q_weight4: R.Tensor((2304, 96), dtype="uint32") = packed_params[134] gpt_neox_layers_8_attention_query_key_value_q_scale4: R.Tensor((2304, 24), dtype="float16") = packed_params[135] gpt_neox_layers_8_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[136] gv395: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv395, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_8_attention_query_key_value_q_weight4, gpt_neox_layers_8_attention_query_key_value_q_scale4, alloc319, gpt_neox_layers_8_attention_query_key_value_bias4, alloc320) R.vm.kill_object(alloc319) R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_q_weight4) R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_q_scale4) R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_bias4) gv396: 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", alloc320, gv396, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc320) 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),)) reshape177: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape176, gv397, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape176) 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),)) alloc321: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv398, 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)), reshape177, alloc321) R.vm.kill_object(reshape177) gv399: 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", alloc321, gv399, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc321) gv400: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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, gv400, 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[132] gpt_neox_layers_8_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[133] gv401: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc322: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv401, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc318, gpt_neox_layers_8_post_attention_layernorm_weight4, gpt_neox_layers_8_post_attention_layernorm_bias4, alloc322) 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_q_weight4: R.Tensor((3072, 96), dtype="uint32") = packed_params[140] gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale4: R.Tensor((3072, 24), dtype="float16") = packed_params[141] gpt_neox_layers_8_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[142] gv402: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc323: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv402, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale4, alloc322, gpt_neox_layers_8_mlp_dense_h_to_4h_bias4, alloc323) R.vm.kill_object(alloc322) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight4) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale4) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_bias4) gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight4: R.Tensor((768, 384), dtype="uint32") = packed_params[143] gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale4: R.Tensor((768, 96), dtype="float16") = packed_params[144] gpt_neox_layers_8_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[145] gv403: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc324: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv403, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale4, alloc323, gpt_neox_layers_8_mlp_dense_4h_to_h_bias4, alloc324) R.vm.kill_object(alloc323) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight4) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale4) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_bias4) gpt_neox_layers_8_attention_dense_q_weight4: R.Tensor((768, 96), dtype="uint32") = packed_params[137] gpt_neox_layers_8_attention_dense_q_scale4: R.Tensor((768, 24), dtype="float16") = packed_params[138] gpt_neox_layers_8_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[139] gv404: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc325: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv404, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_8_attention_dense_q_weight4, gpt_neox_layers_8_attention_dense_q_scale4, reshape179, gpt_neox_layers_8_attention_dense_bias4, alloc324, alloc318, alloc325) R.vm.kill_object(alloc318) R.vm.kill_object(reshape179) R.vm.kill_object(alloc324) R.vm.kill_object(gpt_neox_layers_8_attention_dense_q_weight4) R.vm.kill_object(gpt_neox_layers_8_attention_dense_q_scale4) 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[146] gpt_neox_layers_9_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[147] gv405: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv405, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc325, gpt_neox_layers_9_input_layernorm_weight4, gpt_neox_layers_9_input_layernorm_bias4, alloc326) 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_q_weight4: R.Tensor((2304, 96), dtype="uint32") = packed_params[150] gpt_neox_layers_9_attention_query_key_value_q_scale4: R.Tensor((2304, 24), dtype="float16") = packed_params[151] gpt_neox_layers_9_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[152] gv406: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv406, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_9_attention_query_key_value_q_weight4, gpt_neox_layers_9_attention_query_key_value_q_scale4, alloc326, gpt_neox_layers_9_attention_query_key_value_bias4, alloc327) R.vm.kill_object(alloc326) R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_q_weight4) R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_q_scale4) R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_bias4) gv407: 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", alloc327, gv407, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc327) 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),)) reshape181: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape180, gv408, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape180) 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),)) alloc328: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv409, 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)), reshape181, alloc328) R.vm.kill_object(reshape181) gv410: 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", alloc328, gv410, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc328) gv411: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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, gv411, 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[148] gpt_neox_layers_9_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[149] gv412: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc329: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv412, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc325, gpt_neox_layers_9_post_attention_layernorm_weight4, gpt_neox_layers_9_post_attention_layernorm_bias4, alloc329) 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_q_weight4: R.Tensor((3072, 96), dtype="uint32") = packed_params[156] gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale4: R.Tensor((3072, 24), dtype="float16") = packed_params[157] gpt_neox_layers_9_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[158] gv413: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc330: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv413, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale4, alloc329, gpt_neox_layers_9_mlp_dense_h_to_4h_bias4, alloc330) R.vm.kill_object(alloc329) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight4) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale4) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_bias4) gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight4: R.Tensor((768, 384), dtype="uint32") = packed_params[159] gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale4: R.Tensor((768, 96), dtype="float16") = packed_params[160] gpt_neox_layers_9_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[161] gv414: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc331: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv414, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale4, alloc330, gpt_neox_layers_9_mlp_dense_4h_to_h_bias4, alloc331) R.vm.kill_object(alloc330) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight4) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale4) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_bias4) gpt_neox_layers_9_attention_dense_q_weight4: R.Tensor((768, 96), dtype="uint32") = packed_params[153] gpt_neox_layers_9_attention_dense_q_scale4: R.Tensor((768, 24), dtype="float16") = packed_params[154] gpt_neox_layers_9_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[155] gv415: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc332: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv415, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_9_attention_dense_q_weight4, gpt_neox_layers_9_attention_dense_q_scale4, reshape183, gpt_neox_layers_9_attention_dense_bias4, alloc331, alloc325, alloc332) R.vm.kill_object(alloc325) R.vm.kill_object(reshape183) R.vm.kill_object(alloc331) R.vm.kill_object(gpt_neox_layers_9_attention_dense_q_weight4) R.vm.kill_object(gpt_neox_layers_9_attention_dense_q_scale4) 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[162] gpt_neox_layers_10_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[163] gv416: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv416, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc332, gpt_neox_layers_10_input_layernorm_weight4, gpt_neox_layers_10_input_layernorm_bias4, alloc333) 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_q_weight4: R.Tensor((2304, 96), dtype="uint32") = packed_params[166] gpt_neox_layers_10_attention_query_key_value_q_scale4: R.Tensor((2304, 24), dtype="float16") = packed_params[167] gpt_neox_layers_10_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[168] gv417: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv417, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_10_attention_query_key_value_q_weight4, gpt_neox_layers_10_attention_query_key_value_q_scale4, alloc333, gpt_neox_layers_10_attention_query_key_value_bias4, alloc334) R.vm.kill_object(alloc333) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_q_weight4) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_q_scale4) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_bias4) gv418: 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", alloc334, gv418, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc334) 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),)) reshape185: R.Tensor((batch_size, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape184, gv419, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape184) 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),)) alloc335: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage20, R.prim_value(0), gv420, 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)), reshape185, alloc335) R.vm.kill_object(reshape185) gv421: 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", alloc335, gv421, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc335) gv422: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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, gv422, 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[164] gpt_neox_layers_10_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[165] gv423: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc336: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv423, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc332, gpt_neox_layers_10_post_attention_layernorm_weight4, gpt_neox_layers_10_post_attention_layernorm_bias4, alloc336) 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_q_weight4: R.Tensor((3072, 96), dtype="uint32") = packed_params[172] gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale4: R.Tensor((3072, 24), dtype="float16") = packed_params[173] gpt_neox_layers_10_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[174] gv424: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc337: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv424, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale4, alloc336, gpt_neox_layers_10_mlp_dense_h_to_4h_bias4, alloc337) R.vm.kill_object(alloc336) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight4) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale4) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_bias4) gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight4: R.Tensor((768, 384), dtype="uint32") = packed_params[175] gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale4: R.Tensor((768, 96), dtype="float16") = packed_params[176] gpt_neox_layers_10_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[177] gv425: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc338: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv425, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale4, alloc337, gpt_neox_layers_10_mlp_dense_4h_to_h_bias4, alloc338) R.vm.kill_object(alloc337) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight4) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale4) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_bias4) gpt_neox_layers_10_attention_dense_q_weight4: R.Tensor((768, 96), dtype="uint32") = packed_params[169] gpt_neox_layers_10_attention_dense_q_scale4: R.Tensor((768, 24), dtype="float16") = packed_params[170] gpt_neox_layers_10_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[171] gv426: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc339: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage22, R.prim_value(0), gv426, R.dtype("float16")) R.vm.kill_object(storage22) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_10_attention_dense_q_weight4, gpt_neox_layers_10_attention_dense_q_scale4, reshape187, gpt_neox_layers_10_attention_dense_bias4, alloc338, alloc332, alloc339) R.vm.kill_object(alloc332) R.vm.kill_object(reshape187) R.vm.kill_object(alloc338) R.vm.kill_object(gpt_neox_layers_10_attention_dense_q_weight4) R.vm.kill_object(gpt_neox_layers_10_attention_dense_q_scale4) 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[178] gpt_neox_layers_11_input_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[179] gv427: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv427, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc339, gpt_neox_layers_11_input_layernorm_weight4, gpt_neox_layers_11_input_layernorm_bias4, alloc340) 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_q_weight4: R.Tensor((2304, 96), dtype="uint32") = packed_params[182] gpt_neox_layers_11_attention_query_key_value_q_scale4: R.Tensor((2304, 24), dtype="float16") = packed_params[183] gpt_neox_layers_11_attention_query_key_value_bias4: R.Tensor((2304,), dtype="float16") = packed_params[184] gv428: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv428, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_11_attention_query_key_value_q_weight4, gpt_neox_layers_11_attention_query_key_value_q_scale4, alloc340, gpt_neox_layers_11_attention_query_key_value_bias4, alloc341) R.vm.kill_object(alloc340) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_q_weight4) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_q_scale4) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_bias4) gv429: 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", alloc341, gv429, sinfo_args=(R.Tensor((batch_size, 1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc341) gv430: R.Shape(ndim=3) = R.call_packed("vm.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, gv430, sinfo_args=(R.Tensor((batch_size, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape188) gv431: R.Shape(ndim=3) = R.call_packed("vm.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), gv431, 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)), reshape189, alloc342) R.vm.kill_object(reshape189) gv432: 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", alloc342, gv432, sinfo_args=(R.Tensor((batch_size, 1, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc342) 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),)) reshape191: R.Tensor((batch_size, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape190, gv433, 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[180] gpt_neox_layers_11_post_attention_layernorm_bias4: R.Tensor((768,), dtype="float16") = packed_params[181] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc343: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage21, R.prim_value(0), gv434, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm(alloc339, gpt_neox_layers_11_post_attention_layernorm_weight4, gpt_neox_layers_11_post_attention_layernorm_bias4, alloc343) 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_q_weight4: R.Tensor((3072, 96), dtype="uint32") = packed_params[188] gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale4: R.Tensor((3072, 24), dtype="float16") = packed_params[189] gpt_neox_layers_11_mlp_dense_h_to_4h_bias4: R.Tensor((3072,), dtype="float32") = packed_params[190] gv435: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc344: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage23, R.prim_value(0), gv435, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale4, alloc343, gpt_neox_layers_11_mlp_dense_h_to_4h_bias4, alloc344) R.vm.kill_object(alloc343) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight4) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale4) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_bias4) gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight4: R.Tensor((768, 384), dtype="uint32") = packed_params[191] gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale4: R.Tensor((768, 96), dtype="float16") = packed_params[192] gpt_neox_layers_11_mlp_dense_4h_to_h_bias4: R.Tensor((768,), dtype="float32") = packed_params[193] 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(1), 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), gv436, R.dtype("float16")) R.vm.kill_object(storage21) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale4, alloc344, gpt_neox_layers_11_mlp_dense_4h_to_h_bias4, alloc345) R.vm.kill_object(alloc344) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight4) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale4) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_bias4) gpt_neox_layers_11_attention_dense_q_weight4: R.Tensor((768, 96), dtype="uint32") = packed_params[185] gpt_neox_layers_11_attention_dense_q_scale4: R.Tensor((768, 24), dtype="float16") = packed_params[186] gpt_neox_layers_11_attention_dense_bias4: R.Tensor((768,), dtype="float16") = packed_params[187] 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(1), 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), gv437, R.dtype("float16")) R.vm.kill_object(storage23) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_11_attention_dense_q_weight4, gpt_neox_layers_11_attention_dense_q_scale4, reshape191, gpt_neox_layers_11_attention_dense_bias4, alloc345, alloc339, alloc346) R.vm.kill_object(alloc339) R.vm.kill_object(reshape191) R.vm.kill_object(alloc345) R.vm.kill_object(gpt_neox_layers_11_attention_dense_q_weight4) R.vm.kill_object(gpt_neox_layers_11_attention_dense_q_scale4) R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias4) gpt_neox_final_layer_norm_weight4: R.Tensor((768,), dtype="float16") = packed_params[194] gpt_neox_final_layer_norm_bias4: R.Tensor((768,), dtype="float16") = packed_params[195] gv438: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(1), 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), gv438, R.dtype("float16")) R.vm.kill_object(storage20) cls.gpt_neox_q4f16_1_layer_norm(alloc346, gpt_neox_final_layer_norm_weight4, gpt_neox_final_layer_norm_bias4, alloc347) R.vm.kill_object(alloc346) R.vm.kill_object(gpt_neox_final_layer_norm_weight4) R.vm.kill_object(gpt_neox_final_layer_norm_bias4) embed_out_q_weight4: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[196] embed_out_q_scale4: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[197] gv439: 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),)) storage24: R.Object = R.vm.alloc_storage(gv439, R.prim_value(0), R.dtype("uint8"), R.str("global")) 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(1), R.prim_value(1), sinfo_args=(R.Shape(ndim=3),)) alloc348: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage24, R.prim_value(0), gv440, R.dtype("float32")) R.vm.kill_object(storage24) cls.gpt_neox_q4f16_1_fused_dequantize_fused_NT_matmul4_cast2(embed_out_q_weight4, embed_out_q_scale4, alloc347, alloc348) R.vm.kill_object(alloc347) R.vm.kill_object(embed_out_q_weight4) R.vm.kill_object(embed_out_q_scale4) gv4: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc348, paged_kv_cache R.vm.kill_object(alloc348) gv441: R.Tensor(dtype="float32", ndim=3) = gv4[0] R.call_packed("vm.builtin.match_shape", gv441, 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", 96), dtype="uint32"), R.Tensor(("vocab_size", 24), 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, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor(("vocab_size", 96), dtype="uint32"), R.Tensor(("vocab_size", 24), 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(198), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv21: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[0] R.call_packed("vm.builtin.check_tensor_info", gv21, R.prim_value(2), R.dtype("uint32"), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv22: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[1] R.call_packed("vm.builtin.check_tensor_info", gv22, 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, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv23: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[196] R.call_packed("vm.builtin.check_tensor_info", gv23, R.prim_value(2), R.dtype("uint32"), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv24: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[197] R.call_packed("vm.builtin.check_tensor_info", gv24, 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, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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", gv21, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(2), R.prim_value(0), R.prim_value(96), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv22, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(2), R.prim_value(0), R.prim_value(24), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv23, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(2), R.prim_value(0), R.prim_value(96), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv24, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(2), R.prim_value(0), R.prim_value(24), R.str("ErrorContext(fn=batch_prefill, loc=param[3], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) cls.gpt_neox_q4f16_1_shape_func2(shape_heap) gpt_neox_layers_0_input_layernorm_weight3: R.Tensor((768,), dtype="float16") = packed_params[2] gpt_neox_layers_0_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[3] storage9: R.Object = R.vm.alloc_storage(R.shape([3145728]), R.prim_value(0), R.dtype("uint8"), R.str("global")) 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),)) alloc90: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv25, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(input_embeds, gpt_neox_layers_0_input_layernorm_weight3, gpt_neox_layers_0_input_layernorm_bias3, alloc90) 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_q_weight3: R.Tensor((2304, 96), dtype="uint32") = packed_params[6] gpt_neox_layers_0_attention_query_key_value_q_scale3: R.Tensor((2304, 24), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[8] storage10: R.Object = R.vm.alloc_storage(R.shape([9437184]), R.prim_value(0), R.dtype("uint8"), R.str("global")) 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(2304), sinfo_args=(R.Shape(ndim=3),)) alloc91: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv26, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_0_attention_query_key_value_q_weight3, gpt_neox_layers_0_attention_query_key_value_q_scale3, alloc90, gpt_neox_layers_0_attention_query_key_value_bias3, alloc91) R.vm.kill_object(alloc90) R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_q_weight3) R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_q_scale3) R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_bias3) gv27: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc91, gv27, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc91) gv28: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv28, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape96) gv29: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc92: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv29, R.dtype("float16")) _91: 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, alloc92) R.vm.kill_object(reshape97) gv30: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc92, gv30, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc92) gv31: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape99: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape98, gv31, 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[4] gpt_neox_layers_0_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[5] gv32: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc93: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv32, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(input_embeds, gpt_neox_layers_0_post_attention_layernorm_weight3, gpt_neox_layers_0_post_attention_layernorm_bias3, alloc93) 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_q_weight3: R.Tensor((3072, 96), dtype="uint32") = packed_params[12] gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3: R.Tensor((3072, 24), dtype="float16") = packed_params[13] gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[14] storage11: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) 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(3072), sinfo_args=(R.Shape(ndim=3),)) alloc94: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv33, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3, alloc93, gpt_neox_layers_0_mlp_dense_h_to_4h_bias3, alloc94) R.vm.kill_object(alloc93) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_bias3) gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3: R.Tensor((768, 384), dtype="uint32") = packed_params[15] gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3: R.Tensor((768, 96), dtype="float16") = packed_params[16] gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[17] 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),)) alloc95: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv34, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3, alloc94, gpt_neox_layers_0_mlp_dense_4h_to_h_bias3, alloc95) R.vm.kill_object(alloc94) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_bias3) gpt_neox_layers_0_attention_dense_q_weight3: R.Tensor((768, 96), dtype="uint32") = packed_params[9] gpt_neox_layers_0_attention_dense_q_scale3: R.Tensor((768, 24), dtype="float16") = packed_params[10] gpt_neox_layers_0_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[11] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc96: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv35, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_0_attention_dense_q_weight3, gpt_neox_layers_0_attention_dense_q_scale3, reshape99, gpt_neox_layers_0_attention_dense_bias3, alloc95, input_embeds, alloc96) R.vm.kill_object(reshape99) R.vm.kill_object(alloc95) R.vm.kill_object(gpt_neox_layers_0_attention_dense_q_weight3) R.vm.kill_object(gpt_neox_layers_0_attention_dense_q_scale3) 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[18] gpt_neox_layers_1_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[19] 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),)) alloc97: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv36, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc96, gpt_neox_layers_1_input_layernorm_weight3, gpt_neox_layers_1_input_layernorm_bias3, alloc97) 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_q_weight3: R.Tensor((2304, 96), dtype="uint32") = packed_params[22] gpt_neox_layers_1_attention_query_key_value_q_scale3: R.Tensor((2304, 24), dtype="float16") = packed_params[23] gpt_neox_layers_1_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[24] 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(2304), sinfo_args=(R.Shape(ndim=3),)) alloc98: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv37, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_1_attention_query_key_value_q_weight3, gpt_neox_layers_1_attention_query_key_value_q_scale3, alloc97, gpt_neox_layers_1_attention_query_key_value_bias3, alloc98) R.vm.kill_object(alloc97) R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_q_weight3) R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_q_scale3) R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_bias3) gv38: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc98, gv38, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc98) gv39: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv39, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape100) gv40: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc99: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv40, R.dtype("float16")) _98: 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, alloc99) R.vm.kill_object(reshape101) gv41: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc99, gv41, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc99) gv42: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape103: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape102, gv42, 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[20] gpt_neox_layers_1_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[21] gv43: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc100: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv43, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc96, gpt_neox_layers_1_post_attention_layernorm_weight3, gpt_neox_layers_1_post_attention_layernorm_bias3, alloc100) 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_q_weight3: R.Tensor((3072, 96), dtype="uint32") = packed_params[28] gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale3: R.Tensor((3072, 24), dtype="float16") = packed_params[29] gpt_neox_layers_1_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[30] storage12: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) 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(3072), sinfo_args=(R.Shape(ndim=3),)) alloc101: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv44, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale3, alloc100, gpt_neox_layers_1_mlp_dense_h_to_4h_bias3, alloc101) R.vm.kill_object(alloc100) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight3) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale3) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_bias3) gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight3: R.Tensor((768, 384), dtype="uint32") = packed_params[31] gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale3: R.Tensor((768, 96), dtype="float16") = packed_params[32] gpt_neox_layers_1_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[33] 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),)) alloc102: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv45, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale3, alloc101, gpt_neox_layers_1_mlp_dense_4h_to_h_bias3, alloc102) R.vm.kill_object(alloc101) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight3) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale3) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_bias3) gpt_neox_layers_1_attention_dense_q_weight3: R.Tensor((768, 96), dtype="uint32") = packed_params[25] gpt_neox_layers_1_attention_dense_q_scale3: R.Tensor((768, 24), dtype="float16") = packed_params[26] gpt_neox_layers_1_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[27] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc103: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv46, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_1_attention_dense_q_weight3, gpt_neox_layers_1_attention_dense_q_scale3, reshape103, gpt_neox_layers_1_attention_dense_bias3, alloc102, alloc96, alloc103) R.vm.kill_object(alloc96) R.vm.kill_object(reshape103) R.vm.kill_object(alloc102) R.vm.kill_object(gpt_neox_layers_1_attention_dense_q_weight3) R.vm.kill_object(gpt_neox_layers_1_attention_dense_q_scale3) 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[34] gpt_neox_layers_2_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[35] 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),)) alloc104: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv47, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc103, gpt_neox_layers_2_input_layernorm_weight3, gpt_neox_layers_2_input_layernorm_bias3, alloc104) 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_q_weight3: R.Tensor((2304, 96), dtype="uint32") = packed_params[38] gpt_neox_layers_2_attention_query_key_value_q_scale3: R.Tensor((2304, 24), dtype="float16") = packed_params[39] gpt_neox_layers_2_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[40] 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(2304), sinfo_args=(R.Shape(ndim=3),)) alloc105: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv48, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_2_attention_query_key_value_q_weight3, gpt_neox_layers_2_attention_query_key_value_q_scale3, alloc104, gpt_neox_layers_2_attention_query_key_value_bias3, alloc105) R.vm.kill_object(alloc104) R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_q_weight3) R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_q_scale3) R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_bias3) gv49: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc105, gv49, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc105) gv50: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv50, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape104) gv51: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc106: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv51, R.dtype("float16")) _105: 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, alloc106) R.vm.kill_object(reshape105) gv52: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc106, gv52, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc106) gv53: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape107: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape106, gv53, 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[36] gpt_neox_layers_2_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[37] gv54: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc107: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv54, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc103, gpt_neox_layers_2_post_attention_layernorm_weight3, gpt_neox_layers_2_post_attention_layernorm_bias3, alloc107) 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_q_weight3: R.Tensor((3072, 96), dtype="uint32") = packed_params[44] gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale3: R.Tensor((3072, 24), dtype="float16") = packed_params[45] gpt_neox_layers_2_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[46] 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(3072), sinfo_args=(R.Shape(ndim=3),)) alloc108: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv55, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale3, alloc107, gpt_neox_layers_2_mlp_dense_h_to_4h_bias3, alloc108) R.vm.kill_object(alloc107) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight3) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale3) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_bias3) gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight3: R.Tensor((768, 384), dtype="uint32") = packed_params[47] gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale3: R.Tensor((768, 96), dtype="float16") = packed_params[48] gpt_neox_layers_2_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[49] 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),)) alloc109: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv56, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale3, alloc108, gpt_neox_layers_2_mlp_dense_4h_to_h_bias3, alloc109) R.vm.kill_object(alloc108) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight3) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale3) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_bias3) gpt_neox_layers_2_attention_dense_q_weight3: R.Tensor((768, 96), dtype="uint32") = packed_params[41] gpt_neox_layers_2_attention_dense_q_scale3: R.Tensor((768, 24), dtype="float16") = packed_params[42] gpt_neox_layers_2_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[43] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc110: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv57, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_2_attention_dense_q_weight3, gpt_neox_layers_2_attention_dense_q_scale3, reshape107, gpt_neox_layers_2_attention_dense_bias3, alloc109, alloc103, alloc110) R.vm.kill_object(alloc103) R.vm.kill_object(reshape107) R.vm.kill_object(alloc109) R.vm.kill_object(gpt_neox_layers_2_attention_dense_q_weight3) R.vm.kill_object(gpt_neox_layers_2_attention_dense_q_scale3) 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[50] gpt_neox_layers_3_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[51] 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),)) alloc111: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv58, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc110, gpt_neox_layers_3_input_layernorm_weight3, gpt_neox_layers_3_input_layernorm_bias3, alloc111) 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_q_weight3: R.Tensor((2304, 96), dtype="uint32") = packed_params[54] gpt_neox_layers_3_attention_query_key_value_q_scale3: R.Tensor((2304, 24), dtype="float16") = packed_params[55] gpt_neox_layers_3_attention_query_key_value_bias3: R.Tensor((2304,), 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(2304), sinfo_args=(R.Shape(ndim=3),)) alloc112: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv59, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_3_attention_query_key_value_q_weight3, gpt_neox_layers_3_attention_query_key_value_q_scale3, alloc111, gpt_neox_layers_3_attention_query_key_value_bias3, alloc112) R.vm.kill_object(alloc111) R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_q_weight3) R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_q_scale3) R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_bias3) gv60: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc112, gv60, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc112) gv61: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv61, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape108) gv62: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc113: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv62, R.dtype("float16")) _112: 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, alloc113) R.vm.kill_object(reshape109) gv63: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc113, gv63, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc113) gv64: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape111: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape110, gv64, 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[52] gpt_neox_layers_3_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[53] gv65: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc114: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv65, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc110, gpt_neox_layers_3_post_attention_layernorm_weight3, gpt_neox_layers_3_post_attention_layernorm_bias3, alloc114) 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_q_weight3: R.Tensor((3072, 96), dtype="uint32") = packed_params[60] gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale3: R.Tensor((3072, 24), dtype="float16") = packed_params[61] gpt_neox_layers_3_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[62] 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(3072), sinfo_args=(R.Shape(ndim=3),)) alloc115: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv66, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale3, alloc114, gpt_neox_layers_3_mlp_dense_h_to_4h_bias3, alloc115) R.vm.kill_object(alloc114) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight3) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale3) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_bias3) gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight3: R.Tensor((768, 384), dtype="uint32") = packed_params[63] gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale3: R.Tensor((768, 96), dtype="float16") = packed_params[64] gpt_neox_layers_3_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[65] 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),)) alloc116: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv67, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale3, alloc115, gpt_neox_layers_3_mlp_dense_4h_to_h_bias3, alloc116) R.vm.kill_object(alloc115) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight3) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale3) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_bias3) gpt_neox_layers_3_attention_dense_q_weight3: R.Tensor((768, 96), dtype="uint32") = packed_params[57] gpt_neox_layers_3_attention_dense_q_scale3: R.Tensor((768, 24), dtype="float16") = packed_params[58] gpt_neox_layers_3_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[59] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc117: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv68, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_3_attention_dense_q_weight3, gpt_neox_layers_3_attention_dense_q_scale3, reshape111, gpt_neox_layers_3_attention_dense_bias3, alloc116, alloc110, alloc117) R.vm.kill_object(alloc110) R.vm.kill_object(reshape111) R.vm.kill_object(alloc116) R.vm.kill_object(gpt_neox_layers_3_attention_dense_q_weight3) R.vm.kill_object(gpt_neox_layers_3_attention_dense_q_scale3) 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[66] gpt_neox_layers_4_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[67] 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),)) alloc118: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv69, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc117, gpt_neox_layers_4_input_layernorm_weight3, gpt_neox_layers_4_input_layernorm_bias3, alloc118) 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_q_weight3: R.Tensor((2304, 96), dtype="uint32") = packed_params[70] gpt_neox_layers_4_attention_query_key_value_q_scale3: R.Tensor((2304, 24), dtype="float16") = packed_params[71] gpt_neox_layers_4_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[72] 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(2304), sinfo_args=(R.Shape(ndim=3),)) alloc119: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv70, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_4_attention_query_key_value_q_weight3, gpt_neox_layers_4_attention_query_key_value_q_scale3, alloc118, gpt_neox_layers_4_attention_query_key_value_bias3, alloc119) R.vm.kill_object(alloc118) R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_q_weight3) R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_q_scale3) R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_bias3) gv71: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc119, gv71, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc119) gv72: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv72, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape112) gv73: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc120: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv73, R.dtype("float16")) _119: 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, alloc120) R.vm.kill_object(reshape113) gv74: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc120, gv74, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc120) gv75: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape115: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape114, gv75, 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[68] gpt_neox_layers_4_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[69] gv76: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc121: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv76, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc117, gpt_neox_layers_4_post_attention_layernorm_weight3, gpt_neox_layers_4_post_attention_layernorm_bias3, alloc121) 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_q_weight3: R.Tensor((3072, 96), dtype="uint32") = packed_params[76] gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale3: R.Tensor((3072, 24), dtype="float16") = packed_params[77] gpt_neox_layers_4_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[78] 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(3072), sinfo_args=(R.Shape(ndim=3),)) alloc122: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv77, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale3, alloc121, gpt_neox_layers_4_mlp_dense_h_to_4h_bias3, alloc122) R.vm.kill_object(alloc121) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight3) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale3) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_bias3) gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight3: R.Tensor((768, 384), dtype="uint32") = packed_params[79] gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale3: R.Tensor((768, 96), dtype="float16") = packed_params[80] gpt_neox_layers_4_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[81] 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),)) alloc123: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv78, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale3, alloc122, gpt_neox_layers_4_mlp_dense_4h_to_h_bias3, alloc123) R.vm.kill_object(alloc122) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight3) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale3) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_bias3) gpt_neox_layers_4_attention_dense_q_weight3: R.Tensor((768, 96), dtype="uint32") = packed_params[73] gpt_neox_layers_4_attention_dense_q_scale3: R.Tensor((768, 24), dtype="float16") = packed_params[74] gpt_neox_layers_4_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[75] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc124: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv79, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_4_attention_dense_q_weight3, gpt_neox_layers_4_attention_dense_q_scale3, reshape115, gpt_neox_layers_4_attention_dense_bias3, alloc123, alloc117, alloc124) R.vm.kill_object(alloc117) R.vm.kill_object(reshape115) R.vm.kill_object(alloc123) R.vm.kill_object(gpt_neox_layers_4_attention_dense_q_weight3) R.vm.kill_object(gpt_neox_layers_4_attention_dense_q_scale3) 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[82] gpt_neox_layers_5_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[83] 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),)) alloc125: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv80, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc124, gpt_neox_layers_5_input_layernorm_weight3, gpt_neox_layers_5_input_layernorm_bias3, alloc125) 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_q_weight3: R.Tensor((2304, 96), dtype="uint32") = packed_params[86] gpt_neox_layers_5_attention_query_key_value_q_scale3: R.Tensor((2304, 24), dtype="float16") = packed_params[87] gpt_neox_layers_5_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[88] 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(2304), sinfo_args=(R.Shape(ndim=3),)) alloc126: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv81, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_5_attention_query_key_value_q_weight3, gpt_neox_layers_5_attention_query_key_value_q_scale3, alloc125, gpt_neox_layers_5_attention_query_key_value_bias3, alloc126) R.vm.kill_object(alloc125) R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_q_weight3) R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_q_scale3) R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_bias3) gv82: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc126, gv82, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc126) gv83: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv83, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape116) gv84: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc127: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv84, R.dtype("float16")) _126: 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, alloc127) R.vm.kill_object(reshape117) gv85: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc127, gv85, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc127) gv86: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape119: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape118, gv86, 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[84] gpt_neox_layers_5_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[85] gv87: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc128: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv87, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc124, gpt_neox_layers_5_post_attention_layernorm_weight3, gpt_neox_layers_5_post_attention_layernorm_bias3, alloc128) 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_q_weight3: R.Tensor((3072, 96), dtype="uint32") = packed_params[92] gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale3: R.Tensor((3072, 24), dtype="float16") = packed_params[93] gpt_neox_layers_5_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[94] 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(3072), sinfo_args=(R.Shape(ndim=3),)) alloc129: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv88, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale3, alloc128, gpt_neox_layers_5_mlp_dense_h_to_4h_bias3, alloc129) R.vm.kill_object(alloc128) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight3) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale3) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_bias3) gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight3: R.Tensor((768, 384), dtype="uint32") = packed_params[95] gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale3: R.Tensor((768, 96), dtype="float16") = packed_params[96] gpt_neox_layers_5_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[97] 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),)) alloc130: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv89, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale3, alloc129, gpt_neox_layers_5_mlp_dense_4h_to_h_bias3, alloc130) R.vm.kill_object(alloc129) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight3) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale3) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_bias3) gpt_neox_layers_5_attention_dense_q_weight3: R.Tensor((768, 96), dtype="uint32") = packed_params[89] gpt_neox_layers_5_attention_dense_q_scale3: R.Tensor((768, 24), dtype="float16") = packed_params[90] gpt_neox_layers_5_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[91] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc131: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv90, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_5_attention_dense_q_weight3, gpt_neox_layers_5_attention_dense_q_scale3, reshape119, gpt_neox_layers_5_attention_dense_bias3, alloc130, alloc124, alloc131) R.vm.kill_object(alloc124) R.vm.kill_object(reshape119) R.vm.kill_object(alloc130) R.vm.kill_object(gpt_neox_layers_5_attention_dense_q_weight3) R.vm.kill_object(gpt_neox_layers_5_attention_dense_q_scale3) 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[98] gpt_neox_layers_6_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[99] 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),)) alloc132: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv91, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc131, gpt_neox_layers_6_input_layernorm_weight3, gpt_neox_layers_6_input_layernorm_bias3, alloc132) 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_q_weight3: R.Tensor((2304, 96), dtype="uint32") = packed_params[102] gpt_neox_layers_6_attention_query_key_value_q_scale3: R.Tensor((2304, 24), dtype="float16") = packed_params[103] gpt_neox_layers_6_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[104] 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(2304), sinfo_args=(R.Shape(ndim=3),)) alloc133: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv92, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_6_attention_query_key_value_q_weight3, gpt_neox_layers_6_attention_query_key_value_q_scale3, alloc132, gpt_neox_layers_6_attention_query_key_value_bias3, alloc133) R.vm.kill_object(alloc132) R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_q_weight3) R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_q_scale3) R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_bias3) gv93: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc133, gv93, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc133) gv94: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv94, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape120) gv95: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc134: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv95, R.dtype("float16")) _133: 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, alloc134) R.vm.kill_object(reshape121) gv96: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc134, gv96, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc134) gv97: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape123: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape122, gv97, 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[100] gpt_neox_layers_6_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[101] gv98: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc135: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv98, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc131, gpt_neox_layers_6_post_attention_layernorm_weight3, gpt_neox_layers_6_post_attention_layernorm_bias3, alloc135) 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_q_weight3: R.Tensor((3072, 96), dtype="uint32") = packed_params[108] gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale3: R.Tensor((3072, 24), dtype="float16") = packed_params[109] gpt_neox_layers_6_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[110] 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(3072), sinfo_args=(R.Shape(ndim=3),)) alloc136: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv99, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale3, alloc135, gpt_neox_layers_6_mlp_dense_h_to_4h_bias3, alloc136) R.vm.kill_object(alloc135) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight3) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale3) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_bias3) gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight3: R.Tensor((768, 384), dtype="uint32") = packed_params[111] gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale3: R.Tensor((768, 96), dtype="float16") = packed_params[112] gpt_neox_layers_6_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[113] 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),)) alloc137: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv100, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale3, alloc136, gpt_neox_layers_6_mlp_dense_4h_to_h_bias3, alloc137) R.vm.kill_object(alloc136) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight3) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale3) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_bias3) gpt_neox_layers_6_attention_dense_q_weight3: R.Tensor((768, 96), dtype="uint32") = packed_params[105] gpt_neox_layers_6_attention_dense_q_scale3: R.Tensor((768, 24), dtype="float16") = packed_params[106] gpt_neox_layers_6_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[107] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc138: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv101, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_6_attention_dense_q_weight3, gpt_neox_layers_6_attention_dense_q_scale3, reshape123, gpt_neox_layers_6_attention_dense_bias3, alloc137, alloc131, alloc138) R.vm.kill_object(alloc131) R.vm.kill_object(reshape123) R.vm.kill_object(alloc137) R.vm.kill_object(gpt_neox_layers_6_attention_dense_q_weight3) R.vm.kill_object(gpt_neox_layers_6_attention_dense_q_scale3) 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[114] gpt_neox_layers_7_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[115] 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),)) alloc139: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv102, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc138, gpt_neox_layers_7_input_layernorm_weight3, gpt_neox_layers_7_input_layernorm_bias3, alloc139) 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_q_weight3: R.Tensor((2304, 96), dtype="uint32") = packed_params[118] gpt_neox_layers_7_attention_query_key_value_q_scale3: R.Tensor((2304, 24), dtype="float16") = packed_params[119] gpt_neox_layers_7_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[120] 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(2304), sinfo_args=(R.Shape(ndim=3),)) alloc140: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv103, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_7_attention_query_key_value_q_weight3, gpt_neox_layers_7_attention_query_key_value_q_scale3, alloc139, gpt_neox_layers_7_attention_query_key_value_bias3, alloc140) R.vm.kill_object(alloc139) R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_q_weight3) R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_q_scale3) R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_bias3) gv104: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc140, gv104, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc140) gv105: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv105, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape124) gv106: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc141: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv106, R.dtype("float16")) _140: 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, alloc141) R.vm.kill_object(reshape125) gv107: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc141, gv107, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc141) gv108: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape127: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape126, gv108, 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[116] gpt_neox_layers_7_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[117] gv109: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc142: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv109, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc138, gpt_neox_layers_7_post_attention_layernorm_weight3, gpt_neox_layers_7_post_attention_layernorm_bias3, alloc142) 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_q_weight3: R.Tensor((3072, 96), dtype="uint32") = packed_params[124] gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale3: R.Tensor((3072, 24), dtype="float16") = packed_params[125] gpt_neox_layers_7_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[126] 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(3072), sinfo_args=(R.Shape(ndim=3),)) alloc143: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv110, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale3, alloc142, gpt_neox_layers_7_mlp_dense_h_to_4h_bias3, alloc143) R.vm.kill_object(alloc142) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight3) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale3) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_bias3) gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight3: R.Tensor((768, 384), dtype="uint32") = packed_params[127] gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale3: R.Tensor((768, 96), dtype="float16") = packed_params[128] gpt_neox_layers_7_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[129] 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),)) alloc144: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv111, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale3, alloc143, gpt_neox_layers_7_mlp_dense_4h_to_h_bias3, alloc144) R.vm.kill_object(alloc143) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight3) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale3) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_bias3) gpt_neox_layers_7_attention_dense_q_weight3: R.Tensor((768, 96), dtype="uint32") = packed_params[121] gpt_neox_layers_7_attention_dense_q_scale3: R.Tensor((768, 24), dtype="float16") = packed_params[122] gpt_neox_layers_7_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[123] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc145: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv112, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_7_attention_dense_q_weight3, gpt_neox_layers_7_attention_dense_q_scale3, reshape127, gpt_neox_layers_7_attention_dense_bias3, alloc144, alloc138, alloc145) R.vm.kill_object(alloc138) R.vm.kill_object(reshape127) R.vm.kill_object(alloc144) R.vm.kill_object(gpt_neox_layers_7_attention_dense_q_weight3) R.vm.kill_object(gpt_neox_layers_7_attention_dense_q_scale3) 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[130] gpt_neox_layers_8_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[131] 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),)) alloc146: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv113, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc145, gpt_neox_layers_8_input_layernorm_weight3, gpt_neox_layers_8_input_layernorm_bias3, alloc146) 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_q_weight3: R.Tensor((2304, 96), dtype="uint32") = packed_params[134] gpt_neox_layers_8_attention_query_key_value_q_scale3: R.Tensor((2304, 24), dtype="float16") = packed_params[135] gpt_neox_layers_8_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[136] 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(2304), sinfo_args=(R.Shape(ndim=3),)) alloc147: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv114, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_8_attention_query_key_value_q_weight3, gpt_neox_layers_8_attention_query_key_value_q_scale3, alloc146, gpt_neox_layers_8_attention_query_key_value_bias3, alloc147) R.vm.kill_object(alloc146) R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_q_weight3) R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_q_scale3) R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_bias3) gv115: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc147, gv115, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc147) gv116: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv116, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape128) gv117: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc148: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv117, R.dtype("float16")) _147: 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, alloc148) R.vm.kill_object(reshape129) gv118: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc148, gv118, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc148) gv119: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape131: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape130, gv119, 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[132] gpt_neox_layers_8_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[133] gv120: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc149: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv120, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc145, gpt_neox_layers_8_post_attention_layernorm_weight3, gpt_neox_layers_8_post_attention_layernorm_bias3, alloc149) 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_q_weight3: R.Tensor((3072, 96), dtype="uint32") = packed_params[140] gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale3: R.Tensor((3072, 24), dtype="float16") = packed_params[141] gpt_neox_layers_8_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[142] 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(3072), sinfo_args=(R.Shape(ndim=3),)) alloc150: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv121, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale3, alloc149, gpt_neox_layers_8_mlp_dense_h_to_4h_bias3, alloc150) R.vm.kill_object(alloc149) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight3) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale3) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_bias3) gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight3: R.Tensor((768, 384), dtype="uint32") = packed_params[143] gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale3: R.Tensor((768, 96), dtype="float16") = packed_params[144] gpt_neox_layers_8_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[145] 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),)) alloc151: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv122, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale3, alloc150, gpt_neox_layers_8_mlp_dense_4h_to_h_bias3, alloc151) R.vm.kill_object(alloc150) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight3) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale3) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_bias3) gpt_neox_layers_8_attention_dense_q_weight3: R.Tensor((768, 96), dtype="uint32") = packed_params[137] gpt_neox_layers_8_attention_dense_q_scale3: R.Tensor((768, 24), dtype="float16") = packed_params[138] gpt_neox_layers_8_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[139] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc152: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv123, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_8_attention_dense_q_weight3, gpt_neox_layers_8_attention_dense_q_scale3, reshape131, gpt_neox_layers_8_attention_dense_bias3, alloc151, alloc145, alloc152) R.vm.kill_object(alloc145) R.vm.kill_object(reshape131) R.vm.kill_object(alloc151) R.vm.kill_object(gpt_neox_layers_8_attention_dense_q_weight3) R.vm.kill_object(gpt_neox_layers_8_attention_dense_q_scale3) 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[146] gpt_neox_layers_9_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[147] 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),)) alloc153: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv124, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc152, gpt_neox_layers_9_input_layernorm_weight3, gpt_neox_layers_9_input_layernorm_bias3, alloc153) 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_q_weight3: R.Tensor((2304, 96), dtype="uint32") = packed_params[150] gpt_neox_layers_9_attention_query_key_value_q_scale3: R.Tensor((2304, 24), dtype="float16") = packed_params[151] gpt_neox_layers_9_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[152] 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(2304), sinfo_args=(R.Shape(ndim=3),)) alloc154: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv125, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_9_attention_query_key_value_q_weight3, gpt_neox_layers_9_attention_query_key_value_q_scale3, alloc153, gpt_neox_layers_9_attention_query_key_value_bias3, alloc154) R.vm.kill_object(alloc153) R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_q_weight3) R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_q_scale3) R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_bias3) gv126: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc154, gv126, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc154) gv127: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv127, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape132) gv128: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc155: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv128, R.dtype("float16")) _154: 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, alloc155) R.vm.kill_object(reshape133) gv129: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc155, gv129, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc155) gv130: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape135: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape134, gv130, 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[148] gpt_neox_layers_9_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[149] gv131: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc156: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv131, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc152, gpt_neox_layers_9_post_attention_layernorm_weight3, gpt_neox_layers_9_post_attention_layernorm_bias3, alloc156) 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_q_weight3: R.Tensor((3072, 96), dtype="uint32") = packed_params[156] gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale3: R.Tensor((3072, 24), dtype="float16") = packed_params[157] gpt_neox_layers_9_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[158] 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(3072), sinfo_args=(R.Shape(ndim=3),)) alloc157: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv132, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale3, alloc156, gpt_neox_layers_9_mlp_dense_h_to_4h_bias3, alloc157) R.vm.kill_object(alloc156) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight3) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale3) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_bias3) gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight3: R.Tensor((768, 384), dtype="uint32") = packed_params[159] gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale3: R.Tensor((768, 96), dtype="float16") = packed_params[160] gpt_neox_layers_9_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[161] 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),)) alloc158: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv133, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale3, alloc157, gpt_neox_layers_9_mlp_dense_4h_to_h_bias3, alloc158) R.vm.kill_object(alloc157) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight3) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale3) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_bias3) gpt_neox_layers_9_attention_dense_q_weight3: R.Tensor((768, 96), dtype="uint32") = packed_params[153] gpt_neox_layers_9_attention_dense_q_scale3: R.Tensor((768, 24), dtype="float16") = packed_params[154] gpt_neox_layers_9_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[155] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc159: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv134, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_9_attention_dense_q_weight3, gpt_neox_layers_9_attention_dense_q_scale3, reshape135, gpt_neox_layers_9_attention_dense_bias3, alloc158, alloc152, alloc159) R.vm.kill_object(alloc152) R.vm.kill_object(reshape135) R.vm.kill_object(alloc158) R.vm.kill_object(gpt_neox_layers_9_attention_dense_q_weight3) R.vm.kill_object(gpt_neox_layers_9_attention_dense_q_scale3) 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[162] gpt_neox_layers_10_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[163] 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),)) alloc160: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv135, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc159, gpt_neox_layers_10_input_layernorm_weight3, gpt_neox_layers_10_input_layernorm_bias3, alloc160) 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_q_weight3: R.Tensor((2304, 96), dtype="uint32") = packed_params[166] gpt_neox_layers_10_attention_query_key_value_q_scale3: R.Tensor((2304, 24), dtype="float16") = packed_params[167] gpt_neox_layers_10_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[168] 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(2304), sinfo_args=(R.Shape(ndim=3),)) alloc161: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv136, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_10_attention_query_key_value_q_weight3, gpt_neox_layers_10_attention_query_key_value_q_scale3, alloc160, gpt_neox_layers_10_attention_query_key_value_bias3, alloc161) R.vm.kill_object(alloc160) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_q_weight3) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_q_scale3) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_bias3) gv137: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc161, gv137, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc161) gv138: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv138, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape136) gv139: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc162: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv139, R.dtype("float16")) _161: 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, alloc162) R.vm.kill_object(reshape137) gv140: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc162, gv140, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc162) gv141: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape139: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape138, gv141, 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[164] gpt_neox_layers_10_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[165] gv142: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc163: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv142, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc159, gpt_neox_layers_10_post_attention_layernorm_weight3, gpt_neox_layers_10_post_attention_layernorm_bias3, alloc163) 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_q_weight3: R.Tensor((3072, 96), dtype="uint32") = packed_params[172] gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale3: R.Tensor((3072, 24), dtype="float16") = packed_params[173] gpt_neox_layers_10_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[174] gv143: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) alloc164: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv143, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale3, alloc163, gpt_neox_layers_10_mlp_dense_h_to_4h_bias3, alloc164) R.vm.kill_object(alloc163) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight3) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale3) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_bias3) gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight3: R.Tensor((768, 384), dtype="uint32") = packed_params[175] gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale3: R.Tensor((768, 96), dtype="float16") = packed_params[176] gpt_neox_layers_10_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[177] 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),)) alloc165: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv144, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale3, alloc164, gpt_neox_layers_10_mlp_dense_4h_to_h_bias3, alloc165) R.vm.kill_object(alloc164) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight3) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale3) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_bias3) gpt_neox_layers_10_attention_dense_q_weight3: R.Tensor((768, 96), dtype="uint32") = packed_params[169] gpt_neox_layers_10_attention_dense_q_scale3: R.Tensor((768, 24), dtype="float16") = packed_params[170] gpt_neox_layers_10_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[171] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc166: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage11, R.prim_value(0), gv145, R.dtype("float16")) R.vm.kill_object(storage11) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_10_attention_dense_q_weight3, gpt_neox_layers_10_attention_dense_q_scale3, reshape139, gpt_neox_layers_10_attention_dense_bias3, alloc165, alloc159, alloc166) R.vm.kill_object(alloc159) R.vm.kill_object(reshape139) R.vm.kill_object(alloc165) R.vm.kill_object(gpt_neox_layers_10_attention_dense_q_weight3) R.vm.kill_object(gpt_neox_layers_10_attention_dense_q_scale3) 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[178] gpt_neox_layers_11_input_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[179] gv146: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc167: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv146, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc166, gpt_neox_layers_11_input_layernorm_weight3, gpt_neox_layers_11_input_layernorm_bias3, alloc167) 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_q_weight3: R.Tensor((2304, 96), dtype="uint32") = packed_params[182] gpt_neox_layers_11_attention_query_key_value_q_scale3: R.Tensor((2304, 24), dtype="float16") = packed_params[183] gpt_neox_layers_11_attention_query_key_value_bias3: R.Tensor((2304,), dtype="float16") = packed_params[184] gv147: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc168: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv147, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_11_attention_query_key_value_q_weight3, gpt_neox_layers_11_attention_query_key_value_q_scale3, alloc167, gpt_neox_layers_11_attention_query_key_value_bias3, alloc168) R.vm.kill_object(alloc167) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_q_weight3) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_q_scale3) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_bias3) gv148: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc168, gv148, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc168) gv149: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv149, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape140) gv150: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc169: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv150, R.dtype("float16")) _168: 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, alloc169) R.vm.kill_object(reshape141) gv151: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc169, gv151, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc169) 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(768), sinfo_args=(R.Shape(ndim=3),)) reshape143: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape142, gv152, 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[180] gpt_neox_layers_11_post_attention_layernorm_bias3: R.Tensor((768,), dtype="float16") = packed_params[181] 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),)) alloc170: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv153, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc166, gpt_neox_layers_11_post_attention_layernorm_weight3, gpt_neox_layers_11_post_attention_layernorm_bias3, alloc170) 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_q_weight3: R.Tensor((3072, 96), dtype="uint32") = packed_params[188] gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale3: R.Tensor((3072, 24), dtype="float16") = packed_params[189] gpt_neox_layers_11_mlp_dense_h_to_4h_bias3: R.Tensor((3072,), dtype="float32") = packed_params[190] 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(3072), sinfo_args=(R.Shape(ndim=3),)) alloc171: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv154, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale3, alloc170, gpt_neox_layers_11_mlp_dense_h_to_4h_bias3, alloc171) R.vm.kill_object(alloc170) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight3) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale3) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_bias3) gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight3: R.Tensor((768, 384), dtype="uint32") = packed_params[191] gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale3: R.Tensor((768, 96), dtype="float16") = packed_params[192] gpt_neox_layers_11_mlp_dense_4h_to_h_bias3: R.Tensor((768,), dtype="float32") = packed_params[193] 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),)) alloc172: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage10, R.prim_value(0), gv155, R.dtype("float16")) R.vm.kill_object(storage10) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale3, alloc171, gpt_neox_layers_11_mlp_dense_4h_to_h_bias3, alloc172) R.vm.kill_object(alloc171) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight3) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale3) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_bias3) gpt_neox_layers_11_attention_dense_q_weight3: R.Tensor((768, 96), dtype="uint32") = packed_params[185] gpt_neox_layers_11_attention_dense_q_scale3: R.Tensor((768, 24), dtype="float16") = packed_params[186] gpt_neox_layers_11_attention_dense_bias3: R.Tensor((768,), dtype="float16") = packed_params[187] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc173: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage12, R.prim_value(0), gv156, R.dtype("float16")) R.vm.kill_object(storage12) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_11_attention_dense_q_weight3, gpt_neox_layers_11_attention_dense_q_scale3, reshape143, gpt_neox_layers_11_attention_dense_bias3, alloc172, alloc166, alloc173) R.vm.kill_object(alloc166) R.vm.kill_object(reshape143) R.vm.kill_object(alloc172) R.vm.kill_object(gpt_neox_layers_11_attention_dense_q_weight3) R.vm.kill_object(gpt_neox_layers_11_attention_dense_q_scale3) R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias3) gpt_neox_final_layer_norm_weight3: R.Tensor((768,), dtype="float16") = packed_params[194] gpt_neox_final_layer_norm_bias3: R.Tensor((768,), dtype="float16") = packed_params[195] gv157: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc174: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage9, R.prim_value(0), gv157, R.dtype("float16")) R.vm.kill_object(storage9) cls.gpt_neox_q4f16_1_layer_norm1(alloc173, gpt_neox_final_layer_norm_weight3, gpt_neox_final_layer_norm_bias3, alloc174) R.vm.kill_object(alloc173) R.vm.kill_object(gpt_neox_final_layer_norm_weight3) R.vm.kill_object(gpt_neox_final_layer_norm_bias3) storage13: R.Object = R.vm.alloc_storage(R.shape([196608]), R.prim_value(0), R.dtype("uint8"), R.str("global")) gv158: R.Shape(ndim=3) = R.call_packed("vm.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),)) alloc175: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage13, R.prim_value(0), gv158, R.dtype("float16")) R.vm.kill_object(storage13) cls.gpt_neox_q4f16_1_take(alloc174, logit_positions, alloc175) R.vm.kill_object(alloc174) embed_out_q_weight3: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[196] embed_out_q_scale3: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[197] gv159: 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(gv159, R.prim_value(0), R.dtype("uint8"), R.str("global")) gv160: R.Shape(ndim=3) = R.call_packed("vm.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),)) alloc176: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage14, R.prim_value(0), gv160, R.dtype("float32")) R.vm.kill_object(storage14) cls.gpt_neox_q4f16_1_fused_dequantize_fused_NT_matmul9_cast5(embed_out_q_weight3, embed_out_q_scale3, alloc175, alloc176) R.vm.kill_object(alloc175) R.vm.kill_object(embed_out_q_weight3) R.vm.kill_object(embed_out_q_scale3) gv3: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc176, paged_kv_cache R.vm.kill_object(alloc176) gv161: R.Tensor(dtype="float32", ndim=3) = gv3[0] R.call_packed("vm.builtin.match_shape", gv161, shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(3), R.prim_value(1), R.prim_value(3), R.prim_value(2), R.str("ErrorContext(fn=batch_prefill, loc=return, annotation=R.Tuple(R.Tensor((1, batch_size, vocab_size), dtype=\"float32\"), R.Object)) "), sinfo_args=(R.Tuple,)) return gv3 @R.function def batch_verify(input_embeds: R.Tensor((1, "seq_len", 768), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 96), dtype="uint32"), R.Tensor(("vocab_size", 24), 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, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor(("vocab_size", 96), dtype="uint32"), R.Tensor(("vocab_size", 24), 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(198), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv162: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[0] R.call_packed("vm.builtin.check_tensor_info", gv162, R.prim_value(2), R.dtype("uint32"), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv163: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[1] R.call_packed("vm.builtin.check_tensor_info", gv163, 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, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv164: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[196] R.call_packed("vm.builtin.check_tensor_info", gv164, R.prim_value(2), R.dtype("uint32"), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv165: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[197] R.call_packed("vm.builtin.check_tensor_info", gv165, 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, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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", gv162, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(96), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv163, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(24), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv164, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(96), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv165, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(24), R.str("ErrorContext(fn=batch_verify, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) cls.gpt_neox_q4f16_1_shape_func3(shape_heap) gpt_neox_layers_0_input_layernorm_weight5: R.Tensor((768,), dtype="float16") = packed_params[2] gpt_neox_layers_0_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[3] storage15: R.Object = R.vm.alloc_storage(R.shape([3145728]), R.prim_value(0), R.dtype("uint8"), R.str("global")) 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),)) alloc177: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv166, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(input_embeds, gpt_neox_layers_0_input_layernorm_weight5, gpt_neox_layers_0_input_layernorm_bias5, alloc177) 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_q_weight5: R.Tensor((2304, 96), dtype="uint32") = packed_params[6] gpt_neox_layers_0_attention_query_key_value_q_scale5: R.Tensor((2304, 24), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[8] storage16: R.Object = R.vm.alloc_storage(R.shape([9437184]), R.prim_value(0), R.dtype("uint8"), R.str("global")) 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),)) alloc178: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv167, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_0_attention_query_key_value_q_weight5, gpt_neox_layers_0_attention_query_key_value_q_scale5, alloc177, gpt_neox_layers_0_attention_query_key_value_bias5, alloc178) R.vm.kill_object(alloc177) R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_q_weight5) R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_q_scale5) R.vm.kill_object(gpt_neox_layers_0_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),)) reshape192: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc178, gv168, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc178) 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),)) reshape193: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape192, gv169, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape192) 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),)) alloc179: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv170, 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)), reshape193, alloc179) R.vm.kill_object(reshape193) 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),)) reshape194: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc179, gv171, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc179) 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),)) reshape195: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape194, gv172, 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[4] gpt_neox_layers_0_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[5] 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),)) alloc180: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv173, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(input_embeds, gpt_neox_layers_0_post_attention_layernorm_weight5, gpt_neox_layers_0_post_attention_layernorm_bias5, alloc180) 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_q_weight5: R.Tensor((3072, 96), dtype="uint32") = packed_params[12] gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale5: R.Tensor((3072, 24), dtype="float16") = packed_params[13] gpt_neox_layers_0_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[14] storage17: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) 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),)) alloc181: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv174, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale5, alloc180, gpt_neox_layers_0_mlp_dense_h_to_4h_bias5, alloc181) R.vm.kill_object(alloc180) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight5) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale5) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_bias5) gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight5: R.Tensor((768, 384), dtype="uint32") = packed_params[15] gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale5: R.Tensor((768, 96), dtype="float16") = packed_params[16] gpt_neox_layers_0_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[17] 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),)) alloc182: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv175, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale5, alloc181, gpt_neox_layers_0_mlp_dense_4h_to_h_bias5, alloc182) R.vm.kill_object(alloc181) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight5) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale5) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_bias5) gpt_neox_layers_0_attention_dense_q_weight5: R.Tensor((768, 96), dtype="uint32") = packed_params[9] gpt_neox_layers_0_attention_dense_q_scale5: R.Tensor((768, 24), dtype="float16") = packed_params[10] gpt_neox_layers_0_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[11] 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),)) alloc183: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv176, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_0_attention_dense_q_weight5, gpt_neox_layers_0_attention_dense_q_scale5, reshape195, gpt_neox_layers_0_attention_dense_bias5, alloc182, input_embeds, alloc183) R.vm.kill_object(reshape195) R.vm.kill_object(alloc182) R.vm.kill_object(gpt_neox_layers_0_attention_dense_q_weight5) R.vm.kill_object(gpt_neox_layers_0_attention_dense_q_scale5) 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[18] gpt_neox_layers_1_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[19] 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),)) alloc184: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv177, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc183, gpt_neox_layers_1_input_layernorm_weight5, gpt_neox_layers_1_input_layernorm_bias5, alloc184) 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_q_weight5: R.Tensor((2304, 96), dtype="uint32") = packed_params[22] gpt_neox_layers_1_attention_query_key_value_q_scale5: R.Tensor((2304, 24), dtype="float16") = packed_params[23] gpt_neox_layers_1_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[24] 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),)) alloc185: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv178, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_1_attention_query_key_value_q_weight5, gpt_neox_layers_1_attention_query_key_value_q_scale5, alloc184, gpt_neox_layers_1_attention_query_key_value_bias5, alloc185) R.vm.kill_object(alloc184) R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_q_weight5) R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_q_scale5) R.vm.kill_object(gpt_neox_layers_1_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),)) reshape196: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc185, gv179, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc185) 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),)) reshape197: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape196, gv180, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape196) 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),)) alloc186: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv181, 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)), reshape197, alloc186) R.vm.kill_object(reshape197) 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),)) reshape198: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc186, gv182, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc186) 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),)) reshape199: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape198, gv183, 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[20] gpt_neox_layers_1_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[21] 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),)) alloc187: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv184, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc183, gpt_neox_layers_1_post_attention_layernorm_weight5, gpt_neox_layers_1_post_attention_layernorm_bias5, alloc187) 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_q_weight5: R.Tensor((3072, 96), dtype="uint32") = packed_params[28] gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale5: R.Tensor((3072, 24), dtype="float16") = packed_params[29] gpt_neox_layers_1_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[30] storage18: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) gv185: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) alloc188: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage18, R.prim_value(0), gv185, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale5, alloc187, gpt_neox_layers_1_mlp_dense_h_to_4h_bias5, alloc188) R.vm.kill_object(alloc187) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight5) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale5) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_bias5) gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight5: R.Tensor((768, 384), dtype="uint32") = packed_params[31] gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale5: R.Tensor((768, 96), dtype="float16") = packed_params[32] gpt_neox_layers_1_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[33] 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),)) alloc189: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv186, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale5, alloc188, gpt_neox_layers_1_mlp_dense_4h_to_h_bias5, alloc189) R.vm.kill_object(alloc188) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight5) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale5) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_bias5) gpt_neox_layers_1_attention_dense_q_weight5: R.Tensor((768, 96), dtype="uint32") = packed_params[25] gpt_neox_layers_1_attention_dense_q_scale5: R.Tensor((768, 24), dtype="float16") = packed_params[26] gpt_neox_layers_1_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[27] 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),)) alloc190: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage18, R.prim_value(0), gv187, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_1_attention_dense_q_weight5, gpt_neox_layers_1_attention_dense_q_scale5, reshape199, gpt_neox_layers_1_attention_dense_bias5, alloc189, alloc183, alloc190) R.vm.kill_object(alloc183) R.vm.kill_object(reshape199) R.vm.kill_object(alloc189) R.vm.kill_object(gpt_neox_layers_1_attention_dense_q_weight5) R.vm.kill_object(gpt_neox_layers_1_attention_dense_q_scale5) 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[34] gpt_neox_layers_2_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[35] 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),)) alloc191: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv188, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc190, gpt_neox_layers_2_input_layernorm_weight5, gpt_neox_layers_2_input_layernorm_bias5, alloc191) 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_q_weight5: R.Tensor((2304, 96), dtype="uint32") = packed_params[38] gpt_neox_layers_2_attention_query_key_value_q_scale5: R.Tensor((2304, 24), dtype="float16") = packed_params[39] gpt_neox_layers_2_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[40] 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),)) alloc192: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv189, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_2_attention_query_key_value_q_weight5, gpt_neox_layers_2_attention_query_key_value_q_scale5, alloc191, gpt_neox_layers_2_attention_query_key_value_bias5, alloc192) R.vm.kill_object(alloc191) R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_q_weight5) R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_q_scale5) R.vm.kill_object(gpt_neox_layers_2_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),)) reshape200: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc192, gv190, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc192) 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),)) reshape201: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape200, gv191, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape200) 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),)) alloc193: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv192, 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)), reshape201, alloc193) R.vm.kill_object(reshape201) 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),)) reshape202: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc193, gv193, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc193) 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),)) reshape203: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape202, gv194, 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[36] gpt_neox_layers_2_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[37] 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),)) alloc194: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv195, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc190, gpt_neox_layers_2_post_attention_layernorm_weight5, gpt_neox_layers_2_post_attention_layernorm_bias5, alloc194) 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_q_weight5: R.Tensor((3072, 96), dtype="uint32") = packed_params[44] gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale5: R.Tensor((3072, 24), dtype="float16") = packed_params[45] gpt_neox_layers_2_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[46] 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),)) alloc195: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv196, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale5, alloc194, gpt_neox_layers_2_mlp_dense_h_to_4h_bias5, alloc195) R.vm.kill_object(alloc194) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight5) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale5) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_bias5) gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight5: R.Tensor((768, 384), dtype="uint32") = packed_params[47] gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale5: R.Tensor((768, 96), dtype="float16") = packed_params[48] gpt_neox_layers_2_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[49] 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),)) alloc196: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv197, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale5, alloc195, gpt_neox_layers_2_mlp_dense_4h_to_h_bias5, alloc196) R.vm.kill_object(alloc195) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight5) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale5) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_bias5) gpt_neox_layers_2_attention_dense_q_weight5: R.Tensor((768, 96), dtype="uint32") = packed_params[41] gpt_neox_layers_2_attention_dense_q_scale5: R.Tensor((768, 24), dtype="float16") = packed_params[42] gpt_neox_layers_2_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[43] 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),)) alloc197: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv198, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_2_attention_dense_q_weight5, gpt_neox_layers_2_attention_dense_q_scale5, reshape203, gpt_neox_layers_2_attention_dense_bias5, alloc196, alloc190, alloc197) R.vm.kill_object(alloc190) R.vm.kill_object(reshape203) R.vm.kill_object(alloc196) R.vm.kill_object(gpt_neox_layers_2_attention_dense_q_weight5) R.vm.kill_object(gpt_neox_layers_2_attention_dense_q_scale5) 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[50] gpt_neox_layers_3_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[51] 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),)) alloc198: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv199, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc197, gpt_neox_layers_3_input_layernorm_weight5, gpt_neox_layers_3_input_layernorm_bias5, alloc198) 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_q_weight5: R.Tensor((2304, 96), dtype="uint32") = packed_params[54] gpt_neox_layers_3_attention_query_key_value_q_scale5: R.Tensor((2304, 24), dtype="float16") = packed_params[55] gpt_neox_layers_3_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[56] 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),)) alloc199: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv200, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_3_attention_query_key_value_q_weight5, gpt_neox_layers_3_attention_query_key_value_q_scale5, alloc198, gpt_neox_layers_3_attention_query_key_value_bias5, alloc199) R.vm.kill_object(alloc198) R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_q_weight5) R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_q_scale5) R.vm.kill_object(gpt_neox_layers_3_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),)) reshape204: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc199, gv201, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc199) 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),)) reshape205: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape204, gv202, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape204) 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),)) alloc200: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv203, 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)), reshape205, alloc200) R.vm.kill_object(reshape205) 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),)) reshape206: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc200, gv204, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc200) 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),)) reshape207: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape206, gv205, 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[52] gpt_neox_layers_3_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[53] 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),)) alloc201: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv206, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc197, gpt_neox_layers_3_post_attention_layernorm_weight5, gpt_neox_layers_3_post_attention_layernorm_bias5, alloc201) 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_q_weight5: R.Tensor((3072, 96), dtype="uint32") = packed_params[60] gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale5: R.Tensor((3072, 24), dtype="float16") = packed_params[61] gpt_neox_layers_3_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[62] 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),)) alloc202: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage18, R.prim_value(0), gv207, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale5, alloc201, gpt_neox_layers_3_mlp_dense_h_to_4h_bias5, alloc202) R.vm.kill_object(alloc201) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight5) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale5) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_bias5) gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight5: R.Tensor((768, 384), dtype="uint32") = packed_params[63] gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale5: R.Tensor((768, 96), dtype="float16") = packed_params[64] gpt_neox_layers_3_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[65] 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),)) alloc203: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv208, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale5, alloc202, gpt_neox_layers_3_mlp_dense_4h_to_h_bias5, alloc203) R.vm.kill_object(alloc202) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight5) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale5) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_bias5) gpt_neox_layers_3_attention_dense_q_weight5: R.Tensor((768, 96), dtype="uint32") = packed_params[57] gpt_neox_layers_3_attention_dense_q_scale5: R.Tensor((768, 24), dtype="float16") = packed_params[58] gpt_neox_layers_3_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[59] 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),)) alloc204: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage18, R.prim_value(0), gv209, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_3_attention_dense_q_weight5, gpt_neox_layers_3_attention_dense_q_scale5, reshape207, gpt_neox_layers_3_attention_dense_bias5, alloc203, alloc197, alloc204) R.vm.kill_object(alloc197) R.vm.kill_object(reshape207) R.vm.kill_object(alloc203) R.vm.kill_object(gpt_neox_layers_3_attention_dense_q_weight5) R.vm.kill_object(gpt_neox_layers_3_attention_dense_q_scale5) 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[66] gpt_neox_layers_4_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[67] 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),)) alloc205: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv210, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc204, gpt_neox_layers_4_input_layernorm_weight5, gpt_neox_layers_4_input_layernorm_bias5, alloc205) 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_q_weight5: R.Tensor((2304, 96), dtype="uint32") = packed_params[70] gpt_neox_layers_4_attention_query_key_value_q_scale5: R.Tensor((2304, 24), dtype="float16") = packed_params[71] gpt_neox_layers_4_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[72] 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),)) alloc206: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv211, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_4_attention_query_key_value_q_weight5, gpt_neox_layers_4_attention_query_key_value_q_scale5, alloc205, gpt_neox_layers_4_attention_query_key_value_bias5, alloc206) R.vm.kill_object(alloc205) R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_q_weight5) R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_q_scale5) R.vm.kill_object(gpt_neox_layers_4_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),)) reshape208: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc206, gv212, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc206) 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),)) reshape209: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape208, gv213, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape208) 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),)) alloc207: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv214, 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)), reshape209, alloc207) R.vm.kill_object(reshape209) 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),)) reshape210: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc207, gv215, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc207) 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),)) reshape211: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape210, gv216, 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[68] gpt_neox_layers_4_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[69] 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),)) alloc208: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv217, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc204, gpt_neox_layers_4_post_attention_layernorm_weight5, gpt_neox_layers_4_post_attention_layernorm_bias5, alloc208) 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_q_weight5: R.Tensor((3072, 96), dtype="uint32") = packed_params[76] gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale5: R.Tensor((3072, 24), dtype="float16") = packed_params[77] gpt_neox_layers_4_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[78] 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),)) alloc209: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv218, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale5, alloc208, gpt_neox_layers_4_mlp_dense_h_to_4h_bias5, alloc209) R.vm.kill_object(alloc208) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight5) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale5) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_bias5) gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight5: R.Tensor((768, 384), dtype="uint32") = packed_params[79] gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale5: R.Tensor((768, 96), dtype="float16") = packed_params[80] gpt_neox_layers_4_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[81] 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),)) alloc210: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv219, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale5, alloc209, gpt_neox_layers_4_mlp_dense_4h_to_h_bias5, alloc210) R.vm.kill_object(alloc209) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight5) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale5) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_bias5) gpt_neox_layers_4_attention_dense_q_weight5: R.Tensor((768, 96), dtype="uint32") = packed_params[73] gpt_neox_layers_4_attention_dense_q_scale5: R.Tensor((768, 24), dtype="float16") = packed_params[74] gpt_neox_layers_4_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[75] 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),)) alloc211: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv220, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_4_attention_dense_q_weight5, gpt_neox_layers_4_attention_dense_q_scale5, reshape211, gpt_neox_layers_4_attention_dense_bias5, alloc210, alloc204, alloc211) R.vm.kill_object(alloc204) R.vm.kill_object(reshape211) R.vm.kill_object(alloc210) R.vm.kill_object(gpt_neox_layers_4_attention_dense_q_weight5) R.vm.kill_object(gpt_neox_layers_4_attention_dense_q_scale5) 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[82] gpt_neox_layers_5_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[83] 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),)) alloc212: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv221, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc211, gpt_neox_layers_5_input_layernorm_weight5, gpt_neox_layers_5_input_layernorm_bias5, alloc212) 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_q_weight5: R.Tensor((2304, 96), dtype="uint32") = packed_params[86] gpt_neox_layers_5_attention_query_key_value_q_scale5: R.Tensor((2304, 24), dtype="float16") = packed_params[87] gpt_neox_layers_5_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[88] 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),)) alloc213: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv222, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_5_attention_query_key_value_q_weight5, gpt_neox_layers_5_attention_query_key_value_q_scale5, alloc212, gpt_neox_layers_5_attention_query_key_value_bias5, alloc213) R.vm.kill_object(alloc212) R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_q_weight5) R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_q_scale5) R.vm.kill_object(gpt_neox_layers_5_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),)) reshape212: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc213, gv223, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc213) 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),)) reshape213: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape212, gv224, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape212) 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),)) alloc214: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv225, 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)), reshape213, alloc214) R.vm.kill_object(reshape213) 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),)) reshape214: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc214, gv226, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc214) 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),)) reshape215: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape214, gv227, 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[84] gpt_neox_layers_5_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[85] 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),)) alloc215: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv228, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc211, gpt_neox_layers_5_post_attention_layernorm_weight5, gpt_neox_layers_5_post_attention_layernorm_bias5, alloc215) 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_q_weight5: R.Tensor((3072, 96), dtype="uint32") = packed_params[92] gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale5: R.Tensor((3072, 24), dtype="float16") = packed_params[93] gpt_neox_layers_5_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),)) alloc216: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage18, R.prim_value(0), gv229, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale5, alloc215, gpt_neox_layers_5_mlp_dense_h_to_4h_bias5, alloc216) R.vm.kill_object(alloc215) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight5) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale5) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_bias5) gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight5: R.Tensor((768, 384), dtype="uint32") = packed_params[95] gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale5: R.Tensor((768, 96), dtype="float16") = packed_params[96] gpt_neox_layers_5_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[97] 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),)) alloc217: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv230, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale5, alloc216, gpt_neox_layers_5_mlp_dense_4h_to_h_bias5, alloc217) R.vm.kill_object(alloc216) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight5) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale5) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_bias5) gpt_neox_layers_5_attention_dense_q_weight5: R.Tensor((768, 96), dtype="uint32") = packed_params[89] gpt_neox_layers_5_attention_dense_q_scale5: R.Tensor((768, 24), dtype="float16") = packed_params[90] gpt_neox_layers_5_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[91] 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),)) alloc218: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage18, R.prim_value(0), gv231, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_5_attention_dense_q_weight5, gpt_neox_layers_5_attention_dense_q_scale5, reshape215, gpt_neox_layers_5_attention_dense_bias5, alloc217, alloc211, alloc218) R.vm.kill_object(alloc211) R.vm.kill_object(reshape215) R.vm.kill_object(alloc217) R.vm.kill_object(gpt_neox_layers_5_attention_dense_q_weight5) R.vm.kill_object(gpt_neox_layers_5_attention_dense_q_scale5) 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[98] gpt_neox_layers_6_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[99] 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),)) alloc219: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv232, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc218, gpt_neox_layers_6_input_layernorm_weight5, gpt_neox_layers_6_input_layernorm_bias5, alloc219) 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_q_weight5: R.Tensor((2304, 96), dtype="uint32") = packed_params[102] gpt_neox_layers_6_attention_query_key_value_q_scale5: R.Tensor((2304, 24), dtype="float16") = packed_params[103] gpt_neox_layers_6_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[104] 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),)) alloc220: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv233, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_6_attention_query_key_value_q_weight5, gpt_neox_layers_6_attention_query_key_value_q_scale5, alloc219, gpt_neox_layers_6_attention_query_key_value_bias5, alloc220) R.vm.kill_object(alloc219) R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_q_weight5) R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_q_scale5) R.vm.kill_object(gpt_neox_layers_6_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),)) reshape216: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc220, gv234, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc220) 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),)) reshape217: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape216, gv235, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape216) 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),)) alloc221: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv236, 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)), reshape217, alloc221) R.vm.kill_object(reshape217) 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),)) reshape218: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc221, gv237, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc221) 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),)) reshape219: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape218, gv238, 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[100] gpt_neox_layers_6_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[101] 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),)) alloc222: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv239, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc218, gpt_neox_layers_6_post_attention_layernorm_weight5, gpt_neox_layers_6_post_attention_layernorm_bias5, alloc222) 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_q_weight5: R.Tensor((3072, 96), dtype="uint32") = packed_params[108] gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale5: R.Tensor((3072, 24), dtype="float16") = packed_params[109] gpt_neox_layers_6_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[110] 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),)) alloc223: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv240, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale5, alloc222, gpt_neox_layers_6_mlp_dense_h_to_4h_bias5, alloc223) R.vm.kill_object(alloc222) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight5) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale5) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_bias5) gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight5: R.Tensor((768, 384), dtype="uint32") = packed_params[111] gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale5: R.Tensor((768, 96), dtype="float16") = packed_params[112] gpt_neox_layers_6_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[113] 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),)) alloc224: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv241, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale5, alloc223, gpt_neox_layers_6_mlp_dense_4h_to_h_bias5, alloc224) R.vm.kill_object(alloc223) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight5) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale5) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_bias5) gpt_neox_layers_6_attention_dense_q_weight5: R.Tensor((768, 96), dtype="uint32") = packed_params[105] gpt_neox_layers_6_attention_dense_q_scale5: R.Tensor((768, 24), dtype="float16") = packed_params[106] gpt_neox_layers_6_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[107] 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),)) alloc225: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv242, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_6_attention_dense_q_weight5, gpt_neox_layers_6_attention_dense_q_scale5, reshape219, gpt_neox_layers_6_attention_dense_bias5, alloc224, alloc218, alloc225) R.vm.kill_object(alloc218) R.vm.kill_object(reshape219) R.vm.kill_object(alloc224) R.vm.kill_object(gpt_neox_layers_6_attention_dense_q_weight5) R.vm.kill_object(gpt_neox_layers_6_attention_dense_q_scale5) 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[114] gpt_neox_layers_7_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[115] 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),)) alloc226: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv243, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc225, gpt_neox_layers_7_input_layernorm_weight5, gpt_neox_layers_7_input_layernorm_bias5, alloc226) 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_q_weight5: R.Tensor((2304, 96), dtype="uint32") = packed_params[118] gpt_neox_layers_7_attention_query_key_value_q_scale5: R.Tensor((2304, 24), dtype="float16") = packed_params[119] gpt_neox_layers_7_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[120] 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),)) alloc227: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv244, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_7_attention_query_key_value_q_weight5, gpt_neox_layers_7_attention_query_key_value_q_scale5, alloc226, gpt_neox_layers_7_attention_query_key_value_bias5, alloc227) R.vm.kill_object(alloc226) R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_q_weight5) R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_q_scale5) R.vm.kill_object(gpt_neox_layers_7_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),)) reshape220: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc227, gv245, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc227) 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),)) reshape221: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape220, gv246, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape220) 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),)) alloc228: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv247, 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)), reshape221, alloc228) R.vm.kill_object(reshape221) 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),)) reshape222: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc228, gv248, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc228) 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),)) reshape223: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape222, gv249, 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[116] gpt_neox_layers_7_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[117] 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),)) alloc229: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv250, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc225, gpt_neox_layers_7_post_attention_layernorm_weight5, gpt_neox_layers_7_post_attention_layernorm_bias5, alloc229) 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_q_weight5: R.Tensor((3072, 96), dtype="uint32") = packed_params[124] gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale5: R.Tensor((3072, 24), dtype="float16") = packed_params[125] gpt_neox_layers_7_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[126] 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),)) alloc230: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage18, R.prim_value(0), gv251, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale5, alloc229, gpt_neox_layers_7_mlp_dense_h_to_4h_bias5, alloc230) R.vm.kill_object(alloc229) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight5) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale5) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_bias5) gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight5: R.Tensor((768, 384), dtype="uint32") = packed_params[127] gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale5: R.Tensor((768, 96), dtype="float16") = packed_params[128] gpt_neox_layers_7_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[129] 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),)) alloc231: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv252, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale5, alloc230, gpt_neox_layers_7_mlp_dense_4h_to_h_bias5, alloc231) R.vm.kill_object(alloc230) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight5) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale5) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_bias5) gpt_neox_layers_7_attention_dense_q_weight5: R.Tensor((768, 96), dtype="uint32") = packed_params[121] gpt_neox_layers_7_attention_dense_q_scale5: R.Tensor((768, 24), dtype="float16") = packed_params[122] gpt_neox_layers_7_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[123] 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),)) alloc232: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage18, R.prim_value(0), gv253, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_7_attention_dense_q_weight5, gpt_neox_layers_7_attention_dense_q_scale5, reshape223, gpt_neox_layers_7_attention_dense_bias5, alloc231, alloc225, alloc232) R.vm.kill_object(alloc225) R.vm.kill_object(reshape223) R.vm.kill_object(alloc231) R.vm.kill_object(gpt_neox_layers_7_attention_dense_q_weight5) R.vm.kill_object(gpt_neox_layers_7_attention_dense_q_scale5) 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[130] gpt_neox_layers_8_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[131] 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),)) alloc233: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv254, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc232, gpt_neox_layers_8_input_layernorm_weight5, gpt_neox_layers_8_input_layernorm_bias5, alloc233) 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_q_weight5: R.Tensor((2304, 96), dtype="uint32") = packed_params[134] gpt_neox_layers_8_attention_query_key_value_q_scale5: R.Tensor((2304, 24), dtype="float16") = packed_params[135] gpt_neox_layers_8_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[136] 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),)) alloc234: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv255, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_8_attention_query_key_value_q_weight5, gpt_neox_layers_8_attention_query_key_value_q_scale5, alloc233, gpt_neox_layers_8_attention_query_key_value_bias5, alloc234) R.vm.kill_object(alloc233) R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_q_weight5) R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_q_scale5) R.vm.kill_object(gpt_neox_layers_8_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),)) reshape224: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc234, gv256, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc234) 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),)) reshape225: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape224, gv257, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape224) 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),)) alloc235: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv258, 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)), reshape225, alloc235) R.vm.kill_object(reshape225) 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),)) reshape226: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc235, gv259, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc235) 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),)) reshape227: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape226, gv260, 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[132] gpt_neox_layers_8_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[133] 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),)) alloc236: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv261, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc232, gpt_neox_layers_8_post_attention_layernorm_weight5, gpt_neox_layers_8_post_attention_layernorm_bias5, alloc236) 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_q_weight5: R.Tensor((3072, 96), dtype="uint32") = packed_params[140] gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale5: R.Tensor((3072, 24), dtype="float16") = packed_params[141] gpt_neox_layers_8_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[142] 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),)) alloc237: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv262, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale5, alloc236, gpt_neox_layers_8_mlp_dense_h_to_4h_bias5, alloc237) R.vm.kill_object(alloc236) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight5) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale5) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_bias5) gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight5: R.Tensor((768, 384), dtype="uint32") = packed_params[143] gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale5: R.Tensor((768, 96), dtype="float16") = packed_params[144] gpt_neox_layers_8_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[145] 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),)) alloc238: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv263, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale5, alloc237, gpt_neox_layers_8_mlp_dense_4h_to_h_bias5, alloc238) R.vm.kill_object(alloc237) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight5) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale5) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_bias5) gpt_neox_layers_8_attention_dense_q_weight5: R.Tensor((768, 96), dtype="uint32") = packed_params[137] gpt_neox_layers_8_attention_dense_q_scale5: R.Tensor((768, 24), dtype="float16") = packed_params[138] gpt_neox_layers_8_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[139] 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),)) alloc239: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv264, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_8_attention_dense_q_weight5, gpt_neox_layers_8_attention_dense_q_scale5, reshape227, gpt_neox_layers_8_attention_dense_bias5, alloc238, alloc232, alloc239) R.vm.kill_object(alloc232) R.vm.kill_object(reshape227) R.vm.kill_object(alloc238) R.vm.kill_object(gpt_neox_layers_8_attention_dense_q_weight5) R.vm.kill_object(gpt_neox_layers_8_attention_dense_q_scale5) 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[146] gpt_neox_layers_9_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[147] 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),)) alloc240: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv265, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc239, gpt_neox_layers_9_input_layernorm_weight5, gpt_neox_layers_9_input_layernorm_bias5, alloc240) 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_q_weight5: R.Tensor((2304, 96), dtype="uint32") = packed_params[150] gpt_neox_layers_9_attention_query_key_value_q_scale5: R.Tensor((2304, 24), dtype="float16") = packed_params[151] gpt_neox_layers_9_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[152] 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),)) alloc241: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv266, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_9_attention_query_key_value_q_weight5, gpt_neox_layers_9_attention_query_key_value_q_scale5, alloc240, gpt_neox_layers_9_attention_query_key_value_bias5, alloc241) R.vm.kill_object(alloc240) R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_q_weight5) R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_q_scale5) R.vm.kill_object(gpt_neox_layers_9_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),)) reshape228: R.Tensor((1, seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc241, gv267, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc241) 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),)) reshape229: R.Tensor((seq_len, 36, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape228, gv268, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape228) 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),)) alloc242: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv269, 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)), reshape229, alloc242) R.vm.kill_object(reshape229) 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),)) reshape230: R.Tensor((1, seq_len, 12, 64), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc242, gv270, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc242) 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),)) reshape231: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape230, gv271, 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[148] gpt_neox_layers_9_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[149] 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),)) alloc243: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv272, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc239, gpt_neox_layers_9_post_attention_layernorm_weight5, gpt_neox_layers_9_post_attention_layernorm_bias5, alloc243) 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_q_weight5: R.Tensor((3072, 96), dtype="uint32") = packed_params[156] gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale5: R.Tensor((3072, 24), dtype="float16") = packed_params[157] gpt_neox_layers_9_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[158] 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),)) alloc244: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage18, R.prim_value(0), gv273, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale5, alloc243, gpt_neox_layers_9_mlp_dense_h_to_4h_bias5, alloc244) R.vm.kill_object(alloc243) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight5) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale5) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_bias5) gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight5: R.Tensor((768, 384), dtype="uint32") = packed_params[159] gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale5: R.Tensor((768, 96), dtype="float16") = packed_params[160] gpt_neox_layers_9_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[161] 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),)) alloc245: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv274, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale5, alloc244, gpt_neox_layers_9_mlp_dense_4h_to_h_bias5, alloc245) R.vm.kill_object(alloc244) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight5) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale5) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_bias5) gpt_neox_layers_9_attention_dense_q_weight5: R.Tensor((768, 96), dtype="uint32") = packed_params[153] gpt_neox_layers_9_attention_dense_q_scale5: R.Tensor((768, 24), dtype="float16") = packed_params[154] gpt_neox_layers_9_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[155] 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),)) alloc246: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage18, R.prim_value(0), gv275, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_9_attention_dense_q_weight5, gpt_neox_layers_9_attention_dense_q_scale5, reshape231, gpt_neox_layers_9_attention_dense_bias5, alloc245, alloc239, alloc246) R.vm.kill_object(alloc239) R.vm.kill_object(reshape231) R.vm.kill_object(alloc245) R.vm.kill_object(gpt_neox_layers_9_attention_dense_q_weight5) R.vm.kill_object(gpt_neox_layers_9_attention_dense_q_scale5) 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[162] gpt_neox_layers_10_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[163] 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),)) alloc247: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv276, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc246, gpt_neox_layers_10_input_layernorm_weight5, gpt_neox_layers_10_input_layernorm_bias5, alloc247) 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_q_weight5: R.Tensor((2304, 96), dtype="uint32") = packed_params[166] gpt_neox_layers_10_attention_query_key_value_q_scale5: R.Tensor((2304, 24), dtype="float16") = packed_params[167] gpt_neox_layers_10_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[168] gv277: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc248: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv277, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_10_attention_query_key_value_q_weight5, gpt_neox_layers_10_attention_query_key_value_q_scale5, alloc247, gpt_neox_layers_10_attention_query_key_value_bias5, alloc248) R.vm.kill_object(alloc247) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_q_weight5) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_q_scale5) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_bias5) gv278: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc248, gv278, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc248) gv279: R.Shape(ndim=3) = R.call_packed("vm.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, gv279, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape232) gv280: R.Shape(ndim=3) = R.call_packed("vm.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),)) alloc249: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv280, 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)), reshape233, alloc249) R.vm.kill_object(reshape233) gv281: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc249, gv281, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc249) gv282: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape235: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape234, gv282, 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[164] gpt_neox_layers_10_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[165] gv283: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc250: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv283, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc246, gpt_neox_layers_10_post_attention_layernorm_weight5, gpt_neox_layers_10_post_attention_layernorm_bias5, alloc250) 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_q_weight5: R.Tensor((3072, 96), dtype="uint32") = packed_params[172] gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale5: R.Tensor((3072, 24), dtype="float16") = packed_params[173] gpt_neox_layers_10_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[174] gv284: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) alloc251: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv284, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale5, alloc250, gpt_neox_layers_10_mlp_dense_h_to_4h_bias5, alloc251) R.vm.kill_object(alloc250) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight5) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale5) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_bias5) gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight5: R.Tensor((768, 384), dtype="uint32") = packed_params[175] gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale5: R.Tensor((768, 96), dtype="float16") = packed_params[176] gpt_neox_layers_10_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[177] gv285: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc252: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv285, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale5, alloc251, gpt_neox_layers_10_mlp_dense_4h_to_h_bias5, alloc252) R.vm.kill_object(alloc251) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight5) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale5) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_bias5) gpt_neox_layers_10_attention_dense_q_weight5: R.Tensor((768, 96), dtype="uint32") = packed_params[169] gpt_neox_layers_10_attention_dense_q_scale5: R.Tensor((768, 24), dtype="float16") = packed_params[170] gpt_neox_layers_10_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[171] gv286: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc253: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage17, R.prim_value(0), gv286, R.dtype("float16")) R.vm.kill_object(storage17) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_10_attention_dense_q_weight5, gpt_neox_layers_10_attention_dense_q_scale5, reshape235, gpt_neox_layers_10_attention_dense_bias5, alloc252, alloc246, alloc253) R.vm.kill_object(alloc246) R.vm.kill_object(reshape235) R.vm.kill_object(alloc252) R.vm.kill_object(gpt_neox_layers_10_attention_dense_q_weight5) R.vm.kill_object(gpt_neox_layers_10_attention_dense_q_scale5) 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[178] gpt_neox_layers_11_input_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[179] gv287: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc254: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv287, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc253, gpt_neox_layers_11_input_layernorm_weight5, gpt_neox_layers_11_input_layernorm_bias5, alloc254) 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_q_weight5: R.Tensor((2304, 96), dtype="uint32") = packed_params[182] gpt_neox_layers_11_attention_query_key_value_q_scale5: R.Tensor((2304, 24), dtype="float16") = packed_params[183] gpt_neox_layers_11_attention_query_key_value_bias5: R.Tensor((2304,), dtype="float16") = packed_params[184] gv288: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) alloc255: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv288, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_11_attention_query_key_value_q_weight5, gpt_neox_layers_11_attention_query_key_value_q_scale5, alloc254, gpt_neox_layers_11_attention_query_key_value_bias5, alloc255) R.vm.kill_object(alloc254) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_q_weight5) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_q_scale5) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_bias5) gv289: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc255, gv289, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc255) gv290: R.Shape(ndim=3) = R.call_packed("vm.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, gv290, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape236) gv291: R.Shape(ndim=3) = R.call_packed("vm.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),)) alloc256: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv291, 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)), reshape237, alloc256) R.vm.kill_object(reshape237) gv292: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc256, gv292, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc256) gv293: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape239: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape238, gv293, 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[180] gpt_neox_layers_11_post_attention_layernorm_bias5: R.Tensor((768,), dtype="float16") = packed_params[181] gv294: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc257: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv294, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc253, gpt_neox_layers_11_post_attention_layernorm_weight5, gpt_neox_layers_11_post_attention_layernorm_bias5, alloc257) 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_q_weight5: R.Tensor((3072, 96), dtype="uint32") = packed_params[188] gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale5: R.Tensor((3072, 24), dtype="float16") = packed_params[189] gpt_neox_layers_11_mlp_dense_h_to_4h_bias5: R.Tensor((3072,), dtype="float32") = packed_params[190] 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(3072), sinfo_args=(R.Shape(ndim=3),)) alloc258: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage18, R.prim_value(0), gv295, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale5, alloc257, gpt_neox_layers_11_mlp_dense_h_to_4h_bias5, alloc258) R.vm.kill_object(alloc257) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight5) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale5) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_bias5) gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight5: R.Tensor((768, 384), dtype="uint32") = packed_params[191] gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale5: R.Tensor((768, 96), dtype="float16") = packed_params[192] gpt_neox_layers_11_mlp_dense_4h_to_h_bias5: R.Tensor((768,), dtype="float32") = packed_params[193] 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(768), sinfo_args=(R.Shape(ndim=3),)) alloc259: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage16, R.prim_value(0), gv296, R.dtype("float16")) R.vm.kill_object(storage16) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale5, alloc258, gpt_neox_layers_11_mlp_dense_4h_to_h_bias5, alloc259) R.vm.kill_object(alloc258) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight5) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale5) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_bias5) gpt_neox_layers_11_attention_dense_q_weight5: R.Tensor((768, 96), dtype="uint32") = packed_params[185] gpt_neox_layers_11_attention_dense_q_scale5: R.Tensor((768, 24), dtype="float16") = packed_params[186] gpt_neox_layers_11_attention_dense_bias5: R.Tensor((768,), dtype="float16") = packed_params[187] gv297: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc260: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage18, R.prim_value(0), gv297, R.dtype("float16")) R.vm.kill_object(storage18) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_11_attention_dense_q_weight5, gpt_neox_layers_11_attention_dense_q_scale5, reshape239, gpt_neox_layers_11_attention_dense_bias5, alloc259, alloc253, alloc260) R.vm.kill_object(alloc253) R.vm.kill_object(reshape239) R.vm.kill_object(alloc259) R.vm.kill_object(gpt_neox_layers_11_attention_dense_q_weight5) R.vm.kill_object(gpt_neox_layers_11_attention_dense_q_scale5) R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias5) gpt_neox_final_layer_norm_weight5: R.Tensor((768,), dtype="float16") = packed_params[194] gpt_neox_final_layer_norm_bias5: R.Tensor((768,), dtype="float16") = packed_params[195] gv298: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc261: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage15, R.prim_value(0), gv298, R.dtype("float16")) R.vm.kill_object(storage15) cls.gpt_neox_q4f16_1_layer_norm1(alloc260, gpt_neox_final_layer_norm_weight5, gpt_neox_final_layer_norm_bias5, alloc261) R.vm.kill_object(alloc260) R.vm.kill_object(gpt_neox_final_layer_norm_weight5) R.vm.kill_object(gpt_neox_final_layer_norm_bias5) embed_out_q_weight5: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[196] embed_out_q_scale5: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[197] gv299: 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),)) storage19: R.Object = R.vm.alloc_storage(gv299, R.prim_value(0), R.dtype("uint8"), R.str("global")) gv300: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(1), R.prim_value(1), sinfo_args=(R.Shape(ndim=3),)) alloc262: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage19, R.prim_value(0), gv300, R.dtype("float32")) R.vm.kill_object(storage19) cls.gpt_neox_q4f16_1_fused_dequantize_fused_NT_matmul9_cast5(embed_out_q_weight5, embed_out_q_scale5, alloc261, alloc262) R.vm.kill_object(alloc261) R.vm.kill_object(embed_out_q_weight5) R.vm.kill_object(embed_out_q_scale5) gv5: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc262, paged_kv_cache R.vm.kill_object(alloc262) gv301: R.Tensor(dtype="float32", ndim=3) = gv5[0] R.call_packed("vm.builtin.match_shape", gv301, 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,)) gv7: 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", gv7, 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_q4f16_1_tir_kv_cache_transpose_append, cls.gpt_neox_q4f16_1_batch_prefill_paged_kv, cls.gpt_neox_q4f16_1_batch_decode_paged_kv, cls.gpt_neox_q4f16_1_batch_prefill_paged_kv_sliding_window, cls.gpt_neox_q4f16_1_batch_decode_paged_kv_sliding_window, cls.gpt_neox_q4f16_1_batch_prefill_ragged_kv, cls.gpt_neox_q4f16_1_merge_state_inplace, cls.gpt_neox_q4f16_1_fused_rope, cls.gpt_neox_q4f16_1_copy_single_page, cls.gpt_neox_q4f16_1_tir_kv_cache_debug_get_kv, cls.gpt_neox_q4f16_1_compact_kv_copy, cls.gpt_neox_q4f16_1_batch_tree_attn, cls.gpt_neox_q4f16_1_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", 96), dtype="uint32"), R.Tensor(("vocab_size", 24), 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, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor(("vocab_size", 96), dtype="uint32"), R.Tensor(("vocab_size", 24), 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(198), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[0] R.call_packed("vm.builtin.check_tensor_info", gv, R.prim_value(2), R.dtype("uint32"), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv1: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[1] R.call_packed("vm.builtin.check_tensor_info", gv1, 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, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv2: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[196] R.call_packed("vm.builtin.check_tensor_info", gv2, R.prim_value(2), R.dtype("uint32"), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv3: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[197] R.call_packed("vm.builtin.check_tensor_info", gv3, 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, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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", gv, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(96), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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(0), R.prim_value(0), R.prim_value(24), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv2, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(0), R.prim_value(0), R.prim_value(96), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv3, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(0), R.prim_value(0), R.prim_value(24), R.str("ErrorContext(fn=decode, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) cls.gpt_neox_q4f16_1_shape_func(shape_heap) gpt_neox_layers_0_input_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[2] gpt_neox_layers_0_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[3] storage: R.Object = R.vm.alloc_storage(R.shape([1536]), R.prim_value(0), R.dtype("uint8"), R.str("global")) alloc: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(input_embed, gpt_neox_layers_0_input_layernorm_weight2, gpt_neox_layers_0_input_layernorm_bias2, alloc) 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_q_weight2: R.Tensor((2304, 96), dtype="uint32") = packed_params[6] gpt_neox_layers_0_attention_query_key_value_q_scale2: R.Tensor((2304, 24), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[8] storage1: R.Object = R.vm.alloc_storage(R.shape([4608]), R.prim_value(0), R.dtype("uint8"), R.str("global")) alloc1: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_0_attention_query_key_value_q_weight2, gpt_neox_layers_0_attention_query_key_value_q_scale2, alloc, gpt_neox_layers_0_attention_query_key_value_bias2, alloc1) R.vm.kill_object(alloc) R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_q_weight2) R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_q_scale2) 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", alloc1, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc1) alloc2: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) _2: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), lv148, alloc2) R.vm.kill_object(lv148) lv149: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc2, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) R.vm.kill_object(alloc2) gpt_neox_layers_0_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[4] gpt_neox_layers_0_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[5] alloc3: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(input_embed, gpt_neox_layers_0_post_attention_layernorm_weight2, gpt_neox_layers_0_post_attention_layernorm_bias2, alloc3) 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_q_weight2: R.Tensor((3072, 96), dtype="uint32") = packed_params[12] gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale2: R.Tensor((3072, 24), dtype="float16") = packed_params[13] gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[14] storage2: R.Object = R.vm.alloc_storage(R.shape([6144]), R.prim_value(0), R.dtype("uint8"), R.str("global")) alloc4: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage2, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale2, alloc3, gpt_neox_layers_0_mlp_dense_h_to_4h_bias2, alloc4) R.vm.kill_object(alloc3) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale2) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_bias2) gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2: R.Tensor((768, 384), dtype="uint32") = packed_params[15] gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale2: R.Tensor((768, 96), dtype="float16") = packed_params[16] gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[17] alloc5: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale2, alloc4, gpt_neox_layers_0_mlp_dense_4h_to_h_bias2, alloc5) R.vm.kill_object(alloc4) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale2) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_bias2) gpt_neox_layers_0_attention_dense_q_weight2: R.Tensor((768, 96), dtype="uint32") = packed_params[9] gpt_neox_layers_0_attention_dense_q_scale2: R.Tensor((768, 24), dtype="float16") = packed_params[10] gpt_neox_layers_0_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[11] alloc6: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage2, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_0_attention_dense_q_weight2, gpt_neox_layers_0_attention_dense_q_scale2, lv149, gpt_neox_layers_0_attention_dense_bias2, alloc5, input_embed, alloc6) R.vm.kill_object(lv149) R.vm.kill_object(alloc5) R.vm.kill_object(gpt_neox_layers_0_attention_dense_q_weight2) R.vm.kill_object(gpt_neox_layers_0_attention_dense_q_scale2) 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[18] gpt_neox_layers_1_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[19] alloc7: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc6, gpt_neox_layers_1_input_layernorm_weight2, gpt_neox_layers_1_input_layernorm_bias2, alloc7) 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_q_weight2: R.Tensor((2304, 96), dtype="uint32") = packed_params[22] gpt_neox_layers_1_attention_query_key_value_q_scale2: R.Tensor((2304, 24), dtype="float16") = packed_params[23] gpt_neox_layers_1_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[24] alloc8: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_1_attention_query_key_value_q_weight2, gpt_neox_layers_1_attention_query_key_value_q_scale2, alloc7, gpt_neox_layers_1_attention_query_key_value_bias2, alloc8) R.vm.kill_object(alloc7) R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_q_weight2) R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_q_scale2) 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", alloc8, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc8) alloc9: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) _9: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), lv154, alloc9) R.vm.kill_object(lv154) lv155: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc9, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) R.vm.kill_object(alloc9) gpt_neox_layers_1_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[20] gpt_neox_layers_1_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[21] alloc10: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc6, gpt_neox_layers_1_post_attention_layernorm_weight2, gpt_neox_layers_1_post_attention_layernorm_bias2, alloc10) 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_q_weight2: R.Tensor((3072, 96), dtype="uint32") = packed_params[28] gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale2: R.Tensor((3072, 24), dtype="float16") = packed_params[29] gpt_neox_layers_1_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[30] storage3: R.Object = R.vm.alloc_storage(R.shape([6144]), R.prim_value(0), R.dtype("uint8"), R.str("global")) alloc11: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage3, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale2, alloc10, gpt_neox_layers_1_mlp_dense_h_to_4h_bias2, alloc11) R.vm.kill_object(alloc10) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight2) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale2) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_bias2) gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight2: R.Tensor((768, 384), dtype="uint32") = packed_params[31] gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale2: R.Tensor((768, 96), dtype="float16") = packed_params[32] gpt_neox_layers_1_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[33] alloc12: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale2, alloc11, gpt_neox_layers_1_mlp_dense_4h_to_h_bias2, alloc12) R.vm.kill_object(alloc11) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight2) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale2) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_bias2) gpt_neox_layers_1_attention_dense_q_weight2: R.Tensor((768, 96), dtype="uint32") = packed_params[25] gpt_neox_layers_1_attention_dense_q_scale2: R.Tensor((768, 24), dtype="float16") = packed_params[26] gpt_neox_layers_1_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[27] alloc13: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage3, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_1_attention_dense_q_weight2, gpt_neox_layers_1_attention_dense_q_scale2, lv155, gpt_neox_layers_1_attention_dense_bias2, alloc12, alloc6, alloc13) R.vm.kill_object(alloc6) R.vm.kill_object(lv155) R.vm.kill_object(alloc12) R.vm.kill_object(gpt_neox_layers_1_attention_dense_q_weight2) R.vm.kill_object(gpt_neox_layers_1_attention_dense_q_scale2) 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[34] gpt_neox_layers_2_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[35] alloc14: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc13, gpt_neox_layers_2_input_layernorm_weight2, gpt_neox_layers_2_input_layernorm_bias2, alloc14) 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_q_weight2: R.Tensor((2304, 96), dtype="uint32") = packed_params[38] gpt_neox_layers_2_attention_query_key_value_q_scale2: R.Tensor((2304, 24), dtype="float16") = packed_params[39] gpt_neox_layers_2_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[40] alloc15: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_2_attention_query_key_value_q_weight2, gpt_neox_layers_2_attention_query_key_value_q_scale2, alloc14, gpt_neox_layers_2_attention_query_key_value_bias2, alloc15) R.vm.kill_object(alloc14) R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_q_weight2) R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_q_scale2) 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", alloc15, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc15) alloc16: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) _16: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), lv160, alloc16) R.vm.kill_object(lv160) lv161: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc16, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) R.vm.kill_object(alloc16) gpt_neox_layers_2_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[36] gpt_neox_layers_2_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[37] alloc17: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc13, gpt_neox_layers_2_post_attention_layernorm_weight2, gpt_neox_layers_2_post_attention_layernorm_bias2, alloc17) 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_q_weight2: R.Tensor((3072, 96), dtype="uint32") = packed_params[44] gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale2: R.Tensor((3072, 24), dtype="float16") = packed_params[45] gpt_neox_layers_2_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[46] alloc18: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage2, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale2, alloc17, gpt_neox_layers_2_mlp_dense_h_to_4h_bias2, alloc18) R.vm.kill_object(alloc17) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight2) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale2) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_bias2) gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight2: R.Tensor((768, 384), dtype="uint32") = packed_params[47] gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale2: R.Tensor((768, 96), dtype="float16") = packed_params[48] gpt_neox_layers_2_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[49] alloc19: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale2, alloc18, gpt_neox_layers_2_mlp_dense_4h_to_h_bias2, alloc19) R.vm.kill_object(alloc18) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight2) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale2) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_bias2) gpt_neox_layers_2_attention_dense_q_weight2: R.Tensor((768, 96), dtype="uint32") = packed_params[41] gpt_neox_layers_2_attention_dense_q_scale2: R.Tensor((768, 24), dtype="float16") = packed_params[42] gpt_neox_layers_2_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[43] alloc20: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage2, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_2_attention_dense_q_weight2, gpt_neox_layers_2_attention_dense_q_scale2, lv161, gpt_neox_layers_2_attention_dense_bias2, alloc19, alloc13, alloc20) R.vm.kill_object(alloc13) R.vm.kill_object(lv161) R.vm.kill_object(alloc19) R.vm.kill_object(gpt_neox_layers_2_attention_dense_q_weight2) R.vm.kill_object(gpt_neox_layers_2_attention_dense_q_scale2) 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[50] gpt_neox_layers_3_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[51] alloc21: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc20, gpt_neox_layers_3_input_layernorm_weight2, gpt_neox_layers_3_input_layernorm_bias2, alloc21) 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_q_weight2: R.Tensor((2304, 96), dtype="uint32") = packed_params[54] gpt_neox_layers_3_attention_query_key_value_q_scale2: R.Tensor((2304, 24), dtype="float16") = packed_params[55] gpt_neox_layers_3_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[56] alloc22: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_3_attention_query_key_value_q_weight2, gpt_neox_layers_3_attention_query_key_value_q_scale2, alloc21, gpt_neox_layers_3_attention_query_key_value_bias2, alloc22) R.vm.kill_object(alloc21) R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_q_weight2) R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_q_scale2) 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", alloc22, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc22) alloc23: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) _23: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), lv166, alloc23) R.vm.kill_object(lv166) lv167: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc23, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) R.vm.kill_object(alloc23) gpt_neox_layers_3_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[52] gpt_neox_layers_3_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[53] alloc24: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc20, gpt_neox_layers_3_post_attention_layernorm_weight2, gpt_neox_layers_3_post_attention_layernorm_bias2, alloc24) 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_q_weight2: R.Tensor((3072, 96), dtype="uint32") = packed_params[60] gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale2: R.Tensor((3072, 24), dtype="float16") = packed_params[61] gpt_neox_layers_3_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[62] alloc25: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage3, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale2, alloc24, gpt_neox_layers_3_mlp_dense_h_to_4h_bias2, alloc25) R.vm.kill_object(alloc24) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight2) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale2) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_bias2) gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight2: R.Tensor((768, 384), dtype="uint32") = packed_params[63] gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale2: R.Tensor((768, 96), dtype="float16") = packed_params[64] gpt_neox_layers_3_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[65] alloc26: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale2, alloc25, gpt_neox_layers_3_mlp_dense_4h_to_h_bias2, alloc26) R.vm.kill_object(alloc25) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight2) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale2) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_bias2) gpt_neox_layers_3_attention_dense_q_weight2: R.Tensor((768, 96), dtype="uint32") = packed_params[57] gpt_neox_layers_3_attention_dense_q_scale2: R.Tensor((768, 24), dtype="float16") = packed_params[58] gpt_neox_layers_3_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[59] alloc27: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage3, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_3_attention_dense_q_weight2, gpt_neox_layers_3_attention_dense_q_scale2, lv167, gpt_neox_layers_3_attention_dense_bias2, alloc26, alloc20, alloc27) R.vm.kill_object(alloc20) R.vm.kill_object(lv167) R.vm.kill_object(alloc26) R.vm.kill_object(gpt_neox_layers_3_attention_dense_q_weight2) R.vm.kill_object(gpt_neox_layers_3_attention_dense_q_scale2) 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[66] gpt_neox_layers_4_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[67] alloc28: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc27, gpt_neox_layers_4_input_layernorm_weight2, gpt_neox_layers_4_input_layernorm_bias2, alloc28) 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_q_weight2: R.Tensor((2304, 96), dtype="uint32") = packed_params[70] gpt_neox_layers_4_attention_query_key_value_q_scale2: R.Tensor((2304, 24), dtype="float16") = packed_params[71] gpt_neox_layers_4_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[72] alloc29: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_4_attention_query_key_value_q_weight2, gpt_neox_layers_4_attention_query_key_value_q_scale2, alloc28, gpt_neox_layers_4_attention_query_key_value_bias2, alloc29) R.vm.kill_object(alloc28) R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_q_weight2) R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_q_scale2) 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", alloc29, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc29) alloc30: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) _30: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), lv172, alloc30) R.vm.kill_object(lv172) lv173: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc30, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) R.vm.kill_object(alloc30) gpt_neox_layers_4_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[68] gpt_neox_layers_4_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[69] alloc31: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc27, gpt_neox_layers_4_post_attention_layernorm_weight2, gpt_neox_layers_4_post_attention_layernorm_bias2, alloc31) 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_q_weight2: R.Tensor((3072, 96), dtype="uint32") = packed_params[76] gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale2: R.Tensor((3072, 24), dtype="float16") = packed_params[77] gpt_neox_layers_4_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[78] alloc32: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage2, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale2, alloc31, gpt_neox_layers_4_mlp_dense_h_to_4h_bias2, alloc32) R.vm.kill_object(alloc31) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight2) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale2) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_bias2) gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight2: R.Tensor((768, 384), dtype="uint32") = packed_params[79] gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale2: R.Tensor((768, 96), dtype="float16") = packed_params[80] gpt_neox_layers_4_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[81] alloc33: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale2, alloc32, gpt_neox_layers_4_mlp_dense_4h_to_h_bias2, alloc33) R.vm.kill_object(alloc32) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight2) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale2) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_bias2) gpt_neox_layers_4_attention_dense_q_weight2: R.Tensor((768, 96), dtype="uint32") = packed_params[73] gpt_neox_layers_4_attention_dense_q_scale2: R.Tensor((768, 24), dtype="float16") = packed_params[74] gpt_neox_layers_4_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[75] alloc34: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage2, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_4_attention_dense_q_weight2, gpt_neox_layers_4_attention_dense_q_scale2, lv173, gpt_neox_layers_4_attention_dense_bias2, alloc33, alloc27, alloc34) R.vm.kill_object(alloc27) R.vm.kill_object(lv173) R.vm.kill_object(alloc33) R.vm.kill_object(gpt_neox_layers_4_attention_dense_q_weight2) R.vm.kill_object(gpt_neox_layers_4_attention_dense_q_scale2) 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[82] gpt_neox_layers_5_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[83] alloc35: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc34, gpt_neox_layers_5_input_layernorm_weight2, gpt_neox_layers_5_input_layernorm_bias2, alloc35) 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_q_weight2: R.Tensor((2304, 96), dtype="uint32") = packed_params[86] gpt_neox_layers_5_attention_query_key_value_q_scale2: R.Tensor((2304, 24), dtype="float16") = packed_params[87] gpt_neox_layers_5_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[88] alloc36: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_5_attention_query_key_value_q_weight2, gpt_neox_layers_5_attention_query_key_value_q_scale2, alloc35, gpt_neox_layers_5_attention_query_key_value_bias2, alloc36) R.vm.kill_object(alloc35) R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_q_weight2) R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_q_scale2) 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", alloc36, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc36) alloc37: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) _37: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), lv178, alloc37) R.vm.kill_object(lv178) lv179: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc37, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) R.vm.kill_object(alloc37) gpt_neox_layers_5_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[84] gpt_neox_layers_5_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[85] alloc38: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc34, gpt_neox_layers_5_post_attention_layernorm_weight2, gpt_neox_layers_5_post_attention_layernorm_bias2, alloc38) 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_q_weight2: R.Tensor((3072, 96), dtype="uint32") = packed_params[92] gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale2: R.Tensor((3072, 24), dtype="float16") = packed_params[93] gpt_neox_layers_5_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[94] alloc39: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage3, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale2, alloc38, gpt_neox_layers_5_mlp_dense_h_to_4h_bias2, alloc39) R.vm.kill_object(alloc38) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight2) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale2) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_bias2) gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight2: R.Tensor((768, 384), dtype="uint32") = packed_params[95] gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale2: R.Tensor((768, 96), dtype="float16") = packed_params[96] gpt_neox_layers_5_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[97] alloc40: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale2, alloc39, gpt_neox_layers_5_mlp_dense_4h_to_h_bias2, alloc40) R.vm.kill_object(alloc39) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight2) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale2) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_bias2) gpt_neox_layers_5_attention_dense_q_weight2: R.Tensor((768, 96), dtype="uint32") = packed_params[89] gpt_neox_layers_5_attention_dense_q_scale2: R.Tensor((768, 24), dtype="float16") = packed_params[90] gpt_neox_layers_5_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[91] alloc41: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage3, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_5_attention_dense_q_weight2, gpt_neox_layers_5_attention_dense_q_scale2, lv179, gpt_neox_layers_5_attention_dense_bias2, alloc40, alloc34, alloc41) R.vm.kill_object(alloc34) R.vm.kill_object(lv179) R.vm.kill_object(alloc40) R.vm.kill_object(gpt_neox_layers_5_attention_dense_q_weight2) R.vm.kill_object(gpt_neox_layers_5_attention_dense_q_scale2) 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[98] gpt_neox_layers_6_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[99] alloc42: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc41, gpt_neox_layers_6_input_layernorm_weight2, gpt_neox_layers_6_input_layernorm_bias2, alloc42) 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_q_weight2: R.Tensor((2304, 96), dtype="uint32") = packed_params[102] gpt_neox_layers_6_attention_query_key_value_q_scale2: R.Tensor((2304, 24), dtype="float16") = packed_params[103] gpt_neox_layers_6_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[104] alloc43: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_6_attention_query_key_value_q_weight2, gpt_neox_layers_6_attention_query_key_value_q_scale2, alloc42, gpt_neox_layers_6_attention_query_key_value_bias2, alloc43) R.vm.kill_object(alloc42) R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_q_weight2) R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_q_scale2) 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", alloc43, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc43) alloc44: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) _44: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), lv184, alloc44) R.vm.kill_object(lv184) lv185: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc44, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) R.vm.kill_object(alloc44) gpt_neox_layers_6_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[100] gpt_neox_layers_6_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[101] alloc45: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc41, gpt_neox_layers_6_post_attention_layernorm_weight2, gpt_neox_layers_6_post_attention_layernorm_bias2, alloc45) 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_q_weight2: R.Tensor((3072, 96), dtype="uint32") = packed_params[108] gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale2: R.Tensor((3072, 24), dtype="float16") = packed_params[109] gpt_neox_layers_6_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[110] alloc46: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage2, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale2, alloc45, gpt_neox_layers_6_mlp_dense_h_to_4h_bias2, alloc46) R.vm.kill_object(alloc45) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight2) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale2) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_bias2) gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight2: R.Tensor((768, 384), dtype="uint32") = packed_params[111] gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale2: R.Tensor((768, 96), dtype="float16") = packed_params[112] gpt_neox_layers_6_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[113] alloc47: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale2, alloc46, gpt_neox_layers_6_mlp_dense_4h_to_h_bias2, alloc47) R.vm.kill_object(alloc46) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight2) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale2) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_bias2) gpt_neox_layers_6_attention_dense_q_weight2: R.Tensor((768, 96), dtype="uint32") = packed_params[105] gpt_neox_layers_6_attention_dense_q_scale2: R.Tensor((768, 24), dtype="float16") = packed_params[106] gpt_neox_layers_6_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[107] alloc48: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage2, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_6_attention_dense_q_weight2, gpt_neox_layers_6_attention_dense_q_scale2, lv185, gpt_neox_layers_6_attention_dense_bias2, alloc47, alloc41, alloc48) R.vm.kill_object(alloc41) R.vm.kill_object(lv185) R.vm.kill_object(alloc47) R.vm.kill_object(gpt_neox_layers_6_attention_dense_q_weight2) R.vm.kill_object(gpt_neox_layers_6_attention_dense_q_scale2) 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[114] gpt_neox_layers_7_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[115] alloc49: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc48, gpt_neox_layers_7_input_layernorm_weight2, gpt_neox_layers_7_input_layernorm_bias2, alloc49) 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_q_weight2: R.Tensor((2304, 96), dtype="uint32") = packed_params[118] gpt_neox_layers_7_attention_query_key_value_q_scale2: R.Tensor((2304, 24), dtype="float16") = packed_params[119] gpt_neox_layers_7_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[120] alloc50: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_7_attention_query_key_value_q_weight2, gpt_neox_layers_7_attention_query_key_value_q_scale2, alloc49, gpt_neox_layers_7_attention_query_key_value_bias2, alloc50) R.vm.kill_object(alloc49) R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_q_weight2) R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_q_scale2) 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", alloc50, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc50) alloc51: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) _51: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), lv190, alloc51) R.vm.kill_object(lv190) lv191: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc51, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) R.vm.kill_object(alloc51) gpt_neox_layers_7_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[116] gpt_neox_layers_7_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[117] alloc52: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc48, gpt_neox_layers_7_post_attention_layernorm_weight2, gpt_neox_layers_7_post_attention_layernorm_bias2, alloc52) 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_q_weight2: R.Tensor((3072, 96), dtype="uint32") = packed_params[124] gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale2: R.Tensor((3072, 24), dtype="float16") = packed_params[125] gpt_neox_layers_7_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[126] alloc53: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage3, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale2, alloc52, gpt_neox_layers_7_mlp_dense_h_to_4h_bias2, alloc53) R.vm.kill_object(alloc52) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight2) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale2) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_bias2) gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight2: R.Tensor((768, 384), dtype="uint32") = packed_params[127] gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale2: R.Tensor((768, 96), dtype="float16") = packed_params[128] gpt_neox_layers_7_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[129] alloc54: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale2, alloc53, gpt_neox_layers_7_mlp_dense_4h_to_h_bias2, alloc54) R.vm.kill_object(alloc53) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight2) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale2) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_bias2) gpt_neox_layers_7_attention_dense_q_weight2: R.Tensor((768, 96), dtype="uint32") = packed_params[121] gpt_neox_layers_7_attention_dense_q_scale2: R.Tensor((768, 24), dtype="float16") = packed_params[122] gpt_neox_layers_7_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[123] alloc55: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage3, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_7_attention_dense_q_weight2, gpt_neox_layers_7_attention_dense_q_scale2, lv191, gpt_neox_layers_7_attention_dense_bias2, alloc54, alloc48, alloc55) R.vm.kill_object(alloc48) R.vm.kill_object(lv191) R.vm.kill_object(alloc54) R.vm.kill_object(gpt_neox_layers_7_attention_dense_q_weight2) R.vm.kill_object(gpt_neox_layers_7_attention_dense_q_scale2) 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[130] gpt_neox_layers_8_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[131] alloc56: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc55, gpt_neox_layers_8_input_layernorm_weight2, gpt_neox_layers_8_input_layernorm_bias2, alloc56) 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_q_weight2: R.Tensor((2304, 96), dtype="uint32") = packed_params[134] gpt_neox_layers_8_attention_query_key_value_q_scale2: R.Tensor((2304, 24), dtype="float16") = packed_params[135] gpt_neox_layers_8_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[136] alloc57: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_8_attention_query_key_value_q_weight2, gpt_neox_layers_8_attention_query_key_value_q_scale2, alloc56, gpt_neox_layers_8_attention_query_key_value_bias2, alloc57) R.vm.kill_object(alloc56) R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_q_weight2) R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_q_scale2) 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", alloc57, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc57) alloc58: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) _58: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), lv196, alloc58) R.vm.kill_object(lv196) lv197: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc58, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) R.vm.kill_object(alloc58) gpt_neox_layers_8_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[132] gpt_neox_layers_8_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[133] alloc59: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc55, gpt_neox_layers_8_post_attention_layernorm_weight2, gpt_neox_layers_8_post_attention_layernorm_bias2, alloc59) 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_q_weight2: R.Tensor((3072, 96), dtype="uint32") = packed_params[140] gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale2: R.Tensor((3072, 24), dtype="float16") = packed_params[141] gpt_neox_layers_8_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[142] alloc60: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage2, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale2, alloc59, gpt_neox_layers_8_mlp_dense_h_to_4h_bias2, alloc60) R.vm.kill_object(alloc59) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight2) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale2) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_bias2) gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight2: R.Tensor((768, 384), dtype="uint32") = packed_params[143] gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale2: R.Tensor((768, 96), dtype="float16") = packed_params[144] gpt_neox_layers_8_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[145] alloc61: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale2, alloc60, gpt_neox_layers_8_mlp_dense_4h_to_h_bias2, alloc61) R.vm.kill_object(alloc60) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight2) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale2) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_bias2) gpt_neox_layers_8_attention_dense_q_weight2: R.Tensor((768, 96), dtype="uint32") = packed_params[137] gpt_neox_layers_8_attention_dense_q_scale2: R.Tensor((768, 24), dtype="float16") = packed_params[138] gpt_neox_layers_8_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[139] alloc62: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage2, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_8_attention_dense_q_weight2, gpt_neox_layers_8_attention_dense_q_scale2, lv197, gpt_neox_layers_8_attention_dense_bias2, alloc61, alloc55, alloc62) R.vm.kill_object(alloc55) R.vm.kill_object(lv197) R.vm.kill_object(alloc61) R.vm.kill_object(gpt_neox_layers_8_attention_dense_q_weight2) R.vm.kill_object(gpt_neox_layers_8_attention_dense_q_scale2) 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[146] gpt_neox_layers_9_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[147] alloc63: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc62, gpt_neox_layers_9_input_layernorm_weight2, gpt_neox_layers_9_input_layernorm_bias2, alloc63) 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_q_weight2: R.Tensor((2304, 96), dtype="uint32") = packed_params[150] gpt_neox_layers_9_attention_query_key_value_q_scale2: R.Tensor((2304, 24), dtype="float16") = packed_params[151] gpt_neox_layers_9_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[152] alloc64: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_9_attention_query_key_value_q_weight2, gpt_neox_layers_9_attention_query_key_value_q_scale2, alloc63, gpt_neox_layers_9_attention_query_key_value_bias2, alloc64) R.vm.kill_object(alloc63) R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_q_weight2) R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_q_scale2) 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", alloc64, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc64) alloc65: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) _65: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), lv202, alloc65) R.vm.kill_object(lv202) lv203: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc65, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) R.vm.kill_object(alloc65) gpt_neox_layers_9_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[148] gpt_neox_layers_9_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[149] alloc66: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc62, gpt_neox_layers_9_post_attention_layernorm_weight2, gpt_neox_layers_9_post_attention_layernorm_bias2, alloc66) 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_q_weight2: R.Tensor((3072, 96), dtype="uint32") = packed_params[156] gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale2: R.Tensor((3072, 24), dtype="float16") = packed_params[157] gpt_neox_layers_9_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[158] alloc67: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage3, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale2, alloc66, gpt_neox_layers_9_mlp_dense_h_to_4h_bias2, alloc67) R.vm.kill_object(alloc66) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight2) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale2) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_bias2) gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight2: R.Tensor((768, 384), dtype="uint32") = packed_params[159] gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale2: R.Tensor((768, 96), dtype="float16") = packed_params[160] gpt_neox_layers_9_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[161] alloc68: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale2, alloc67, gpt_neox_layers_9_mlp_dense_4h_to_h_bias2, alloc68) R.vm.kill_object(alloc67) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight2) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale2) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_bias2) gpt_neox_layers_9_attention_dense_q_weight2: R.Tensor((768, 96), dtype="uint32") = packed_params[153] gpt_neox_layers_9_attention_dense_q_scale2: R.Tensor((768, 24), dtype="float16") = packed_params[154] gpt_neox_layers_9_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[155] alloc69: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage3, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_9_attention_dense_q_weight2, gpt_neox_layers_9_attention_dense_q_scale2, lv203, gpt_neox_layers_9_attention_dense_bias2, alloc68, alloc62, alloc69) R.vm.kill_object(alloc62) R.vm.kill_object(lv203) R.vm.kill_object(alloc68) R.vm.kill_object(gpt_neox_layers_9_attention_dense_q_weight2) R.vm.kill_object(gpt_neox_layers_9_attention_dense_q_scale2) 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[162] gpt_neox_layers_10_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[163] alloc70: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc69, gpt_neox_layers_10_input_layernorm_weight2, gpt_neox_layers_10_input_layernorm_bias2, alloc70) 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_q_weight2: R.Tensor((2304, 96), dtype="uint32") = packed_params[166] gpt_neox_layers_10_attention_query_key_value_q_scale2: R.Tensor((2304, 24), dtype="float16") = packed_params[167] gpt_neox_layers_10_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[168] alloc71: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_10_attention_query_key_value_q_weight2, gpt_neox_layers_10_attention_query_key_value_q_scale2, alloc70, gpt_neox_layers_10_attention_query_key_value_bias2, alloc71) R.vm.kill_object(alloc70) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_q_weight2) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_q_scale2) 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", alloc71, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc71) alloc72: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) _72: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), lv208, alloc72) R.vm.kill_object(lv208) lv209: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc72, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) R.vm.kill_object(alloc72) gpt_neox_layers_10_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[164] gpt_neox_layers_10_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[165] alloc73: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc69, gpt_neox_layers_10_post_attention_layernorm_weight2, gpt_neox_layers_10_post_attention_layernorm_bias2, alloc73) 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_q_weight2: R.Tensor((3072, 96), dtype="uint32") = packed_params[172] gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale2: R.Tensor((3072, 24), dtype="float16") = packed_params[173] gpt_neox_layers_10_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[174] alloc74: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage2, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale2, alloc73, gpt_neox_layers_10_mlp_dense_h_to_4h_bias2, alloc74) R.vm.kill_object(alloc73) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight2) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale2) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_bias2) gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight2: R.Tensor((768, 384), dtype="uint32") = packed_params[175] gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale2: R.Tensor((768, 96), dtype="float16") = packed_params[176] gpt_neox_layers_10_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[177] alloc75: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale2, alloc74, gpt_neox_layers_10_mlp_dense_4h_to_h_bias2, alloc75) R.vm.kill_object(alloc74) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight2) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale2) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_bias2) gpt_neox_layers_10_attention_dense_q_weight2: R.Tensor((768, 96), dtype="uint32") = packed_params[169] gpt_neox_layers_10_attention_dense_q_scale2: R.Tensor((768, 24), dtype="float16") = packed_params[170] gpt_neox_layers_10_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[171] alloc76: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage2, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) R.vm.kill_object(storage2) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_10_attention_dense_q_weight2, gpt_neox_layers_10_attention_dense_q_scale2, lv209, gpt_neox_layers_10_attention_dense_bias2, alloc75, alloc69, alloc76) R.vm.kill_object(alloc69) R.vm.kill_object(lv209) R.vm.kill_object(alloc75) R.vm.kill_object(gpt_neox_layers_10_attention_dense_q_weight2) R.vm.kill_object(gpt_neox_layers_10_attention_dense_q_scale2) 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[178] gpt_neox_layers_11_input_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[179] alloc77: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc76, gpt_neox_layers_11_input_layernorm_weight2, gpt_neox_layers_11_input_layernorm_bias2, alloc77) 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_q_weight2: R.Tensor((2304, 96), dtype="uint32") = packed_params[182] gpt_neox_layers_11_attention_query_key_value_q_scale2: R.Tensor((2304, 24), dtype="float16") = packed_params[183] gpt_neox_layers_11_attention_query_key_value_bias2: R.Tensor((2304,), dtype="float16") = packed_params[184] alloc78: R.Tensor((1, 1, 2304), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 2304]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_11_attention_query_key_value_q_weight2, gpt_neox_layers_11_attention_query_key_value_q_scale2, alloc77, gpt_neox_layers_11_attention_query_key_value_bias2, alloc78) R.vm.kill_object(alloc77) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_q_weight2) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_q_scale2) 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", alloc78, R.shape([1, 36, 64]), sinfo_args=(R.Tensor((1, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc78) alloc79: R.Tensor((1, 12, 64), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 12, 64]), R.dtype("float16")) _79: R.Object = R.call_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), lv214, alloc79) R.vm.kill_object(lv214) lv215: R.Tensor((1, 1, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", alloc79, R.shape([1, 1, 768]), sinfo_args=(R.Tensor((1, 1, 768), dtype="float16"),)) R.vm.kill_object(alloc79) gpt_neox_layers_11_post_attention_layernorm_weight2: R.Tensor((768,), dtype="float16") = packed_params[180] gpt_neox_layers_11_post_attention_layernorm_bias2: R.Tensor((768,), dtype="float16") = packed_params[181] alloc80: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm2(alloc76, gpt_neox_layers_11_post_attention_layernorm_weight2, gpt_neox_layers_11_post_attention_layernorm_bias2, alloc80) 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_q_weight2: R.Tensor((3072, 96), dtype="uint32") = packed_params[188] gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale2: R.Tensor((3072, 24), dtype="float16") = packed_params[189] gpt_neox_layers_11_mlp_dense_h_to_4h_bias2: R.Tensor((3072,), dtype="float32") = packed_params[190] alloc81: R.Tensor((1, 1, 3072), dtype="float16") = R.vm.alloc_tensor(storage3, R.prim_value(0), R.shape([1, 1, 3072]), R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale2, alloc80, gpt_neox_layers_11_mlp_dense_h_to_4h_bias2, alloc81) R.vm.kill_object(alloc80) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight2) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale2) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_bias2) gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight2: R.Tensor((768, 384), dtype="uint32") = packed_params[191] gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale2: R.Tensor((768, 96), dtype="float16") = packed_params[192] gpt_neox_layers_11_mlp_dense_4h_to_h_bias2: R.Tensor((768,), dtype="float32") = packed_params[193] alloc82: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage1, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) R.vm.kill_object(storage1) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale2, alloc81, gpt_neox_layers_11_mlp_dense_4h_to_h_bias2, alloc82) R.vm.kill_object(alloc81) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight2) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale2) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_bias2) gpt_neox_layers_11_attention_dense_q_weight2: R.Tensor((768, 96), dtype="uint32") = packed_params[185] gpt_neox_layers_11_attention_dense_q_scale2: R.Tensor((768, 24), dtype="float16") = packed_params[186] gpt_neox_layers_11_attention_dense_bias2: R.Tensor((768,), dtype="float16") = packed_params[187] alloc83: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage3, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) R.vm.kill_object(storage3) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_11_attention_dense_q_weight2, gpt_neox_layers_11_attention_dense_q_scale2, lv215, gpt_neox_layers_11_attention_dense_bias2, alloc82, alloc76, alloc83) R.vm.kill_object(alloc76) R.vm.kill_object(lv215) R.vm.kill_object(alloc82) R.vm.kill_object(gpt_neox_layers_11_attention_dense_q_weight2) R.vm.kill_object(gpt_neox_layers_11_attention_dense_q_scale2) R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias2) gpt_neox_final_layer_norm_weight2: R.Tensor((768,), dtype="float16") = packed_params[194] gpt_neox_final_layer_norm_bias2: R.Tensor((768,), dtype="float16") = packed_params[195] alloc84: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) R.vm.kill_object(storage) cls.gpt_neox_q4f16_1_layer_norm2(alloc83, gpt_neox_final_layer_norm_weight2, gpt_neox_final_layer_norm_bias2, alloc84) R.vm.kill_object(alloc83) R.vm.kill_object(gpt_neox_final_layer_norm_weight2) R.vm.kill_object(gpt_neox_final_layer_norm_bias2) embed_out_q_weight2: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[196] embed_out_q_scale2: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[197] gv4: 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),)) storage4: R.Object = R.vm.alloc_storage(gv4, 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(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), sinfo_args=(R.Shape(ndim=3),)) alloc85: R.Tensor(dtype="float32", ndim=3) = R.vm.alloc_tensor(storage4, R.prim_value(0), gv5, R.dtype("float32")) R.vm.kill_object(storage4) cls.gpt_neox_q4f16_1_fused_dequantize_fused_NT_matmul14_cast8(embed_out_q_weight2, embed_out_q_scale2, alloc84, alloc85) R.vm.kill_object(alloc84) R.vm.kill_object(embed_out_q_weight2) R.vm.kill_object(embed_out_q_scale2) gv2_1: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc85, paged_kv_cache R.vm.kill_object(alloc85) gv6: R.Tensor(dtype="float32", ndim=3) = gv2_1[0] R.call_packed("vm.builtin.match_shape", gv6, 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_1 @R.function def embed(input_ids: R.Tensor(("seq_len",), dtype="int32"), packed_params: R.Tuple(R.Tensor(("vocab_size", 96), dtype="uint32"), R.Tensor(("vocab_size", 24), 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, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor(("vocab_size", 96), dtype="uint32"), R.Tensor(("vocab_size", 24), 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(198), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv16: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[0] R.call_packed("vm.builtin.check_tensor_info", gv16, R.prim_value(2), R.dtype("uint32"), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv17: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[1] R.call_packed("vm.builtin.check_tensor_info", gv17, 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, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv18: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[196] R.call_packed("vm.builtin.check_tensor_info", gv18, R.prim_value(2), R.dtype("uint32"), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv19: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[197] R.call_packed("vm.builtin.check_tensor_info", gv19, 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, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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", gv16, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(96), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv17, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(24), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv18, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(96), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv19, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(24), R.str("ErrorContext(fn=embed, loc=param[1], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gpt_neox_embed_in_q_weight: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[0] gpt_neox_embed_in_q_scale: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[1] storage8: R.Object = R.vm.alloc_storage(R.shape([3145728]), R.prim_value(0), R.dtype("uint8"), R.str("global")) gv20: 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),)) alloc89: R.Tensor(dtype="float16", ndim=2) = R.vm.alloc_tensor(storage8, R.prim_value(0), gv20, R.dtype("float16")) R.vm.kill_object(storage8) cls.gpt_neox_q4f16_1_fused_dequantize_take1(gpt_neox_embed_in_q_weight, gpt_neox_embed_in_q_scale, input_ids, alloc89) R.vm.kill_object(gpt_neox_embed_in_q_weight) R.vm.kill_object(gpt_neox_embed_in_q_scale) R.call_packed("vm.builtin.match_shape", alloc89, 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 alloc89 @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", 96), dtype="uint32"), R.Tensor(("vocab_size", 24), 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, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((2304, 96), dtype="uint32"), R.Tensor((2304, 24), dtype="float16"), R.Tensor((2304,), dtype="float16"), R.Tensor((768, 96), dtype="uint32"), R.Tensor((768, 24), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor((3072, 96), dtype="uint32"), R.Tensor((3072, 24), dtype="float16"), R.Tensor((3072,), dtype="float32"), R.Tensor((768, 384), dtype="uint32"), R.Tensor((768, 96), dtype="float16"), R.Tensor((768,), dtype="float32"), R.Tensor((768,), dtype="float16"), R.Tensor((768,), dtype="float16"), R.Tensor(("vocab_size", 96), dtype="uint32"), R.Tensor(("vocab_size", 24), 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(198), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv442: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[0] R.call_packed("vm.builtin.check_tensor_info", gv442, R.prim_value(2), R.dtype("uint32"), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv443: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[1] R.call_packed("vm.builtin.check_tensor_info", gv443, 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, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv444: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[196] R.call_packed("vm.builtin.check_tensor_info", gv444, R.prim_value(2), R.dtype("uint32"), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) gv445: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[197] R.call_packed("vm.builtin.check_tensor_info", gv445, 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, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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", gv442, shape_heap, R.prim_value(2), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(96), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv443, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(24), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv444, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(96), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) R.call_packed("vm.builtin.match_shape", gv445, shape_heap, R.prim_value(2), R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(24), R.str("ErrorContext(fn=prefill, loc=param[2], param=packed_params, annotation=R.Tuple(R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), 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, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((2304, 96), dtype=\"uint32\"), R.Tensor((2304, 24), dtype=\"float16\"), R.Tensor((2304,), dtype=\"float16\"), R.Tensor((768, 96), dtype=\"uint32\"), R.Tensor((768, 24), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((3072, 96), dtype=\"uint32\"), R.Tensor((3072, 24), dtype=\"float16\"), R.Tensor((3072,), dtype=\"float32\"), R.Tensor((768, 384), dtype=\"uint32\"), R.Tensor((768, 96), dtype=\"float16\"), R.Tensor((768,), dtype=\"float32\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((768,), dtype=\"float16\"), R.Tensor((vocab_size, 96), dtype=\"uint32\"), R.Tensor((vocab_size, 24), dtype=\"float16\"))) "), sinfo_args=(R.Tuple,)) cls.gpt_neox_q4f16_1_shape_func5(shape_heap) gpt_neox_layers_0_input_layernorm_weight1: R.Tensor((768,), dtype="float16") = packed_params[2] gpt_neox_layers_0_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[3] storage25: R.Object = R.vm.alloc_storage(R.shape([3145728]), R.prim_value(0), R.dtype("uint8"), R.str("global")) gv446: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc349: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv446, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(input_embed, gpt_neox_layers_0_input_layernorm_weight1, gpt_neox_layers_0_input_layernorm_bias1, alloc349) 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_q_weight1: R.Tensor((2304, 96), dtype="uint32") = packed_params[6] gpt_neox_layers_0_attention_query_key_value_q_scale1: R.Tensor((2304, 24), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[8] storage26: R.Object = R.vm.alloc_storage(R.shape([9437184]), R.prim_value(0), R.dtype("uint8"), R.str("global")) gv447: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) alloc350: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv447, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_0_attention_query_key_value_q_weight1, gpt_neox_layers_0_attention_query_key_value_q_scale1, alloc349, gpt_neox_layers_0_attention_query_key_value_bias1, alloc350) R.vm.kill_object(alloc349) R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_q_weight1) R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_q_scale1) R.vm.kill_object(gpt_neox_layers_0_attention_query_key_value_bias1) gv448: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc350, gv448, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc350) gv449: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv449, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape) 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc351: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv450, R.dtype("float16")) _350: 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, alloc351) R.vm.kill_object(reshape1) gv451: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc351, gv451, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc351) gv452: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape3: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape2, gv452, 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[4] gpt_neox_layers_0_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[5] gv453: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc352: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv453, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(input_embed, gpt_neox_layers_0_post_attention_layernorm_weight1, gpt_neox_layers_0_post_attention_layernorm_bias1, alloc352) 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_q_weight1: R.Tensor((3072, 96), dtype="uint32") = packed_params[12] gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale1: R.Tensor((3072, 24), dtype="float16") = packed_params[13] gpt_neox_layers_0_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[14] storage27: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) gv454: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) alloc353: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv454, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale1, alloc352, gpt_neox_layers_0_mlp_dense_h_to_4h_bias1, alloc353) R.vm.kill_object(alloc352) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight1) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale1) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_h_to_4h_bias1) gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight1: R.Tensor((768, 384), dtype="uint32") = packed_params[15] gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale1: R.Tensor((768, 96), dtype="float16") = packed_params[16] gpt_neox_layers_0_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[17] gv455: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc354: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv455, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale1, alloc353, gpt_neox_layers_0_mlp_dense_4h_to_h_bias1, alloc354) R.vm.kill_object(alloc353) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight1) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale1) R.vm.kill_object(gpt_neox_layers_0_mlp_dense_4h_to_h_bias1) gpt_neox_layers_0_attention_dense_q_weight1: R.Tensor((768, 96), dtype="uint32") = packed_params[9] gpt_neox_layers_0_attention_dense_q_scale1: R.Tensor((768, 24), dtype="float16") = packed_params[10] gpt_neox_layers_0_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[11] gv456: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc355: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv456, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_0_attention_dense_q_weight1, gpt_neox_layers_0_attention_dense_q_scale1, reshape3, gpt_neox_layers_0_attention_dense_bias1, alloc354, input_embed, alloc355) R.vm.kill_object(reshape3) R.vm.kill_object(alloc354) R.vm.kill_object(gpt_neox_layers_0_attention_dense_q_weight1) R.vm.kill_object(gpt_neox_layers_0_attention_dense_q_scale1) 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[18] gpt_neox_layers_1_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[19] gv457: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc356: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv457, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc355, gpt_neox_layers_1_input_layernorm_weight1, gpt_neox_layers_1_input_layernorm_bias1, alloc356) 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_q_weight1: R.Tensor((2304, 96), dtype="uint32") = packed_params[22] gpt_neox_layers_1_attention_query_key_value_q_scale1: R.Tensor((2304, 24), dtype="float16") = packed_params[23] gpt_neox_layers_1_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[24] gv458: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) alloc357: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv458, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_1_attention_query_key_value_q_weight1, gpt_neox_layers_1_attention_query_key_value_q_scale1, alloc356, gpt_neox_layers_1_attention_query_key_value_bias1, alloc357) R.vm.kill_object(alloc356) R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_q_weight1) R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_q_scale1) R.vm.kill_object(gpt_neox_layers_1_attention_query_key_value_bias1) gv459: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc357, gv459, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc357) gv460: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv460, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape4) 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc358: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv461, R.dtype("float16")) _357: 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, alloc358) R.vm.kill_object(reshape5) gv462: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc358, gv462, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc358) gv463: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape7: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape6, gv463, 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[20] gpt_neox_layers_1_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[21] gv464: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc359: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv464, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc355, gpt_neox_layers_1_post_attention_layernorm_weight1, gpt_neox_layers_1_post_attention_layernorm_bias1, alloc359) 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_q_weight1: R.Tensor((3072, 96), dtype="uint32") = packed_params[28] gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale1: R.Tensor((3072, 24), dtype="float16") = packed_params[29] gpt_neox_layers_1_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[30] storage28: R.Object = R.vm.alloc_storage(R.shape([12582912]), R.prim_value(0), R.dtype("uint8"), R.str("global")) gv465: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) alloc360: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv465, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale1, alloc359, gpt_neox_layers_1_mlp_dense_h_to_4h_bias1, alloc360) R.vm.kill_object(alloc359) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight1) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale1) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_h_to_4h_bias1) gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight1: R.Tensor((768, 384), dtype="uint32") = packed_params[31] gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale1: R.Tensor((768, 96), dtype="float16") = packed_params[32] gpt_neox_layers_1_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[33] gv466: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc361: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv466, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale1, alloc360, gpt_neox_layers_1_mlp_dense_4h_to_h_bias1, alloc361) R.vm.kill_object(alloc360) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight1) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale1) R.vm.kill_object(gpt_neox_layers_1_mlp_dense_4h_to_h_bias1) gpt_neox_layers_1_attention_dense_q_weight1: R.Tensor((768, 96), dtype="uint32") = packed_params[25] gpt_neox_layers_1_attention_dense_q_scale1: R.Tensor((768, 24), dtype="float16") = packed_params[26] gpt_neox_layers_1_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[27] gv467: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc362: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv467, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_1_attention_dense_q_weight1, gpt_neox_layers_1_attention_dense_q_scale1, reshape7, gpt_neox_layers_1_attention_dense_bias1, alloc361, alloc355, alloc362) R.vm.kill_object(alloc355) R.vm.kill_object(reshape7) R.vm.kill_object(alloc361) R.vm.kill_object(gpt_neox_layers_1_attention_dense_q_weight1) R.vm.kill_object(gpt_neox_layers_1_attention_dense_q_scale1) 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[34] gpt_neox_layers_2_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[35] gv468: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc363: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv468, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc362, gpt_neox_layers_2_input_layernorm_weight1, gpt_neox_layers_2_input_layernorm_bias1, alloc363) 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_q_weight1: R.Tensor((2304, 96), dtype="uint32") = packed_params[38] gpt_neox_layers_2_attention_query_key_value_q_scale1: R.Tensor((2304, 24), dtype="float16") = packed_params[39] gpt_neox_layers_2_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[40] gv469: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) alloc364: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv469, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_2_attention_query_key_value_q_weight1, gpt_neox_layers_2_attention_query_key_value_q_scale1, alloc363, gpt_neox_layers_2_attention_query_key_value_bias1, alloc364) R.vm.kill_object(alloc363) R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_q_weight1) R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_q_scale1) R.vm.kill_object(gpt_neox_layers_2_attention_query_key_value_bias1) gv470: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc364, gv470, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc364) gv471: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv471, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape8) 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc365: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv472, R.dtype("float16")) _364: 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, alloc365) R.vm.kill_object(reshape9) gv473: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc365, gv473, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc365) gv474: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape11: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape10, gv474, 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[36] gpt_neox_layers_2_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[37] gv475: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc366: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv475, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc362, gpt_neox_layers_2_post_attention_layernorm_weight1, gpt_neox_layers_2_post_attention_layernorm_bias1, alloc366) 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_q_weight1: R.Tensor((3072, 96), dtype="uint32") = packed_params[44] gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale1: R.Tensor((3072, 24), dtype="float16") = packed_params[45] gpt_neox_layers_2_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[46] gv476: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) alloc367: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv476, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale1, alloc366, gpt_neox_layers_2_mlp_dense_h_to_4h_bias1, alloc367) R.vm.kill_object(alloc366) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight1) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale1) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_h_to_4h_bias1) gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight1: R.Tensor((768, 384), dtype="uint32") = packed_params[47] gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale1: R.Tensor((768, 96), dtype="float16") = packed_params[48] gpt_neox_layers_2_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[49] gv477: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc368: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv477, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale1, alloc367, gpt_neox_layers_2_mlp_dense_4h_to_h_bias1, alloc368) R.vm.kill_object(alloc367) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight1) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale1) R.vm.kill_object(gpt_neox_layers_2_mlp_dense_4h_to_h_bias1) gpt_neox_layers_2_attention_dense_q_weight1: R.Tensor((768, 96), dtype="uint32") = packed_params[41] gpt_neox_layers_2_attention_dense_q_scale1: R.Tensor((768, 24), dtype="float16") = packed_params[42] gpt_neox_layers_2_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[43] gv478: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc369: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv478, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_2_attention_dense_q_weight1, gpt_neox_layers_2_attention_dense_q_scale1, reshape11, gpt_neox_layers_2_attention_dense_bias1, alloc368, alloc362, alloc369) R.vm.kill_object(alloc362) R.vm.kill_object(reshape11) R.vm.kill_object(alloc368) R.vm.kill_object(gpt_neox_layers_2_attention_dense_q_weight1) R.vm.kill_object(gpt_neox_layers_2_attention_dense_q_scale1) 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[50] gpt_neox_layers_3_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[51] gv479: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc370: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv479, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc369, gpt_neox_layers_3_input_layernorm_weight1, gpt_neox_layers_3_input_layernorm_bias1, alloc370) 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_q_weight1: R.Tensor((2304, 96), dtype="uint32") = packed_params[54] gpt_neox_layers_3_attention_query_key_value_q_scale1: R.Tensor((2304, 24), dtype="float16") = packed_params[55] gpt_neox_layers_3_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[56] gv480: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) alloc371: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv480, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_3_attention_query_key_value_q_weight1, gpt_neox_layers_3_attention_query_key_value_q_scale1, alloc370, gpt_neox_layers_3_attention_query_key_value_bias1, alloc371) R.vm.kill_object(alloc370) R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_q_weight1) R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_q_scale1) R.vm.kill_object(gpt_neox_layers_3_attention_query_key_value_bias1) gv481: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc371, gv481, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc371) gv482: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv482, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape12) 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc372: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv483, R.dtype("float16")) _371: 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, alloc372) R.vm.kill_object(reshape13) gv484: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc372, gv484, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc372) gv485: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape15: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape14, gv485, 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[52] gpt_neox_layers_3_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[53] gv486: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc373: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv486, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc369, gpt_neox_layers_3_post_attention_layernorm_weight1, gpt_neox_layers_3_post_attention_layernorm_bias1, alloc373) 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_q_weight1: R.Tensor((3072, 96), dtype="uint32") = packed_params[60] gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale1: R.Tensor((3072, 24), dtype="float16") = packed_params[61] gpt_neox_layers_3_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[62] gv487: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) alloc374: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv487, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale1, alloc373, gpt_neox_layers_3_mlp_dense_h_to_4h_bias1, alloc374) R.vm.kill_object(alloc373) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight1) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale1) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_h_to_4h_bias1) gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight1: R.Tensor((768, 384), dtype="uint32") = packed_params[63] gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale1: R.Tensor((768, 96), dtype="float16") = packed_params[64] gpt_neox_layers_3_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[65] gv488: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc375: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv488, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale1, alloc374, gpt_neox_layers_3_mlp_dense_4h_to_h_bias1, alloc375) R.vm.kill_object(alloc374) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight1) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale1) R.vm.kill_object(gpt_neox_layers_3_mlp_dense_4h_to_h_bias1) gpt_neox_layers_3_attention_dense_q_weight1: R.Tensor((768, 96), dtype="uint32") = packed_params[57] gpt_neox_layers_3_attention_dense_q_scale1: R.Tensor((768, 24), dtype="float16") = packed_params[58] gpt_neox_layers_3_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[59] gv489: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc376: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv489, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_3_attention_dense_q_weight1, gpt_neox_layers_3_attention_dense_q_scale1, reshape15, gpt_neox_layers_3_attention_dense_bias1, alloc375, alloc369, alloc376) R.vm.kill_object(alloc369) R.vm.kill_object(reshape15) R.vm.kill_object(alloc375) R.vm.kill_object(gpt_neox_layers_3_attention_dense_q_weight1) R.vm.kill_object(gpt_neox_layers_3_attention_dense_q_scale1) 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[66] gpt_neox_layers_4_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[67] gv490: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc377: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv490, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc376, gpt_neox_layers_4_input_layernorm_weight1, gpt_neox_layers_4_input_layernorm_bias1, alloc377) 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_q_weight1: R.Tensor((2304, 96), dtype="uint32") = packed_params[70] gpt_neox_layers_4_attention_query_key_value_q_scale1: R.Tensor((2304, 24), dtype="float16") = packed_params[71] gpt_neox_layers_4_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[72] gv491: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) alloc378: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv491, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_4_attention_query_key_value_q_weight1, gpt_neox_layers_4_attention_query_key_value_q_scale1, alloc377, gpt_neox_layers_4_attention_query_key_value_bias1, alloc378) R.vm.kill_object(alloc377) R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_q_weight1) R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_q_scale1) R.vm.kill_object(gpt_neox_layers_4_attention_query_key_value_bias1) gv492: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc378, gv492, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc378) gv493: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv493, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape16) 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc379: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv494, R.dtype("float16")) _378: 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, alloc379) R.vm.kill_object(reshape17) gv495: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc379, gv495, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc379) gv496: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape19: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape18, gv496, 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[68] gpt_neox_layers_4_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[69] gv497: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc380: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv497, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc376, gpt_neox_layers_4_post_attention_layernorm_weight1, gpt_neox_layers_4_post_attention_layernorm_bias1, alloc380) 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_q_weight1: R.Tensor((3072, 96), dtype="uint32") = packed_params[76] gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale1: R.Tensor((3072, 24), dtype="float16") = packed_params[77] gpt_neox_layers_4_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[78] gv498: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) alloc381: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv498, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale1, alloc380, gpt_neox_layers_4_mlp_dense_h_to_4h_bias1, alloc381) R.vm.kill_object(alloc380) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight1) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale1) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_h_to_4h_bias1) gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight1: R.Tensor((768, 384), dtype="uint32") = packed_params[79] gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale1: R.Tensor((768, 96), dtype="float16") = packed_params[80] gpt_neox_layers_4_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[81] gv499: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc382: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv499, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale1, alloc381, gpt_neox_layers_4_mlp_dense_4h_to_h_bias1, alloc382) R.vm.kill_object(alloc381) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight1) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale1) R.vm.kill_object(gpt_neox_layers_4_mlp_dense_4h_to_h_bias1) gpt_neox_layers_4_attention_dense_q_weight1: R.Tensor((768, 96), dtype="uint32") = packed_params[73] gpt_neox_layers_4_attention_dense_q_scale1: R.Tensor((768, 24), dtype="float16") = packed_params[74] gpt_neox_layers_4_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[75] gv500: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc383: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv500, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_4_attention_dense_q_weight1, gpt_neox_layers_4_attention_dense_q_scale1, reshape19, gpt_neox_layers_4_attention_dense_bias1, alloc382, alloc376, alloc383) R.vm.kill_object(alloc376) R.vm.kill_object(reshape19) R.vm.kill_object(alloc382) R.vm.kill_object(gpt_neox_layers_4_attention_dense_q_weight1) R.vm.kill_object(gpt_neox_layers_4_attention_dense_q_scale1) 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[82] gpt_neox_layers_5_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[83] gv501: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc384: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv501, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc383, gpt_neox_layers_5_input_layernorm_weight1, gpt_neox_layers_5_input_layernorm_bias1, alloc384) 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_q_weight1: R.Tensor((2304, 96), dtype="uint32") = packed_params[86] gpt_neox_layers_5_attention_query_key_value_q_scale1: R.Tensor((2304, 24), dtype="float16") = packed_params[87] gpt_neox_layers_5_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[88] gv502: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) alloc385: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv502, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_5_attention_query_key_value_q_weight1, gpt_neox_layers_5_attention_query_key_value_q_scale1, alloc384, gpt_neox_layers_5_attention_query_key_value_bias1, alloc385) R.vm.kill_object(alloc384) R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_q_weight1) R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_q_scale1) R.vm.kill_object(gpt_neox_layers_5_attention_query_key_value_bias1) gv503: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc385, gv503, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc385) gv504: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv504, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape20) 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc386: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv505, R.dtype("float16")) _385: 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, alloc386) R.vm.kill_object(reshape21) gv506: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc386, gv506, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc386) gv507: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape23: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape22, gv507, 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[84] gpt_neox_layers_5_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[85] gv508: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc387: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv508, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc383, gpt_neox_layers_5_post_attention_layernorm_weight1, gpt_neox_layers_5_post_attention_layernorm_bias1, alloc387) 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_q_weight1: R.Tensor((3072, 96), dtype="uint32") = packed_params[92] gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale1: R.Tensor((3072, 24), dtype="float16") = packed_params[93] gpt_neox_layers_5_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[94] gv509: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) alloc388: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv509, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale1, alloc387, gpt_neox_layers_5_mlp_dense_h_to_4h_bias1, alloc388) R.vm.kill_object(alloc387) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight1) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale1) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_h_to_4h_bias1) gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight1: R.Tensor((768, 384), dtype="uint32") = packed_params[95] gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale1: R.Tensor((768, 96), dtype="float16") = packed_params[96] gpt_neox_layers_5_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[97] gv510: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc389: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv510, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale1, alloc388, gpt_neox_layers_5_mlp_dense_4h_to_h_bias1, alloc389) R.vm.kill_object(alloc388) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight1) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale1) R.vm.kill_object(gpt_neox_layers_5_mlp_dense_4h_to_h_bias1) gpt_neox_layers_5_attention_dense_q_weight1: R.Tensor((768, 96), dtype="uint32") = packed_params[89] gpt_neox_layers_5_attention_dense_q_scale1: R.Tensor((768, 24), dtype="float16") = packed_params[90] gpt_neox_layers_5_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[91] gv511: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc390: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv511, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_5_attention_dense_q_weight1, gpt_neox_layers_5_attention_dense_q_scale1, reshape23, gpt_neox_layers_5_attention_dense_bias1, alloc389, alloc383, alloc390) R.vm.kill_object(alloc383) R.vm.kill_object(reshape23) R.vm.kill_object(alloc389) R.vm.kill_object(gpt_neox_layers_5_attention_dense_q_weight1) R.vm.kill_object(gpt_neox_layers_5_attention_dense_q_scale1) 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[98] gpt_neox_layers_6_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[99] gv512: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc391: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv512, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc390, gpt_neox_layers_6_input_layernorm_weight1, gpt_neox_layers_6_input_layernorm_bias1, alloc391) 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_q_weight1: R.Tensor((2304, 96), dtype="uint32") = packed_params[102] gpt_neox_layers_6_attention_query_key_value_q_scale1: R.Tensor((2304, 24), dtype="float16") = packed_params[103] gpt_neox_layers_6_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[104] gv513: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) alloc392: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv513, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_6_attention_query_key_value_q_weight1, gpt_neox_layers_6_attention_query_key_value_q_scale1, alloc391, gpt_neox_layers_6_attention_query_key_value_bias1, alloc392) R.vm.kill_object(alloc391) R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_q_weight1) R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_q_scale1) R.vm.kill_object(gpt_neox_layers_6_attention_query_key_value_bias1) gv514: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc392, gv514, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc392) gv515: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv515, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape24) 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc393: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv516, R.dtype("float16")) _392: 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, alloc393) R.vm.kill_object(reshape25) gv517: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc393, gv517, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc393) gv518: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape27: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape26, gv518, 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[100] gpt_neox_layers_6_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[101] gv519: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc394: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv519, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc390, gpt_neox_layers_6_post_attention_layernorm_weight1, gpt_neox_layers_6_post_attention_layernorm_bias1, alloc394) 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_q_weight1: R.Tensor((3072, 96), dtype="uint32") = packed_params[108] gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale1: R.Tensor((3072, 24), dtype="float16") = packed_params[109] gpt_neox_layers_6_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[110] gv520: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) alloc395: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv520, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale1, alloc394, gpt_neox_layers_6_mlp_dense_h_to_4h_bias1, alloc395) R.vm.kill_object(alloc394) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight1) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale1) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_h_to_4h_bias1) gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight1: R.Tensor((768, 384), dtype="uint32") = packed_params[111] gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale1: R.Tensor((768, 96), dtype="float16") = packed_params[112] gpt_neox_layers_6_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[113] gv521: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc396: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv521, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale1, alloc395, gpt_neox_layers_6_mlp_dense_4h_to_h_bias1, alloc396) R.vm.kill_object(alloc395) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight1) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale1) R.vm.kill_object(gpt_neox_layers_6_mlp_dense_4h_to_h_bias1) gpt_neox_layers_6_attention_dense_q_weight1: R.Tensor((768, 96), dtype="uint32") = packed_params[105] gpt_neox_layers_6_attention_dense_q_scale1: R.Tensor((768, 24), dtype="float16") = packed_params[106] gpt_neox_layers_6_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[107] gv522: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc397: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv522, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_6_attention_dense_q_weight1, gpt_neox_layers_6_attention_dense_q_scale1, reshape27, gpt_neox_layers_6_attention_dense_bias1, alloc396, alloc390, alloc397) R.vm.kill_object(alloc390) R.vm.kill_object(reshape27) R.vm.kill_object(alloc396) R.vm.kill_object(gpt_neox_layers_6_attention_dense_q_weight1) R.vm.kill_object(gpt_neox_layers_6_attention_dense_q_scale1) 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[114] gpt_neox_layers_7_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[115] gv523: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc398: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv523, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc397, gpt_neox_layers_7_input_layernorm_weight1, gpt_neox_layers_7_input_layernorm_bias1, alloc398) 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_q_weight1: R.Tensor((2304, 96), dtype="uint32") = packed_params[118] gpt_neox_layers_7_attention_query_key_value_q_scale1: R.Tensor((2304, 24), dtype="float16") = packed_params[119] gpt_neox_layers_7_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[120] gv524: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) alloc399: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv524, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_7_attention_query_key_value_q_weight1, gpt_neox_layers_7_attention_query_key_value_q_scale1, alloc398, gpt_neox_layers_7_attention_query_key_value_bias1, alloc399) R.vm.kill_object(alloc398) R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_q_weight1) R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_q_scale1) R.vm.kill_object(gpt_neox_layers_7_attention_query_key_value_bias1) gv525: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc399, gv525, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc399) gv526: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv526, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape28) 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc400: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv527, R.dtype("float16")) _399: 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, alloc400) R.vm.kill_object(reshape29) gv528: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc400, gv528, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc400) gv529: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape31: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape30, gv529, 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[116] gpt_neox_layers_7_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[117] gv530: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc401: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv530, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc397, gpt_neox_layers_7_post_attention_layernorm_weight1, gpt_neox_layers_7_post_attention_layernorm_bias1, alloc401) 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_q_weight1: R.Tensor((3072, 96), dtype="uint32") = packed_params[124] gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale1: R.Tensor((3072, 24), dtype="float16") = packed_params[125] gpt_neox_layers_7_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[126] gv531: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) alloc402: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv531, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale1, alloc401, gpt_neox_layers_7_mlp_dense_h_to_4h_bias1, alloc402) R.vm.kill_object(alloc401) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight1) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale1) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_h_to_4h_bias1) gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight1: R.Tensor((768, 384), dtype="uint32") = packed_params[127] gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale1: R.Tensor((768, 96), dtype="float16") = packed_params[128] gpt_neox_layers_7_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[129] gv532: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc403: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv532, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale1, alloc402, gpt_neox_layers_7_mlp_dense_4h_to_h_bias1, alloc403) R.vm.kill_object(alloc402) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight1) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale1) R.vm.kill_object(gpt_neox_layers_7_mlp_dense_4h_to_h_bias1) gpt_neox_layers_7_attention_dense_q_weight1: R.Tensor((768, 96), dtype="uint32") = packed_params[121] gpt_neox_layers_7_attention_dense_q_scale1: R.Tensor((768, 24), dtype="float16") = packed_params[122] gpt_neox_layers_7_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[123] gv533: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc404: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv533, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_7_attention_dense_q_weight1, gpt_neox_layers_7_attention_dense_q_scale1, reshape31, gpt_neox_layers_7_attention_dense_bias1, alloc403, alloc397, alloc404) R.vm.kill_object(alloc397) R.vm.kill_object(reshape31) R.vm.kill_object(alloc403) R.vm.kill_object(gpt_neox_layers_7_attention_dense_q_weight1) R.vm.kill_object(gpt_neox_layers_7_attention_dense_q_scale1) 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[130] gpt_neox_layers_8_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[131] gv534: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc405: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv534, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc404, gpt_neox_layers_8_input_layernorm_weight1, gpt_neox_layers_8_input_layernorm_bias1, alloc405) 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_q_weight1: R.Tensor((2304, 96), dtype="uint32") = packed_params[134] gpt_neox_layers_8_attention_query_key_value_q_scale1: R.Tensor((2304, 24), dtype="float16") = packed_params[135] gpt_neox_layers_8_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[136] gv535: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) alloc406: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv535, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_8_attention_query_key_value_q_weight1, gpt_neox_layers_8_attention_query_key_value_q_scale1, alloc405, gpt_neox_layers_8_attention_query_key_value_bias1, alloc406) R.vm.kill_object(alloc405) R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_q_weight1) R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_q_scale1) R.vm.kill_object(gpt_neox_layers_8_attention_query_key_value_bias1) gv536: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc406, gv536, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc406) gv537: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv537, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape32) 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc407: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv538, R.dtype("float16")) _406: 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, alloc407) R.vm.kill_object(reshape33) gv539: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(4), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc407, gv539, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc407) gv540: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape35: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape34, gv540, 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[132] gpt_neox_layers_8_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[133] gv541: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc408: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv541, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc404, gpt_neox_layers_8_post_attention_layernorm_weight1, gpt_neox_layers_8_post_attention_layernorm_bias1, alloc408) 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_q_weight1: R.Tensor((3072, 96), dtype="uint32") = packed_params[140] gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale1: R.Tensor((3072, 24), dtype="float16") = packed_params[141] gpt_neox_layers_8_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[142] gv542: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(3072), sinfo_args=(R.Shape(ndim=3),)) alloc409: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv542, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale1, alloc408, gpt_neox_layers_8_mlp_dense_h_to_4h_bias1, alloc409) R.vm.kill_object(alloc408) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight1) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale1) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_h_to_4h_bias1) gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight1: R.Tensor((768, 384), dtype="uint32") = packed_params[143] gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale1: R.Tensor((768, 96), dtype="float16") = packed_params[144] gpt_neox_layers_8_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[145] gv543: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc410: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv543, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale1, alloc409, gpt_neox_layers_8_mlp_dense_4h_to_h_bias1, alloc410) R.vm.kill_object(alloc409) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight1) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale1) R.vm.kill_object(gpt_neox_layers_8_mlp_dense_4h_to_h_bias1) gpt_neox_layers_8_attention_dense_q_weight1: R.Tensor((768, 96), dtype="uint32") = packed_params[137] gpt_neox_layers_8_attention_dense_q_scale1: R.Tensor((768, 24), dtype="float16") = packed_params[138] gpt_neox_layers_8_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[139] gv544: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc411: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv544, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_8_attention_dense_q_weight1, gpt_neox_layers_8_attention_dense_q_scale1, reshape35, gpt_neox_layers_8_attention_dense_bias1, alloc410, alloc404, alloc411) R.vm.kill_object(alloc404) R.vm.kill_object(reshape35) R.vm.kill_object(alloc410) R.vm.kill_object(gpt_neox_layers_8_attention_dense_q_weight1) R.vm.kill_object(gpt_neox_layers_8_attention_dense_q_scale1) 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[146] gpt_neox_layers_9_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[147] gv545: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc412: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv545, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc411, gpt_neox_layers_9_input_layernorm_weight1, gpt_neox_layers_9_input_layernorm_bias1, alloc412) 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_q_weight1: R.Tensor((2304, 96), dtype="uint32") = packed_params[150] gpt_neox_layers_9_attention_query_key_value_q_scale1: R.Tensor((2304, 24), dtype="float16") = packed_params[151] gpt_neox_layers_9_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[152] gv546: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) alloc413: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv546, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_9_attention_query_key_value_q_weight1, gpt_neox_layers_9_attention_query_key_value_q_scale1, alloc412, gpt_neox_layers_9_attention_query_key_value_bias1, alloc413) R.vm.kill_object(alloc412) R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_q_weight1) R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_q_scale1) R.vm.kill_object(gpt_neox_layers_9_attention_query_key_value_bias1) gv547: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc413, gv547, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc413) gv548: R.Shape(ndim=3) = R.call_packed("vm.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, gv548, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape36) 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc414: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv549, R.dtype("float16")) _413: 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, alloc414) R.vm.kill_object(reshape37) gv550: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc414, gv550, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc414) gv551: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape39: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape38, gv551, 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[148] gpt_neox_layers_9_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[149] gv552: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc415: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv552, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc411, gpt_neox_layers_9_post_attention_layernorm_weight1, gpt_neox_layers_9_post_attention_layernorm_bias1, alloc415) 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_q_weight1: R.Tensor((3072, 96), dtype="uint32") = packed_params[156] gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale1: R.Tensor((3072, 24), dtype="float16") = packed_params[157] gpt_neox_layers_9_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[158] gv553: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc416: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv553, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale1, alloc415, gpt_neox_layers_9_mlp_dense_h_to_4h_bias1, alloc416) R.vm.kill_object(alloc415) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight1) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale1) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_h_to_4h_bias1) gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight1: R.Tensor((768, 384), dtype="uint32") = packed_params[159] gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale1: R.Tensor((768, 96), dtype="float16") = packed_params[160] gpt_neox_layers_9_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[161] gv554: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc417: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv554, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale1, alloc416, gpt_neox_layers_9_mlp_dense_4h_to_h_bias1, alloc417) R.vm.kill_object(alloc416) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight1) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale1) R.vm.kill_object(gpt_neox_layers_9_mlp_dense_4h_to_h_bias1) gpt_neox_layers_9_attention_dense_q_weight1: R.Tensor((768, 96), dtype="uint32") = packed_params[153] gpt_neox_layers_9_attention_dense_q_scale1: R.Tensor((768, 24), dtype="float16") = packed_params[154] gpt_neox_layers_9_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[155] gv555: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc418: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv555, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_9_attention_dense_q_weight1, gpt_neox_layers_9_attention_dense_q_scale1, reshape39, gpt_neox_layers_9_attention_dense_bias1, alloc417, alloc411, alloc418) R.vm.kill_object(alloc411) R.vm.kill_object(reshape39) R.vm.kill_object(alloc417) R.vm.kill_object(gpt_neox_layers_9_attention_dense_q_weight1) R.vm.kill_object(gpt_neox_layers_9_attention_dense_q_scale1) 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[162] gpt_neox_layers_10_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[163] gv556: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc419: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv556, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc418, gpt_neox_layers_10_input_layernorm_weight1, gpt_neox_layers_10_input_layernorm_bias1, alloc419) 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_q_weight1: R.Tensor((2304, 96), dtype="uint32") = packed_params[166] gpt_neox_layers_10_attention_query_key_value_q_scale1: R.Tensor((2304, 24), dtype="float16") = packed_params[167] gpt_neox_layers_10_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[168] gv557: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) alloc420: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv557, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_10_attention_query_key_value_q_weight1, gpt_neox_layers_10_attention_query_key_value_q_scale1, alloc419, gpt_neox_layers_10_attention_query_key_value_bias1, alloc420) R.vm.kill_object(alloc419) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_q_weight1) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_q_scale1) R.vm.kill_object(gpt_neox_layers_10_attention_query_key_value_bias1) gv558: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc420, gv558, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc420) gv559: R.Shape(ndim=3) = R.call_packed("vm.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, gv559, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape40) 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(12), R.prim_value(0), R.prim_value(64), sinfo_args=(R.Shape(ndim=3),)) alloc421: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv560, R.dtype("float16")) _420: 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, alloc421) R.vm.kill_object(reshape41) gv561: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc421, gv561, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc421) gv562: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape43: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape42, gv562, 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[164] gpt_neox_layers_10_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[165] gv563: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc422: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv563, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc418, gpt_neox_layers_10_post_attention_layernorm_weight1, gpt_neox_layers_10_post_attention_layernorm_bias1, alloc422) 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_q_weight1: R.Tensor((3072, 96), dtype="uint32") = packed_params[172] gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale1: R.Tensor((3072, 24), dtype="float16") = packed_params[173] gpt_neox_layers_10_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[174] gv564: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc423: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv564, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale1, alloc422, gpt_neox_layers_10_mlp_dense_h_to_4h_bias1, alloc423) R.vm.kill_object(alloc422) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight1) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale1) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_h_to_4h_bias1) gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight1: R.Tensor((768, 384), dtype="uint32") = packed_params[175] gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale1: R.Tensor((768, 96), dtype="float16") = packed_params[176] gpt_neox_layers_10_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[177] gv565: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc424: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv565, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale1, alloc423, gpt_neox_layers_10_mlp_dense_4h_to_h_bias1, alloc424) R.vm.kill_object(alloc423) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight1) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale1) R.vm.kill_object(gpt_neox_layers_10_mlp_dense_4h_to_h_bias1) gpt_neox_layers_10_attention_dense_q_weight1: R.Tensor((768, 96), dtype="uint32") = packed_params[169] gpt_neox_layers_10_attention_dense_q_scale1: R.Tensor((768, 24), dtype="float16") = packed_params[170] gpt_neox_layers_10_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[171] gv566: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc425: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage27, R.prim_value(0), gv566, R.dtype("float16")) R.vm.kill_object(storage27) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_10_attention_dense_q_weight1, gpt_neox_layers_10_attention_dense_q_scale1, reshape43, gpt_neox_layers_10_attention_dense_bias1, alloc424, alloc418, alloc425) R.vm.kill_object(alloc418) R.vm.kill_object(reshape43) R.vm.kill_object(alloc424) R.vm.kill_object(gpt_neox_layers_10_attention_dense_q_weight1) R.vm.kill_object(gpt_neox_layers_10_attention_dense_q_scale1) 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[178] gpt_neox_layers_11_input_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[179] gv567: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc426: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv567, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc425, gpt_neox_layers_11_input_layernorm_weight1, gpt_neox_layers_11_input_layernorm_bias1, alloc426) 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_q_weight1: R.Tensor((2304, 96), dtype="uint32") = packed_params[182] gpt_neox_layers_11_attention_query_key_value_q_scale1: R.Tensor((2304, 24), dtype="float16") = packed_params[183] gpt_neox_layers_11_attention_query_key_value_bias1: R.Tensor((2304,), dtype="float16") = packed_params[184] gv568: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(2304), sinfo_args=(R.Shape(ndim=3),)) alloc427: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv568, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_11_attention_query_key_value_q_weight1, gpt_neox_layers_11_attention_query_key_value_q_scale1, alloc426, gpt_neox_layers_11_attention_query_key_value_bias1, alloc427) R.vm.kill_object(alloc426) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_q_weight1) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_q_scale1) R.vm.kill_object(gpt_neox_layers_11_attention_query_key_value_bias1) gv569: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc427, gv569, sinfo_args=(R.Tensor((1, seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(alloc427) gv570: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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, gv570, sinfo_args=(R.Tensor((seq_len, 36, 64), dtype="float16"),)) R.vm.kill_object(reshape44) gv571: R.Shape(ndim=3) = R.call_packed("vm.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),)) alloc428: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv571, R.dtype("float16")) _427: 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, alloc428) R.vm.kill_object(reshape45) gv572: R.Shape(ndim=4) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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", alloc428, gv572, sinfo_args=(R.Tensor((1, seq_len, 12, 64), dtype="float16"),)) R.vm.kill_object(alloc428) gv573: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) reshape47: R.Tensor((1, seq_len, 768), dtype="float16") = R.call_packed("vm.builtin.reshape", reshape46, gv573, 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[180] gpt_neox_layers_11_post_attention_layernorm_bias1: R.Tensor((768,), dtype="float16") = packed_params[181] gv574: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(768), sinfo_args=(R.Shape(ndim=3),)) alloc429: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv574, R.dtype("float16")) cls.gpt_neox_q4f16_1_layer_norm1(alloc425, gpt_neox_layers_11_post_attention_layernorm_weight1, gpt_neox_layers_11_post_attention_layernorm_bias1, alloc429) 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_q_weight1: R.Tensor((3072, 96), dtype="uint32") = packed_params[188] gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale1: R.Tensor((3072, 24), dtype="float16") = packed_params[189] gpt_neox_layers_11_mlp_dense_h_to_4h_bias1: R.Tensor((3072,), dtype="float32") = packed_params[190] gv575: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc430: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv575, R.dtype("float16")) cls.gpt_neox_q4f16_1_fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale1, alloc429, gpt_neox_layers_11_mlp_dense_h_to_4h_bias1, alloc430) R.vm.kill_object(alloc429) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight1) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale1) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_h_to_4h_bias1) gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight1: R.Tensor((768, 384), dtype="uint32") = packed_params[191] gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale1: R.Tensor((768, 96), dtype="float16") = packed_params[192] gpt_neox_layers_11_mlp_dense_4h_to_h_bias1: R.Tensor((768,), dtype="float32") = packed_params[193] gv576: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc431: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage26, R.prim_value(0), gv576, R.dtype("float16")) R.vm.kill_object(storage26) cls.gpt_neox_q4f16_1_fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale1, alloc430, gpt_neox_layers_11_mlp_dense_4h_to_h_bias1, alloc431) R.vm.kill_object(alloc430) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight1) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale1) R.vm.kill_object(gpt_neox_layers_11_mlp_dense_4h_to_h_bias1) gpt_neox_layers_11_attention_dense_q_weight1: R.Tensor((768, 96), dtype="uint32") = packed_params[185] gpt_neox_layers_11_attention_dense_q_scale1: R.Tensor((768, 24), dtype="float16") = packed_params[186] gpt_neox_layers_11_attention_dense_bias1: R.Tensor((768,), dtype="float16") = packed_params[187] gv577: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc432: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage28, R.prim_value(0), gv577, R.dtype("float16")) R.vm.kill_object(storage28) cls.gpt_neox_q4f16_1_fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_11_attention_dense_q_weight1, gpt_neox_layers_11_attention_dense_q_scale1, reshape47, gpt_neox_layers_11_attention_dense_bias1, alloc431, alloc425, alloc432) R.vm.kill_object(alloc425) R.vm.kill_object(reshape47) R.vm.kill_object(alloc431) R.vm.kill_object(gpt_neox_layers_11_attention_dense_q_weight1) R.vm.kill_object(gpt_neox_layers_11_attention_dense_q_scale1) R.vm.kill_object(gpt_neox_layers_11_attention_dense_bias1) gpt_neox_final_layer_norm_weight1: R.Tensor((768,), dtype="float16") = packed_params[194] gpt_neox_final_layer_norm_bias1: R.Tensor((768,), dtype="float16") = packed_params[195] gv578: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(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),)) alloc433: R.Tensor(dtype="float16", ndim=3) = R.vm.alloc_tensor(storage25, R.prim_value(0), gv578, R.dtype("float16")) R.vm.kill_object(storage25) cls.gpt_neox_q4f16_1_layer_norm1(alloc432, gpt_neox_final_layer_norm_weight1, gpt_neox_final_layer_norm_bias1, alloc433) R.vm.kill_object(alloc432) R.vm.kill_object(gpt_neox_final_layer_norm_weight1) R.vm.kill_object(gpt_neox_final_layer_norm_bias1) storage29: R.Object = R.vm.alloc_storage(R.shape([1536]), R.prim_value(0), R.dtype("uint8"), R.str("global")) alloc434: R.Tensor((1, 1, 768), dtype="float16") = R.vm.alloc_tensor(storage29, R.prim_value(0), R.shape([1, 1, 768]), R.dtype("float16")) R.vm.kill_object(storage29) cls.gpt_neox_q4f16_1_index(alloc433, alloc434) R.vm.kill_object(alloc433) embed_out_q_weight1: R.Tensor((vocab_size, 96), dtype="uint32") = packed_params[196] embed_out_q_scale1: R.Tensor((vocab_size, 24), dtype="float16") = packed_params[197] gv579: 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(gv579, R.prim_value(0), R.dtype("uint8"), R.str("global")) gv580: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(0), R.prim_value(1), R.prim_value(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), gv580, R.dtype("float32")) R.vm.kill_object(storage30) cls.gpt_neox_q4f16_1_fused_dequantize_fused_NT_matmul14_cast8(embed_out_q_weight1, embed_out_q_scale1, alloc434, alloc435) R.vm.kill_object(alloc434) R.vm.kill_object(embed_out_q_weight1) R.vm.kill_object(embed_out_q_scale1) gv1: R.Tuple(R.Tensor(dtype="float32", ndim=3), R.Object) = alloc435, paged_kv_cache R.vm.kill_object(alloc435) gv581: R.Tensor(dtype="float32", ndim=3) = gv1[0] R.call_packed("vm.builtin.match_shape", gv581, 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_q4f16_1_shape_func1(shape_heap) gv8: 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, gv8, sinfo_args=(R.Tensor((batch_size, vocab_size), dtype="float32"),)) gv9: 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),)) storage5: R.Object = R.vm.alloc_storage(gv9, R.prim_value(0), R.dtype("uint8"), R.str("global")) gv10: 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),)) alloc86: R.Tensor(dtype="float32", ndim=2) = R.vm.alloc_tensor(storage5, R.prim_value(0), gv10, R.dtype("float32")) R.vm.kill_object(storage5) gv11: 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),)) storage6: R.Object = R.vm.alloc_storage(gv11, R.prim_value(0), R.dtype("uint8"), R.str("global")) gv12: 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),)) alloc87: R.Tensor(dtype="float32", ndim=2) = R.vm.alloc_tensor(storage6, R.prim_value(0), gv12, R.dtype("float32")) R.vm.kill_object(storage6) cls.gpt_neox_q4f16_1_chunk_lse(lv, temperature, alloc86, alloc87) lv1: R.Tuple(R.Tensor(dtype="float32", ndim=2), R.Tensor(dtype="float32", ndim=2)) = alloc86, alloc87 gv13: 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),)) storage7: R.Object = R.vm.alloc_storage(gv13, R.prim_value(0), R.dtype("uint8"), R.str("global")) gv14: 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),)) alloc88: R.Tensor(dtype="float32", ndim=2) = R.vm.alloc_tensor(storage7, R.prim_value(0), gv14, R.dtype("float32")) R.vm.kill_object(storage7) cls.gpt_neox_q4f16_1_softmax_with_chunked_sum(lv, temperature, alloc86, alloc87, alloc88) R.vm.kill_object(lv) R.vm.kill_object(alloc86) R.vm.kill_object(alloc87) gv15: R.Shape(ndim=3) = R.call_packed("vm.builtin.make_shape", shape_heap, R.prim_value(3), R.prim_value(1), R.prim_value(0), R.prim_value(0), R.prim_value(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", alloc88, gv15, sinfo_args=(R.Tensor((batch_size, 1, vocab_size), dtype="float32"),)) R.vm.kill_object(alloc88) return gv