# 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 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 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 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 batch_decode_paged_kv(_0: T.int32, Q_handle: T.handle, pages_handle: T.handle, page_table_indptr_handle: T.handle, page_table_values_handle: T.handle, var_length_info: T.handle, k_rope_pos_offset_handle: T.handle, q_rope_position_handle: T.handle, output_handle: T.handle, lse_handle: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) B = T.int32(is_size_var=True) Q = T.match_buffer(Q_handle, (B, 16, 64), "float16") max_num_pages = T.int32(is_size_var=True) pages = T.match_buffer(pages_handle, (max_num_pages, 2, 16, 16, 64), "float16", offset_factor=1) page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) nnz_pages = T.int32(is_size_var=True) page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) length_info = T.match_buffer(var_length_info, (B,), "int32", offset_factor=1) k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) output = T.match_buffer(output_handle, (B, 16, 64), "float16") lse = T.match_buffer(lse_handle, (B, 16)) # with T.block("root"): sm_scale: T.float32 = T.float32(0.18033688011112042) for bx in T.thread_binding(B, thread="blockIdx.x"): for fused_by_bz in T.thread_binding(16, thread="blockIdx.y"): for ty in T.thread_binding(1, thread="threadIdx.y"): for tx in T.thread_binding(16, thread="threadIdx.x"): for tz in T.thread_binding(16, thread="threadIdx.z"): with T.block("attn"): T.reads(page_table_indptr[bx:bx + 2], length_info[bx], q_rope_position[bx], Q[bx, fused_by_bz // 16 + ty + fused_by_bz % 16, tx * 4 - 32:tx * 4 - 32 + 68]) T.writes(output[bx, fused_by_bz % 16 + fused_by_bz // 16 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 16 + fused_by_bz // 16 + ty]) Q_local = T.alloc_buffer((4,), "float16", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") O_allreduce = T.alloc_buffer((16, 1, 64), scope="shared") md_allreduce = T.alloc_buffer((16, 1, 2), scope="shared") S_reduce_local = T.alloc_buffer((1,), scope="local") t0 = T.alloc_buffer((1,), scope="local") S_local = T.alloc_buffer((1,), scope="local") QK_local = T.alloc_buffer((4,), scope="local") V_local = T.alloc_buffer((4,), "float16", scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_prev = T.alloc_buffer((1,), scope="local") other_m = T.alloc_buffer((1,), scope="local") other_d = T.alloc_buffer((1,), scope="local") exp_mprev = T.alloc_buffer((1,), scope="local") exp_otherm = T.alloc_buffer((1,), scope="local") other_o = T.alloc_buffer((4,), scope="local") st_m = T.alloc_buffer((1,), scope="local") st_d = T.alloc_buffer((1,), scope="local") O_local = T.alloc_buffer((4,), scope="local") by: T.int32 = fused_by_bz % 16 bz: T.int32 = fused_by_bz // 16 batch_idx: T.int32 = bx cur_page_indptr_begin: T.int32 = page_table_indptr[batch_idx] cur_page_indptr_end: T.int32 = page_table_indptr[batch_idx + 1] kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[batch_idx], 0) st_m[0] = T.float32(-50000.0) st_d[0] = T.float32(1.0) for vec in T.vectorized(4): O_local[vec] = T.float32(0.0) for vec in T.vectorized(4): freq = T.float32() Q_local[vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", Q[bx, by + bz + ty, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, Q[bx, by + bz + ty, tx * 4 + vec + 32] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) for iterator in range((kv_chunk_len[0] + 15) // 16): tile_start_s: T.int32 = tz + ty tile_start_g: T.int32 = iterator * 16 + tz + ty for j in range(1): with T.block("KV_load"): T.reads() T.writes() row_g: T.int32 = tile_start_g + j if row_g < kv_chunk_len[0]: seq_offset: T.int32 = row_g page_no: T.int32 = page_table_values[cur_page_indptr_begin + seq_offset // 16] page_offset: T.int32 = seq_offset % 16 for vec in T.vectorized(4): freq = T.float32() K_smem[tile_start_s + j, tx * 4 + vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, pages[page_no, 0, by, page_offset, tx * 4 + vec + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[batch_idx] + row_g) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, tx * 4 + vec]) V_smem[tile_start_s + j, tx * 4 + vec] = pages[page_no, 1, by, page_offset, tx * 4 + vec] else: for vec in T.vectorized(4): K_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) V_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) T.tvm_storage_sync("shared") m_prev[0] = st_m[0] for j in range(1): for vec in T.vectorized(4): QK_local[vec] = T.Cast("float32", Q_local[vec]) * T.Cast("float32", K_smem[tz + j, tx * 4 + vec]) * attn_score_scaling_factor * sm_scale S_reduce_local[0] = T.float32(0.0) for vec in T.unroll(4): S_reduce_local[0] = S_reduce_local[0] + QK_local[vec] with T.block("block_cross_thread"): T.reads(S_reduce_local[0]) T.writes(t0[0]) T.attr(T.comm_reducer(lambda x0, y0: x0 + y0, [T.float32(0.0)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))) T.tvm_thread_allreduce(T.uint32(1), S_reduce_local[0], T.bool(True), t0[0], tx) S_local[j] = T.float32(-50000.0) if iterator * 16 + tz + j < kv_chunk_len[0]: S_local[j] = t0[0] st_m[0] = T.max(st_m[0], S_local[j]) o_scale: T.float32 = T.exp2(m_prev[0] - st_m[0]) st_d[0] = st_d[0] * o_scale for j in range(1): S_local[j] = T.exp2(S_local[j] - st_m[0]) st_d[0] = st_d[0] + S_local[j] for j in T.vectorized(4): O_local[j] = O_local[j] * o_scale for j in range(1): for vec in T.vectorized(4): V_local[vec] = V_smem[tz + j, tx * 4 + vec] for vec in T.vectorized(4): O_local[vec] = O_local[vec] + T.Cast("float32", V_local[vec]) * S_local[j] for vec in T.vectorized(4): O_allreduce[tz, ty, tx * 4 + vec] = O_local[vec] md_allreduce[tz, ty, 0] = st_m[0] md_allreduce[tz, ty, 1] = st_d[0] T.tvm_storage_sync("shared") st_m[0] = T.float32(-50000.0) st_d[0] = T.float32(1.0) for vec in T.vectorized(4): O_local[vec] = T.float32(0.0) for j in range(16): m_prev[0] = st_m[0] d_prev[0] = st_d[0] other_m[0] = md_allreduce[j, ty, 0] other_d[0] = md_allreduce[j, ty, 1] for vec in T.vectorized(4): other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] st_m[0] = T.max(st_m[0], other_m[0]) st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) for vec in T.vectorized(4): O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] for vec in T.vectorized(4): O_local[vec] = O_local[vec] / st_d[0] for vec in T.vectorized(4): output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) @T.prim_func def batch_decode_paged_kv_sliding_window(_0: T.int32, Q_handle: T.handle, pages_handle: T.handle, page_table_indptr_handle: T.handle, page_table_values_handle: T.handle, var_length_info: T.handle, k_rope_pos_offset_handle: T.handle, q_rope_position_handle: T.handle, output_handle: T.handle, lse_handle: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) B = T.int32(is_size_var=True) Q = T.match_buffer(Q_handle, (B, 16, 64), "float16") max_num_pages = T.int32(is_size_var=True) pages = T.match_buffer(pages_handle, (max_num_pages, 2, 16, 16, 64), "float16", offset_factor=1) page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) nnz_pages = T.int32(is_size_var=True) page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) length_info = T.match_buffer(var_length_info, (3, B), "int32", offset_factor=1) k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) output = T.match_buffer(output_handle, (B, 16, 64), "float16") lse = T.match_buffer(lse_handle, (B, 16)) # with T.block("root"): sm_scale: T.float32 = T.float32(0.18033688011112042) for bx in T.thread_binding(B, thread="blockIdx.x"): for fused_by_bz in T.thread_binding(16, thread="blockIdx.y"): for ty in T.thread_binding(1, thread="threadIdx.y"): for tx in T.thread_binding(16, thread="threadIdx.x"): for tz in T.thread_binding(16, thread="threadIdx.z"): with T.block("attn"): T.reads(page_table_indptr[bx:bx + 2], length_info[0:3, bx], q_rope_position[bx], Q[bx, fused_by_bz // 16 + ty + fused_by_bz % 16, tx * 4 - 32:tx * 4 - 32 + 68]) T.writes(output[bx, fused_by_bz % 16 + fused_by_bz // 16 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 16 + fused_by_bz // 16 + ty]) Q_local = T.alloc_buffer((4,), "float16", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") O_allreduce = T.alloc_buffer((16, 1, 64), scope="shared") md_allreduce = T.alloc_buffer((16, 1, 2), scope="shared") S_reduce_local = T.alloc_buffer((1,), scope="local") t0 = T.alloc_buffer((1,), scope="local") S_local = T.alloc_buffer((1,), scope="local") QK_local = T.alloc_buffer((4,), scope="local") V_local = T.alloc_buffer((4,), "float16", scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_prev = T.alloc_buffer((1,), scope="local") other_m = T.alloc_buffer((1,), scope="local") other_d = T.alloc_buffer((1,), scope="local") exp_mprev = T.alloc_buffer((1,), scope="local") exp_otherm = T.alloc_buffer((1,), scope="local") other_o = T.alloc_buffer((4,), scope="local") st_m = T.alloc_buffer((1,), scope="local") st_d = T.alloc_buffer((1,), scope="local") O_local = T.alloc_buffer((4,), scope="local") by: T.int32 = fused_by_bz % 16 bz: T.int32 = fused_by_bz // 16 batch_idx: T.int32 = bx cur_page_indptr_begin: T.int32 = page_table_indptr[batch_idx] cur_page_indptr_end: T.int32 = page_table_indptr[batch_idx + 1] kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[0, batch_idx] - length_info[1, batch_idx] + length_info[2, batch_idx], 0) st_m[0] = T.float32(-50000.0) st_d[0] = T.float32(1.0) for vec in T.vectorized(4): O_local[vec] = T.float32(0.0) for vec in T.vectorized(4): freq = T.float32() Q_local[vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", Q[bx, by + bz + ty, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, Q[bx, by + bz + ty, tx * 4 + vec + 32] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) for iterator in range((kv_chunk_len[0] + 15) // 16): tile_start_s: T.int32 = tz + ty tile_start_g: T.int32 = iterator * 16 + tz + ty for j in range(1): with T.block("KV_load"): T.reads() T.writes() row_g: T.int32 = tile_start_g + j if row_g < kv_chunk_len[0]: seq_offset: T.int32 = T.if_then_else(row_g < length_info[2, batch_idx], row_g, row_g - length_info[2, batch_idx] + length_info[1, batch_idx]) page_no: T.int32 = page_table_values[cur_page_indptr_begin + seq_offset // 16] page_offset: T.int32 = seq_offset % 16 for vec in T.vectorized(4): freq = T.float32() K_smem[tile_start_s + j, tx * 4 + vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, pages[page_no, 0, by, page_offset, tx * 4 + vec + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[batch_idx] + row_g) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, tx * 4 + vec]) V_smem[tile_start_s + j, tx * 4 + vec] = pages[page_no, 1, by, page_offset, tx * 4 + vec] else: for vec in T.vectorized(4): K_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) V_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) T.tvm_storage_sync("shared") m_prev[0] = st_m[0] for j in range(1): for vec in T.vectorized(4): QK_local[vec] = T.Cast("float32", Q_local[vec]) * T.Cast("float32", K_smem[tz + j, tx * 4 + vec]) * attn_score_scaling_factor * sm_scale S_reduce_local[0] = T.float32(0.0) for vec in T.unroll(4): S_reduce_local[0] = S_reduce_local[0] + QK_local[vec] with T.block("block_cross_thread"): T.reads(S_reduce_local[0]) T.writes(t0[0]) T.attr(T.comm_reducer(lambda x0, y0: x0 + y0, [T.float32(0.0)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))) T.tvm_thread_allreduce(T.uint32(1), S_reduce_local[0], T.bool(True), t0[0], tx) S_local[j] = T.float32(-50000.0) if iterator * 16 + tz + j < kv_chunk_len[0]: S_local[j] = t0[0] st_m[0] = T.max(st_m[0], S_local[j]) o_scale: T.float32 = T.exp2(m_prev[0] - st_m[0]) st_d[0] = st_d[0] * o_scale for j in range(1): S_local[j] = T.exp2(S_local[j] - st_m[0]) st_d[0] = st_d[0] + S_local[j] for j in T.vectorized(4): O_local[j] = O_local[j] * o_scale for j in range(1): for vec in T.vectorized(4): V_local[vec] = V_smem[tz + j, tx * 4 + vec] for vec in T.vectorized(4): O_local[vec] = O_local[vec] + T.Cast("float32", V_local[vec]) * S_local[j] for vec in T.vectorized(4): O_allreduce[tz, ty, tx * 4 + vec] = O_local[vec] md_allreduce[tz, ty, 0] = st_m[0] md_allreduce[tz, ty, 1] = st_d[0] T.tvm_storage_sync("shared") st_m[0] = T.float32(-50000.0) st_d[0] = T.float32(1.0) for vec in T.vectorized(4): O_local[vec] = T.float32(0.0) for j in range(16): m_prev[0] = st_m[0] d_prev[0] = st_d[0] other_m[0] = md_allreduce[j, ty, 0] other_d[0] = md_allreduce[j, ty, 1] for vec in T.vectorized(4): other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] st_m[0] = T.max(st_m[0], other_m[0]) st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) for vec in T.vectorized(4): O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] for vec in T.vectorized(4): O_local[vec] = O_local[vec] / st_d[0] for vec in T.vectorized(4): output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) @T.prim_func def batch_prefill_paged_kv(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) total_len = T.int32(is_size_var=True) q = T.match_buffer(var_q, (total_len, 16, 64), "float16") batch_size = T.int32(is_size_var=True) q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) max_num_pages = T.int32(is_size_var=True) pages = T.match_buffer(var_pages, (max_num_pages, 2, 16, 16, 64), "float16", offset_factor=1) page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) nnz_pages = T.int32(is_size_var=True) page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) output = T.match_buffer(var_output, (total_len, 16, 64), "float16") lse = T.match_buffer(var_lse, (total_len, 16)) # with T.block("root"): for lbx in T.thread_binding(16, thread="blockIdx.x"): for lby in T.thread_binding(16, thread="blockIdx.y"): for lty in T.thread_binding(4, thread="threadIdx.y"): for ltx in T.thread_binding(32, thread="threadIdx.x"): with T.block("attn"): bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) T.reads() T.writes() tile_id = T.alloc_buffer((1,), "int32", scope="local") batch_idx = T.alloc_buffer((1,), "int32", scope="local") batch_tiles = T.alloc_buffer((1,), "int32", scope="local") batch_rows = T.alloc_buffer((1,), "int32", scope="local") iterator = T.alloc_buffer((1,), "int32", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") S_smem = T.alloc_buffer((32, 16), scope="shared") S_local = T.alloc_buffer((32, 16), scope="local") O_local = T.alloc_buffer((32, 64), scope="local") m_smem = T.alloc_buffer((32,), scope="shared") m_prev_smem = T.alloc_buffer((32,), scope="shared") d_smem = T.alloc_buffer((32,), scope="shared") m_new = T.alloc_buffer((1,), scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_new = T.alloc_buffer((1,), scope="local") tile_id[0] = bx batch_idx[0] = 0 batch_rows[0] = q_indptr[1] - q_indptr[0] batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 while T.tvm_thread_invariant(batch_idx[0] < batch_size): while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: tile_id[0] = tile_id[0] - batch_tiles[0] batch_idx[0] = batch_idx[0] + 1 if batch_idx[0] < batch_size: b_idx: T.int32 = batch_idx[0] batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 if T.tvm_thread_invariant(batch_idx[0] < batch_size): b_idx: T.int32 = batch_idx[0] LH_start: T.int32 = tile_id[0] * 32 q_indptr_val: T.int32 = q_indptr[b_idx] cur_page_indptr_begin: T.int32 = page_indptr[b_idx] cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[b_idx], 0) T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: m_smem[row] = T.float32(-50000.0) d_smem[row] = T.float32(1.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 4): with T.block("O_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) T.reads() T.writes(O_local[i, j]) O_local[i, j] = T.float32(0.0) T.tvm_storage_sync("shared") for li_lj_fused_0 in range(4): for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for li_lj_fused_3 in T.vectorized(4): with T.block("Q_load"): i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) T.reads() T.writes() cur_L: T.int32 = q_indptr_val + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: freq = T.float32() Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) else: Q_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for iterator_1 in range((kv_chunk_len[0] + 15) // 16): L_kv_start: T.int32 = iterator_1 * 16 for lz_ly_fused_0 in range(2): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("K_load"): i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: seq_offset: T.int32 = cur_L page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] page_offset: T.int32 = seq_offset % 16 freq = T.float32() K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, pages[page_no, 0, by, page_offset, j + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, j]) else: K_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for lz_ly_fused_0 in range(2): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("V_load"): i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: seq_offset: T.int32 = cur_L page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] page_offset: T.int32 = seq_offset % 16 V_smem[i, j] = pages[page_no, 1, by, page_offset, j] else: V_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") with T.block(""): T.reads(Q_smem[0:32, 0:64], K_smem[0:16, 0:64]) T.writes(S_local[0:32, 0:16]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(2, 2): with T.block("S_gemm_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) T.reads() T.writes(S_local[i, j]) S_local[i, j] = T.float32(0.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 2, 8): with T.block("S_gemm_update"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) k = T.axis.reduce(64, lk_0 * 8 + lk_1) T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) T.writes(S_local[i, j]) S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.18033688011112042) T.tvm_storage_sync("shared") for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(2, 2): with T.block("S_store"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) T.reads(S_local[i, j]) T.writes(S_smem[i, j]) S_smem[i, j] = S_local[i, j] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: with T.block("update1"): T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) T.writes(m_prev[i], m_new[i], d_new[i]) m_prev[i] = m_smem[row] m_new[i] = m_smem[row] row_: T.int32 = LH_start + row for j in range(16): if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): m_new[i] = T.max(m_new[i], S_smem[row, j]) d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx with T.block("update"): T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) T.writes(S_smem[row, 0:16]) for j in range(16): if row < 32: row_: T.int32 = LH_start + row if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) else: S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: with T.block("update"): T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) for j in range(16): d_new[i] = d_new[i] + S_smem[row, j] m_smem[row] = m_new[i] d_smem[row] = d_new[i] m_prev_smem[row] = m_prev[i] T.tvm_storage_sync("shared") with T.block(""): T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:16], V_smem[0:16, 0:64]) T.writes(O_local[0:32, 0:64]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(4, 4): with T.block("O_gemm_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) T.reads() T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, lk_1, li_1, lj_1 in T.grid(2, 8, 4, 4): with T.block("O_gemm_update"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) k = T.axis.reduce(16, lk_0 * 8 + lk_1) T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 4): with T.block("O_store"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) for li_0 in range(1): for li_1 in T.thread_binding(4, thread="threadIdx.y"): for li_2 in T.thread_binding(32, thread="threadIdx.x"): with T.block("lse_store"): i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) tile_id[0] = tile_id[0] + 16 @T.prim_func def batch_prefill_paged_kv_sliding_window(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) total_len = T.int32(is_size_var=True) q = T.match_buffer(var_q, (total_len, 16, 64), "float16") batch_size = T.int32(is_size_var=True) q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) max_num_pages = T.int32(is_size_var=True) pages = T.match_buffer(var_pages, (max_num_pages, 2, 16, 16, 64), "float16", offset_factor=1) page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) nnz_pages = T.int32(is_size_var=True) page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) length_info = T.match_buffer(var_length_info, (3, batch_size), "int32", offset_factor=1) k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) output = T.match_buffer(var_output, (total_len, 16, 64), "float16") lse = T.match_buffer(var_lse, (total_len, 16)) # with T.block("root"): for lbx in T.thread_binding(16, thread="blockIdx.x"): for lby in T.thread_binding(16, thread="blockIdx.y"): for lty in T.thread_binding(4, thread="threadIdx.y"): for ltx in T.thread_binding(32, thread="threadIdx.x"): with T.block("attn"): bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) T.reads() T.writes() tile_id = T.alloc_buffer((1,), "int32", scope="local") batch_idx = T.alloc_buffer((1,), "int32", scope="local") batch_tiles = T.alloc_buffer((1,), "int32", scope="local") batch_rows = T.alloc_buffer((1,), "int32", scope="local") iterator = T.alloc_buffer((1,), "int32", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") S_smem = T.alloc_buffer((32, 16), scope="shared") S_local = T.alloc_buffer((32, 16), scope="local") O_local = T.alloc_buffer((32, 64), scope="local") m_smem = T.alloc_buffer((32,), scope="shared") m_prev_smem = T.alloc_buffer((32,), scope="shared") d_smem = T.alloc_buffer((32,), scope="shared") m_new = T.alloc_buffer((1,), scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_new = T.alloc_buffer((1,), scope="local") tile_id[0] = bx batch_idx[0] = 0 batch_rows[0] = q_indptr[1] - q_indptr[0] batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 while T.tvm_thread_invariant(batch_idx[0] < batch_size): while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: tile_id[0] = tile_id[0] - batch_tiles[0] batch_idx[0] = batch_idx[0] + 1 if batch_idx[0] < batch_size: b_idx: T.int32 = batch_idx[0] batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 if T.tvm_thread_invariant(batch_idx[0] < batch_size): b_idx: T.int32 = batch_idx[0] LH_start: T.int32 = tile_id[0] * 32 q_indptr_val: T.int32 = q_indptr[b_idx] cur_page_indptr_begin: T.int32 = page_indptr[b_idx] cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[0, b_idx] - length_info[1, b_idx] + length_info[2, b_idx], 0) T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: m_smem[row] = T.float32(-50000.0) d_smem[row] = T.float32(1.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 4): with T.block("O_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) T.reads() T.writes(O_local[i, j]) O_local[i, j] = T.float32(0.0) T.tvm_storage_sync("shared") for li_lj_fused_0 in range(4): for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for li_lj_fused_3 in T.vectorized(4): with T.block("Q_load"): i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) T.reads() T.writes() cur_L: T.int32 = q_indptr_val + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: freq = T.float32() Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) else: Q_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for iterator_1 in range((kv_chunk_len[0] + 15) // 16): L_kv_start: T.int32 = iterator_1 * 16 for lz_ly_fused_0 in range(2): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("K_load"): i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: seq_offset: T.int32 = T.if_then_else(cur_L < length_info[2, b_idx], cur_L, cur_L - length_info[2, b_idx] + length_info[1, b_idx]) page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] page_offset: T.int32 = seq_offset % 16 freq = T.float32() K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, pages[page_no, 0, by, page_offset, j + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, j]) else: K_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for lz_ly_fused_0 in range(2): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("V_load"): i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: seq_offset: T.int32 = T.if_then_else(cur_L < length_info[2, b_idx], cur_L, cur_L - length_info[2, b_idx] + length_info[1, b_idx]) page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] page_offset: T.int32 = seq_offset % 16 V_smem[i, j] = pages[page_no, 1, by, page_offset, j] else: V_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") with T.block(""): T.reads(Q_smem[0:32, 0:64], K_smem[0:16, 0:64]) T.writes(S_local[0:32, 0:16]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(2, 2): with T.block("S_gemm_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) T.reads() T.writes(S_local[i, j]) S_local[i, j] = T.float32(0.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 2, 8): with T.block("S_gemm_update"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) k = T.axis.reduce(64, lk_0 * 8 + lk_1) T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) T.writes(S_local[i, j]) S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.18033688011112042) T.tvm_storage_sync("shared") for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(2, 2): with T.block("S_store"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) T.reads(S_local[i, j]) T.writes(S_smem[i, j]) S_smem[i, j] = S_local[i, j] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: with T.block("update1"): T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) T.writes(m_prev[i], m_new[i], d_new[i]) m_prev[i] = m_smem[row] m_new[i] = m_smem[row] row_: T.int32 = LH_start + row for j in range(16): if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): m_new[i] = T.max(m_new[i], S_smem[row, j]) d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx with T.block("update"): T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) T.writes(S_smem[row, 0:16]) for j in range(16): if row < 32: row_: T.int32 = LH_start + row if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) else: S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: with T.block("update"): T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) for j in range(16): d_new[i] = d_new[i] + S_smem[row, j] m_smem[row] = m_new[i] d_smem[row] = d_new[i] m_prev_smem[row] = m_prev[i] T.tvm_storage_sync("shared") with T.block(""): T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:16], V_smem[0:16, 0:64]) T.writes(O_local[0:32, 0:64]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(4, 4): with T.block("O_gemm_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) T.reads() T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, lk_1, li_1, lj_1 in T.grid(2, 8, 4, 4): with T.block("O_gemm_update"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) k = T.axis.reduce(16, lk_0 * 8 + lk_1) T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 4): with T.block("O_store"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) for li_0 in range(1): for li_1 in T.thread_binding(4, thread="threadIdx.y"): for li_2 in T.thread_binding(32, thread="threadIdx.x"): with T.block("lse_store"): i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) tile_id[0] = tile_id[0] + 16 @T.prim_func def batch_prefill_ragged_kv(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_k_rope_pos_offset: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) qo_len = T.int32(is_size_var=True) q = T.match_buffer(var_q, (qo_len, 16, 64), "float16") batch_size = T.int32(is_size_var=True) q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) kv_len = T.int32(is_size_var=True) k = T.match_buffer(var_k, (kv_len, 16, 64), "float16") v = T.match_buffer(var_v, (kv_len, 16, 64), "float16") kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) output = T.match_buffer(var_output, (qo_len, 16, 64), "float16") lse = T.match_buffer(var_lse, (qo_len, 16)) # with T.block("root"): for lbx in T.thread_binding(8, thread="blockIdx.x"): for lby in T.thread_binding(16, thread="blockIdx.y"): for lty in T.thread_binding(4, thread="threadIdx.y"): for ltx in T.thread_binding(32, thread="threadIdx.x"): with T.block("attn"): bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) T.reads() T.writes() tile_id = T.alloc_buffer((1,), "int32", scope="local") batch_idx = T.alloc_buffer((1,), "int32", scope="local") batch_tiles = T.alloc_buffer((1,), "int32", scope="local") batch_rows = T.alloc_buffer((1,), "int32", scope="local") iterator = T.alloc_buffer((1,), "int32", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") K_smem = T.alloc_buffer((64, 32), "float16", scope="shared") V_smem = T.alloc_buffer((32, 64), "float16", scope="shared") S_smem = T.alloc_buffer((32, 32), scope="shared") S_local = T.alloc_buffer((32, 32), scope="local") O_local = T.alloc_buffer((32, 64), scope="local") m_smem = T.alloc_buffer((32,), scope="shared") m_prev_smem = T.alloc_buffer((32,), scope="shared") d_smem = T.alloc_buffer((32,), scope="shared") m_new = T.alloc_buffer((1,), scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_new = T.alloc_buffer((1,), scope="local") tile_id[0] = bx batch_idx[0] = 0 batch_rows[0] = q_indptr[1] - q_indptr[0] batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 while T.tvm_thread_invariant(batch_idx[0] < batch_size): while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: tile_id[0] = tile_id[0] - batch_tiles[0] batch_idx[0] = batch_idx[0] + 1 if batch_idx[0] < batch_size: b_idx: T.int32 = batch_idx[0] batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 if T.tvm_thread_invariant(batch_idx[0] < batch_size): b_idx: T.int32 = batch_idx[0] q_indptr_val: T.int32 = q_indptr[b_idx] LH_start: T.int32 = tile_id[0] * 32 kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: m_smem[row] = T.float32(-50000.0) d_smem[row] = T.float32(1.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1 in range(4): for lj_1_0 in T.unroll(1): for lj_1_1 in T.vectorized(4): with T.block("O_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1_0 * 4 + lj_1_1) T.reads() T.writes(O_local[i, j]) O_local[i, j] = T.float32(0.0) T.tvm_storage_sync("shared") for li_1_lj_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_1_lj_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_0, lj_0_0 in T.grid(1, 2): for lj_1 in T.vectorized(8): with T.block("Q_load"): i = T.axis.spatial(32, li_0 * 32 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) // 4) j = T.axis.spatial(64, lj_0_0 * 32 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) % 4 * 8 + lj_1) T.reads() T.writes() cur_L: T.int32 = q_indptr_val + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: freq = T.float32() Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) else: Q_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for iterator_1 in range((kv_chunk_len[0] + 31) // 32): L_kv_start: T.int32 = iterator_1 * 32 L_kv_base: T.int32 = kv_indptr[b_idx] for lz_1_ly_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for lz_1_ly_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lz_0, ly_0_0 in T.grid(1, 2): for ly_1 in T.vectorized(8): with T.block("K_load"): i = T.axis.spatial(32, lz_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 4) j = T.axis.spatial(64, ly_0_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 4 * 8 + ly_1) T.reads(kv_chunk_len[0], k_rope_pos_offset[b_idx], k[L_kv_base + L_kv_start + i, by, j - 32:j - 32 + 65]) T.writes(K_smem[j, i]) cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: freq = T.float32() K_smem[j, i] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", k[L_kv_base + cur_L, by, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, k[L_kv_base + cur_L, by, j + 32] * T.float16(-1.0), k[L_kv_base + cur_L, by, j - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), k[L_kv_base + cur_L, by, j]) else: K_smem[j, i] = T.float16(0.0) T.tvm_storage_sync("shared") for lz_1_ly_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for lz_1_ly_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lz_0, ly_0_0 in T.grid(1, 2): for ly_1 in T.vectorized(8): with T.block("V_load"): i = T.axis.spatial(32, lz_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 4) j = T.axis.spatial(64, ly_0_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 4 * 8 + ly_1) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: V_smem[i, j] = v[L_kv_base + cur_L, by, j] else: V_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") with T.block(""): T.reads(Q_smem[0:32, 0:64], K_smem[0:64, 0:32]) T.writes(S_local[0:32, 0:32]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init in T.unroll(2): for lj_1_0_init in T.unroll(1): for lj_1_1_init in T.vectorized(4): with T.block("S_gemm_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_0_init * 4 + lj_1_1_init) T.reads() T.writes(S_local[i, j]) S_local[i, j] = T.float32(0.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0 in range(4): for li_1 in T.unroll(2): for lj_1_0 in T.unroll(1): for lj_1_1 in T.vectorized(4): for lk_1 in range(16): with T.block("S_gemm_update"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1_0 * 4 + lj_1_1) k_1 = T.axis.reduce(64, lk_0 * 16 + lk_1) T.reads(S_local[i, j], Q_smem[i, k_1], K_smem[k_1, j]) T.writes(S_local[i, j]) S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k_1]) * T.Cast("float32", K_smem[k_1, j]) * attn_score_scaling_factor * T.float32(0.18033688011112042) T.tvm_storage_sync("shared") for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1 in range(2): for lj_1_0 in T.unroll(1): for lj_1_1 in T.vectorized(4): with T.block("S_store"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1_0 * 4 + lj_1_1) T.reads(S_local[i, j]) T.writes(S_smem[i, j]) S_smem[i, j] = S_local[i, j] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: with T.block("update1"): T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) T.writes(m_prev[i], m_new[i], d_new[i]) m_prev[i] = m_smem[row] m_new[i] = m_smem[row] row_: T.int32 = LH_start + row for j in range(32): if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): m_new[i] = T.max(m_new[i], S_smem[row, j]) d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx with T.block("update"): T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) T.writes(S_smem[row, 0:32]) for j in range(32): if row < 32: row_: T.int32 = LH_start + row if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) else: S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: with T.block("update"): T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) for j in range(32): d_new[i] = d_new[i] + S_smem[row, j] m_smem[row] = m_new[i] d_smem[row] = d_new[i] m_prev_smem[row] = m_prev[i] T.tvm_storage_sync("shared") with T.block(""): T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:64]) T.writes(O_local[0:32, 0:64]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init in T.unroll(4): for lj_1_0_init in T.unroll(1): for lj_1_1_init in T.vectorized(4): with T.block("O_gemm_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_0_init * 4 + lj_1_1_init) T.reads() T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, lk_1 in T.grid(2, 16): for li_1 in T.unroll(4): for lj_1_0 in T.unroll(1): for lj_1_1 in T.vectorized(4): with T.block("O_gemm_update"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1_0 * 4 + lj_1_1) k_1 = T.axis.reduce(32, lk_0 * 16 + lk_1) T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k_1], V_smem[k_1, j]) T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] + S_smem[i, k_1] * T.Cast("float32", V_smem[k_1, j]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1 in range(4): for lj_1_0 in T.unroll(1): for lj_1_1 in T.vectorized(4): with T.block("O_store"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1_0 * 4 + lj_1_1) T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) for li_0 in range(1): for li_1 in T.thread_binding(4, thread="threadIdx.y"): for li_2 in T.thread_binding(32, thread="threadIdx.x"): with T.block("lse_store"): i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) tile_id[0] = tile_id[0] + 8 @T.prim_func def batch_tree_attn(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_mn_indptr: T.handle, var_mask: T.handle, var_output: T.handle, var_lse: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32, batch_size: T.int32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) qo_len = T.int32(is_size_var=True) q = T.match_buffer(var_q, (qo_len, 16, 64), "float16") q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) kv_len = T.int32(is_size_var=True) k = T.match_buffer(var_k, (kv_len, 16, 64), "float16") v = T.match_buffer(var_v, (kv_len, 16, 64), "float16") kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) mn_indptr = T.match_buffer(var_mn_indptr, (batch_size + 1,), "int32", offset_factor=1) tree_size = T.int32(is_size_var=True) mask = T.match_buffer(var_mask, (tree_size, 2), "int32", offset_factor=1) output = T.match_buffer(var_output, (qo_len, 16, 64), "float16") lse = T.match_buffer(var_lse, (qo_len, 16)) # with T.block("root"): for lbx in T.thread_binding(16, thread="blockIdx.x"): for lby in T.thread_binding(16, thread="blockIdx.y"): for lty in T.thread_binding(4, thread="threadIdx.y"): for ltx in T.thread_binding(32, thread="threadIdx.x"): with T.block("attn"): bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) T.reads() T.writes() tile_id = T.alloc_buffer((1,), "int32", scope="local") batch_idx = T.alloc_buffer((1,), "int32", scope="local") batch_tiles = T.alloc_buffer((1,), "int32", scope="local") batch_rows = T.alloc_buffer((1,), "int32", scope="local") iterator = T.alloc_buffer((1,), "int32", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") K_smem = T.alloc_buffer((32, 64), "float16", scope="shared") V_smem = T.alloc_buffer((32, 64), "float16", scope="shared") S_smem = T.alloc_buffer((32, 32), scope="shared") S_local = T.alloc_buffer((32, 32), scope="local") O_local = T.alloc_buffer((32, 64), scope="local") m_smem = T.alloc_buffer((32,), scope="shared") m_prev_smem = T.alloc_buffer((32,), scope="shared") d_smem = T.alloc_buffer((32,), scope="shared") m_new = T.alloc_buffer((1,), scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_new = T.alloc_buffer((1,), scope="local") tile_id[0] = bx batch_idx[0] = 0 batch_rows[0] = q_indptr[1] - q_indptr[0] batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 while T.tvm_thread_invariant(batch_idx[0] < batch_size): while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: tile_id[0] = tile_id[0] - batch_tiles[0] batch_idx[0] = batch_idx[0] + 1 if batch_idx[0] < batch_size: b_idx: T.int32 = batch_idx[0] batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 if T.tvm_thread_invariant(batch_idx[0] < batch_size): b_idx: T.int32 = batch_idx[0] LH_start: T.int32 = tile_id[0] * 32 q_indptr_val: T.int32 = q_indptr[b_idx] kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: m_smem[row] = T.float32(-50000.0) d_smem[row] = T.float32(1.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 4): with T.block("O_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) T.reads() T.writes(O_local[i, j]) O_local[i, j] = T.float32(0.0) T.tvm_storage_sync("shared") for li_lj_fused_0 in range(4): for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for li_lj_fused_3 in T.vectorized(4): with T.block("Q_load"): i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) T.reads() T.writes() cur_L: T.int32 = q_indptr_val + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: freq = T.float32() Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) else: Q_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for iterator_1 in range((kv_chunk_len[0] + 31) // 32): L_kv_start: T.int32 = iterator_1 * 32 L_kv_base: T.int32 = kv_indptr[b_idx] for lz_ly_fused_0 in range(4): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("KV_load"): i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) T.reads() T.writes() cur_L: T.int32 = L_kv_base + L_kv_start + i if L_kv_start + i < kv_chunk_len[0]: freq = T.float32() K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", k[cur_L, by, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, k[cur_L, by, j + 32] * T.float16(-1.0), k[cur_L, by, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), k[cur_L, by, j]) V_smem[i, j] = v[cur_L, by, j] else: K_smem[i, j] = T.float16(0.0) V_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") with T.block(""): T.reads(Q_smem[0:32, 0:64], K_smem[0:32, 0:64]) T.writes(S_local[0:32, 0:32]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(2, 4): with T.block("S_gemm_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_init) T.reads() T.writes(S_local[i, j]) S_local[i, j] = T.float32(0.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 4, 8): with T.block("S_gemm_update"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) k_1 = T.axis.reduce(64, lk_0 * 8 + lk_1) T.reads(S_local[i, j], Q_smem[i, k_1], K_smem[j, k_1]) T.writes(S_local[i, j]) S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k_1]) * T.Cast("float32", K_smem[j, k_1]) * attn_score_scaling_factor * T.float32(0.18033688011112042) T.tvm_storage_sync("shared") for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(2, 4): with T.block("S_store"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) T.reads(S_local[i, j]) T.writes(S_smem[i, j]) S_smem[i, j] = S_local[i, j] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: with T.block("update1"): T.reads(m_smem[row], kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) T.writes(m_prev[i], m_new[i], d_new[i]) m_prev[i] = m_smem[row] m_new[i] = m_smem[row] row_: T.int32 = LH_start + row for j in range(32): if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): m_new[i] = T.max(m_new[i], S_smem[row, j]) d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx with T.block("update"): T.reads(kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) T.writes(S_smem[row, 0:32]) for j in range(32): if row < 32: row_: T.int32 = LH_start + row if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) else: S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: with T.block("update"): T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) for j in range(32): d_new[i] = d_new[i] + S_smem[row, j] m_smem[row] = m_new[i] d_smem[row] = d_new[i] m_prev_smem[row] = m_prev[i] T.tvm_storage_sync("shared") with T.block(""): T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:64]) T.writes(O_local[0:32, 0:64]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(4, 4): with T.block("O_gemm_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) T.reads() T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, lk_1, li_1, lj_1 in T.grid(4, 8, 4, 4): with T.block("O_gemm_update"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) k_1 = T.axis.reduce(32, lk_0 * 8 + lk_1) T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k_1], V_smem[k_1, j]) T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] + S_smem[i, k_1] * T.Cast("float32", V_smem[k_1, j]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 4): with T.block("O_store"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) for li_0 in range(1): for li_1 in T.thread_binding(4, thread="threadIdx.y"): for li_2 in T.thread_binding(32, thread="threadIdx.x"): with T.block("lse_store"): i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) tile_id[0] = tile_id[0] + 16 @T.prim_func def 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.int64(is_size_var=True), T.int64(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.int64(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(T.int64(1), T.int64(1)): for ax2_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): for ax2_fused_0 in T.serial(T.int64(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(T.int64(4096), ax2_fused_0 * T.int64(256) + ax2_fused_1) T.reads(temperature[v0], A[v0, v1 * T.int64(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 * T.int64(4096) + v2 < vocab_size, T.if_then_else(temperature[v0] > T.float32(1.0000000000000001e-05), A[v0, v1 * T.int64(4096) + v2] / temperature[v0], A[v0, v1 * T.int64(4096) + v2]), T.float32(-340282346638528859811704183484516925440.0))) for ax0, ax1 in T.grid(T.int64(1), T.int64(1)): for ax2_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): for ax2_fused_0 in T.serial(T.int64(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(T.int64(4096), ax2_fused_0 * T.int64(256) + ax2_fused_1) T.reads(temperature[v0], A[v0, v1 * T.int64(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 * T.int64(4096) + v2 < vocab_size, T.Select(temperature[v0] > T.float32(1.0000000000000001e-05), T.exp(T.if_then_else(v1 * T.int64(4096) + v2 < vocab_size, T.if_then_else(temperature[v0] > T.float32(1.0000000000000001e-05), A[v0, v1 * T.int64(4096) + v2] / temperature[v0], A[v0, v1 * T.int64(4096) + v2]), T.float32(-340282346638528859811704183484516925440.0)) - temp_max_shared[v0, v1]), T.Cast("float32", T.if_then_else(v1 * T.int64(4096) + v2 < vocab_size, T.if_then_else(temperature[v0] > T.float32(1.0000000000000001e-05), A[v0, v1 * T.int64(4096) + v2] / temperature[v0], A[v0, v1 * T.int64(4096) + v2]), T.float32(-340282346638528859811704183484516925440.0)) == temp_max_shared[v0, v1])), T.float32(0.0)) for ax2_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): for ax2_0 in T.serial(T.int64(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(T.int64(1), ax2_0 * T.int64(256) + ax2_1) T.where(ax2_0 * T.int64(256) + ax2_1 < T.int64(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 compact_kv_copy(var_pages: T.handle, var_copy_length_indptr: T.handle, var_copy_src_dst_pos: T.handle, batch_size: T.int32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) num_pages = T.int32() pages = T.match_buffer(var_pages, (num_pages, 2, 16, 16, 64), "float16", offset_factor=1) copy_length_indptr = T.match_buffer(var_copy_length_indptr, (batch_size + 1,), "int32", offset_factor=1) total_copy_length = T.int32() copy_src_dst_pos = T.match_buffer(var_copy_src_dst_pos, (2, total_copy_length), "int32", offset_factor=1) with T.block("root"): T.reads() T.writes() for bhd_o in T.thread_binding(batch_size * 4, thread="blockIdx.x"): for bhd_i in T.thread_binding(256, thread="threadIdx.x"): b: T.int32 = (bhd_o * 256 + bhd_i) // 1024 h: T.int32 = (bhd_o * 256 + bhd_i) // 64 % 16 d: T.int32 = (bhd_o * 256 + bhd_i) % 64 if bhd_o * 256 + bhd_i < batch_size * 16 * 64: for i in range(copy_length_indptr[b + 1] - copy_length_indptr[b]): src_pos: T.int32 = copy_src_dst_pos[0, copy_length_indptr[b] + i] dst_pos: T.int32 = copy_src_dst_pos[1, copy_length_indptr[b] + i] pages[dst_pos // 16, 0, h, dst_pos % 16, d] = pages[src_pos // 16, 0, h, src_pos % 16, d] pages[dst_pos // 16, 1, h, dst_pos % 16, d] = pages[src_pos // 16, 1, h, src_pos % 16, d] @T.prim_func def copy_single_page(var_pages: T.handle, src_page_id: T.int64, tgt_page_id: T.int64, copy_length: T.int64): 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.int64() pages = T.match_buffer(var_pages, (num_pages, 2, 16, page_size, 64), "float16", offset_factor=1) # with T.block("root"): for b in T.thread_binding(copy_length * T.int64(4), thread="blockIdx.x"): for t in T.thread_binding(256, thread="threadIdx.x"): with T.block("copy"): vh = T.axis.spatial(16, T.Cast("int32", (b * T.int64(256) + T.Cast("int64", t)) // (copy_length * T.int64(64)))) vp = T.axis.spatial(copy_length, (b * T.int64(256) + T.Cast("int64", t)) % (copy_length * T.int64(64)) // T.int64(64)) vd = T.axis.spatial(64, T.Cast("int32", (b * T.int64(256) + T.Cast("int64", t)) % T.int64(64))) T.where(b * T.int64(256) + T.Cast("int64", t) < copy_length * T.int64(16) * T.int64(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(private=True) def fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_0_attention_query_key_value_q_weight2: T.Buffer((T.int64(3072), T.int64(128)), "uint32"), gpt_neox_layers_0_attention_query_key_value_q_scale2: T.Buffer((T.int64(3072), T.int64(32)), "float16"), layer_norm49: T.Buffer((T.int64(1), T.int64(1), T.int64(1024)), "float16"), gpt_neox_layers_0_attention_query_key_value_bias2: T.Buffer((T.int64(3072),), "float16"), T_add_intermediate_intermediate: T.Buffer((T.int64(1), T.int64(1), T.int64(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((T.int64(1), T.int64(1), T.int64(3072)), "float16", scope="local") NT_matmul_intermediate_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), T.int64(1), T.int64(3072)), "float16", scope="local") NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), T.int64(1), T.int64(3072)), "float16", scope="local") gpt_neox_layers_0_attention_query_key_value_q_weight2_local = T.alloc_buffer((T.int64(3072), T.int64(128)), "uint32", scope="local") for u_fused_ax0_fused_fused_0 in T.thread_binding(T.int64(1536), thread="blockIdx.x"): for u_fused_ax0_fused_fused_1 in T.thread_binding(T.int64(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(T.int64(32), thread="threadIdx.y"): for u_fused_ax0_fused_fused_2_init in range(T.int64(1)): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init in T.vectorized(T.int64(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(T.int64(256), ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init) v0 = T.axis.spatial(T.int64(3072), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] = T.float16(0.0) for ax1_0_fused_ax1_1_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_ax1_fused_0 in range(T.int64(1)): for ax0_ax1_fused_1 in T.vectorized(T.int64(1)): with T.block("gpt_neox_layers_0_attention_query_key_value_q_weight2_local"): v0 = T.axis.spatial(T.int64(3072), u_fused_ax0_fused_fused_0 * T.int64(2) + u_fused_ax0_fused_fused_1) v1 = T.axis.spatial(T.int64(128), ax1_0_fused_ax1_1_fused_0 * T.int64(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(T.int64(1), T.int64(1)): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1 in T.vectorized(T.int64(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(T.int64(256), ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1) v0 = T.axis.spatial(T.int64(3072), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0], layer_norm49[T.int64(0), T.int64(0), vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(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 * T.int64(32) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // T.int64(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 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // T.int64(32)]) T.writes(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] + layer_norm49[T.int64(0), T.int64(0), vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(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 * T.int64(32) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // T.int64(8) + vax1_0_fused_ax1_1_fused_2], T.Cast("uint32", (vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) % T.int64(8) * T.int64(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 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // T.int64(32)]) for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax2_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2_fused_2_1 in T.vectorized(T.int64(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(T.int64(32), ax0) v0 = T.axis.spatial(T.int64(3072), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] = T.float16(0.0) for ax1 in range(T.int64(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(T.int64(3072), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0], NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, T.int64(0), T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] + NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, T.int64(0), T.int64(0), v0] for ax1_fused_2 in range(T.int64(1)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(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(T.int64(32), ax0) v0 = T.axis.spatial(T.int64(3072), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0]) T.writes(NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0]) with T.init(): NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] = T.float16(0.0) NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] + NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] for ax0_fused_0_ax0_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0_fused_2 in range(T.int64(1)): with T.block("T_add"): v0 = T.axis.spatial(T.int64(3072), u_fused_ax0_fused_fused_0 * T.int64(2) + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) T.reads(NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0], gpt_neox_layers_0_attention_query_key_value_bias2[v0]) T.writes(T_add_intermediate_intermediate[T.int64(0), T.int64(0), v0]) T_add_intermediate_intermediate[T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] + gpt_neox_layers_0_attention_query_key_value_bias2[v0] @T.prim_func(private=True) def fused_dequantize1_fused_NT_matmul5_add5(gpt_neox_layers_0_attention_query_key_value_q_weight3: T.Buffer((T.int64(3072), T.int64(128)), "uint32"), gpt_neox_layers_0_attention_query_key_value_q_scale3: T.Buffer((T.int64(3072), T.int64(32)), "float16"), p_layer_norm98: T.handle, gpt_neox_layers_0_attention_query_key_value_bias3: T.Buffer((T.int64(3072),), "float16"), p_output0: T.handle): T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) seq_len = T.int64() layer_norm98 = T.match_buffer(p_layer_norm98, (T.int64(1), seq_len, T.int64(1024)), "float16") T_add_intermediate_intermediate = T.match_buffer(p_output0, (T.int64(1), seq_len, T.int64(3072)), "float16") # with T.block("root"): if T.tvm_thread_invariant(seq_len <= T.int64(2)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(3072), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer((T.int64(1), (seq_len + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(3072)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), (seq_len + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(3072)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), (seq_len + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(3072)), "float16", scope="local") for ax0_0 in T.thread_binding((seq_len + T.int64(1)) // T.int64(2), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(768), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(2), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1_init) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = T.float16(0.0) for ax2_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_attention_query_key_value_q_weight3[v0, v1 // T.int64(8)], gpt_neox_layers_0_attention_query_key_value_q_scale3[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_query_key_value_q_scale3[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(2), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v0, v1], layer_norm98[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] + T.if_then_else(v0 < seq_len, layer_norm98[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused] for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(2)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = T.float16(0.0) for ax1 in range(T.int64(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 + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(2)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax1) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1]) with T.init(): NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = T.float16(0.0) NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] for ax0 in range(T.int64(2)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(seq_len, ax0_0 * T.int64(2) + ax0) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (seq_len + T.int64(1)) // T.int64(2) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(2) + ax0 < seq_len) T.reads(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1], gpt_neox_layers_0_attention_query_key_value_bias3[v1]) T.writes(T_add_intermediate_intermediate[T.int64(0), v0, v1]) T_add_intermediate_intermediate[T.int64(0), v0, v1] = NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] + gpt_neox_layers_0_attention_query_key_value_bias3[v1] else: if T.tvm_thread_invariant(seq_len <= T.int64(8)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(3072), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer((T.int64(1), (seq_len + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(3072)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), (seq_len + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(3072)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), (seq_len + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(3072)), "float16", scope="local") for ax0_0 in T.thread_binding((seq_len + T.int64(3)) // T.int64(4), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(768), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(4), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1_init) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = T.float16(0.0) for ax2_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_attention_query_key_value_q_weight3[v0, v1 // T.int64(8)], gpt_neox_layers_0_attention_query_key_value_q_scale3[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_query_key_value_q_scale3[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(4), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v0, v1], layer_norm98[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] + T.if_then_else(v0 < seq_len, layer_norm98[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused] for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(4)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = T.float16(0.0) for ax1 in range(T.int64(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 + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(4)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax1) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1]) with T.init(): NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = T.float16(0.0) NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] for ax0 in range(T.int64(4)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(seq_len, ax0_0 * T.int64(4) + ax0) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (seq_len + T.int64(3)) // T.int64(4) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(4) + ax0 < seq_len) T.reads(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1], gpt_neox_layers_0_attention_query_key_value_bias3[v1]) T.writes(T_add_intermediate_intermediate[T.int64(0), v0, v1]) T_add_intermediate_intermediate[T.int64(0), v0, v1] = NT_matmul_intermediate_pad_local[T.int64(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((T.int64(1), (seq_len + T.int64(7)) // T.int64(8) * T.int64(8), T.int64(3072)), "float16", scope="local") for ax0_ax1_0_fused in T.thread_binding((seq_len + T.int64(7)) // T.int64(8), thread="blockIdx.y"): for ax2_0 in T.thread_binding(T.int64(12), thread="blockIdx.x"): for ax1_1 in T.thread_binding(T.int64(1), thread="vthread.y"): for ax2_1 in T.thread_binding(T.int64(1), thread="vthread.x"): for ax1_2 in T.thread_binding(T.int64(4), thread="threadIdx.y"): for ax2_2 in T.thread_binding(T.int64(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(T.int64(2), T.int64(1)): for ax2_3_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_init"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((seq_len + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3_init) v2 = T.axis.spatial(T.int64(3072), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0_init * T.int64(8) + ax2_3_1_init) T.reads() T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = T.float16(0.0) for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(T.int64(64), T.int64(16), T.int64(2), T.int64(1)): for ax2_3_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_update"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((seq_len + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3) v2 = T.axis.spatial(T.int64(3072), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0 * T.int64(8) + ax2_3_1) v3 = T.axis.reduce(T.int64(1024), ax3_0 * T.int64(16) + ax3_1) T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], layer_norm98[T.int64(0), v1, v3], gpt_neox_layers_0_attention_query_key_value_q_weight3[v2, v3 // T.int64(8)], gpt_neox_layers_0_attention_query_key_value_q_scale3[v2, v3 // T.int64(32)]) T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] + T.if_then_else(v1 < seq_len, layer_norm98[T.int64(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 // T.int64(8)], T.Cast("uint32", v3 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_query_key_value_q_scale3[v2, v3 // T.int64(32)]) for ax0, ax1, ax2_0_1 in T.grid(T.int64(1), T.int64(2), T.int64(1)): for ax2_1_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_intermediate_reindex_pad_local"): v0 = T.axis.spatial(T.int64(1), ax0) v1 = T.axis.spatial((seq_len + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1) v2 = T.axis.spatial(T.int64(3072), ax2_0 * T.int64(256) + ax2_2 * T.int64(8) + ax2_0_1 * T.int64(8) + ax2_1_1) T.where(ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(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[T.int64(0), v1, v2]) T_add_intermediate_intermediate[T.int64(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(private=True) def fused_dequantize1_fused_NT_matmul_add(gpt_neox_layers_0_attention_query_key_value_q_weight4: T.Buffer((T.int64(3072), T.int64(128)), "uint32"), gpt_neox_layers_0_attention_query_key_value_q_scale4: T.Buffer((T.int64(3072), T.int64(32)), "float16"), p_layer_norm147: T.handle, gpt_neox_layers_0_attention_query_key_value_bias4: T.Buffer((T.int64(3072),), "float16"), p_output0: T.handle): T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) batch_size = T.int64() layer_norm147 = T.match_buffer(p_layer_norm147, (batch_size, T.int64(1), T.int64(1024)), "float16") T_add_intermediate_intermediate = T.match_buffer(p_output0, (batch_size, T.int64(1), T.int64(3072)), "float16") # with T.block("root"): if T.tvm_thread_invariant(batch_size <= T.int64(2)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(3072), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1), T.int64(3072)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), (batch_size + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1), T.int64(3072)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), (batch_size + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1), T.int64(3072)), "float16", scope="local") for ax0_0 in T.thread_binding((batch_size + T.int64(1)) // T.int64(2), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(768), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(2), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1_init) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = T.float16(0.0) for ax2_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_attention_query_key_value_q_weight4[v0, v1 // T.int64(8)], gpt_neox_layers_0_attention_query_key_value_q_scale4[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_query_key_value_q_scale4[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(2), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v1], layer_norm147[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] + T.if_then_else(v0 < batch_size, layer_norm147[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused] for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(2)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(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, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = T.float16(0.0) for ax1 in range(T.int64(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 + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(2)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax1) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1]) with T.init(): NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = T.float16(0.0) NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] for ax0 in range(T.int64(2)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(batch_size, ax0_0 * T.int64(2) + ax0) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (batch_size + T.int64(1)) // T.int64(2) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(2) + ax0 < batch_size) T.reads(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1], gpt_neox_layers_0_attention_query_key_value_bias4[v1]) T.writes(T_add_intermediate_intermediate[v0, T.int64(0), v1]) T_add_intermediate_intermediate[v0, T.int64(0), v1] = NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] + gpt_neox_layers_0_attention_query_key_value_bias4[v1] else: if T.tvm_thread_invariant(batch_size <= T.int64(8)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(3072), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1), T.int64(3072)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), (batch_size + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1), T.int64(3072)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), (batch_size + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1), T.int64(3072)), "float16", scope="local") for ax0_0 in T.thread_binding((batch_size + T.int64(3)) // T.int64(4), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(768), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(4), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1_init) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = T.float16(0.0) for ax2_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_attention_query_key_value_q_weight4[v0, v1 // T.int64(8)], gpt_neox_layers_0_attention_query_key_value_q_scale4[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_query_key_value_q_scale4[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(4), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v1], layer_norm147[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] + T.if_then_else(v0 < batch_size, layer_norm147[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused] for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(4)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(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, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = T.float16(0.0) for ax1 in range(T.int64(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 + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(4)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax1) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1]) with T.init(): NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = T.float16(0.0) NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] for ax0 in range(T.int64(4)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(batch_size, ax0_0 * T.int64(4) + ax0) v1 = T.axis.spatial(T.int64(3072), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (batch_size + T.int64(3)) // T.int64(4) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(4) + ax0 < batch_size) T.reads(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1], gpt_neox_layers_0_attention_query_key_value_bias4[v1]) T.writes(T_add_intermediate_intermediate[v0, T.int64(0), v1]) T_add_intermediate_intermediate[v0, T.int64(0), v1] = NT_matmul_intermediate_pad_local[v0, T.int64(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((T.int64(1), (batch_size + T.int64(7)) // T.int64(8) * T.int64(8), T.int64(3072)), "float16", scope="local") for ax0_ax1_0_fused in T.thread_binding((batch_size + T.int64(7)) // T.int64(8), thread="blockIdx.y"): for ax2_0 in T.thread_binding(T.int64(12), thread="blockIdx.x"): for ax1_1 in T.thread_binding(T.int64(1), thread="vthread.y"): for ax2_1 in T.thread_binding(T.int64(1), thread="vthread.x"): for ax1_2 in T.thread_binding(T.int64(4), thread="threadIdx.y"): for ax2_2 in T.thread_binding(T.int64(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(T.int64(2), T.int64(1)): for ax2_3_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_init"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3_init) v2 = T.axis.spatial(T.int64(3072), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0_init * T.int64(8) + ax2_3_1_init) T.reads() T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = T.float16(0.0) for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(T.int64(64), T.int64(16), T.int64(2), T.int64(1)): for ax2_3_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_update"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3) v2 = T.axis.spatial(T.int64(3072), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0 * T.int64(8) + ax2_3_1) v3 = T.axis.reduce(T.int64(1024), ax3_0 * T.int64(16) + ax3_1) T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], layer_norm147[v1, T.int64(0), v3], gpt_neox_layers_0_attention_query_key_value_q_weight4[v2, v3 // T.int64(8)], gpt_neox_layers_0_attention_query_key_value_q_scale4[v2, v3 // T.int64(32)]) T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] + T.if_then_else(v1 < batch_size, layer_norm147[v1, T.int64(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 // T.int64(8)], T.Cast("uint32", v3 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_query_key_value_q_scale4[v2, v3 // T.int64(32)]) for ax0, ax1, ax2_0_1 in T.grid(T.int64(1), T.int64(2), T.int64(1)): for ax2_1_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_intermediate_reindex_pad_local"): v0 = T.axis.spatial(T.int64(1), ax0) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1) v2 = T.axis.spatial(T.int64(3072), ax2_0 * T.int64(256) + ax2_2 * T.int64(8) + ax2_0_1 * T.int64(8) + ax2_1_1) T.where(ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(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, T.int64(0), v2]) T_add_intermediate_intermediate[v1, T.int64(0), v2] = NT_matmul_intermediate_reindex_pad_local[v0, v1, v2] + gpt_neox_layers_0_attention_query_key_value_bias4[v2] @T.prim_func(private=True) def fused_dequantize2_fused_NT_matmul11_add11_add14_add14(gpt_neox_layers_0_attention_dense_q_weight2: T.Buffer((T.int64(1024), T.int64(128)), "uint32"), gpt_neox_layers_0_attention_dense_q_scale2: T.Buffer((T.int64(1024), T.int64(32)), "float16"), lv293: T.Buffer((T.int64(1), T.int64(1), T.int64(1024)), "float16"), gpt_neox_layers_0_attention_dense_bias2: T.Buffer((T.int64(1024),), "float16"), lv218: T.Buffer((T.int64(1), T.int64(1), T.int64(1024)), "float16"), input_embed: T.Buffer((T.int64(1), T.int64(1), T.int64(1024)), "float16"), T_add_intermediate_1_2_intermediate: T.Buffer((T.int64(1), T.int64(1), T.int64(1024)), "float16")): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) # with T.block("root"): NT_matmul_intermediate_local = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), T.int64(1), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), T.int64(1), T.int64(1024)), "float16", scope="local") gpt_neox_layers_0_attention_dense_q_weight2_local = T.alloc_buffer((T.int64(1024), T.int64(128)), "uint32", scope="local") for u_fused_ax0_fused_fused_0 in T.thread_binding(T.int64(512), thread="blockIdx.x"): for u_fused_ax0_fused_fused_1 in T.thread_binding(T.int64(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(T.int64(32), thread="threadIdx.y"): for u_fused_ax0_fused_fused_2_init in range(T.int64(1)): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init in T.vectorized(T.int64(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(T.int64(256), ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init) v0 = T.axis.spatial(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] = T.float16(0.0) for ax1_0_fused_ax1_1_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_ax1_fused_0 in range(T.int64(1)): for ax0_ax1_fused_1 in T.vectorized(T.int64(1)): with T.block("gpt_neox_layers_0_attention_dense_q_weight2_local"): v0 = T.axis.spatial(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(2) + u_fused_ax0_fused_fused_1) v1 = T.axis.spatial(T.int64(128), ax1_0_fused_ax1_1_fused_0 * T.int64(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(T.int64(1), T.int64(1)): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1 in T.vectorized(T.int64(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(T.int64(256), ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1) v0 = T.axis.spatial(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0], lv293[T.int64(0), T.int64(0), vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(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 * T.int64(32) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // T.int64(8) + vax1_0_fused_ax1_1_fused_2], gpt_neox_layers_0_attention_dense_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // T.int64(32)]) T.writes(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] + lv293[T.int64(0), T.int64(0), vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(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 * T.int64(32) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // T.int64(8) + vax1_0_fused_ax1_1_fused_2], T.Cast("uint32", (vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // T.int64(32)]) for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax2_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2_fused_2_1 in T.vectorized(T.int64(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(T.int64(32), ax0) v0 = T.axis.spatial(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] = T.float16(0.0) for ax1 in range(T.int64(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(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0], NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, T.int64(0), T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] + NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, T.int64(0), T.int64(0), v0] for ax1_fused_2 in range(T.int64(1)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(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(T.int64(32), ax0) v0 = T.axis.spatial(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0]) T.writes(NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0]) with T.init(): NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] = T.float16(0.0) NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] + NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] for ax0_fused_0_ax0_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0_fused_2 in range(T.int64(1)): with T.block("T_add_2"): v0 = T.axis.spatial(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(2) + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) T.reads(lv218[T.int64(0), T.int64(0), v0], NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0], gpt_neox_layers_0_attention_dense_bias2[v0], input_embed[T.int64(0), T.int64(0), v0]) T.writes(T_add_intermediate_1_2_intermediate[T.int64(0), T.int64(0), v0]) T_add_intermediate_1_2_intermediate[T.int64(0), T.int64(0), v0] = lv218[T.int64(0), T.int64(0), v0] + (NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] + gpt_neox_layers_0_attention_dense_bias2[v0]) + input_embed[T.int64(0), T.int64(0), v0] @T.prim_func(private=True) def fused_dequantize2_fused_NT_matmul1_add1_add4_add4(gpt_neox_layers_0_attention_dense_q_weight4: T.Buffer((T.int64(1024), T.int64(128)), "uint32"), gpt_neox_layers_0_attention_dense_q_scale4: T.Buffer((T.int64(1024), T.int64(32)), "float16"), p_reshape291: T.handle, gpt_neox_layers_0_attention_dense_bias4: T.Buffer((T.int64(1024),), "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.int64() reshape291 = T.match_buffer(p_reshape291, (batch_size, T.int64(1), T.int64(1024)), "float16") lv2 = T.match_buffer(p_lv2, (batch_size, T.int64(1), T.int64(1024)), "float16") input_embeds = T.match_buffer(p_input_embeds, (batch_size, T.int64(1), T.int64(1024)), "float16") T_add_intermediate_1_2_intermediate = T.match_buffer(p_output0, (batch_size, T.int64(1), T.int64(1024)), "float16") # with T.block("root"): if T.tvm_thread_invariant(batch_size <= T.int64(2)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(1024), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), (batch_size + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), (batch_size + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1), T.int64(1024)), "float16", scope="local") for ax0_0 in T.thread_binding((batch_size + T.int64(1)) // T.int64(2), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(256), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(2), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1_init) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = T.float16(0.0) for ax2_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_attention_dense_q_weight4[v0, v1 // T.int64(8)], gpt_neox_layers_0_attention_dense_q_scale4[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale4[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(2), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v1], reshape291[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] + T.if_then_else(v0 < batch_size, reshape291[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused] for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(2)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(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, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = T.float16(0.0) for ax1 in range(T.int64(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 + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(2)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1]) with T.init(): NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = T.float16(0.0) NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] for ax0 in range(T.int64(2)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(batch_size, ax0_0 * T.int64(2) + ax0) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (batch_size + T.int64(1)) // T.int64(2) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(2) + ax0 < batch_size) T.reads(lv2[v0, T.int64(0), v1], NT_matmul_intermediate_pad_local[v0, T.int64(0), v1], gpt_neox_layers_0_attention_dense_bias4[v1], input_embeds[v0, T.int64(0), v1]) T.writes(T_add_intermediate_1_2_intermediate[v0, T.int64(0), v1]) T_add_intermediate_1_2_intermediate[v0, T.int64(0), v1] = lv2[v0, T.int64(0), v1] + (NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] + gpt_neox_layers_0_attention_dense_bias4[v1]) + input_embeds[v0, T.int64(0), v1] else: if T.tvm_thread_invariant(batch_size <= T.int64(8)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(1024), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), (batch_size + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), (batch_size + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1), T.int64(1024)), "float16", scope="local") for ax0_0 in T.thread_binding((batch_size + T.int64(3)) // T.int64(4), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(256), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(4), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1_init) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = T.float16(0.0) for ax2_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_attention_dense_q_weight4[v0, v1 // T.int64(8)], gpt_neox_layers_0_attention_dense_q_scale4[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale4[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(4), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v1], reshape291[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] + T.if_then_else(v0 < batch_size, reshape291[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused] for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(4)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(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, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = T.float16(0.0) for ax1 in range(T.int64(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 + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(4)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1]) with T.init(): NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = T.float16(0.0) NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] for ax0 in range(T.int64(4)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(batch_size, ax0_0 * T.int64(4) + ax0) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (batch_size + T.int64(3)) // T.int64(4) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(4) + ax0 < batch_size) T.reads(lv2[v0, T.int64(0), v1], NT_matmul_intermediate_pad_local[v0, T.int64(0), v1], gpt_neox_layers_0_attention_dense_bias4[v1], input_embeds[v0, T.int64(0), v1]) T.writes(T_add_intermediate_1_2_intermediate[v0, T.int64(0), v1]) T_add_intermediate_1_2_intermediate[v0, T.int64(0), v1] = lv2[v0, T.int64(0), v1] + (NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] + gpt_neox_layers_0_attention_dense_bias4[v1]) + input_embeds[v0, T.int64(0), v1] else: with T.block("root"): T.reads() T.writes() NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((T.int64(1), (batch_size + T.int64(7)) // T.int64(8) * T.int64(8), T.int64(1024)), "float16", scope="local") for ax0_ax1_0_fused in T.thread_binding((batch_size + T.int64(7)) // T.int64(8), thread="blockIdx.y"): for ax2_0 in T.thread_binding(T.int64(4), thread="blockIdx.x"): for ax1_1 in T.thread_binding(T.int64(1), thread="vthread.y"): for ax2_1 in T.thread_binding(T.int64(1), thread="vthread.x"): for ax1_2 in T.thread_binding(T.int64(4), thread="threadIdx.y"): for ax2_2 in T.thread_binding(T.int64(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(T.int64(2), T.int64(1)): for ax2_3_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_init"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3_init) v2 = T.axis.spatial(T.int64(1024), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0_init * T.int64(8) + ax2_3_1_init) T.reads() T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = T.float16(0.0) for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(T.int64(64), T.int64(16), T.int64(2), T.int64(1)): for ax2_3_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_update"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3) v2 = T.axis.spatial(T.int64(1024), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0 * T.int64(8) + ax2_3_1) v3 = T.axis.reduce(T.int64(1024), ax3_0 * T.int64(16) + ax3_1) T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], reshape291[v1, T.int64(0), v3], gpt_neox_layers_0_attention_dense_q_weight4[v2, v3 // T.int64(8)], gpt_neox_layers_0_attention_dense_q_scale4[v2, v3 // T.int64(32)]) T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] + T.if_then_else(v1 < batch_size, reshape291[v1, T.int64(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 // T.int64(8)], T.Cast("uint32", v3 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale4[v2, v3 // T.int64(32)]) for ax0, ax1, ax2_0_1 in T.grid(T.int64(1), T.int64(2), T.int64(1)): for ax2_1_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_intermediate_reindex_pad_local"): v0 = T.axis.spatial(T.int64(1), ax0) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1) v2 = T.axis.spatial(T.int64(1024), ax2_0 * T.int64(256) + ax2_2 * T.int64(8) + ax2_0_1 * T.int64(8) + ax2_1_1) T.where(ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1 < batch_size) T.reads(lv2[v1, T.int64(0), v2], NT_matmul_intermediate_reindex_pad_local[v0, v1, v2], gpt_neox_layers_0_attention_dense_bias4[v2], input_embeds[v1, T.int64(0), v2]) T.writes(T_add_intermediate_1_2_intermediate[v1, T.int64(0), v2]) T_add_intermediate_1_2_intermediate[v1, T.int64(0), v2] = lv2[v1, T.int64(0), v2] + (NT_matmul_intermediate_reindex_pad_local[v0, v1, v2] + gpt_neox_layers_0_attention_dense_bias4[v2]) + input_embeds[v1, T.int64(0), v2] @T.prim_func(private=True) def fused_dequantize2_fused_NT_matmul6_add6_add9_add9(gpt_neox_layers_0_attention_dense_q_weight3: T.Buffer((T.int64(1024), T.int64(128)), "uint32"), gpt_neox_layers_0_attention_dense_q_scale3: T.Buffer((T.int64(1024), T.int64(32)), "float16"), p_reshape195: T.handle, gpt_neox_layers_0_attention_dense_bias3: T.Buffer((T.int64(1024),), "float16"), p_lv74: 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.int64() reshape195 = T.match_buffer(p_reshape195, (T.int64(1), seq_len, T.int64(1024)), "float16") lv74 = T.match_buffer(p_lv74, (T.int64(1), seq_len, T.int64(1024)), "float16") input_embeds = T.match_buffer(p_input_embeds, (T.int64(1), seq_len, T.int64(1024)), "float16") T_add_intermediate_1_2_intermediate = T.match_buffer(p_output0, (T.int64(1), seq_len, T.int64(1024)), "float16") # with T.block("root"): if T.tvm_thread_invariant(seq_len <= T.int64(2)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(1024), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer((T.int64(1), (seq_len + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), (seq_len + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), (seq_len + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1024)), "float16", scope="local") for ax0_0 in T.thread_binding((seq_len + T.int64(1)) // T.int64(2), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(256), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(2), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1_init) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = T.float16(0.0) for ax2_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_attention_dense_q_weight3[v0, v1 // T.int64(8)], gpt_neox_layers_0_attention_dense_q_scale3[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale3[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(2), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v0, v1], reshape195[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] + T.if_then_else(v0 < seq_len, reshape195[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused] for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(2)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = T.float16(0.0) for ax1 in range(T.int64(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 + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(2)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1]) with T.init(): NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = T.float16(0.0) NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] for ax0 in range(T.int64(2)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(seq_len, ax0_0 * T.int64(2) + ax0) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (seq_len + T.int64(1)) // T.int64(2) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(2) + ax0 < seq_len) T.reads(lv74[T.int64(0), v0, v1], NT_matmul_intermediate_pad_local[T.int64(0), v0, v1], gpt_neox_layers_0_attention_dense_bias3[v1], input_embeds[T.int64(0), v0, v1]) T.writes(T_add_intermediate_1_2_intermediate[T.int64(0), v0, v1]) T_add_intermediate_1_2_intermediate[T.int64(0), v0, v1] = lv74[T.int64(0), v0, v1] + (NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] + gpt_neox_layers_0_attention_dense_bias3[v1]) + input_embeds[T.int64(0), v0, v1] else: if T.tvm_thread_invariant(seq_len <= T.int64(8)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(1024), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer((T.int64(1), (seq_len + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), (seq_len + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), (seq_len + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1024)), "float16", scope="local") for ax0_0 in T.thread_binding((seq_len + T.int64(3)) // T.int64(4), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(256), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(4), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1_init) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = T.float16(0.0) for ax2_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_attention_dense_q_weight3[v0, v1 // T.int64(8)], gpt_neox_layers_0_attention_dense_q_scale3[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale3[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(4), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v0, v1], reshape195[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] + T.if_then_else(v0 < seq_len, reshape195[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0)) * dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused] for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(4)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = T.float16(0.0) for ax1 in range(T.int64(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 + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(4)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1]) with T.init(): NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = T.float16(0.0) NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] for ax0 in range(T.int64(4)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(seq_len, ax0_0 * T.int64(4) + ax0) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (seq_len + T.int64(3)) // T.int64(4) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(4) + ax0 < seq_len) T.reads(lv74[T.int64(0), v0, v1], NT_matmul_intermediate_pad_local[T.int64(0), v0, v1], gpt_neox_layers_0_attention_dense_bias3[v1], input_embeds[T.int64(0), v0, v1]) T.writes(T_add_intermediate_1_2_intermediate[T.int64(0), v0, v1]) T_add_intermediate_1_2_intermediate[T.int64(0), v0, v1] = lv74[T.int64(0), v0, v1] + (NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] + gpt_neox_layers_0_attention_dense_bias3[v1]) + input_embeds[T.int64(0), v0, v1] else: with T.block("root"): T.reads() T.writes() NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((T.int64(1), (seq_len + T.int64(7)) // T.int64(8) * T.int64(8), T.int64(1024)), "float16", scope="local") for ax0_ax1_0_fused in T.thread_binding((seq_len + T.int64(7)) // T.int64(8), thread="blockIdx.y"): for ax2_0 in T.thread_binding(T.int64(4), thread="blockIdx.x"): for ax1_1 in T.thread_binding(T.int64(1), thread="vthread.y"): for ax2_1 in T.thread_binding(T.int64(1), thread="vthread.x"): for ax1_2 in T.thread_binding(T.int64(4), thread="threadIdx.y"): for ax2_2 in T.thread_binding(T.int64(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(T.int64(2), T.int64(1)): for ax2_3_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_init"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((seq_len + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3_init) v2 = T.axis.spatial(T.int64(1024), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0_init * T.int64(8) + ax2_3_1_init) T.reads() T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = T.float16(0.0) for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(T.int64(64), T.int64(16), T.int64(2), T.int64(1)): for ax2_3_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_update"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((seq_len + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3) v2 = T.axis.spatial(T.int64(1024), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0 * T.int64(8) + ax2_3_1) v3 = T.axis.reduce(T.int64(1024), ax3_0 * T.int64(16) + ax3_1) T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], reshape195[T.int64(0), v1, v3], gpt_neox_layers_0_attention_dense_q_weight3[v2, v3 // T.int64(8)], gpt_neox_layers_0_attention_dense_q_scale3[v2, v3 // T.int64(32)]) T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] + T.if_then_else(v1 < seq_len, reshape195[T.int64(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 // T.int64(8)], T.Cast("uint32", v3 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_attention_dense_q_scale3[v2, v3 // T.int64(32)]) for ax0, ax1, ax2_0_1 in T.grid(T.int64(1), T.int64(2), T.int64(1)): for ax2_1_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_intermediate_reindex_pad_local"): v0 = T.axis.spatial(T.int64(1), ax0) v1 = T.axis.spatial((seq_len + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1) v2 = T.axis.spatial(T.int64(1024), ax2_0 * T.int64(256) + ax2_2 * T.int64(8) + ax2_0_1 * T.int64(8) + ax2_1_1) T.where(ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1 < seq_len) T.reads(lv74[T.int64(0), v1, v2], NT_matmul_intermediate_reindex_pad_local[v0, v1, v2], gpt_neox_layers_0_attention_dense_bias3[v2], input_embeds[T.int64(0), v1, v2]) T.writes(T_add_intermediate_1_2_intermediate[T.int64(0), v1, v2]) T_add_intermediate_1_2_intermediate[T.int64(0), v1, v2] = lv74[T.int64(0), v1, v2] + (NT_matmul_intermediate_reindex_pad_local[v0, v1, v2] + gpt_neox_layers_0_attention_dense_bias3[v2]) + input_embeds[T.int64(0), v1, v2] @T.prim_func(private=True) def fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2: T.Buffer((T.int64(4096), T.int64(128)), "uint32"), gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale2: T.Buffer((T.int64(4096), T.int64(32)), "float16"), layer_norm50: T.Buffer((T.int64(1), T.int64(1), T.int64(1024)), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: T.Buffer((T.int64(4096),), "float32"), compute_intermediate_intermediate: T.Buffer((T.int64(1), T.int64(1), T.int64(4096)), "float16")): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) # with T.block("root"): NT_matmul_intermediate_local = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(4096)), scope="local") NT_matmul_intermediate_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), T.int64(1), T.int64(4096)), scope="local") NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), T.int64(1), T.int64(4096)), scope="local") gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2_local = T.alloc_buffer((T.int64(4096), T.int64(128)), "uint32", scope="local") for u_fused_ax0_fused_fused_0 in T.thread_binding(T.int64(2048), thread="blockIdx.x"): for u_fused_ax0_fused_fused_1 in T.thread_binding(T.int64(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(T.int64(32), thread="threadIdx.y"): for u_fused_ax0_fused_fused_2_init in range(T.int64(1)): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init in T.vectorized(T.int64(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(T.int64(256), ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init) v0 = T.axis.spatial(T.int64(4096), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] = T.float32(0.0) for ax1_0_fused_ax1_1_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_ax1_fused_0 in range(T.int64(1)): for ax0_ax1_fused_1 in T.vectorized(T.int64(1)): with T.block("gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2_local"): v0 = T.axis.spatial(T.int64(4096), u_fused_ax0_fused_fused_0 * T.int64(2) + u_fused_ax0_fused_fused_1) v1 = T.axis.spatial(T.int64(128), ax1_0_fused_ax1_1_fused_0 * T.int64(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(T.int64(1), T.int64(1)): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1 in T.vectorized(T.int64(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(T.int64(256), ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1) v0 = T.axis.spatial(T.int64(4096), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0], layer_norm50[T.int64(0), T.int64(0), vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(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 * T.int64(32) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // T.int64(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 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // T.int64(32)]) T.writes(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] + T.Cast("float32", layer_norm50[T.int64(0), T.int64(0), vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(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 * T.int64(32) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // T.int64(8) + vax1_0_fused_ax1_1_fused_2], T.Cast("uint32", (vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) % T.int64(8) * T.int64(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 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // T.int64(32)]) for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax2_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2_fused_2_1 in T.vectorized(T.int64(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(T.int64(32), ax0) v0 = T.axis.spatial(T.int64(4096), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] = T.float32(0.0) for ax1 in range(T.int64(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(T.int64(4096), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0], NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, T.int64(0), T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] + NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, T.int64(0), T.int64(0), v0] for ax1_fused_2 in range(T.int64(1)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(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(T.int64(32), ax0) v0 = T.axis.spatial(T.int64(4096), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0]) T.writes(NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0]) with T.init(): NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] = T.float32(0.0) NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] + NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] for ax0_fused_0_ax0_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0_fused_2 in range(T.int64(1)): with T.block("compute_1_1"): v0 = T.axis.spatial(T.int64(4096), u_fused_ax0_fused_fused_0 * T.int64(2) + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) T.reads(NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0], gpt_neox_layers_0_mlp_dense_h_to_4h_bias2[v0]) T.writes(compute_intermediate_intermediate[T.int64(0), T.int64(0), v0]) compute_intermediate_intermediate[T.int64(0), T.int64(0), v0] = T.Cast("float16", (NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias2[v0]) * (T.float32(0.5) + T.erf((NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias2[v0]) * T.float32(0.70710678118654757)) * T.float32(0.5))) @T.prim_func(private=True) def fused_dequantize3_fused_NT_matmul2_add2_gelu_cast(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4: T.Buffer((T.int64(4096), T.int64(128)), "uint32"), gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4: T.Buffer((T.int64(4096), T.int64(32)), "float16"), p_layer_norm148: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: T.Buffer((T.int64(4096),), "float32"), p_output0: T.handle): T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) batch_size = T.int64() layer_norm148 = T.match_buffer(p_layer_norm148, (batch_size, T.int64(1), T.int64(1024)), "float16") compute_intermediate_intermediate = T.match_buffer(p_output0, (batch_size, T.int64(1), T.int64(4096)), "float16") # with T.block("root"): if T.tvm_thread_invariant(batch_size <= T.int64(2)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(4096), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1), T.int64(4096)), scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), (batch_size + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1), T.int64(4096)), scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), (batch_size + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1), T.int64(4096)), scope="local") for ax0_0 in T.thread_binding((batch_size + T.int64(1)) // T.int64(2), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(1024), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(2), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1_init) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = T.float32(0.0) for ax2_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4[v0, v1 // T.int64(8)], gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(2), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v1], layer_norm148[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] + T.Cast("float32", T.if_then_else(v0 < batch_size, layer_norm148[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(2)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(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, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = T.float32(0.0) for ax1 in range(T.int64(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 + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(2)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax1) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1]) with T.init(): NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = T.float32(0.0) NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] for ax0 in range(T.int64(2)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(batch_size, ax0_0 * T.int64(2) + ax0) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (batch_size + T.int64(1)) // T.int64(2) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(2) + ax0 < batch_size) T.reads(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1], gpt_neox_layers_0_mlp_dense_h_to_4h_bias4[v1]) T.writes(compute_intermediate_intermediate[v0, T.int64(0), v1]) compute_intermediate_intermediate[v0, T.int64(0), v1] = T.Cast("float16", (NT_matmul_intermediate_pad_local[v0, T.int64(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, T.int64(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 <= T.int64(8)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(4096), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1), T.int64(4096)), scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), (batch_size + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1), T.int64(4096)), scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), (batch_size + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1), T.int64(4096)), scope="local") for ax0_0 in T.thread_binding((batch_size + T.int64(3)) // T.int64(4), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(1024), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(4), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1_init) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = T.float32(0.0) for ax2_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4[v0, v1 // T.int64(8)], gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(4), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v1], layer_norm148[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] + T.Cast("float32", T.if_then_else(v0 < batch_size, layer_norm148[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(4)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(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, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = T.float32(0.0) for ax1 in range(T.int64(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 + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(4)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax1) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1]) with T.init(): NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = T.float32(0.0) NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] for ax0 in range(T.int64(4)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(batch_size, ax0_0 * T.int64(4) + ax0) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (batch_size + T.int64(3)) // T.int64(4) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(4) + ax0 < batch_size) T.reads(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1], gpt_neox_layers_0_mlp_dense_h_to_4h_bias4[v1]) T.writes(compute_intermediate_intermediate[v0, T.int64(0), v1]) compute_intermediate_intermediate[v0, T.int64(0), v1] = T.Cast("float16", (NT_matmul_intermediate_pad_local[v0, T.int64(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, T.int64(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((T.int64(1), (batch_size + T.int64(7)) // T.int64(8) * T.int64(8), T.int64(4096)), scope="local") for ax0_ax1_0_fused in T.thread_binding((batch_size + T.int64(7)) // T.int64(8), thread="blockIdx.y"): for ax2_0 in T.thread_binding(T.int64(16), thread="blockIdx.x"): for ax1_1 in T.thread_binding(T.int64(1), thread="vthread.y"): for ax2_1 in T.thread_binding(T.int64(1), thread="vthread.x"): for ax1_2 in T.thread_binding(T.int64(4), thread="threadIdx.y"): for ax2_2 in T.thread_binding(T.int64(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(T.int64(2), T.int64(1)): for ax2_3_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_init"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3_init) v2 = T.axis.spatial(T.int64(4096), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0_init * T.int64(8) + ax2_3_1_init) T.reads() T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = T.float32(0.0) for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(T.int64(64), T.int64(16), T.int64(2), T.int64(1)): for ax2_3_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_update"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3) v2 = T.axis.spatial(T.int64(4096), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0 * T.int64(8) + ax2_3_1) v3 = T.axis.reduce(T.int64(1024), ax3_0 * T.int64(16) + ax3_1) T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], layer_norm148[v1, T.int64(0), v3], gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4[v2, v3 // T.int64(8)], gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4[v2, v3 // T.int64(32)]) T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] + T.Cast("float32", T.if_then_else(v1 < batch_size, layer_norm148[v1, T.int64(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 // T.int64(8)], T.Cast("uint32", v3 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4[v2, v3 // T.int64(32)]) for ax0, ax1, ax2_0_1 in T.grid(T.int64(1), T.int64(2), T.int64(1)): for ax2_1_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_intermediate_reindex_pad_local"): v0 = T.axis.spatial(T.int64(1), ax0) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1) v2 = T.axis.spatial(T.int64(4096), ax2_0 * T.int64(256) + ax2_2 * T.int64(8) + ax2_0_1 * T.int64(8) + ax2_1_1) T.where(ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(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, T.int64(0), v2]) compute_intermediate_intermediate[v1, T.int64(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(private=True) def fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3: T.Buffer((T.int64(4096), T.int64(128)), "uint32"), gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3: T.Buffer((T.int64(4096), T.int64(32)), "float16"), p_layer_norm99: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: T.Buffer((T.int64(4096),), "float32"), p_output0: T.handle): T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) seq_len = T.int64() layer_norm99 = T.match_buffer(p_layer_norm99, (T.int64(1), seq_len, T.int64(1024)), "float16") compute_intermediate_intermediate = T.match_buffer(p_output0, (T.int64(1), seq_len, T.int64(4096)), "float16") # with T.block("root"): if T.tvm_thread_invariant(seq_len <= T.int64(2)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(4096), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer((T.int64(1), (seq_len + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(4096)), scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), (seq_len + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(4096)), scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), (seq_len + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(4096)), scope="local") for ax0_0 in T.thread_binding((seq_len + T.int64(1)) // T.int64(2), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(1024), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(2), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1_init) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = T.float32(0.0) for ax2_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3[v0, v1 // T.int64(8)], gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(2), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v0, v1], layer_norm99[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] + T.Cast("float32", T.if_then_else(v0 < seq_len, layer_norm99[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(2)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = T.float32(0.0) for ax1 in range(T.int64(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 + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(2)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax1) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1]) with T.init(): NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = T.float32(0.0) NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] for ax0 in range(T.int64(2)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(seq_len, ax0_0 * T.int64(2) + ax0) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (seq_len + T.int64(1)) // T.int64(2) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(2) + ax0 < seq_len) T.reads(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1], gpt_neox_layers_0_mlp_dense_h_to_4h_bias3[v1]) T.writes(compute_intermediate_intermediate[T.int64(0), v0, v1]) compute_intermediate_intermediate[T.int64(0), v0, v1] = T.Cast("float16", (NT_matmul_intermediate_pad_local[T.int64(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[T.int64(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 <= T.int64(8)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(4096), T.int64(1024)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer((T.int64(1), (seq_len + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(4096)), scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), (seq_len + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(4096)), scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), (seq_len + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(4096)), scope="local") for ax0_0 in T.thread_binding((seq_len + T.int64(3)) // T.int64(4), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(1024), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(4), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1_init) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = T.float32(0.0) for ax2_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3[v0, v1 // T.int64(8)], gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(4), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v0, v1], layer_norm99[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] + T.Cast("float32", T.if_then_else(v0 < seq_len, layer_norm99[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(4)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = T.float32(0.0) for ax1 in range(T.int64(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 + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(4)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax1) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1]) with T.init(): NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = T.float32(0.0) NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] for ax0 in range(T.int64(4)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(seq_len, ax0_0 * T.int64(4) + ax0) v1 = T.axis.spatial(T.int64(4096), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (seq_len + T.int64(3)) // T.int64(4) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(4) + ax0 < seq_len) T.reads(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1], gpt_neox_layers_0_mlp_dense_h_to_4h_bias3[v1]) T.writes(compute_intermediate_intermediate[T.int64(0), v0, v1]) compute_intermediate_intermediate[T.int64(0), v0, v1] = T.Cast("float16", (NT_matmul_intermediate_pad_local[T.int64(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[T.int64(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((T.int64(1), (seq_len + T.int64(7)) // T.int64(8) * T.int64(8), T.int64(4096)), scope="local") for ax0_ax1_0_fused in T.thread_binding((seq_len + T.int64(7)) // T.int64(8), thread="blockIdx.y"): for ax2_0 in T.thread_binding(T.int64(16), thread="blockIdx.x"): for ax1_1 in T.thread_binding(T.int64(1), thread="vthread.y"): for ax2_1 in T.thread_binding(T.int64(1), thread="vthread.x"): for ax1_2 in T.thread_binding(T.int64(4), thread="threadIdx.y"): for ax2_2 in T.thread_binding(T.int64(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(T.int64(2), T.int64(1)): for ax2_3_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_init"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((seq_len + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3_init) v2 = T.axis.spatial(T.int64(4096), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0_init * T.int64(8) + ax2_3_1_init) T.reads() T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = T.float32(0.0) for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(T.int64(64), T.int64(16), T.int64(2), T.int64(1)): for ax2_3_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_update"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((seq_len + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3) v2 = T.axis.spatial(T.int64(4096), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0 * T.int64(8) + ax2_3_1) v3 = T.axis.reduce(T.int64(1024), ax3_0 * T.int64(16) + ax3_1) T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], layer_norm99[T.int64(0), v1, v3], gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3[v2, v3 // T.int64(8)], gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3[v2, v3 // T.int64(32)]) T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] + T.Cast("float32", T.if_then_else(v1 < seq_len, layer_norm99[T.int64(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 // T.int64(8)], T.Cast("uint32", v3 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3[v2, v3 // T.int64(32)]) for ax0, ax1, ax2_0_1 in T.grid(T.int64(1), T.int64(2), T.int64(1)): for ax2_1_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_intermediate_reindex_pad_local"): v0 = T.axis.spatial(T.int64(1), ax0) v1 = T.axis.spatial((seq_len + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1) v2 = T.axis.spatial(T.int64(4096), ax2_0 * T.int64(256) + ax2_2 * T.int64(8) + ax2_0_1 * T.int64(8) + ax2_1_1) T.where(ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(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[T.int64(0), v1, v2]) compute_intermediate_intermediate[T.int64(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(private=True) def fused_dequantize4_fused_NT_matmul13_add13_cast7(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2: T.Buffer((T.int64(1024), T.int64(512)), "uint32"), gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale2: T.Buffer((T.int64(1024), T.int64(128)), "float16"), lv294: T.Buffer((T.int64(1), T.int64(1), T.int64(4096)), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: T.Buffer((T.int64(1024),), "float32"), compute_intermediate_intermediate: T.Buffer((T.int64(1), T.int64(1), T.int64(1024)), "float16")): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) # with T.block("root"): NT_matmul_intermediate_local = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(1024)), scope="local") NT_matmul_intermediate_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), T.int64(1), T.int64(1024)), scope="local") NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), T.int64(1), T.int64(1024)), scope="local") gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2_local = T.alloc_buffer((T.int64(1024), T.int64(512)), "uint32", scope="local") for u_fused_ax0_fused_fused_0 in T.thread_binding(T.int64(512), thread="blockIdx.x"): for u_fused_ax0_fused_fused_1 in T.thread_binding(T.int64(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(T.int64(32), thread="threadIdx.y"): for u_fused_ax0_fused_fused_2_init in range(T.int64(1)): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init in T.vectorized(T.int64(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(T.int64(256), ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init) v0 = T.axis.spatial(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] = T.float32(0.0) for ax1_0_fused_ax1_1_fused_0 in T.serial(T.int64(16), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_ax1_fused_0 in range(T.int64(1)): for ax0_ax1_fused_1 in T.vectorized(T.int64(1)): with T.block("gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2_local"): v0 = T.axis.spatial(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(2) + u_fused_ax0_fused_fused_1) v1 = T.axis.spatial(T.int64(512), ax1_0_fused_ax1_1_fused_0 * T.int64(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(T.int64(1), T.int64(1)): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1 in T.vectorized(T.int64(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(T.int64(256), ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1) v0 = T.axis.spatial(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0], lv294[T.int64(0), T.int64(0), vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(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 * T.int64(32) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // T.int64(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 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // T.int64(32)]) T.writes(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] + T.Cast("float32", lv294[T.int64(0), T.int64(0), vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(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 * T.int64(32) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // T.int64(8) + vax1_0_fused_ax1_1_fused_2], T.Cast("uint32", (vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) % T.int64(8) * T.int64(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 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // T.int64(32)]) for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax2_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2_fused_2_1 in T.vectorized(T.int64(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(T.int64(32), ax0) v0 = T.axis.spatial(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] = T.float32(0.0) for ax1 in range(T.int64(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(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0], NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, T.int64(0), T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] + NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, T.int64(0), T.int64(0), v0] for ax1_fused_2 in range(T.int64(1)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(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(T.int64(32), ax0) v0 = T.axis.spatial(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0]) T.writes(NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0]) with T.init(): NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] = T.float32(0.0) NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] + NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] for ax0_fused_0_ax0_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0_fused_2 in range(T.int64(1)): with T.block("compute_1"): v0 = T.axis.spatial(T.int64(1024), u_fused_ax0_fused_fused_0 * T.int64(2) + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) T.reads(NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0], gpt_neox_layers_0_mlp_dense_4h_to_h_bias2[v0]) T.writes(compute_intermediate_intermediate[T.int64(0), T.int64(0), v0]) compute_intermediate_intermediate[T.int64(0), T.int64(0), v0] = T.Cast("float16", NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias2[v0]) @T.prim_func(private=True) def fused_dequantize4_fused_NT_matmul3_add3_cast1(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4: T.Buffer((T.int64(1024), T.int64(512)), "uint32"), gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4: T.Buffer((T.int64(1024), T.int64(128)), "float16"), p_lv1: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: T.Buffer((T.int64(1024),), "float32"), p_output0: T.handle): T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) batch_size = T.int64() lv1 = T.match_buffer(p_lv1, (batch_size, T.int64(1), T.int64(4096)), "float16") compute_intermediate_intermediate = T.match_buffer(p_output0, (batch_size, T.int64(1), T.int64(1024)), "float16") # with T.block("root"): if T.tvm_thread_invariant(batch_size <= T.int64(2)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(1024), T.int64(4096)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1), T.int64(1024)), scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), (batch_size + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1), T.int64(1024)), scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), (batch_size + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1), T.int64(1024)), scope="local") for ax0_0 in T.thread_binding((batch_size + T.int64(1)) // T.int64(2), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(256), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(2), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1_init) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = T.float32(0.0) for ax2_fused_0 in T.serial(T.int64(16), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(4096), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4[v0, v1 // T.int64(8)], gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(2), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v1], lv1[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] + T.Cast("float32", T.if_then_else(v0 < batch_size, lv1[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(2)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(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, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = T.float32(0.0) for ax1 in range(T.int64(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 + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(2)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1]) with T.init(): NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = T.float32(0.0) NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] for ax0 in range(T.int64(2)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(batch_size, ax0_0 * T.int64(2) + ax0) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (batch_size + T.int64(1)) // T.int64(2) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(2) + ax0 < batch_size) T.reads(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1], gpt_neox_layers_0_mlp_dense_4h_to_h_bias4[v1]) T.writes(compute_intermediate_intermediate[v0, T.int64(0), v1]) compute_intermediate_intermediate[v0, T.int64(0), v1] = T.Cast("float16", NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias4[v1]) else: if T.tvm_thread_invariant(batch_size <= T.int64(8)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(1024), T.int64(4096)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer(((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1), T.int64(1024)), scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), (batch_size + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1), T.int64(1024)), scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), (batch_size + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1), T.int64(1024)), scope="local") for ax0_0 in T.thread_binding((batch_size + T.int64(3)) // T.int64(4), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(256), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(4), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1_init) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = T.float32(0.0) for ax2_fused_0 in T.serial(T.int64(16), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(4096), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4[v0, v1 // T.int64(8)], gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(4), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v1], lv1[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, v0, T.int64(0), v1] + T.Cast("float32", T.if_then_else(v0 < batch_size, lv1[v0, T.int64(0), vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(4)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(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, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = T.float32(0.0) for ax1 in range(T.int64(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 + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, v0, T.int64(0), v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(4)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((batch_size + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1]) T.writes(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1]) with T.init(): NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = T.float32(0.0) NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] = NT_matmul_intermediate_pad_local[v0, T.int64(0), v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, v0, T.int64(0), v1] for ax0 in range(T.int64(4)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(batch_size, ax0_0 * T.int64(4) + ax0) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (batch_size + T.int64(3)) // T.int64(4) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(4) + ax0 < batch_size) T.reads(NT_matmul_intermediate_pad_local[v0, T.int64(0), v1], gpt_neox_layers_0_mlp_dense_4h_to_h_bias4[v1]) T.writes(compute_intermediate_intermediate[v0, T.int64(0), v1]) compute_intermediate_intermediate[v0, T.int64(0), v1] = T.Cast("float16", NT_matmul_intermediate_pad_local[v0, T.int64(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((T.int64(1), (batch_size + T.int64(7)) // T.int64(8) * T.int64(8), T.int64(1024)), scope="local") for ax0_ax1_0_fused in T.thread_binding((batch_size + T.int64(7)) // T.int64(8), thread="blockIdx.y"): for ax2_0 in T.thread_binding(T.int64(4), thread="blockIdx.x"): for ax1_1 in T.thread_binding(T.int64(1), thread="vthread.y"): for ax2_1 in T.thread_binding(T.int64(1), thread="vthread.x"): for ax1_2 in T.thread_binding(T.int64(4), thread="threadIdx.y"): for ax2_2 in T.thread_binding(T.int64(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(T.int64(2), T.int64(1)): for ax2_3_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_init"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3_init) v2 = T.axis.spatial(T.int64(1024), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0_init * T.int64(8) + ax2_3_1_init) T.reads() T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = T.float32(0.0) for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(T.int64(256), T.int64(16), T.int64(2), T.int64(1)): for ax2_3_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_update"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3) v2 = T.axis.spatial(T.int64(1024), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0 * T.int64(8) + ax2_3_1) v3 = T.axis.reduce(T.int64(4096), ax3_0 * T.int64(16) + ax3_1) T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], lv1[v1, T.int64(0), v3], gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4[v2, v3 // T.int64(8)], gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4[v2, v3 // T.int64(32)]) T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] + T.Cast("float32", T.if_then_else(v1 < batch_size, lv1[v1, T.int64(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 // T.int64(8)], T.Cast("uint32", v3 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4[v2, v3 // T.int64(32)]) for ax0, ax1, ax2_0_1 in T.grid(T.int64(1), T.int64(2), T.int64(1)): for ax2_1_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_intermediate_reindex_pad_local"): v0 = T.axis.spatial(T.int64(1), ax0) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1) v2 = T.axis.spatial(T.int64(1024), ax2_0 * T.int64(256) + ax2_2 * T.int64(8) + ax2_0_1 * T.int64(8) + ax2_1_1) T.where(ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(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, T.int64(0), v2]) compute_intermediate_intermediate[v1, T.int64(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(private=True) def fused_dequantize4_fused_NT_matmul8_add8_cast4(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3: T.Buffer((T.int64(1024), T.int64(512)), "uint32"), gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3: T.Buffer((T.int64(1024), T.int64(128)), "float16"), p_lv98: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: T.Buffer((T.int64(1024),), "float32"), p_output0: T.handle): T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) seq_len = T.int64() lv98 = T.match_buffer(p_lv98, (T.int64(1), seq_len, T.int64(4096)), "float16") compute_intermediate_intermediate = T.match_buffer(p_output0, (T.int64(1), seq_len, T.int64(1024)), "float16") # with T.block("root"): if T.tvm_thread_invariant(seq_len <= T.int64(2)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(1024), T.int64(4096)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer((T.int64(1), (seq_len + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1024)), scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), (seq_len + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1024)), scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), (seq_len + T.int64(1)) // T.int64(2) * T.int64(2), T.int64(1024)), scope="local") for ax0_0 in T.thread_binding((seq_len + T.int64(1)) // T.int64(2), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(256), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(2), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1_init) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = T.float32(0.0) for ax2_fused_0 in T.serial(T.int64(16), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(4096), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3[v0, v1 // T.int64(8)], gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(2), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v0, v1], lv98[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] + T.Cast("float32", T.if_then_else(v0 < seq_len, lv98[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(2)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = T.float32(0.0) for ax1 in range(T.int64(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 + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(2)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(1)) // T.int64(2) * T.int64(2), ax0_0 * T.int64(2) + ax1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1]) with T.init(): NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = T.float32(0.0) NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] for ax0 in range(T.int64(2)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(seq_len, ax0_0 * T.int64(2) + ax0) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (seq_len + T.int64(1)) // T.int64(2) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(2) + ax0 < seq_len) T.reads(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1], gpt_neox_layers_0_mlp_dense_4h_to_h_bias3[v1]) T.writes(compute_intermediate_intermediate[T.int64(0), v0, v1]) compute_intermediate_intermediate[T.int64(0), v0, v1] = T.Cast("float16", NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias3[v1]) else: if T.tvm_thread_invariant(seq_len <= T.int64(8)): with T.block("root"): T.reads() T.writes() dequantize_intermediate_local = T.alloc_buffer((T.int64(1024), T.int64(4096)), "float16", scope="local") NT_matmul_intermediate_pad_local = T.alloc_buffer((T.int64(1), (seq_len + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1024)), scope="local") NT_matmul_intermediate_pad_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), (seq_len + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1024)), scope="local") NT_matmul_intermediate_pad_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), (seq_len + T.int64(3)) // T.int64(4) * T.int64(4), T.int64(1024)), scope="local") for ax0_0 in T.thread_binding((seq_len + T.int64(3)) // T.int64(4), thread="blockIdx.y"): for ax1_fused_0 in T.thread_binding(T.int64(256), thread="blockIdx.x"): for ax1_fused_1 in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax2_fused_1_ax2_fused_3_fused_0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax0_1_init, ax1_fused_2_init in T.grid(T.int64(4), T.int64(2)): for ax2_fused_1_ax2_fused_3_fused_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1_init) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1_init) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax1_fused_2_init) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = T.float32(0.0) for ax2_fused_0 in T.serial(T.int64(16), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_0_1, ax1 in T.grid(T.int64(2), T.int64(8)): for ax0_1 in T.vectorized(T.int64(1)): with T.block("dequantize"): v0 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) v1 = T.axis.spatial(T.int64(4096), ax2_fused_0 * T.int64(256) + ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax1) T.reads(gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3[v0, v1 // T.int64(8)], gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3[v0, v1 // T.int64(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 // T.int64(8)], T.Cast("uint32", v1 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3[v0, v1 // T.int64(32)] for ax0_1, ax1_fused_2, ax2_fused_2 in T.grid(T.int64(4), T.int64(2), T.int64(1)): for ax2_fused_1_ax2_fused_3_fused_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_rf_update"): vax2_fused_1_ax2_fused_3_fused = T.axis.spatial(T.int64(256), ax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + ax2_fused_1_ax2_fused_3_fused_1) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax0_1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(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, T.int64(0), v0, v1], lv98[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) T.writes(NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused, T.int64(0), v0, v1] + T.Cast("float32", T.if_then_else(v0 < seq_len, lv98[T.int64(0), v0, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused], T.float16(0.0))) * T.Cast("float32", dequantize_intermediate_local[v1, vax2_fused_2 * T.int64(8) + vax2_fused_0 * T.int64(256) + vax2_fused_1_ax2_fused_3_fused]) for ax3_fused_0_ax3_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax3_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2 in range(T.int64(4)): for ax3_fused_2_1 in T.vectorized(T.int64(2)): with T.block("NT_matmul_rf_init"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.spatial(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads() T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = T.float32(0.0) for ax1 in range(T.int64(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 + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax2) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax3_fused_0_ax3_fused_1_fused * T.int64(2) + ax3_fused_2_0 * T.int64(2) + ax3_fused_2_1) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1], NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] = NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local[vax2_fused_1_ax2_fused_3_fused_0 * T.int64(8) + vax2_fused_1_ax2_fused_3_fused_1, T.int64(0), v0, v1] for ax2_fused_2, ax1 in T.grid(T.int64(2), T.int64(4)): for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): with T.block("NT_matmul"): vax2_fused_1_ax2_fused_3_fused_0 = T.axis.reduce(T.int64(32), ax0) v0 = T.axis.spatial((seq_len + T.int64(3)) // T.int64(4) * T.int64(4), ax0_0 * T.int64(4) + ax1) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax2_fused_0_ax2_fused_1_fused * T.int64(2) + ax2_fused_2) T.reads(NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1]) T.writes(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1]) with T.init(): NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = T.float32(0.0) NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] = NT_matmul_intermediate_pad_local[T.int64(0), v0, v1] + NT_matmul_intermediate_pad_rf_local_1[vax2_fused_1_ax2_fused_3_fused_0, T.int64(0), v0, v1] for ax0 in range(T.int64(4)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax1_fused_2 in range(T.int64(2)): with T.block("NT_matmul_intermediate_pad"): v0 = T.axis.spatial(seq_len, ax0_0 * T.int64(4) + ax0) v1 = T.axis.spatial(T.int64(1024), ax1_fused_0 * T.int64(4) + ax1_fused_0_ax1_fused_1_fused * T.int64(2) + ax1_fused_2) T.where((ax0_0 - (seq_len + T.int64(3)) // T.int64(4) < T.int64(0) or ax0_0 == T.int64(0)) and ax0_0 * T.int64(4) + ax0 < seq_len) T.reads(NT_matmul_intermediate_pad_local[T.int64(0), v0, v1], gpt_neox_layers_0_mlp_dense_4h_to_h_bias3[v1]) T.writes(compute_intermediate_intermediate[T.int64(0), v0, v1]) compute_intermediate_intermediate[T.int64(0), v0, v1] = T.Cast("float16", NT_matmul_intermediate_pad_local[T.int64(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((T.int64(1), (seq_len + T.int64(7)) // T.int64(8) * T.int64(8), T.int64(1024)), scope="local") for ax0_ax1_0_fused in T.thread_binding((seq_len + T.int64(7)) // T.int64(8), thread="blockIdx.y"): for ax2_0 in T.thread_binding(T.int64(4), thread="blockIdx.x"): for ax1_1 in T.thread_binding(T.int64(1), thread="vthread.y"): for ax2_1 in T.thread_binding(T.int64(1), thread="vthread.x"): for ax1_2 in T.thread_binding(T.int64(4), thread="threadIdx.y"): for ax2_2 in T.thread_binding(T.int64(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(T.int64(2), T.int64(1)): for ax2_3_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_init"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((seq_len + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3_init) v2 = T.axis.spatial(T.int64(1024), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0_init * T.int64(8) + ax2_3_1_init) T.reads() T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = T.float32(0.0) for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(T.int64(256), T.int64(16), T.int64(2), T.int64(1)): for ax2_3_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_update"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((seq_len + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3) v2 = T.axis.spatial(T.int64(1024), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0 * T.int64(8) + ax2_3_1) v3 = T.axis.reduce(T.int64(4096), ax3_0 * T.int64(16) + ax3_1) T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], lv98[T.int64(0), v1, v3], gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3[v2, v3 // T.int64(8)], gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3[v2, v3 // T.int64(32)]) T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] + T.Cast("float32", T.if_then_else(v1 < seq_len, lv98[T.int64(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 // T.int64(8)], T.Cast("uint32", v3 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3[v2, v3 // T.int64(32)]) for ax0, ax1, ax2_0_1 in T.grid(T.int64(1), T.int64(2), T.int64(1)): for ax2_1_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_intermediate_reindex_pad_local"): v0 = T.axis.spatial(T.int64(1), ax0) v1 = T.axis.spatial((seq_len + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1) v2 = T.axis.spatial(T.int64(1024), ax2_0 * T.int64(256) + ax2_2 * T.int64(8) + ax2_0_1 * T.int64(8) + ax2_1_1) T.where(ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(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[T.int64(0), v1, v2]) compute_intermediate_intermediate[T.int64(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(private=True) def fused_dequantize_fused_NT_matmul14_cast8(p_embed_out_q_weight2: T.handle, p_embed_out_q_scale2: T.handle, layer_norm97: T.Buffer((T.int64(1), T.int64(1), T.int64(1024)), "float16"), p_output0: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) vocab_size = T.int64() embed_out_q_weight2 = T.match_buffer(p_embed_out_q_weight2, (vocab_size, T.int64(128)), "uint32") embed_out_q_scale2 = T.match_buffer(p_embed_out_q_scale2, (vocab_size, T.int64(32)), "float16") compute_intermediate_intermediate = T.match_buffer(p_output0, (T.int64(1), T.int64(1), vocab_size)) # with T.block("root"): NT_matmul_intermediate_local = T.alloc_buffer((T.int64(1), T.int64(1), vocab_size), "float16", scope="local") NT_matmul_intermediate_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), T.int64(1), vocab_size), "float16", scope="local") NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), T.int64(1), vocab_size), "float16", scope="local") embed_out_q_weight2_local = T.alloc_buffer((vocab_size, T.int64(128)), "uint32", scope="local") for u_fused_ax0_fused_fused_0 in T.thread_binding((vocab_size + T.int64(1)) // T.int64(2), thread="blockIdx.x"): for u_fused_ax0_fused_fused_1 in T.thread_binding(T.int64(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(T.int64(32), thread="threadIdx.y"): for u_fused_ax0_fused_fused_2_init in range(T.int64(1)): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1_init in T.vectorized(T.int64(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(T.int64(256), ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(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 * T.int64(2) + u_fused_ax0_fused_fused_1 + u_fused_ax0_fused_fused_2_init) T.where(u_fused_ax0_fused_fused_0 * T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] = T.float16(0.0) for ax1_0_fused_ax1_1_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax0_ax1_fused_0 in range(T.int64(1)): for ax0_ax1_fused_1 in T.vectorized(T.int64(1)): with T.block("embed_out_q_weight2_local"): v0 = T.axis.spatial(vocab_size, u_fused_ax0_fused_fused_0 * T.int64(2) + u_fused_ax0_fused_fused_1) v1 = T.axis.spatial(T.int64(128), ax1_0_fused_ax1_1_fused_0 * T.int64(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 * T.int64(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(T.int64(1), T.int64(1)): for ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1 in T.vectorized(T.int64(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(T.int64(256), ax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(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 * T.int64(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 * T.int64(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, T.int64(0), T.int64(0), v0], layer_norm97[T.int64(0), T.int64(0), vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(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 * T.int64(32) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // T.int64(8) + vax1_0_fused_ax1_1_fused_2], embed_out_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // T.int64(32)]) T.writes(NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused, T.int64(0), T.int64(0), v0] + layer_norm97[T.int64(0), T.int64(0), vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(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 * T.int64(32) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused // T.int64(8) + vax1_0_fused_ax1_1_fused_2], T.Cast("uint32", (vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * embed_out_q_scale2[v0, (vax1_0_fused_ax1_1_fused_2 * T.int64(8) + vax1_0_fused_ax1_1_fused_0 * T.int64(256) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused) // T.int64(32)]) for ax2_fused_0_ax2_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(32), thread="threadIdx.y"): for ax2_fused_2_0 in T.serial(T.int64(1), annotations={"pragma_auto_unroll_max_step": 8, "pragma_unroll_explicit": 1}): for ax2_fused_2_1 in T.vectorized(T.int64(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(T.int64(32), ax0) v0 = T.axis.spatial(vocab_size, u_fused_ax0_fused_fused_0 * T.int64(2) + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) T.where(u_fused_ax0_fused_fused_0 * T.int64(2) + (T.Mul(T.int64(0), T.int64(2)) + ax2_fused_0_ax2_fused_1_fused % T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] = T.float16(0.0) for ax1 in range(T.int64(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 * T.int64(2) + ax2_fused_0_ax2_fused_1_fused + ax2_fused_2_0 + ax2_fused_2_1) T.where(u_fused_ax0_fused_fused_0 * T.int64(2) + (T.Mul(T.int64(0), T.int64(2)) + ax2_fused_0_ax2_fused_1_fused % T.int64(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, T.int64(0), T.int64(0), v0], NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, T.int64(0), T.int64(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, T.int64(0), T.int64(0), v0]) NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] + NT_matmul_intermediate_rf_local[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0 * T.int64(8) + vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_1, T.int64(0), T.int64(0), v0] for ax1_fused_2 in range(T.int64(1)): for ax1_fused_0_ax1_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0 in T.thread_binding(T.int64(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(T.int64(32), ax0) v0 = T.axis.spatial(vocab_size, u_fused_ax0_fused_fused_0 * T.int64(2) + ax1_fused_0_ax1_fused_1_fused + ax1_fused_2) T.where(u_fused_ax0_fused_fused_0 * T.int64(2) + (T.Mul(T.int64(0), T.int64(2)) + ax1_fused_0_ax1_fused_1_fused % T.int64(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, T.int64(0), T.int64(0), v0]) T.writes(NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0]) with T.init(): NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] = T.float16(0.0) NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] = NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0] + NT_matmul_intermediate_rf_local_1[vax1_0_fused_ax1_1_fused_1_ax1_0_fused_ax1_1_fused_3_fused_0, T.int64(0), T.int64(0), v0] for ax0_fused_0_ax0_fused_1_fused in T.thread_binding(T.int64(2), thread="threadIdx.x"): for ax0_fused_2 in range(T.int64(1)): with T.block("compute_1"): v0 = T.axis.spatial(vocab_size, u_fused_ax0_fused_fused_0 * T.int64(2) + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) T.where(u_fused_ax0_fused_fused_0 * T.int64(2) + (T.Mul(T.int64(0), T.int64(2)) + ax0_fused_0_ax0_fused_1_fused % T.int64(2) + ax0_fused_2) < vocab_size) T.reads(NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0]) T.writes(compute_intermediate_intermediate[T.int64(0), T.int64(0), v0]) compute_intermediate_intermediate[T.int64(0), T.int64(0), v0] = T.Cast("float32", NT_matmul_intermediate_local[T.int64(0), T.int64(0), v0]) @T.prim_func(private=True) def fused_dequantize_fused_NT_matmul4_cast2(p_embed_out_q_weight4: T.handle, p_embed_out_q_scale4: T.handle, p_layer_norm195: T.handle, p_output0: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) vocab_size = T.int64() embed_out_q_weight4 = T.match_buffer(p_embed_out_q_weight4, (vocab_size, T.int64(128)), "uint32") embed_out_q_scale4 = T.match_buffer(p_embed_out_q_scale4, (vocab_size, T.int64(32)), "float16") batch_size = T.int64() layer_norm195 = T.match_buffer(p_layer_norm195, (batch_size, T.int64(1), T.int64(1024)), "float16") compute_intermediate_intermediate = T.match_buffer(p_output0, (batch_size, T.int64(1), vocab_size)) # with T.block("root"): NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((T.int64(1), (batch_size + T.int64(7)) // T.int64(8) * T.int64(8), (vocab_size + T.int64(255)) // T.int64(256) * T.int64(256)), "float16", scope="local") for ax0_ax1_0_fused in T.thread_binding((batch_size + T.int64(7)) // T.int64(8), thread="blockIdx.y"): for ax2_0 in T.thread_binding((vocab_size + T.int64(255)) // T.int64(256), thread="blockIdx.x"): for ax1_1 in T.thread_binding(T.int64(1), thread="vthread.y"): for ax2_1 in T.thread_binding(T.int64(1), thread="vthread.x"): for ax1_2 in T.thread_binding(T.int64(4), thread="threadIdx.y"): for ax2_2 in T.thread_binding(T.int64(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(T.int64(2), T.int64(1)): for ax2_3_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_init"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3_init) v2 = T.axis.spatial((vocab_size + T.int64(255)) // T.int64(256) * T.int64(256), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0_init * T.int64(8) + ax2_3_1_init) T.reads() T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = T.float16(0.0) for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(T.int64(64), T.int64(16), T.int64(2), T.int64(1)): for ax2_3_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_update"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3) v2 = T.axis.spatial((vocab_size + T.int64(255)) // T.int64(256) * T.int64(256), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0 * T.int64(8) + ax2_3_1) v3 = T.axis.reduce(T.int64(1024), ax3_0 * T.int64(16) + ax3_1) T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], layer_norm195[v1, T.int64(0), v3], embed_out_q_weight4[v2, v3 // T.int64(8)], embed_out_q_scale4[v2, v3 // T.int64(32)]) T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] + T.if_then_else(v1 < batch_size, layer_norm195[v1, T.int64(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 // T.int64(8)], T.Cast("uint32", v3 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * embed_out_q_scale4[v2, v3 // T.int64(32)], T.float16(0.0)) for ax0, ax1, ax2_0_1 in T.grid(T.int64(1), T.int64(2), T.int64(1)): for ax2_1_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_intermediate_reindex_pad_local"): v0 = T.axis.spatial(T.int64(1), ax0) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1) v2 = T.axis.spatial((vocab_size + T.int64(255)) // T.int64(256) * T.int64(256), ax2_0 * T.int64(256) + ax2_2 * T.int64(8) + ax2_0_1 * T.int64(8) + ax2_1_1) T.where(ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1 < batch_size and ax2_0 * T.int64(256) + ax2_2 * T.int64(8) + ax2_0_1 * T.int64(8) + ax2_1_1 < vocab_size) T.reads(NT_matmul_intermediate_reindex_pad_local[v0, v1, v2]) T.writes(compute_intermediate_intermediate[v1, T.int64(0), v2]) compute_intermediate_intermediate[v1, T.int64(0), v2] = T.Cast("float32", NT_matmul_intermediate_reindex_pad_local[v0, v1, v2]) @T.prim_func(private=True) def 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.int64() embed_out_q_weight3 = T.match_buffer(p_embed_out_q_weight3, (vocab_size, T.int64(128)), "uint32") embed_out_q_scale3 = T.match_buffer(p_embed_out_q_scale3, (vocab_size, T.int64(32)), "float16") batch_size = T.int64() take1 = T.match_buffer(p_take1, (T.int64(1), batch_size, T.int64(1024)), "float16") compute_intermediate_intermediate = T.match_buffer(p_output0, (T.int64(1), batch_size, vocab_size)) # with T.block("root"): NT_matmul_intermediate_reindex_pad_local = T.alloc_buffer((T.int64(1), (batch_size + T.int64(7)) // T.int64(8) * T.int64(8), (vocab_size + T.int64(255)) // T.int64(256) * T.int64(256)), "float16", scope="local") for ax0_ax1_0_fused in T.thread_binding((batch_size + T.int64(7)) // T.int64(8), thread="blockIdx.y"): for ax2_0 in T.thread_binding((vocab_size + T.int64(255)) // T.int64(256), thread="blockIdx.x"): for ax1_1 in T.thread_binding(T.int64(1), thread="vthread.y"): for ax2_1 in T.thread_binding(T.int64(1), thread="vthread.x"): for ax1_2 in T.thread_binding(T.int64(4), thread="threadIdx.y"): for ax2_2 in T.thread_binding(T.int64(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(T.int64(2), T.int64(1)): for ax2_3_1_init in T.vectorized(T.int64(8)): with T.block("NT_matmul_init"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3_init) v2 = T.axis.spatial((vocab_size + T.int64(255)) // T.int64(256) * T.int64(256), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0_init * T.int64(8) + ax2_3_1_init) T.reads() T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = T.float16(0.0) for ax3_0, ax3_1, ax1_3, ax2_3_0 in T.grid(T.int64(64), T.int64(16), T.int64(2), T.int64(1)): for ax2_3_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_update"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_1 * T.int64(8) + ax1_2 * T.int64(2) + ax1_3) v2 = T.axis.spatial((vocab_size + T.int64(255)) // T.int64(256) * T.int64(256), ax2_0 * T.int64(256) + ax2_1 * T.int64(256) + ax2_2 * T.int64(8) + ax2_3_0 * T.int64(8) + ax2_3_1) v3 = T.axis.reduce(T.int64(1024), ax3_0 * T.int64(16) + ax3_1) T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], take1[T.int64(0), v1, v3], embed_out_q_weight3[v2, v3 // T.int64(8)], embed_out_q_scale3[v2, v3 // T.int64(32)]) T.writes(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2]) NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] = NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2] + T.if_then_else(v1 < batch_size, take1[T.int64(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 // T.int64(8)], T.Cast("uint32", v3 % T.int64(8) * T.int64(4))), T.uint32(15))) - T.float16(7.0)) * embed_out_q_scale3[v2, v3 // T.int64(32)], T.float16(0.0)) for ax0, ax1, ax2_0_1 in T.grid(T.int64(1), T.int64(2), T.int64(1)): for ax2_1_1 in T.vectorized(T.int64(8)): with T.block("NT_matmul_intermediate_reindex_pad_local"): v0 = T.axis.spatial(T.int64(1), ax0) v1 = T.axis.spatial((batch_size + T.int64(7)) // T.int64(8) * T.int64(8), ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1) v2 = T.axis.spatial((vocab_size + T.int64(255)) // T.int64(256) * T.int64(256), ax2_0 * T.int64(256) + ax2_2 * T.int64(8) + ax2_0_1 * T.int64(8) + ax2_1_1) T.where(ax0_ax1_0_fused * T.int64(8) + ax1_2 * T.int64(2) + ax1 < batch_size and ax2_0 * T.int64(256) + ax2_2 * T.int64(8) + ax2_0_1 * T.int64(8) + ax2_1_1 < vocab_size) T.reads(NT_matmul_intermediate_reindex_pad_local[v0, v1, v2]) T.writes(compute_intermediate_intermediate[T.int64(0), v1, v2]) compute_intermediate_intermediate[T.int64(0), v1, v2] = T.Cast("float32", NT_matmul_intermediate_reindex_pad_local[v0, v1, v2]) @T.prim_func(private=True) def 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, 128), "uint32") gpt_neox_embed_in_q_scale = T.match_buffer(p_gpt_neox_embed_in_q_scale, (vocab_size, 32), "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, 1024), "float16") # with T.block("root"): for ax0_ax1_fused_0 in T.thread_binding(seq_len * 4, thread="blockIdx.x"): for ax0_ax1_fused_1 in T.thread_binding(256, thread="threadIdx.x"): with T.block("T_take"): v0 = T.axis.spatial(seq_len, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) // 1024) v1 = T.axis.spatial(1024, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 1024) T.reads(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(private=True) def fused_reshape10_reshape11(lv124: T.Buffer((T.int64(1), T.int64(16), T.int64(64)), "float16"), T_reshape_intermediate_1: T.Buffer((T.int64(1), T.int64(1), T.int64(1024)), "float16")): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) # with T.block("root"): for ax0_fused_0 in T.thread_binding(T.int64(4), thread="blockIdx.x"): for ax0_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): with T.block("T_reshape_1"): v0 = T.axis.spatial(T.int64(1024), ax0_fused_0 * T.int64(256) + ax0_fused_1) T.reads(lv124[T.int64(0), v0 // T.int64(64), v0 % T.int64(64)]) T.writes(T_reshape_intermediate_1[T.int64(0), T.int64(0), v0]) T_reshape_intermediate_1[T.int64(0), T.int64(0), v0] = lv124[T.int64(0), v0 // T.int64(64), v0 % T.int64(64)] @T.prim_func(private=True) def fused_reshape8_reshape9(add144: T.Buffer((T.int64(1), T.int64(1), T.int64(3072)), "float16"), T_reshape_intermediate_1: T.Buffer((T.int64(1), T.int64(48), T.int64(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(T.int64(12), thread="blockIdx.x"): for ax0_ax1_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): with T.block("T_reshape_1"): v0 = T.axis.spatial(T.int64(48), (ax0_ax1_fused_0 * T.int64(256) + ax0_ax1_fused_1) // T.int64(64)) v1 = T.axis.spatial(T.int64(64), (ax0_ax1_fused_0 * T.int64(256) + ax0_ax1_fused_1) % T.int64(64)) T.reads(add144[T.int64(0), T.int64(0), v0 * T.int64(64) + v1]) T.writes(T_reshape_intermediate_1[T.int64(0), v0, v1]) T_reshape_intermediate_1[T.int64(0), v0, v1] = add144[T.int64(0), T.int64(0), v0 * T.int64(64) + v1] @T.prim_func def fused_rope(var_qkv: T.handle, var_position_map: T.handle, var_q: T.handle, var_k: T.handle, var_v: T.handle, apply_rope: T.int32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) seq_len = T.int32() qkv = T.match_buffer(var_qkv, (seq_len, 48, 64), "float16") position_map = T.match_buffer(var_position_map, (seq_len,), "int32", offset_factor=1) q = T.match_buffer(var_q, (seq_len, 16, 64), "float16") k = T.match_buffer(var_k, (seq_len, 16, 64), "float16") v = T.match_buffer(var_v, (seq_len, 16, 64), "float16") # with T.block("root"): for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 12, thread="blockIdx.x"): for ax0_ax1_ax2_fused_1 in T.thread_binding(256, thread="threadIdx.x"): with T.block("llama_fused_rope"): v0 = T.axis.spatial(seq_len, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) // 3072) v1 = T.axis.spatial(48, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 3072 // 64) v2 = T.axis.spatial(64, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 64) T.reads(position_map[v0], qkv[v0, v1, v2 + -8:v2 + -8 + 17]) T.writes(q[v0, v1, v2], k[v0, v1 + -16, v2], v[v0, v1 + -32, v2]) if v1 < 16: freq = T.float32() q[v0, v1, v2] = T.if_then_else(apply_rope > 0 and v2 < 16, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", qkv[v0, v1, v2]) + T.sin(freq) * T.Cast("float32", T.if_then_else(v2 < 8, qkv[v0, v1, v2 + 8] * T.float16(-1.0), qkv[v0, v1, v2 + -8]))), where={freq: T.Cast("float32", position_map[v0]) / T.pow(T.float32(10000.0), T.Cast("float32", v2 * 2 % 16) / T.float32(16.0))}), qkv[v0, v1, v2]) else: if v1 < 32: freq = T.float32() k[v0, v1 + -16, v2] = T.if_then_else(apply_rope > 0 and v2 < 16, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", qkv[v0, v1, v2]) + T.sin(freq) * T.Cast("float32", T.if_then_else(v2 < 8, qkv[v0, v1, v2 + 8] * T.float16(-1.0), qkv[v0, v1, v2 + -8]))), where={freq: T.Cast("float32", position_map[v0]) / T.pow(T.float32(10000.0), T.Cast("float32", v2 * 2 % 16) / T.float32(16.0))}), qkv[v0, v1, v2]) else: v[v0, v1 + -32, v2] = qkv[v0, v1, v2] @T.prim_func def 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(private=True) def index(var_layer_norm48: T.handle, index: T.Buffer((T.int64(1), T.int64(1), T.int64(1024)), "float16")): T.func_attr({"target": T.target({"keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) seq_len = T.int64() layer_norm48 = T.match_buffer(var_layer_norm48, (T.int64(1), seq_len, T.int64(1024)), "float16") # with T.block("root"): for ax0_fused_0 in T.thread_binding(T.int64(4), thread="blockIdx.x"): for ax0_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): with T.block("index"): v0 = T.axis.spatial(T.int64(1024), ax0_fused_0 * T.int64(256) + ax0_fused_1) T.reads(layer_norm48[T.int64(0), seq_len - T.int64(1), v0]) T.writes(index[T.int64(0), T.int64(0), v0]) index[T.int64(0), T.int64(0), v0] = layer_norm48[T.int64(0), seq_len - T.int64(1), v0] @T.prim_func(private=True) def layer_norm(var_input_embeds: T.handle, gpt_neox_layers_0_input_layernorm_weight4: T.Buffer((T.int64(1024),), "float16"), gpt_neox_layers_0_input_layernorm_bias4: T.Buffer((T.int64(1024),), "float16"), var_T_layer_norm: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) batch_size = T.int64() input_embeds = T.match_buffer(var_input_embeds, (batch_size, T.int64(1), T.int64(1024)), "float16") T_layer_norm = T.match_buffer(var_T_layer_norm, (batch_size, T.int64(1), T.int64(1024)), "float16") # with T.block("root"): input_embeds_red_temp_v0_shared = T.alloc_buffer((batch_size, T.int64(1)), scope="shared") input_embeds_red_temp_v1_shared = T.alloc_buffer((batch_size, T.int64(1)), scope="shared") for ax0_fused in T.thread_binding(batch_size, thread="blockIdx.x"): for ax0 in range(T.int64(1)): for ax1_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): for ax1_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): with T.block("input_embeds_red_temp"): v0 = T.axis.spatial(batch_size, ax0_fused + ax0) v1 = T.axis.reduce(T.int64(1024), ax1_fused_0 * T.int64(256) + ax1_fused_1) T.reads(input_embeds[v0, T.int64(0), v1]) T.writes(input_embeds_red_temp_v0_shared[v0, T.int64(0)], input_embeds_red_temp_v1_shared[v0, T.int64(0)]) with T.init(): input_embeds_red_temp_v0_shared[v0, T.int64(0)] = T.float32(0.0) input_embeds_red_temp_v1_shared[v0, T.int64(0)] = T.float32(0.0) v_input_embeds_red_temp_v0: T.float32 = input_embeds_red_temp_v0_shared[v0, T.int64(0)] + T.Cast("float32", input_embeds[v0, T.int64(0), v1]) v_input_embeds_red_temp_v1: T.float32 = input_embeds_red_temp_v1_shared[v0, T.int64(0)] + T.Cast("float32", input_embeds[v0, T.int64(0), v1]) * T.Cast("float32", input_embeds[v0, T.int64(0), v1]) input_embeds_red_temp_v0_shared[v0, T.int64(0)] = v_input_embeds_red_temp_v0 input_embeds_red_temp_v1_shared[v0, T.int64(0)] = v_input_embeds_red_temp_v1 for ax1_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): for ax1_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): with T.block("T_layer_norm"): v0 = T.axis.spatial(batch_size, ax0_fused) v1 = T.axis.spatial(T.int64(1024), ax1_0 * T.int64(256) + ax1_1) T.reads(input_embeds[v0, T.int64(0), v1], input_embeds_red_temp_v0_shared[v0, T.int64(0)], input_embeds_red_temp_v1_shared[v0, T.int64(0)], gpt_neox_layers_0_input_layernorm_weight4[v1], gpt_neox_layers_0_input_layernorm_bias4[v1]) T.writes(T_layer_norm[v0, T.int64(0), v1]) T_layer_norm[v0, T.int64(0), v1] = T.Cast("float16", (T.Cast("float32", input_embeds[v0, T.int64(0), v1]) - input_embeds_red_temp_v0_shared[v0, T.int64(0)] * T.float32(0.0009765625)) * T.rsqrt(input_embeds_red_temp_v1_shared[v0, T.int64(0)] * T.float32(0.0009765625) - input_embeds_red_temp_v0_shared[v0, T.int64(0)] * T.float32(0.0009765625) * (input_embeds_red_temp_v0_shared[v0, T.int64(0)] * T.float32(0.0009765625)) + T.float32(1.0000000000000001e-05))) * gpt_neox_layers_0_input_layernorm_weight4[v1] + gpt_neox_layers_0_input_layernorm_bias4[v1] @T.prim_func(private=True) def layer_norm1(var_input_embeds: T.handle, gpt_neox_layers_0_input_layernorm_weight3: T.Buffer((T.int64(1024),), "float16"), gpt_neox_layers_0_input_layernorm_bias3: T.Buffer((T.int64(1024),), "float16"), var_T_layer_norm: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) seq_len = T.int64() input_embeds = T.match_buffer(var_input_embeds, (T.int64(1), seq_len, T.int64(1024)), "float16") T_layer_norm = T.match_buffer(var_T_layer_norm, (T.int64(1), seq_len, T.int64(1024)), "float16") # with T.block("root"): input_embeds_red_temp_v0_shared = T.alloc_buffer((T.int64(1), seq_len), scope="shared") input_embeds_red_temp_v1_shared = T.alloc_buffer((T.int64(1), seq_len), scope="shared") for ax0_fused in T.thread_binding(seq_len, thread="blockIdx.x"): for ax0 in range(T.int64(1)): for ax1_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): for ax1_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): with T.block("input_embeds_red_temp"): v0 = T.axis.spatial(seq_len, ax0_fused + ax0) v1 = T.axis.reduce(T.int64(1024), ax1_fused_0 * T.int64(256) + ax1_fused_1) T.reads(input_embeds[T.int64(0), v0, v1]) T.writes(input_embeds_red_temp_v0_shared[T.int64(0), v0], input_embeds_red_temp_v1_shared[T.int64(0), v0]) with T.init(): input_embeds_red_temp_v0_shared[T.int64(0), v0] = T.float32(0.0) input_embeds_red_temp_v1_shared[T.int64(0), v0] = T.float32(0.0) v_input_embeds_red_temp_v0: T.float32 = input_embeds_red_temp_v0_shared[T.int64(0), v0] + T.Cast("float32", input_embeds[T.int64(0), v0, v1]) v_input_embeds_red_temp_v1: T.float32 = input_embeds_red_temp_v1_shared[T.int64(0), v0] + T.Cast("float32", input_embeds[T.int64(0), v0, v1]) * T.Cast("float32", input_embeds[T.int64(0), v0, v1]) input_embeds_red_temp_v0_shared[T.int64(0), v0] = v_input_embeds_red_temp_v0 input_embeds_red_temp_v1_shared[T.int64(0), v0] = v_input_embeds_red_temp_v1 for ax1_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): for ax1_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): with T.block("T_layer_norm"): v0 = T.axis.spatial(seq_len, ax0_fused) v1 = T.axis.spatial(T.int64(1024), ax1_0 * T.int64(256) + ax1_1) T.reads(input_embeds[T.int64(0), v0, v1], input_embeds_red_temp_v0_shared[T.int64(0), v0], input_embeds_red_temp_v1_shared[T.int64(0), v0], gpt_neox_layers_0_input_layernorm_weight3[v1], gpt_neox_layers_0_input_layernorm_bias3[v1]) T.writes(T_layer_norm[T.int64(0), v0, v1]) T_layer_norm[T.int64(0), v0, v1] = T.Cast("float16", (T.Cast("float32", input_embeds[T.int64(0), v0, v1]) - input_embeds_red_temp_v0_shared[T.int64(0), v0] * T.float32(0.0009765625)) * T.rsqrt(input_embeds_red_temp_v1_shared[T.int64(0), v0] * T.float32(0.0009765625) - input_embeds_red_temp_v0_shared[T.int64(0), v0] * T.float32(0.0009765625) * (input_embeds_red_temp_v0_shared[T.int64(0), v0] * T.float32(0.0009765625)) + T.float32(1.0000000000000001e-05))) * gpt_neox_layers_0_input_layernorm_weight3[v1] + gpt_neox_layers_0_input_layernorm_bias3[v1] @T.prim_func(private=True) def layer_norm2(input_embed: T.Buffer((T.int64(1), T.int64(1), T.int64(1024)), "float16"), gpt_neox_layers_0_input_layernorm_weight2: T.Buffer((T.int64(1024),), "float16"), gpt_neox_layers_0_input_layernorm_bias2: T.Buffer((T.int64(1024),), "float16"), T_layer_norm: T.Buffer((T.int64(1), T.int64(1), T.int64(1024)), "float16")): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) # with T.block("root"): input_embed_red_temp_v0_shared = T.alloc_buffer((T.int64(1), T.int64(1)), scope="shared") input_embed_red_temp_v1_shared = T.alloc_buffer((T.int64(1), T.int64(1)), scope="shared") for ax0_fused in T.thread_binding(T.int64(1), thread="blockIdx.x"): for ax0 in range(T.int64(1)): for ax1_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): for ax1_fused_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): with T.block("input_embed_red_temp"): v0 = T.axis.spatial(T.int64(1), ax0) v1 = T.axis.reduce(T.int64(1024), ax1_fused_0 * T.int64(256) + ax1_fused_1) T.reads(input_embed[T.int64(0), T.int64(0), v1]) T.writes(input_embed_red_temp_v0_shared[T.int64(0), T.int64(0)], input_embed_red_temp_v1_shared[T.int64(0), T.int64(0)]) with T.init(): input_embed_red_temp_v0_shared[T.int64(0), T.int64(0)] = T.float32(0.0) input_embed_red_temp_v1_shared[T.int64(0), T.int64(0)] = T.float32(0.0) v_input_embed_red_temp_v0: T.float32 = input_embed_red_temp_v0_shared[T.int64(0), T.int64(0)] + T.Cast("float32", input_embed[T.int64(0), T.int64(0), v1]) v_input_embed_red_temp_v1: T.float32 = input_embed_red_temp_v1_shared[T.int64(0), T.int64(0)] + T.Cast("float32", input_embed[T.int64(0), T.int64(0), v1]) * T.Cast("float32", input_embed[T.int64(0), T.int64(0), v1]) input_embed_red_temp_v0_shared[T.int64(0), T.int64(0)] = v_input_embed_red_temp_v0 input_embed_red_temp_v1_shared[T.int64(0), T.int64(0)] = v_input_embed_red_temp_v1 for ax1_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): for ax1_0 in T.serial(T.int64(4), annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): with T.block("T_layer_norm"): v0 = T.axis.spatial(T.int64(1), T.int64(0)) v1 = T.axis.spatial(T.int64(1024), ax1_0 * T.int64(256) + ax1_1) T.reads(input_embed[T.int64(0), T.int64(0), v1], input_embed_red_temp_v0_shared[T.int64(0), T.int64(0)], input_embed_red_temp_v1_shared[T.int64(0), T.int64(0)], gpt_neox_layers_0_input_layernorm_weight2[v1], gpt_neox_layers_0_input_layernorm_bias2[v1]) T.writes(T_layer_norm[T.int64(0), T.int64(0), v1]) T_layer_norm[T.int64(0), T.int64(0), v1] = T.Cast("float16", (T.Cast("float32", input_embed[T.int64(0), T.int64(0), v1]) - input_embed_red_temp_v0_shared[T.int64(0), T.int64(0)] * T.float32(0.0009765625)) * T.rsqrt(input_embed_red_temp_v1_shared[T.int64(0), T.int64(0)] * T.float32(0.0009765625) - input_embed_red_temp_v0_shared[T.int64(0), T.int64(0)] * T.float32(0.0009765625) * (input_embed_red_temp_v0_shared[T.int64(0), T.int64(0)] * T.float32(0.0009765625)) + T.float32(1.0000000000000001e-05))) * gpt_neox_layers_0_input_layernorm_weight2[v1] + gpt_neox_layers_0_input_layernorm_bias2[v1] @T.prim_func def merge_state_inplace(v: T.handle, s: T.handle, v_other: T.handle, s_other: T.handle): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) N, H, D = T.int32(is_size_var=True), T.int32(is_size_var=True), T.int32(is_size_var=True) V = T.match_buffer(v, (N, H, D), "float16") S = T.match_buffer(s, (N, H)) V_other = T.match_buffer(v_other, (N, H, D), "float16") S_other = T.match_buffer(s_other, (N, H)) # with T.block("root"): for bx in T.thread_binding(N, thread="blockIdx.x"): for by in T.thread_binding(1, thread="blockIdx.y"): for ty in T.thread_binding(16, thread="threadIdx.y"): for tx in T.thread_binding(16, thread="threadIdx.x"): with T.block("merge"): T.reads(S[bx, ty + by * 16], S_other[bx, ty + by * 16], V[bx, ty + by * 16, tx * 4:tx * 4 + 4], V_other[bx, ty + by * 16, tx * 4:tx * 4 + 4]) T.writes(V[bx, ty + by * 16, tx * 4:tx * 4 + 4], S[bx, ty + by * 16]) s_val = T.alloc_buffer((1,), scope="local") s_other_val = T.alloc_buffer((1,), scope="local") s_max = T.alloc_buffer((1,), scope="local") scale = T.alloc_buffer((1,), scope="local") other_scale = T.alloc_buffer((1,), scope="local") v_vec = T.alloc_buffer((4,), "float16", scope="local") v_other_vec = T.alloc_buffer((4,), "float16", scope="local") s_val[0] = S[bx, ty + by * 16] s_other_val[0] = S_other[bx, ty + by * 16] s_max[0] = T.max(s_val[0], s_other_val[0]) s_val[0] = T.exp2(s_val[0] - s_max[0]) s_other_val[0] = T.exp2(s_other_val[0] - s_max[0]) scale[0] = s_val[0] / (s_val[0] + s_other_val[0]) other_scale[0] = s_other_val[0] / (s_val[0] + s_other_val[0]) for vec in T.vectorized(4): v_vec[vec] = V[bx, ty + by * 16, tx * 4 + vec] for vec in T.vectorized(4): v_other_vec[vec] = V_other[bx, ty + by * 16, tx * 4 + vec] for vec in range(4): v_vec[vec] = T.Cast("float16", T.Cast("float32", v_vec[vec]) * scale[0] + T.Cast("float32", v_other_vec[vec]) * other_scale[0]) for vec in T.vectorized(4): V[bx, ty + by * 16, tx * 4 + vec] = v_vec[vec] S[bx, ty + by * 16] = T.log2(s_val[0] + s_other_val[0]) + s_max[0] @T.prim_func(private=True) def reshape(var_add432: T.handle, var_T_reshape: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) batch_size = T.int64() add432 = T.match_buffer(var_add432, (batch_size, T.int64(1), T.int64(3072)), "float16") T_reshape = T.match_buffer(var_T_reshape, (batch_size, T.int64(1), T.int64(48), T.int64(64)), "float16") # with T.block("root"): for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * T.int64(12), thread="blockIdx.x"): for ax0_ax1_ax2_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): with T.block("T_reshape"): v0 = T.axis.spatial(batch_size, (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) // T.int64(3072)) v1 = T.axis.spatial(T.int64(48), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(3072) // T.int64(64)) v2 = T.axis.spatial(T.int64(64), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(64)) T.reads(add432[v0, T.int64(0), v1 * T.int64(64) + v2]) T.writes(T_reshape[v0, T.int64(0), v1, v2]) T_reshape[v0, T.int64(0), v1, v2] = add432[v0, T.int64(0), v1 * T.int64(64) + v2] @T.prim_func(private=True) def reshape1(var_reshape288: T.handle, var_T_reshape: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) batch_size = T.int64() reshape288 = T.match_buffer(var_reshape288, (batch_size, T.int64(1), T.int64(48), T.int64(64)), "float16") T_reshape = T.match_buffer(var_T_reshape, (batch_size, T.int64(48), T.int64(64)), "float16") # with T.block("root"): for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * T.int64(12), thread="blockIdx.x"): for ax0_ax1_ax2_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): with T.block("T_reshape"): v0 = T.axis.spatial(batch_size, (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) // T.int64(3072)) v1 = T.axis.spatial(T.int64(48), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(3072) // T.int64(64)) v2 = T.axis.spatial(T.int64(64), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(64)) T.reads(reshape288[v0, T.int64(0), v1, v2]) T.writes(T_reshape[v0, v1, v2]) T_reshape[v0, v1, v2] = reshape288[v0, T.int64(0), v1, v2] @T.prim_func(private=True) def reshape2(var_lv366: T.handle, var_T_reshape: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) batch_size = T.int64() lv366 = T.match_buffer(var_lv366, (batch_size, T.int64(16), T.int64(64)), "float16") T_reshape = T.match_buffer(var_T_reshape, (batch_size, T.int64(1), T.int64(16), T.int64(64)), "float16") # with T.block("root"): for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * T.int64(4), thread="blockIdx.x"): for ax0_ax1_ax2_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): with T.block("T_reshape"): v0 = T.axis.spatial(batch_size, (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) // T.int64(1024)) v1 = T.axis.spatial(T.int64(16), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(1024) // T.int64(64)) v2 = T.axis.spatial(T.int64(64), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(64)) T.reads(lv366[v0, v1, v2]) T.writes(T_reshape[v0, T.int64(0), v1, v2]) T_reshape[v0, T.int64(0), v1, v2] = lv366[v0, v1, v2] @T.prim_func(private=True) def reshape3(var_reshape290: T.handle, var_T_reshape: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) batch_size = T.int64() reshape290 = T.match_buffer(var_reshape290, (batch_size, T.int64(1), T.int64(16), T.int64(64)), "float16") T_reshape = T.match_buffer(var_T_reshape, (batch_size, T.int64(1), T.int64(1024)), "float16") # with T.block("root"): for ax0_ax1_fused_0 in T.thread_binding(batch_size * T.int64(4), thread="blockIdx.x"): for ax0_ax1_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): with T.block("T_reshape"): v0 = T.axis.spatial(batch_size, (ax0_ax1_fused_0 * T.int64(256) + ax0_ax1_fused_1) // T.int64(1024)) v1 = T.axis.spatial(T.int64(1024), (ax0_ax1_fused_0 * T.int64(256) + ax0_ax1_fused_1) % T.int64(1024)) T.reads(reshape290[v0, T.int64(0), v1 // T.int64(64), v1 % T.int64(64)]) T.writes(T_reshape[v0, T.int64(0), v1]) T_reshape[v0, T.int64(0), v1] = reshape290[v0, T.int64(0), v1 // T.int64(64), v1 % T.int64(64)] @T.prim_func(private=True) def reshape4(var_add288: T.handle, var_T_reshape: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) seq_len = T.int64() add288 = T.match_buffer(var_add288, (T.int64(1), seq_len, T.int64(3072)), "float16") T_reshape = T.match_buffer(var_T_reshape, (T.int64(1), seq_len, T.int64(48), T.int64(64)), "float16") # with T.block("root"): for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * T.int64(12), thread="blockIdx.x"): for ax0_ax1_ax2_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): with T.block("T_reshape"): v0 = T.axis.spatial(seq_len, (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) // T.int64(3072)) v1 = T.axis.spatial(T.int64(48), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(3072) // T.int64(64)) v2 = T.axis.spatial(T.int64(64), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(64)) T.reads(add288[T.int64(0), v0, v1 * T.int64(64) + v2]) T.writes(T_reshape[T.int64(0), v0, v1, v2]) T_reshape[T.int64(0), v0, v1, v2] = add288[T.int64(0), v0, v1 * T.int64(64) + v2] @T.prim_func(private=True) def reshape5(var_reshape192: T.handle, var_T_reshape: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) seq_len = T.int64() reshape192 = T.match_buffer(var_reshape192, (T.int64(1), seq_len, T.int64(48), T.int64(64)), "float16") T_reshape = T.match_buffer(var_T_reshape, (seq_len, T.int64(48), T.int64(64)), "float16") # with T.block("root"): for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * T.int64(12), thread="blockIdx.x"): for ax0_ax1_ax2_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): with T.block("T_reshape"): v0 = T.axis.spatial(seq_len, (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) // T.int64(3072)) v1 = T.axis.spatial(T.int64(48), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(3072) // T.int64(64)) v2 = T.axis.spatial(T.int64(64), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(64)) T.reads(reshape192[T.int64(0), v0, v1, v2]) T.writes(T_reshape[v0, v1, v2]) T_reshape[v0, v1, v2] = reshape192[T.int64(0), v0, v1, v2] @T.prim_func(private=True) def reshape6(var_lv245: T.handle, var_T_reshape: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) seq_len = T.int64() lv245 = T.match_buffer(var_lv245, (seq_len, T.int64(16), T.int64(64)), "float16") T_reshape = T.match_buffer(var_T_reshape, (T.int64(1), seq_len, T.int64(16), T.int64(64)), "float16") # with T.block("root"): for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * T.int64(4), thread="blockIdx.x"): for ax0_ax1_ax2_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): with T.block("T_reshape"): v0 = T.axis.spatial(seq_len, (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) // T.int64(1024)) v1 = T.axis.spatial(T.int64(16), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(1024) // T.int64(64)) v2 = T.axis.spatial(T.int64(64), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(64)) T.reads(lv245[v0, v1, v2]) T.writes(T_reshape[T.int64(0), v0, v1, v2]) T_reshape[T.int64(0), v0, v1, v2] = lv245[v0, v1, v2] @T.prim_func(private=True) def reshape7(var_reshape194: T.handle, var_T_reshape: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) seq_len = T.int64() reshape194 = T.match_buffer(var_reshape194, (T.int64(1), seq_len, T.int64(16), T.int64(64)), "float16") T_reshape = T.match_buffer(var_T_reshape, (T.int64(1), seq_len, T.int64(1024)), "float16") # with T.block("root"): for ax0_ax1_fused_0 in T.thread_binding(seq_len * T.int64(4), thread="blockIdx.x"): for ax0_ax1_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): with T.block("T_reshape"): v0 = T.axis.spatial(seq_len, (ax0_ax1_fused_0 * T.int64(256) + ax0_ax1_fused_1) // T.int64(1024)) v1 = T.axis.spatial(T.int64(1024), (ax0_ax1_fused_0 * T.int64(256) + ax0_ax1_fused_1) % T.int64(1024)) T.reads(reshape194[T.int64(0), v0, v1 // T.int64(64), v1 % T.int64(64)]) T.writes(T_reshape[T.int64(0), v0, v1]) T_reshape[T.int64(0), v0, v1] = reshape194[T.int64(0), v0, v1 // T.int64(64), v1 % T.int64(64)] @T.prim_func def 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 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.int64(is_size_var=True), T.int64(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.int64(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(T.int64(32), thread="threadIdx.x"): for ax0_0 in T.serial((num_chunks + T.int64(31)) // T.int64(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 * T.int64(32) + ax0_1) T.where(ax0_0 * T.int64(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(T.int64(32), thread="threadIdx.x"): for ax0_0 in T.serial((num_chunks + T.int64(31)) // T.int64(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 * T.int64(32) + ax0_1) T.where(ax0_0 * T.int64(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(T.int64(16), annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): for l2_1 in T.thread_binding(T.int64(8), thread="threadIdx.y"): for l2_2 in T.thread_binding(T.int64(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(T.int64(4096), l2_0 * T.int64(256) + l2_1 * T.int64(32) + l2_2) T.reads(temperature[v0], A[v0, v1 * T.int64(4096) + v2], temp_sum_shared[v0], temp_max_shared[v0]) T.writes(softmax[v0, v1 * T.int64(4096) + v2]) if v1 * T.int64(4096) + v2 < vocab_size: softmax[v0, v1 * T.int64(4096) + v2] = T.if_then_else(temperature[v0] > T.float32(1.0000000000000001e-05), T.exp(A[v0, v1 * T.int64(4096) + v2] / temperature[v0] - (T.log(temp_sum_shared[v0]) + temp_max_shared[v0])), T.Cast("float32", A[v0, v1 * T.int64(4096) + v2] == temp_max_shared[v0]) / temp_sum_shared[v0]) @T.prim_func(private=True) def take(var_layer_norm146: T.handle, var_logit_positions: T.handle, var_T_take: T.handle): T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) seq_len = T.int64() layer_norm146 = T.match_buffer(var_layer_norm146, (T.int64(1), seq_len, T.int64(1024)), "float16") batch_size = T.int64() logit_positions = T.match_buffer(var_logit_positions, (batch_size,), "int32") T_take = T.match_buffer(var_T_take, (T.int64(1), batch_size, T.int64(1024)), "float16") # with T.block("root"): for ax0_ax1_fused_0 in T.thread_binding(batch_size * T.int64(4), thread="blockIdx.x"): for ax0_ax1_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): with T.block("T_take"): v0 = T.axis.spatial(batch_size, (ax0_ax1_fused_0 * T.int64(256) + ax0_ax1_fused_1) // T.int64(1024)) v1 = T.axis.spatial(T.int64(1024), (ax0_ax1_fused_0 * T.int64(256) + ax0_ax1_fused_1) % T.int64(1024)) T.reads(layer_norm146[T.int64(0), logit_positions[v0], v1], logit_positions[v0]) T.writes(T_take[T.int64(0), v0, v1]) T_take[T.int64(0), v0, v1] = layer_norm146[T.int64(0), logit_positions[v0], v1] @T.prim_func def 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.int64): 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.int64(), T.int64(is_size_var=True) pages = T.match_buffer(var_pages, (num_pages, 2, 16, page_size, 64), "float16", offset_factor=1) seqlen = T.int64(is_size_var=True) position_map = T.match_buffer(var_position_map, (seqlen,), "int32", offset_factor=1) k_data = T.match_buffer(var_k_data, (24, seqlen, 16, 64), "float16") v_data = T.match_buffer(var_v_data, (24, seqlen, 16, 64), "float16") # with T.block("root"): for p_h_d_fused_0 in T.thread_binding(seqlen * T.int64(4), thread="blockIdx.x"): for p_h_d_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): with T.block("copy0"): vp = T.axis.spatial(seqlen, (p_h_d_fused_0 * T.int64(256) + p_h_d_fused_1) // T.int64(1024)) vh = T.axis.spatial(16, T.Cast("int32", (p_h_d_fused_0 * T.int64(256) + p_h_d_fused_1) % T.int64(1024) // T.int64(64))) vd = T.axis.spatial(64, T.Cast("int32", (p_h_d_fused_0 * T.int64(256) + p_h_d_fused_1) % T.int64(64))) T.reads(position_map[vp], pages[T.Cast("int64", position_map[vp]) // page_size, 0:2, vh, T.Cast("int64", 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[T.Cast("int64", position) // page_size, 0, vh, T.Cast("int64", position) % page_size, vd] v_data[layer_id, vp, vh, vd] = pages[T.Cast("int64", position) // page_size, 1, vh, T.Cast("int64", position) % page_size, vd] @T.prim_func def 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.int64() pages = T.match_buffer(var_pages, (num_pages, 2, 16, 16, 64), "float16", offset_factor=1) ntoken = T.int64(is_size_var=True) k_data = T.match_buffer(var_k_data, (ntoken, 16, 64), "float16") v_data = T.match_buffer(var_v_data, (ntoken, 16, 64), "float16") position_map = T.match_buffer(var_position_map, (ntoken,), "int32", offset_factor=1) # with T.block("root"): for global_pos_h_f_fused_0 in T.thread_binding(ntoken * T.int64(4), thread="blockIdx.x"): for global_pos_h_f_fused_1 in T.thread_binding(T.int64(256), thread="threadIdx.x"): if position_map[(global_pos_h_f_fused_0 * T.int64(256) + global_pos_h_f_fused_1) // T.int64(1024)] != -1: with T.block("k_transpose_append"): vgpos = T.axis.spatial(ntoken, (global_pos_h_f_fused_0 * T.int64(256) + global_pos_h_f_fused_1) // T.int64(1024)) vh = T.axis.spatial(16, T.Cast("int32", (global_pos_h_f_fused_0 * T.int64(256) + global_pos_h_f_fused_1) % T.int64(1024) // T.int64(64))) vf = T.axis.spatial(64, T.Cast("int32", (global_pos_h_f_fused_0 * T.int64(256) + global_pos_h_f_fused_1) % T.int64(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 * T.int64(256) + global_pos_h_f_fused_1) // T.int64(1024)) vh = T.axis.spatial(16, T.Cast("int32", (global_pos_h_f_fused_0 * T.int64(256) + global_pos_h_f_fused_1) % T.int64(1024) // T.int64(64))) vf = T.axis.spatial(64, T.Cast("int32", (global_pos_h_f_fused_0 * T.int64(256) + global_pos_h_f_fused_1) % T.int64(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 tree_attn_paged_kv(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32, tree_order_indptr_handle: T.handle, tree_order_handle: T.handle): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) total_len = T.int32(is_size_var=True) q = T.match_buffer(var_q, (total_len, 16, 64), "float16") batch_size = T.int32(is_size_var=True) q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) max_num_pages = T.int32(is_size_var=True) pages = T.match_buffer(var_pages, (max_num_pages, 2, 16, 32, 64), "float16") page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) nnz_pages = T.int32(is_size_var=True) page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) output = T.match_buffer(var_output, (total_len, 16, 64), "float16") lse = T.match_buffer(var_lse, (total_len, 16)) tree_order_indptr = T.match_buffer(tree_order_indptr_handle, (batch_size + 1,), "int32", offset_factor=1) total_tree_order_len = T.int32(is_size_var=True) tree_order = T.match_buffer(tree_order_handle, (total_tree_order_len, 2), "int32", offset_factor=1) # with T.block("root"): assert rotary_mode == 0, "Inline rotary mode is not supported in tree attention." for lbx in T.thread_binding(16, thread="blockIdx.x"): for lby in T.thread_binding(16, thread="blockIdx.y"): for lty in T.thread_binding(4, thread="threadIdx.y"): for ltx in T.thread_binding(32, thread="threadIdx.x"): with T.block("attn"): bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) T.reads() T.writes() tile_id = T.alloc_buffer((1,), "int32", scope="local") batch_idx = T.alloc_buffer((1,), "int32", scope="local") batch_tiles = T.alloc_buffer((1,), "int32", scope="local") batch_rows = T.alloc_buffer((1,), "int32", scope="local") iterator = T.alloc_buffer((1,), "int32", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") Q_smem = T.alloc_buffer((32, 64), "float16", scope="shared") K_smem = T.alloc_buffer((32, 64), "float16", scope="shared") V_smem = T.alloc_buffer((32, 64), "float16", scope="shared") S_smem = T.alloc_buffer((32, 32), scope="shared") S_local = T.alloc_buffer((32, 32), scope="local") O_local = T.alloc_buffer((32, 64), scope="local") m_smem = T.alloc_buffer((32,), scope="shared") m_prev_smem = T.alloc_buffer((32,), scope="shared") d_smem = T.alloc_buffer((32,), scope="shared") m_new = T.alloc_buffer((1,), scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_new = T.alloc_buffer((1,), scope="local") tile_id[0] = bx batch_idx[0] = 0 batch_rows[0] = q_indptr[1] - q_indptr[0] batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 while T.tvm_thread_invariant(batch_idx[0] < batch_size): while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: tile_id[0] = tile_id[0] - batch_tiles[0] batch_idx[0] = batch_idx[0] + 1 if batch_idx[0] < batch_size: b_idx: T.int32 = batch_idx[0] batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 if T.tvm_thread_invariant(batch_idx[0] < batch_size): b_idx: T.int32 = batch_idx[0] LH_start: T.int32 = tile_id[0] * 32 q_indptr_val: T.int32 = q_indptr[b_idx] cur_page_indptr_begin: T.int32 = page_indptr[b_idx] cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 32 + length_info[b_idx], 0) T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: m_smem[row] = T.float32(-50000.0) d_smem[row] = T.float32(1.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 4): with T.block("O_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) T.reads() T.writes(O_local[i, j]) O_local[i, j] = T.float32(0.0) T.tvm_storage_sync("shared") for li_lj_fused_0 in range(4): for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for li_lj_fused_3 in T.vectorized(4): with T.block("Q_load"): i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) T.reads() T.writes() cur_L: T.int32 = q_indptr_val + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: freq = T.float32() Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) else: Q_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for iterator_1 in range((kv_chunk_len[0] + 31) // 32): L_kv_start: T.int32 = iterator_1 * 32 for lz_ly_fused_0 in range(4): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("K_load"): i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: seq_offset: T.int32 = cur_L page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 32] page_offset: T.int32 = seq_offset % 32 K_smem[i, j] = pages[page_no, 0, by, page_offset, j] else: K_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for lz_ly_fused_0 in range(4): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("V_load"): i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: seq_offset: T.int32 = cur_L page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 32] page_offset: T.int32 = seq_offset % 32 V_smem[i, j] = pages[page_no, 1, by, page_offset, j] else: V_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") with T.block(""): T.reads(Q_smem[0:32, 0:64], K_smem[0:32, 0:64]) T.writes(S_local[0:32, 0:32]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(2, 4): with T.block("S_gemm_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_init) T.reads() T.writes(S_local[i, j]) S_local[i, j] = T.float32(0.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 4, 8): with T.block("S_gemm_update"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) k = T.axis.reduce(64, lk_0 * 8 + lk_1) T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) T.writes(S_local[i, j]) S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.18033688011112042) T.tvm_storage_sync("shared") for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(2, 4): with T.block("S_store"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) T.reads(S_local[i, j]) T.writes(S_smem[i, j]) S_smem[i, j] = S_local[i, j] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: with T.block("update1"): T.reads(m_smem[row], kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) T.writes(m_prev[i], m_new[i], d_new[i]) m_prev[i] = m_smem[row] m_new[i] = m_smem[row] row_: T.int32 = LH_start + row for j in range(32): if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): m_new[i] = T.max(m_new[i], S_smem[row, j]) d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx with T.block("update"): T.reads(kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) T.writes(S_smem[row, 0:32]) for j in range(32): if row < 32: row_: T.int32 = LH_start + row if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) else: S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 32: with T.block("update"): T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) for j in range(32): d_new[i] = d_new[i] + S_smem[row, j] m_smem[row] = m_new[i] d_smem[row] = d_new[i] m_prev_smem[row] = m_prev[i] T.tvm_storage_sync("shared") with T.block(""): T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:64]) T.writes(O_local[0:32, 0:64]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(4, 4): with T.block("O_gemm_init"): i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) T.reads() T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, lk_1, li_1, lj_1 in T.grid(4, 8, 4, 4): with T.block("O_gemm_update"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) k = T.axis.reduce(32, lk_0 * 8 + lk_1) T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 4): with T.block("O_store"): i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) for li_0 in range(1): for li_1 in T.thread_binding(4, thread="threadIdx.y"): for li_2 in T.thread_binding(32, thread="threadIdx.x"): with T.block("lse_store"): i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) tile_id[0] = tile_id[0] + 16 @R.function def alloc_embedding_tensor() -> R.Tensor((2048, 1024), dtype="float16"): R.func_attr({"relax.memory_plan_dynamic_func_output": True}) gv: R.Tensor((2048, 1024), dtype="float16") = R.builtin.alloc_tensor(R.shape([2048, 1024]), R.dtype("float16"), R.prim_value(0), R.str("global")) return gv @R.function def batch_decode(input_embeds: R.Tensor(("batch_size", 1, 1024), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 128), dtype="uint32"), R.Tensor(("vocab_size", 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor(("vocab_size", 128), dtype="uint32"), R.Tensor(("vocab_size", 32), 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.memory_plan_dynamic_func_output": 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 with R.dataflow(): gpt_neox_layers_0_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[2] gpt_neox_layers_0_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[3] gpt_neox_layers_0_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[4] gpt_neox_layers_0_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[5] gpt_neox_layers_0_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[6] gpt_neox_layers_0_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[8] gpt_neox_layers_0_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[9] gpt_neox_layers_0_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[10] gpt_neox_layers_0_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[11] gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[12] gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[13] gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[14] gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[15] gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[16] gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[17] gpt_neox_layers_1_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[18] gpt_neox_layers_1_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[19] gpt_neox_layers_1_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[20] gpt_neox_layers_1_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[21] gpt_neox_layers_1_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[22] gpt_neox_layers_1_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[23] gpt_neox_layers_1_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[24] gpt_neox_layers_1_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[25] gpt_neox_layers_1_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[26] gpt_neox_layers_1_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[27] gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[28] gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[29] gpt_neox_layers_1_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[30] gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[31] gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[32] gpt_neox_layers_1_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[33] gpt_neox_layers_2_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[34] gpt_neox_layers_2_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[35] gpt_neox_layers_2_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[36] gpt_neox_layers_2_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[37] gpt_neox_layers_2_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[38] gpt_neox_layers_2_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[39] gpt_neox_layers_2_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[40] gpt_neox_layers_2_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[41] gpt_neox_layers_2_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[42] gpt_neox_layers_2_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[43] gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[44] gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[45] gpt_neox_layers_2_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[46] gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[47] gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[48] gpt_neox_layers_2_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[49] gpt_neox_layers_3_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[50] gpt_neox_layers_3_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[51] gpt_neox_layers_3_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[52] gpt_neox_layers_3_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[53] gpt_neox_layers_3_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[54] gpt_neox_layers_3_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[55] gpt_neox_layers_3_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[56] gpt_neox_layers_3_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[57] gpt_neox_layers_3_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[58] gpt_neox_layers_3_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[59] gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[60] gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[61] gpt_neox_layers_3_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[62] gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[63] gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[64] gpt_neox_layers_3_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[65] gpt_neox_layers_4_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[66] gpt_neox_layers_4_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[67] gpt_neox_layers_4_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[68] gpt_neox_layers_4_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[69] gpt_neox_layers_4_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[70] gpt_neox_layers_4_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[71] gpt_neox_layers_4_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[72] gpt_neox_layers_4_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[73] gpt_neox_layers_4_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[74] gpt_neox_layers_4_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[75] gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[76] gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[77] gpt_neox_layers_4_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[78] gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[79] gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[80] gpt_neox_layers_4_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[81] gpt_neox_layers_5_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[82] gpt_neox_layers_5_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[83] gpt_neox_layers_5_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[84] gpt_neox_layers_5_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[85] gpt_neox_layers_5_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[86] gpt_neox_layers_5_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[87] gpt_neox_layers_5_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[88] gpt_neox_layers_5_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[89] gpt_neox_layers_5_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[90] gpt_neox_layers_5_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[91] gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[92] gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[93] gpt_neox_layers_5_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[94] gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[95] gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[96] gpt_neox_layers_5_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[97] gpt_neox_layers_6_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[98] gpt_neox_layers_6_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[99] gpt_neox_layers_6_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[100] gpt_neox_layers_6_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[101] gpt_neox_layers_6_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[102] gpt_neox_layers_6_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[103] gpt_neox_layers_6_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[104] gpt_neox_layers_6_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[105] gpt_neox_layers_6_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[106] gpt_neox_layers_6_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[107] gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[108] gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[109] gpt_neox_layers_6_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[110] gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[111] gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[112] gpt_neox_layers_6_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[113] gpt_neox_layers_7_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[114] gpt_neox_layers_7_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[115] gpt_neox_layers_7_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[116] gpt_neox_layers_7_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[117] gpt_neox_layers_7_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[118] gpt_neox_layers_7_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[119] gpt_neox_layers_7_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[120] gpt_neox_layers_7_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[121] gpt_neox_layers_7_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[122] gpt_neox_layers_7_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[123] gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[124] gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[125] gpt_neox_layers_7_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[126] gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[127] gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[128] gpt_neox_layers_7_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[129] gpt_neox_layers_8_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[130] gpt_neox_layers_8_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[131] gpt_neox_layers_8_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[132] gpt_neox_layers_8_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[133] gpt_neox_layers_8_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[134] gpt_neox_layers_8_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[135] gpt_neox_layers_8_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[136] gpt_neox_layers_8_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[137] gpt_neox_layers_8_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[138] gpt_neox_layers_8_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[139] gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[140] gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[141] gpt_neox_layers_8_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[142] gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[143] gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[144] gpt_neox_layers_8_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[145] gpt_neox_layers_9_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[146] gpt_neox_layers_9_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[147] gpt_neox_layers_9_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[148] gpt_neox_layers_9_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[149] gpt_neox_layers_9_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[150] gpt_neox_layers_9_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[151] gpt_neox_layers_9_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[152] gpt_neox_layers_9_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[153] gpt_neox_layers_9_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[154] gpt_neox_layers_9_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[155] gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[156] gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[157] gpt_neox_layers_9_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[158] gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[159] gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[160] gpt_neox_layers_9_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[161] gpt_neox_layers_10_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[162] gpt_neox_layers_10_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[163] gpt_neox_layers_10_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[164] gpt_neox_layers_10_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[165] gpt_neox_layers_10_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[166] gpt_neox_layers_10_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[167] gpt_neox_layers_10_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[168] gpt_neox_layers_10_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[169] gpt_neox_layers_10_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[170] gpt_neox_layers_10_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[171] gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[172] gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[173] gpt_neox_layers_10_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[174] gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[175] gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[176] gpt_neox_layers_10_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[177] gpt_neox_layers_11_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[178] gpt_neox_layers_11_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[179] gpt_neox_layers_11_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[180] gpt_neox_layers_11_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[181] gpt_neox_layers_11_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[182] gpt_neox_layers_11_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[183] gpt_neox_layers_11_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[184] gpt_neox_layers_11_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[185] gpt_neox_layers_11_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[186] gpt_neox_layers_11_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[187] gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[188] gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[189] gpt_neox_layers_11_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[190] gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[191] gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[192] gpt_neox_layers_11_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[193] gpt_neox_layers_12_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[194] gpt_neox_layers_12_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[195] gpt_neox_layers_12_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[196] gpt_neox_layers_12_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[197] gpt_neox_layers_12_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[198] gpt_neox_layers_12_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[199] gpt_neox_layers_12_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[200] gpt_neox_layers_12_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[201] gpt_neox_layers_12_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[202] gpt_neox_layers_12_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[203] gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[204] gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[205] gpt_neox_layers_12_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[206] gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[207] gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[208] gpt_neox_layers_12_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[209] gpt_neox_layers_13_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[210] gpt_neox_layers_13_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[211] gpt_neox_layers_13_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[212] gpt_neox_layers_13_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[213] gpt_neox_layers_13_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[214] gpt_neox_layers_13_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[215] gpt_neox_layers_13_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[216] gpt_neox_layers_13_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[217] gpt_neox_layers_13_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[218] gpt_neox_layers_13_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[219] gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[220] gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[221] gpt_neox_layers_13_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[222] gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[223] gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[224] gpt_neox_layers_13_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[225] gpt_neox_layers_14_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[226] gpt_neox_layers_14_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[227] gpt_neox_layers_14_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[228] gpt_neox_layers_14_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[229] gpt_neox_layers_14_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[230] gpt_neox_layers_14_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[231] gpt_neox_layers_14_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[232] gpt_neox_layers_14_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[233] gpt_neox_layers_14_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[234] gpt_neox_layers_14_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[235] gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[236] gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[237] gpt_neox_layers_14_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[238] gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[239] gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[240] gpt_neox_layers_14_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[241] gpt_neox_layers_15_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[242] gpt_neox_layers_15_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[243] gpt_neox_layers_15_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[244] gpt_neox_layers_15_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[245] gpt_neox_layers_15_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[246] gpt_neox_layers_15_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[247] gpt_neox_layers_15_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[248] gpt_neox_layers_15_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[249] gpt_neox_layers_15_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[250] gpt_neox_layers_15_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[251] gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[252] gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[253] gpt_neox_layers_15_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[254] gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[255] gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[256] gpt_neox_layers_15_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[257] gpt_neox_layers_16_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[258] gpt_neox_layers_16_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[259] gpt_neox_layers_16_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[260] gpt_neox_layers_16_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[261] gpt_neox_layers_16_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[262] gpt_neox_layers_16_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[263] gpt_neox_layers_16_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[264] gpt_neox_layers_16_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[265] gpt_neox_layers_16_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[266] gpt_neox_layers_16_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[267] gpt_neox_layers_16_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[268] gpt_neox_layers_16_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[269] gpt_neox_layers_16_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[270] gpt_neox_layers_16_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[271] gpt_neox_layers_16_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[272] gpt_neox_layers_16_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[273] gpt_neox_layers_17_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[274] gpt_neox_layers_17_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[275] gpt_neox_layers_17_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[276] gpt_neox_layers_17_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[277] gpt_neox_layers_17_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[278] gpt_neox_layers_17_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[279] gpt_neox_layers_17_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[280] gpt_neox_layers_17_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[281] gpt_neox_layers_17_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[282] gpt_neox_layers_17_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[283] gpt_neox_layers_17_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[284] gpt_neox_layers_17_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[285] gpt_neox_layers_17_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[286] gpt_neox_layers_17_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[287] gpt_neox_layers_17_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[288] gpt_neox_layers_17_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[289] gpt_neox_layers_18_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[290] gpt_neox_layers_18_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[291] gpt_neox_layers_18_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[292] gpt_neox_layers_18_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[293] gpt_neox_layers_18_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[294] gpt_neox_layers_18_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[295] gpt_neox_layers_18_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[296] gpt_neox_layers_18_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[297] gpt_neox_layers_18_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[298] gpt_neox_layers_18_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[299] gpt_neox_layers_18_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[300] gpt_neox_layers_18_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[301] gpt_neox_layers_18_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[302] gpt_neox_layers_18_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[303] gpt_neox_layers_18_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[304] gpt_neox_layers_18_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[305] gpt_neox_layers_19_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[306] gpt_neox_layers_19_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[307] gpt_neox_layers_19_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[308] gpt_neox_layers_19_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[309] gpt_neox_layers_19_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[310] gpt_neox_layers_19_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[311] gpt_neox_layers_19_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[312] gpt_neox_layers_19_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[313] gpt_neox_layers_19_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[314] gpt_neox_layers_19_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[315] gpt_neox_layers_19_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[316] gpt_neox_layers_19_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[317] gpt_neox_layers_19_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[318] gpt_neox_layers_19_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[319] gpt_neox_layers_19_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[320] gpt_neox_layers_19_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[321] gpt_neox_layers_20_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[322] gpt_neox_layers_20_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[323] gpt_neox_layers_20_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[324] gpt_neox_layers_20_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[325] gpt_neox_layers_20_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[326] gpt_neox_layers_20_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[327] gpt_neox_layers_20_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[328] gpt_neox_layers_20_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[329] gpt_neox_layers_20_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[330] gpt_neox_layers_20_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[331] gpt_neox_layers_20_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[332] gpt_neox_layers_20_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[333] gpt_neox_layers_20_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[334] gpt_neox_layers_20_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[335] gpt_neox_layers_20_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[336] gpt_neox_layers_20_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[337] gpt_neox_layers_21_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[338] gpt_neox_layers_21_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[339] gpt_neox_layers_21_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[340] gpt_neox_layers_21_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[341] gpt_neox_layers_21_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[342] gpt_neox_layers_21_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[343] gpt_neox_layers_21_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[344] gpt_neox_layers_21_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[345] gpt_neox_layers_21_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[346] gpt_neox_layers_21_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[347] gpt_neox_layers_21_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[348] gpt_neox_layers_21_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[349] gpt_neox_layers_21_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[350] gpt_neox_layers_21_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[351] gpt_neox_layers_21_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[352] gpt_neox_layers_21_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[353] gpt_neox_layers_22_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[354] gpt_neox_layers_22_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[355] gpt_neox_layers_22_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[356] gpt_neox_layers_22_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[357] gpt_neox_layers_22_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[358] gpt_neox_layers_22_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[359] gpt_neox_layers_22_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[360] gpt_neox_layers_22_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[361] gpt_neox_layers_22_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[362] gpt_neox_layers_22_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[363] gpt_neox_layers_22_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[364] gpt_neox_layers_22_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[365] gpt_neox_layers_22_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[366] gpt_neox_layers_22_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[367] gpt_neox_layers_22_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[368] gpt_neox_layers_22_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[369] gpt_neox_layers_23_input_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[370] gpt_neox_layers_23_input_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[371] gpt_neox_layers_23_post_attention_layernorm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[372] gpt_neox_layers_23_post_attention_layernorm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[373] gpt_neox_layers_23_attention_query_key_value_q_weight4: R.Tensor((3072, 128), dtype="uint32") = packed_params[374] gpt_neox_layers_23_attention_query_key_value_q_scale4: R.Tensor((3072, 32), dtype="float16") = packed_params[375] gpt_neox_layers_23_attention_query_key_value_bias4: R.Tensor((3072,), dtype="float16") = packed_params[376] gpt_neox_layers_23_attention_dense_q_weight4: R.Tensor((1024, 128), dtype="uint32") = packed_params[377] gpt_neox_layers_23_attention_dense_q_scale4: R.Tensor((1024, 32), dtype="float16") = packed_params[378] gpt_neox_layers_23_attention_dense_bias4: R.Tensor((1024,), dtype="float16") = packed_params[379] gpt_neox_layers_23_mlp_dense_h_to_4h_q_weight4: R.Tensor((4096, 128), dtype="uint32") = packed_params[380] gpt_neox_layers_23_mlp_dense_h_to_4h_q_scale4: R.Tensor((4096, 32), dtype="float16") = packed_params[381] gpt_neox_layers_23_mlp_dense_h_to_4h_bias4: R.Tensor((4096,), dtype="float32") = packed_params[382] gpt_neox_layers_23_mlp_dense_4h_to_h_q_weight4: R.Tensor((1024, 512), dtype="uint32") = packed_params[383] gpt_neox_layers_23_mlp_dense_4h_to_h_q_scale4: R.Tensor((1024, 128), dtype="float16") = packed_params[384] gpt_neox_layers_23_mlp_dense_4h_to_h_bias4: R.Tensor((1024,), dtype="float32") = packed_params[385] gpt_neox_final_layer_norm_weight4: R.Tensor((1024,), dtype="float16") = packed_params[386] gpt_neox_final_layer_norm_bias4: R.Tensor((1024,), dtype="float16") = packed_params[387] embed_out_q_weight4: R.Tensor((vocab_size, 128), dtype="uint32") = packed_params[388] embed_out_q_scale4: R.Tensor((vocab_size, 32), dtype="float16") = packed_params[389] layer_norm147 = R.call_tir(cls.layer_norm, (input_embeds, gpt_neox_layers_0_input_layernorm_weight4, gpt_neox_layers_0_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv = R.call_tir(cls.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, layer_norm147, gpt_neox_layers_0_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape288 = R.call_tir(cls.reshape, (lv,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape289 = R.call_tir(cls.reshape1, (reshape288,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv366 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), reshape289), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape290 = R.call_tir(cls.reshape2, (lv366,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape291 = R.call_tir(cls.reshape3, (reshape290,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm148 = R.call_tir(cls.layer_norm, (input_embeds, gpt_neox_layers_0_post_attention_layernorm_weight4, gpt_neox_layers_0_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv1 = R.call_tir(cls.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, layer_norm148, gpt_neox_layers_0_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv2 = R.call_tir(cls.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, lv1, gpt_neox_layers_0_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_0_attention_dense_q_weight4, gpt_neox_layers_0_attention_dense_q_scale4, reshape291, gpt_neox_layers_0_attention_dense_bias4, lv2, input_embeds), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm149 = R.call_tir(cls.layer_norm, (lv_1, gpt_neox_layers_1_input_layernorm_weight4, gpt_neox_layers_1_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv3 = R.call_tir(cls.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, layer_norm149, gpt_neox_layers_1_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape292 = R.call_tir(cls.reshape, (lv3,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape293 = R.call_tir(cls.reshape1, (reshape292,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv371 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), reshape293), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape294 = R.call_tir(cls.reshape2, (lv371,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape295 = R.call_tir(cls.reshape3, (reshape294,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm150 = R.call_tir(cls.layer_norm, (lv_1, gpt_neox_layers_1_post_attention_layernorm_weight4, gpt_neox_layers_1_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv4 = R.call_tir(cls.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, layer_norm150, gpt_neox_layers_1_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv5 = R.call_tir(cls.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, lv4, gpt_neox_layers_1_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv1_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_1_attention_dense_q_weight4, gpt_neox_layers_1_attention_dense_q_scale4, reshape295, gpt_neox_layers_1_attention_dense_bias4, lv5, lv_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm151 = R.call_tir(cls.layer_norm, (lv1_1, gpt_neox_layers_2_input_layernorm_weight4, gpt_neox_layers_2_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv6 = R.call_tir(cls.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, layer_norm151, gpt_neox_layers_2_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape296 = R.call_tir(cls.reshape, (lv6,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape297 = R.call_tir(cls.reshape1, (reshape296,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv376 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), reshape297), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape298 = R.call_tir(cls.reshape2, (lv376,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape299 = R.call_tir(cls.reshape3, (reshape298,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm152 = R.call_tir(cls.layer_norm, (lv1_1, gpt_neox_layers_2_post_attention_layernorm_weight4, gpt_neox_layers_2_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv7 = R.call_tir(cls.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, layer_norm152, gpt_neox_layers_2_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv8 = R.call_tir(cls.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, lv7, gpt_neox_layers_2_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv2_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_2_attention_dense_q_weight4, gpt_neox_layers_2_attention_dense_q_scale4, reshape299, gpt_neox_layers_2_attention_dense_bias4, lv8, lv1_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm153 = R.call_tir(cls.layer_norm, (lv2_1, gpt_neox_layers_3_input_layernorm_weight4, gpt_neox_layers_3_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv9 = R.call_tir(cls.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, layer_norm153, gpt_neox_layers_3_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape300 = R.call_tir(cls.reshape, (lv9,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape301 = R.call_tir(cls.reshape1, (reshape300,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv381 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), reshape301), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape302 = R.call_tir(cls.reshape2, (lv381,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape303 = R.call_tir(cls.reshape3, (reshape302,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm154 = R.call_tir(cls.layer_norm, (lv2_1, gpt_neox_layers_3_post_attention_layernorm_weight4, gpt_neox_layers_3_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv10 = R.call_tir(cls.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, layer_norm154, gpt_neox_layers_3_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv11 = R.call_tir(cls.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, lv10, gpt_neox_layers_3_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv3_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_3_attention_dense_q_weight4, gpt_neox_layers_3_attention_dense_q_scale4, reshape303, gpt_neox_layers_3_attention_dense_bias4, lv11, lv2_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm155 = R.call_tir(cls.layer_norm, (lv3_1, gpt_neox_layers_4_input_layernorm_weight4, gpt_neox_layers_4_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv12 = R.call_tir(cls.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, layer_norm155, gpt_neox_layers_4_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape304 = R.call_tir(cls.reshape, (lv12,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape305 = R.call_tir(cls.reshape1, (reshape304,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv386 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), reshape305), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape306 = R.call_tir(cls.reshape2, (lv386,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape307 = R.call_tir(cls.reshape3, (reshape306,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm156 = R.call_tir(cls.layer_norm, (lv3_1, gpt_neox_layers_4_post_attention_layernorm_weight4, gpt_neox_layers_4_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv13 = R.call_tir(cls.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, layer_norm156, gpt_neox_layers_4_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv14 = R.call_tir(cls.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, lv13, gpt_neox_layers_4_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv4_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_4_attention_dense_q_weight4, gpt_neox_layers_4_attention_dense_q_scale4, reshape307, gpt_neox_layers_4_attention_dense_bias4, lv14, lv3_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm157 = R.call_tir(cls.layer_norm, (lv4_1, gpt_neox_layers_5_input_layernorm_weight4, gpt_neox_layers_5_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv15 = R.call_tir(cls.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, layer_norm157, gpt_neox_layers_5_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape308 = R.call_tir(cls.reshape, (lv15,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape309 = R.call_tir(cls.reshape1, (reshape308,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv391 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), reshape309), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape310 = R.call_tir(cls.reshape2, (lv391,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape311 = R.call_tir(cls.reshape3, (reshape310,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm158 = R.call_tir(cls.layer_norm, (lv4_1, gpt_neox_layers_5_post_attention_layernorm_weight4, gpt_neox_layers_5_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv16 = R.call_tir(cls.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, layer_norm158, gpt_neox_layers_5_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv17 = R.call_tir(cls.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, lv16, gpt_neox_layers_5_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv5_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_5_attention_dense_q_weight4, gpt_neox_layers_5_attention_dense_q_scale4, reshape311, gpt_neox_layers_5_attention_dense_bias4, lv17, lv4_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm159 = R.call_tir(cls.layer_norm, (lv5_1, gpt_neox_layers_6_input_layernorm_weight4, gpt_neox_layers_6_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv18 = R.call_tir(cls.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, layer_norm159, gpt_neox_layers_6_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape312 = R.call_tir(cls.reshape, (lv18,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape313 = R.call_tir(cls.reshape1, (reshape312,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv396 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), reshape313), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape314 = R.call_tir(cls.reshape2, (lv396,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape315 = R.call_tir(cls.reshape3, (reshape314,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm160 = R.call_tir(cls.layer_norm, (lv5_1, gpt_neox_layers_6_post_attention_layernorm_weight4, gpt_neox_layers_6_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv19 = R.call_tir(cls.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, layer_norm160, gpt_neox_layers_6_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv20 = R.call_tir(cls.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, lv19, gpt_neox_layers_6_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv6_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_6_attention_dense_q_weight4, gpt_neox_layers_6_attention_dense_q_scale4, reshape315, gpt_neox_layers_6_attention_dense_bias4, lv20, lv5_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm161 = R.call_tir(cls.layer_norm, (lv6_1, gpt_neox_layers_7_input_layernorm_weight4, gpt_neox_layers_7_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv21 = R.call_tir(cls.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, layer_norm161, gpt_neox_layers_7_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape316 = R.call_tir(cls.reshape, (lv21,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape317 = R.call_tir(cls.reshape1, (reshape316,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv401 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), reshape317), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape318 = R.call_tir(cls.reshape2, (lv401,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape319 = R.call_tir(cls.reshape3, (reshape318,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm162 = R.call_tir(cls.layer_norm, (lv6_1, gpt_neox_layers_7_post_attention_layernorm_weight4, gpt_neox_layers_7_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv22 = R.call_tir(cls.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, layer_norm162, gpt_neox_layers_7_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv23 = R.call_tir(cls.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, lv22, gpt_neox_layers_7_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv7_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_7_attention_dense_q_weight4, gpt_neox_layers_7_attention_dense_q_scale4, reshape319, gpt_neox_layers_7_attention_dense_bias4, lv23, lv6_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm163 = R.call_tir(cls.layer_norm, (lv7_1, gpt_neox_layers_8_input_layernorm_weight4, gpt_neox_layers_8_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv24 = R.call_tir(cls.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, layer_norm163, gpt_neox_layers_8_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape320 = R.call_tir(cls.reshape, (lv24,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape321 = R.call_tir(cls.reshape1, (reshape320,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv406 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), reshape321), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape322 = R.call_tir(cls.reshape2, (lv406,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape323 = R.call_tir(cls.reshape3, (reshape322,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm164 = R.call_tir(cls.layer_norm, (lv7_1, gpt_neox_layers_8_post_attention_layernorm_weight4, gpt_neox_layers_8_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv25 = R.call_tir(cls.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, layer_norm164, gpt_neox_layers_8_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv26 = R.call_tir(cls.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, lv25, gpt_neox_layers_8_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv8_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_8_attention_dense_q_weight4, gpt_neox_layers_8_attention_dense_q_scale4, reshape323, gpt_neox_layers_8_attention_dense_bias4, lv26, lv7_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm165 = R.call_tir(cls.layer_norm, (lv8_1, gpt_neox_layers_9_input_layernorm_weight4, gpt_neox_layers_9_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv27 = R.call_tir(cls.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, layer_norm165, gpt_neox_layers_9_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape324 = R.call_tir(cls.reshape, (lv27,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape325 = R.call_tir(cls.reshape1, (reshape324,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv411 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), reshape325), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape326 = R.call_tir(cls.reshape2, (lv411,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape327 = R.call_tir(cls.reshape3, (reshape326,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm166 = R.call_tir(cls.layer_norm, (lv8_1, gpt_neox_layers_9_post_attention_layernorm_weight4, gpt_neox_layers_9_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv28 = R.call_tir(cls.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, layer_norm166, gpt_neox_layers_9_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv29 = R.call_tir(cls.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, lv28, gpt_neox_layers_9_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv9_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_9_attention_dense_q_weight4, gpt_neox_layers_9_attention_dense_q_scale4, reshape327, gpt_neox_layers_9_attention_dense_bias4, lv29, lv8_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm167 = R.call_tir(cls.layer_norm, (lv9_1, gpt_neox_layers_10_input_layernorm_weight4, gpt_neox_layers_10_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv30 = R.call_tir(cls.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, layer_norm167, gpt_neox_layers_10_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape328 = R.call_tir(cls.reshape, (lv30,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape329 = R.call_tir(cls.reshape1, (reshape328,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv416 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), reshape329), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape330 = R.call_tir(cls.reshape2, (lv416,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape331 = R.call_tir(cls.reshape3, (reshape330,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm168 = R.call_tir(cls.layer_norm, (lv9_1, gpt_neox_layers_10_post_attention_layernorm_weight4, gpt_neox_layers_10_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv31 = R.call_tir(cls.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, layer_norm168, gpt_neox_layers_10_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv32 = R.call_tir(cls.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, lv31, gpt_neox_layers_10_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv10_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_10_attention_dense_q_weight4, gpt_neox_layers_10_attention_dense_q_scale4, reshape331, gpt_neox_layers_10_attention_dense_bias4, lv32, lv9_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm169 = R.call_tir(cls.layer_norm, (lv10_1, gpt_neox_layers_11_input_layernorm_weight4, gpt_neox_layers_11_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv33 = R.call_tir(cls.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, layer_norm169, gpt_neox_layers_11_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape332 = R.call_tir(cls.reshape, (lv33,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape333 = R.call_tir(cls.reshape1, (reshape332,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv421 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), reshape333), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape334 = R.call_tir(cls.reshape2, (lv421,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape335 = R.call_tir(cls.reshape3, (reshape334,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm170 = R.call_tir(cls.layer_norm, (lv10_1, gpt_neox_layers_11_post_attention_layernorm_weight4, gpt_neox_layers_11_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv34 = R.call_tir(cls.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, layer_norm170, gpt_neox_layers_11_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv35 = R.call_tir(cls.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, lv34, gpt_neox_layers_11_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv11_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_11_attention_dense_q_weight4, gpt_neox_layers_11_attention_dense_q_scale4, reshape335, gpt_neox_layers_11_attention_dense_bias4, lv35, lv10_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm171 = R.call_tir(cls.layer_norm, (lv11_1, gpt_neox_layers_12_input_layernorm_weight4, gpt_neox_layers_12_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv36 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul_add, (gpt_neox_layers_12_attention_query_key_value_q_weight4, gpt_neox_layers_12_attention_query_key_value_q_scale4, layer_norm171, gpt_neox_layers_12_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape336 = R.call_tir(cls.reshape, (lv36,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape337 = R.call_tir(cls.reshape1, (reshape336,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv426 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape337), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape338 = R.call_tir(cls.reshape2, (lv426,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape339 = R.call_tir(cls.reshape3, (reshape338,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm172 = R.call_tir(cls.layer_norm, (lv11_1, gpt_neox_layers_12_post_attention_layernorm_weight4, gpt_neox_layers_12_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv37 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul2_add2_gelu_cast, (gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale4, layer_norm172, gpt_neox_layers_12_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv38 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul3_add3_cast1, (gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale4, lv37, gpt_neox_layers_12_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv12_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_12_attention_dense_q_weight4, gpt_neox_layers_12_attention_dense_q_scale4, reshape339, gpt_neox_layers_12_attention_dense_bias4, lv38, lv11_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm173 = R.call_tir(cls.layer_norm, (lv12_1, gpt_neox_layers_13_input_layernorm_weight4, gpt_neox_layers_13_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv39 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul_add, (gpt_neox_layers_13_attention_query_key_value_q_weight4, gpt_neox_layers_13_attention_query_key_value_q_scale4, layer_norm173, gpt_neox_layers_13_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape340 = R.call_tir(cls.reshape, (lv39,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape341 = R.call_tir(cls.reshape1, (reshape340,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv431 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape341), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape342 = R.call_tir(cls.reshape2, (lv431,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape343 = R.call_tir(cls.reshape3, (reshape342,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm174 = R.call_tir(cls.layer_norm, (lv12_1, gpt_neox_layers_13_post_attention_layernorm_weight4, gpt_neox_layers_13_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv40 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul2_add2_gelu_cast, (gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale4, layer_norm174, gpt_neox_layers_13_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv41 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul3_add3_cast1, (gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale4, lv40, gpt_neox_layers_13_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv13_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_13_attention_dense_q_weight4, gpt_neox_layers_13_attention_dense_q_scale4, reshape343, gpt_neox_layers_13_attention_dense_bias4, lv41, lv12_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm175 = R.call_tir(cls.layer_norm, (lv13_1, gpt_neox_layers_14_input_layernorm_weight4, gpt_neox_layers_14_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv42 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul_add, (gpt_neox_layers_14_attention_query_key_value_q_weight4, gpt_neox_layers_14_attention_query_key_value_q_scale4, layer_norm175, gpt_neox_layers_14_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape344 = R.call_tir(cls.reshape, (lv42,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape345 = R.call_tir(cls.reshape1, (reshape344,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv436 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape345), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape346 = R.call_tir(cls.reshape2, (lv436,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape347 = R.call_tir(cls.reshape3, (reshape346,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm176 = R.call_tir(cls.layer_norm, (lv13_1, gpt_neox_layers_14_post_attention_layernorm_weight4, gpt_neox_layers_14_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv43 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul2_add2_gelu_cast, (gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale4, layer_norm176, gpt_neox_layers_14_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv44 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul3_add3_cast1, (gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale4, lv43, gpt_neox_layers_14_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv14_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_14_attention_dense_q_weight4, gpt_neox_layers_14_attention_dense_q_scale4, reshape347, gpt_neox_layers_14_attention_dense_bias4, lv44, lv13_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm177 = R.call_tir(cls.layer_norm, (lv14_1, gpt_neox_layers_15_input_layernorm_weight4, gpt_neox_layers_15_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv45 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul_add, (gpt_neox_layers_15_attention_query_key_value_q_weight4, gpt_neox_layers_15_attention_query_key_value_q_scale4, layer_norm177, gpt_neox_layers_15_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape348 = R.call_tir(cls.reshape, (lv45,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape349 = R.call_tir(cls.reshape1, (reshape348,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv441 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape349), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape350 = R.call_tir(cls.reshape2, (lv441,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape351 = R.call_tir(cls.reshape3, (reshape350,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm178 = R.call_tir(cls.layer_norm, (lv14_1, gpt_neox_layers_15_post_attention_layernorm_weight4, gpt_neox_layers_15_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv46 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul2_add2_gelu_cast, (gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale4, layer_norm178, gpt_neox_layers_15_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv47 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul3_add3_cast1, (gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale4, lv46, gpt_neox_layers_15_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv15_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_15_attention_dense_q_weight4, gpt_neox_layers_15_attention_dense_q_scale4, reshape351, gpt_neox_layers_15_attention_dense_bias4, lv47, lv14_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm179 = R.call_tir(cls.layer_norm, (lv15_1, gpt_neox_layers_16_input_layernorm_weight4, gpt_neox_layers_16_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv48 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul_add, (gpt_neox_layers_16_attention_query_key_value_q_weight4, gpt_neox_layers_16_attention_query_key_value_q_scale4, layer_norm179, gpt_neox_layers_16_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape352 = R.call_tir(cls.reshape, (lv48,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape353 = R.call_tir(cls.reshape1, (reshape352,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv446 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), reshape353), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape354 = R.call_tir(cls.reshape2, (lv446,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape355 = R.call_tir(cls.reshape3, (reshape354,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm180 = R.call_tir(cls.layer_norm, (lv15_1, gpt_neox_layers_16_post_attention_layernorm_weight4, gpt_neox_layers_16_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv49 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul2_add2_gelu_cast, (gpt_neox_layers_16_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_16_mlp_dense_h_to_4h_q_scale4, layer_norm180, gpt_neox_layers_16_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv50 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul3_add3_cast1, (gpt_neox_layers_16_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_16_mlp_dense_4h_to_h_q_scale4, lv49, gpt_neox_layers_16_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv16_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_16_attention_dense_q_weight4, gpt_neox_layers_16_attention_dense_q_scale4, reshape355, gpt_neox_layers_16_attention_dense_bias4, lv50, lv15_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm181 = R.call_tir(cls.layer_norm, (lv16_1, gpt_neox_layers_17_input_layernorm_weight4, gpt_neox_layers_17_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv51 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul_add, (gpt_neox_layers_17_attention_query_key_value_q_weight4, gpt_neox_layers_17_attention_query_key_value_q_scale4, layer_norm181, gpt_neox_layers_17_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape356 = R.call_tir(cls.reshape, (lv51,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape357 = R.call_tir(cls.reshape1, (reshape356,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv451 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), reshape357), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape358 = R.call_tir(cls.reshape2, (lv451,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape359 = R.call_tir(cls.reshape3, (reshape358,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm182 = R.call_tir(cls.layer_norm, (lv16_1, gpt_neox_layers_17_post_attention_layernorm_weight4, gpt_neox_layers_17_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv52 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul2_add2_gelu_cast, (gpt_neox_layers_17_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_17_mlp_dense_h_to_4h_q_scale4, layer_norm182, gpt_neox_layers_17_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv53 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul3_add3_cast1, (gpt_neox_layers_17_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_17_mlp_dense_4h_to_h_q_scale4, lv52, gpt_neox_layers_17_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv17_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_17_attention_dense_q_weight4, gpt_neox_layers_17_attention_dense_q_scale4, reshape359, gpt_neox_layers_17_attention_dense_bias4, lv53, lv16_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm183 = R.call_tir(cls.layer_norm, (lv17_1, gpt_neox_layers_18_input_layernorm_weight4, gpt_neox_layers_18_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv54 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul_add, (gpt_neox_layers_18_attention_query_key_value_q_weight4, gpt_neox_layers_18_attention_query_key_value_q_scale4, layer_norm183, gpt_neox_layers_18_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape360 = R.call_tir(cls.reshape, (lv54,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape361 = R.call_tir(cls.reshape1, (reshape360,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv456 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), reshape361), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape362 = R.call_tir(cls.reshape2, (lv456,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape363 = R.call_tir(cls.reshape3, (reshape362,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm184 = R.call_tir(cls.layer_norm, (lv17_1, gpt_neox_layers_18_post_attention_layernorm_weight4, gpt_neox_layers_18_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv55 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul2_add2_gelu_cast, (gpt_neox_layers_18_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_18_mlp_dense_h_to_4h_q_scale4, layer_norm184, gpt_neox_layers_18_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv56 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul3_add3_cast1, (gpt_neox_layers_18_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_18_mlp_dense_4h_to_h_q_scale4, lv55, gpt_neox_layers_18_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv18_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_18_attention_dense_q_weight4, gpt_neox_layers_18_attention_dense_q_scale4, reshape363, gpt_neox_layers_18_attention_dense_bias4, lv56, lv17_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm185 = R.call_tir(cls.layer_norm, (lv18_1, gpt_neox_layers_19_input_layernorm_weight4, gpt_neox_layers_19_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv57 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul_add, (gpt_neox_layers_19_attention_query_key_value_q_weight4, gpt_neox_layers_19_attention_query_key_value_q_scale4, layer_norm185, gpt_neox_layers_19_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape364 = R.call_tir(cls.reshape, (lv57,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape365 = R.call_tir(cls.reshape1, (reshape364,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv461 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), reshape365), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape366 = R.call_tir(cls.reshape2, (lv461,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape367 = R.call_tir(cls.reshape3, (reshape366,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm186 = R.call_tir(cls.layer_norm, (lv18_1, gpt_neox_layers_19_post_attention_layernorm_weight4, gpt_neox_layers_19_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv58 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul2_add2_gelu_cast, (gpt_neox_layers_19_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_19_mlp_dense_h_to_4h_q_scale4, layer_norm186, gpt_neox_layers_19_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv59 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul3_add3_cast1, (gpt_neox_layers_19_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_19_mlp_dense_4h_to_h_q_scale4, lv58, gpt_neox_layers_19_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv19_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_19_attention_dense_q_weight4, gpt_neox_layers_19_attention_dense_q_scale4, reshape367, gpt_neox_layers_19_attention_dense_bias4, lv59, lv18_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm187 = R.call_tir(cls.layer_norm, (lv19_1, gpt_neox_layers_20_input_layernorm_weight4, gpt_neox_layers_20_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv60 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul_add, (gpt_neox_layers_20_attention_query_key_value_q_weight4, gpt_neox_layers_20_attention_query_key_value_q_scale4, layer_norm187, gpt_neox_layers_20_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape368 = R.call_tir(cls.reshape, (lv60,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape369 = R.call_tir(cls.reshape1, (reshape368,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv466 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), reshape369), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape370 = R.call_tir(cls.reshape2, (lv466,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape371 = R.call_tir(cls.reshape3, (reshape370,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm188 = R.call_tir(cls.layer_norm, (lv19_1, gpt_neox_layers_20_post_attention_layernorm_weight4, gpt_neox_layers_20_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv61 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul2_add2_gelu_cast, (gpt_neox_layers_20_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_20_mlp_dense_h_to_4h_q_scale4, layer_norm188, gpt_neox_layers_20_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv62 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul3_add3_cast1, (gpt_neox_layers_20_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_20_mlp_dense_4h_to_h_q_scale4, lv61, gpt_neox_layers_20_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv20_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_20_attention_dense_q_weight4, gpt_neox_layers_20_attention_dense_q_scale4, reshape371, gpt_neox_layers_20_attention_dense_bias4, lv62, lv19_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm189 = R.call_tir(cls.layer_norm, (lv20_1, gpt_neox_layers_21_input_layernorm_weight4, gpt_neox_layers_21_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv63 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul_add, (gpt_neox_layers_21_attention_query_key_value_q_weight4, gpt_neox_layers_21_attention_query_key_value_q_scale4, layer_norm189, gpt_neox_layers_21_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape372 = R.call_tir(cls.reshape, (lv63,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape373 = R.call_tir(cls.reshape1, (reshape372,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv471 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), reshape373), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape374 = R.call_tir(cls.reshape2, (lv471,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape375 = R.call_tir(cls.reshape3, (reshape374,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm190 = R.call_tir(cls.layer_norm, (lv20_1, gpt_neox_layers_21_post_attention_layernorm_weight4, gpt_neox_layers_21_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv64 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul2_add2_gelu_cast, (gpt_neox_layers_21_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_21_mlp_dense_h_to_4h_q_scale4, layer_norm190, gpt_neox_layers_21_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv65 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul3_add3_cast1, (gpt_neox_layers_21_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_21_mlp_dense_4h_to_h_q_scale4, lv64, gpt_neox_layers_21_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv21_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_21_attention_dense_q_weight4, gpt_neox_layers_21_attention_dense_q_scale4, reshape375, gpt_neox_layers_21_attention_dense_bias4, lv65, lv20_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm191 = R.call_tir(cls.layer_norm, (lv21_1, gpt_neox_layers_22_input_layernorm_weight4, gpt_neox_layers_22_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv66 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul_add, (gpt_neox_layers_22_attention_query_key_value_q_weight4, gpt_neox_layers_22_attention_query_key_value_q_scale4, layer_norm191, gpt_neox_layers_22_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape376 = R.call_tir(cls.reshape, (lv66,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape377 = R.call_tir(cls.reshape1, (reshape376,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv476 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), reshape377), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape378 = R.call_tir(cls.reshape2, (lv476,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape379 = R.call_tir(cls.reshape3, (reshape378,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm192 = R.call_tir(cls.layer_norm, (lv21_1, gpt_neox_layers_22_post_attention_layernorm_weight4, gpt_neox_layers_22_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv67 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul2_add2_gelu_cast, (gpt_neox_layers_22_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_22_mlp_dense_h_to_4h_q_scale4, layer_norm192, gpt_neox_layers_22_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv68 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul3_add3_cast1, (gpt_neox_layers_22_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_22_mlp_dense_4h_to_h_q_scale4, lv67, gpt_neox_layers_22_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv22_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_22_attention_dense_q_weight4, gpt_neox_layers_22_attention_dense_q_scale4, reshape379, gpt_neox_layers_22_attention_dense_bias4, lv68, lv21_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm193 = R.call_tir(cls.layer_norm, (lv22_1, gpt_neox_layers_23_input_layernorm_weight4, gpt_neox_layers_23_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv69 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul_add, (gpt_neox_layers_23_attention_query_key_value_q_weight4, gpt_neox_layers_23_attention_query_key_value_q_scale4, layer_norm193, gpt_neox_layers_23_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 3072), dtype="float16")) reshape380 = R.call_tir(cls.reshape, (lv69,), out_sinfo=R.Tensor((batch_size, 1, 48, 64), dtype="float16")) reshape381 = R.call_tir(cls.reshape1, (reshape380,), out_sinfo=R.Tensor((batch_size, 48, 64), dtype="float16")) lv481 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), reshape381), out_sinfo=R.Tensor((batch_size, 16, 64), dtype="float16")) reshape382 = R.call_tir(cls.reshape2, (lv481,), out_sinfo=R.Tensor((batch_size, 1, 16, 64), dtype="float16")) reshape383 = R.call_tir(cls.reshape3, (reshape382,), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm194 = R.call_tir(cls.layer_norm, (lv22_1, gpt_neox_layers_23_post_attention_layernorm_weight4, gpt_neox_layers_23_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv70 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul2_add2_gelu_cast, (gpt_neox_layers_23_mlp_dense_h_to_4h_q_weight4, gpt_neox_layers_23_mlp_dense_h_to_4h_q_scale4, layer_norm194, gpt_neox_layers_23_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 4096), dtype="float16")) lv71 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul3_add3_cast1, (gpt_neox_layers_23_mlp_dense_4h_to_h_q_weight4, gpt_neox_layers_23_mlp_dense_4h_to_h_q_scale4, lv70, gpt_neox_layers_23_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv23_1 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul1_add1_add4_add4, (gpt_neox_layers_23_attention_dense_q_weight4, gpt_neox_layers_23_attention_dense_q_scale4, reshape383, gpt_neox_layers_23_attention_dense_bias4, lv71, lv22_1), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) layer_norm195 = R.call_tir(cls.layer_norm, (lv23_1, gpt_neox_final_layer_norm_weight4, gpt_neox_final_layer_norm_bias4), out_sinfo=R.Tensor((batch_size, 1, 1024), dtype="float16")) lv_2 = R.call_tir(cls.fused_dequantize_fused_NT_matmul4_cast2, (embed_out_q_weight4, embed_out_q_scale4, layer_norm195), out_sinfo=R.Tensor((batch_size, 1, vocab_size), dtype="float32")) gv4: R.Tuple(R.Tensor((batch_size, 1, vocab_size), dtype="float32"), R.Object) = lv_2, paged_kv_cache R.output(gv4) return gv4 @R.function def batch_prefill(input_embeds: R.Tensor((1, "seq_len", 1024), dtype="float16"), logit_positions: R.Tensor(("batch_size",), dtype="int32"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 128), dtype="uint32"), R.Tensor(("vocab_size", 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor(("vocab_size", 128), dtype="uint32"), R.Tensor(("vocab_size", 32), 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.memory_plan_dynamic_func_output": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) cls = Module with R.dataflow(): gpt_neox_layers_0_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[2] gpt_neox_layers_0_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[3] gpt_neox_layers_0_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[4] gpt_neox_layers_0_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[5] gpt_neox_layers_0_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[6] gpt_neox_layers_0_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[8] gpt_neox_layers_0_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[9] gpt_neox_layers_0_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[10] gpt_neox_layers_0_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[11] gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[12] gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[13] gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[14] gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[15] gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[16] gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[17] gpt_neox_layers_1_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[18] gpt_neox_layers_1_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[19] gpt_neox_layers_1_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[20] gpt_neox_layers_1_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[21] gpt_neox_layers_1_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[22] gpt_neox_layers_1_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[23] gpt_neox_layers_1_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[24] gpt_neox_layers_1_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[25] gpt_neox_layers_1_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[26] gpt_neox_layers_1_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[27] gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[28] gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[29] gpt_neox_layers_1_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[30] gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[31] gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[32] gpt_neox_layers_1_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[33] gpt_neox_layers_2_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[34] gpt_neox_layers_2_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[35] gpt_neox_layers_2_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[36] gpt_neox_layers_2_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[37] gpt_neox_layers_2_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[38] gpt_neox_layers_2_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[39] gpt_neox_layers_2_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[40] gpt_neox_layers_2_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[41] gpt_neox_layers_2_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[42] gpt_neox_layers_2_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[43] gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[44] gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[45] gpt_neox_layers_2_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[46] gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[47] gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[48] gpt_neox_layers_2_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[49] gpt_neox_layers_3_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[50] gpt_neox_layers_3_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[51] gpt_neox_layers_3_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[52] gpt_neox_layers_3_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[53] gpt_neox_layers_3_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[54] gpt_neox_layers_3_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[55] gpt_neox_layers_3_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[56] gpt_neox_layers_3_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[57] gpt_neox_layers_3_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[58] gpt_neox_layers_3_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[59] gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[60] gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[61] gpt_neox_layers_3_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[62] gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[63] gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[64] gpt_neox_layers_3_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[65] gpt_neox_layers_4_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[66] gpt_neox_layers_4_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[67] gpt_neox_layers_4_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[68] gpt_neox_layers_4_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[69] gpt_neox_layers_4_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[70] gpt_neox_layers_4_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[71] gpt_neox_layers_4_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[72] gpt_neox_layers_4_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[73] gpt_neox_layers_4_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[74] gpt_neox_layers_4_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[75] gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[76] gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[77] gpt_neox_layers_4_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[78] gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[79] gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[80] gpt_neox_layers_4_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[81] gpt_neox_layers_5_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[82] gpt_neox_layers_5_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[83] gpt_neox_layers_5_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[84] gpt_neox_layers_5_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[85] gpt_neox_layers_5_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[86] gpt_neox_layers_5_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[87] gpt_neox_layers_5_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[88] gpt_neox_layers_5_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[89] gpt_neox_layers_5_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[90] gpt_neox_layers_5_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[91] gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[92] gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[93] gpt_neox_layers_5_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[94] gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[95] gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[96] gpt_neox_layers_5_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[97] gpt_neox_layers_6_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[98] gpt_neox_layers_6_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[99] gpt_neox_layers_6_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[100] gpt_neox_layers_6_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[101] gpt_neox_layers_6_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[102] gpt_neox_layers_6_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[103] gpt_neox_layers_6_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[104] gpt_neox_layers_6_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[105] gpt_neox_layers_6_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[106] gpt_neox_layers_6_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[107] gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[108] gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[109] gpt_neox_layers_6_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[110] gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[111] gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[112] gpt_neox_layers_6_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[113] gpt_neox_layers_7_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[114] gpt_neox_layers_7_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[115] gpt_neox_layers_7_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[116] gpt_neox_layers_7_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[117] gpt_neox_layers_7_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[118] gpt_neox_layers_7_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[119] gpt_neox_layers_7_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[120] gpt_neox_layers_7_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[121] gpt_neox_layers_7_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[122] gpt_neox_layers_7_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[123] gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[124] gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[125] gpt_neox_layers_7_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[126] gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[127] gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[128] gpt_neox_layers_7_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[129] gpt_neox_layers_8_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[130] gpt_neox_layers_8_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[131] gpt_neox_layers_8_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[132] gpt_neox_layers_8_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[133] gpt_neox_layers_8_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[134] gpt_neox_layers_8_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[135] gpt_neox_layers_8_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[136] gpt_neox_layers_8_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[137] gpt_neox_layers_8_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[138] gpt_neox_layers_8_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[139] gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[140] gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[141] gpt_neox_layers_8_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[142] gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[143] gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[144] gpt_neox_layers_8_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[145] gpt_neox_layers_9_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[146] gpt_neox_layers_9_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[147] gpt_neox_layers_9_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[148] gpt_neox_layers_9_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[149] gpt_neox_layers_9_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[150] gpt_neox_layers_9_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[151] gpt_neox_layers_9_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[152] gpt_neox_layers_9_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[153] gpt_neox_layers_9_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[154] gpt_neox_layers_9_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[155] gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[156] gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[157] gpt_neox_layers_9_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[158] gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[159] gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[160] gpt_neox_layers_9_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[161] gpt_neox_layers_10_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[162] gpt_neox_layers_10_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[163] gpt_neox_layers_10_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[164] gpt_neox_layers_10_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[165] gpt_neox_layers_10_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[166] gpt_neox_layers_10_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[167] gpt_neox_layers_10_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[168] gpt_neox_layers_10_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[169] gpt_neox_layers_10_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[170] gpt_neox_layers_10_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[171] gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[172] gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[173] gpt_neox_layers_10_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[174] gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[175] gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[176] gpt_neox_layers_10_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[177] gpt_neox_layers_11_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[178] gpt_neox_layers_11_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[179] gpt_neox_layers_11_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[180] gpt_neox_layers_11_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[181] gpt_neox_layers_11_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[182] gpt_neox_layers_11_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[183] gpt_neox_layers_11_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[184] gpt_neox_layers_11_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[185] gpt_neox_layers_11_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[186] gpt_neox_layers_11_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[187] gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[188] gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[189] gpt_neox_layers_11_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[190] gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[191] gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[192] gpt_neox_layers_11_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[193] gpt_neox_layers_12_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[194] gpt_neox_layers_12_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[195] gpt_neox_layers_12_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[196] gpt_neox_layers_12_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[197] gpt_neox_layers_12_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[198] gpt_neox_layers_12_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[199] gpt_neox_layers_12_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[200] gpt_neox_layers_12_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[201] gpt_neox_layers_12_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[202] gpt_neox_layers_12_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[203] gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[204] gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[205] gpt_neox_layers_12_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[206] gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[207] gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[208] gpt_neox_layers_12_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[209] gpt_neox_layers_13_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[210] gpt_neox_layers_13_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[211] gpt_neox_layers_13_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[212] gpt_neox_layers_13_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[213] gpt_neox_layers_13_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[214] gpt_neox_layers_13_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[215] gpt_neox_layers_13_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[216] gpt_neox_layers_13_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[217] gpt_neox_layers_13_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[218] gpt_neox_layers_13_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[219] gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[220] gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[221] gpt_neox_layers_13_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[222] gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[223] gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[224] gpt_neox_layers_13_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[225] gpt_neox_layers_14_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[226] gpt_neox_layers_14_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[227] gpt_neox_layers_14_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[228] gpt_neox_layers_14_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[229] gpt_neox_layers_14_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[230] gpt_neox_layers_14_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[231] gpt_neox_layers_14_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[232] gpt_neox_layers_14_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[233] gpt_neox_layers_14_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[234] gpt_neox_layers_14_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[235] gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[236] gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[237] gpt_neox_layers_14_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[238] gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[239] gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[240] gpt_neox_layers_14_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[241] gpt_neox_layers_15_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[242] gpt_neox_layers_15_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[243] gpt_neox_layers_15_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[244] gpt_neox_layers_15_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[245] gpt_neox_layers_15_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[246] gpt_neox_layers_15_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[247] gpt_neox_layers_15_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[248] gpt_neox_layers_15_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[249] gpt_neox_layers_15_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[250] gpt_neox_layers_15_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[251] gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[252] gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[253] gpt_neox_layers_15_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[254] gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[255] gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[256] gpt_neox_layers_15_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[257] gpt_neox_layers_16_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[258] gpt_neox_layers_16_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[259] gpt_neox_layers_16_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[260] gpt_neox_layers_16_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[261] gpt_neox_layers_16_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[262] gpt_neox_layers_16_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[263] gpt_neox_layers_16_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[264] gpt_neox_layers_16_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[265] gpt_neox_layers_16_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[266] gpt_neox_layers_16_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[267] gpt_neox_layers_16_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[268] gpt_neox_layers_16_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[269] gpt_neox_layers_16_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[270] gpt_neox_layers_16_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[271] gpt_neox_layers_16_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[272] gpt_neox_layers_16_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[273] gpt_neox_layers_17_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[274] gpt_neox_layers_17_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[275] gpt_neox_layers_17_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[276] gpt_neox_layers_17_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[277] gpt_neox_layers_17_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[278] gpt_neox_layers_17_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[279] gpt_neox_layers_17_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[280] gpt_neox_layers_17_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[281] gpt_neox_layers_17_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[282] gpt_neox_layers_17_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[283] gpt_neox_layers_17_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[284] gpt_neox_layers_17_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[285] gpt_neox_layers_17_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[286] gpt_neox_layers_17_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[287] gpt_neox_layers_17_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[288] gpt_neox_layers_17_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[289] gpt_neox_layers_18_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[290] gpt_neox_layers_18_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[291] gpt_neox_layers_18_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[292] gpt_neox_layers_18_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[293] gpt_neox_layers_18_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[294] gpt_neox_layers_18_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[295] gpt_neox_layers_18_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[296] gpt_neox_layers_18_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[297] gpt_neox_layers_18_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[298] gpt_neox_layers_18_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[299] gpt_neox_layers_18_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[300] gpt_neox_layers_18_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[301] gpt_neox_layers_18_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[302] gpt_neox_layers_18_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[303] gpt_neox_layers_18_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[304] gpt_neox_layers_18_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[305] gpt_neox_layers_19_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[306] gpt_neox_layers_19_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[307] gpt_neox_layers_19_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[308] gpt_neox_layers_19_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[309] gpt_neox_layers_19_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[310] gpt_neox_layers_19_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[311] gpt_neox_layers_19_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[312] gpt_neox_layers_19_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[313] gpt_neox_layers_19_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[314] gpt_neox_layers_19_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[315] gpt_neox_layers_19_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[316] gpt_neox_layers_19_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[317] gpt_neox_layers_19_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[318] gpt_neox_layers_19_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[319] gpt_neox_layers_19_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[320] gpt_neox_layers_19_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[321] gpt_neox_layers_20_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[322] gpt_neox_layers_20_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[323] gpt_neox_layers_20_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[324] gpt_neox_layers_20_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[325] gpt_neox_layers_20_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[326] gpt_neox_layers_20_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[327] gpt_neox_layers_20_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[328] gpt_neox_layers_20_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[329] gpt_neox_layers_20_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[330] gpt_neox_layers_20_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[331] gpt_neox_layers_20_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[332] gpt_neox_layers_20_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[333] gpt_neox_layers_20_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[334] gpt_neox_layers_20_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[335] gpt_neox_layers_20_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[336] gpt_neox_layers_20_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[337] gpt_neox_layers_21_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[338] gpt_neox_layers_21_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[339] gpt_neox_layers_21_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[340] gpt_neox_layers_21_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[341] gpt_neox_layers_21_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[342] gpt_neox_layers_21_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[343] gpt_neox_layers_21_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[344] gpt_neox_layers_21_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[345] gpt_neox_layers_21_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[346] gpt_neox_layers_21_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[347] gpt_neox_layers_21_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[348] gpt_neox_layers_21_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[349] gpt_neox_layers_21_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[350] gpt_neox_layers_21_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[351] gpt_neox_layers_21_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[352] gpt_neox_layers_21_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[353] gpt_neox_layers_22_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[354] gpt_neox_layers_22_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[355] gpt_neox_layers_22_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[356] gpt_neox_layers_22_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[357] gpt_neox_layers_22_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[358] gpt_neox_layers_22_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[359] gpt_neox_layers_22_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[360] gpt_neox_layers_22_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[361] gpt_neox_layers_22_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[362] gpt_neox_layers_22_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[363] gpt_neox_layers_22_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[364] gpt_neox_layers_22_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[365] gpt_neox_layers_22_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[366] gpt_neox_layers_22_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[367] gpt_neox_layers_22_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[368] gpt_neox_layers_22_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[369] gpt_neox_layers_23_input_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[370] gpt_neox_layers_23_input_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[371] gpt_neox_layers_23_post_attention_layernorm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[372] gpt_neox_layers_23_post_attention_layernorm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[373] gpt_neox_layers_23_attention_query_key_value_q_weight3: R.Tensor((3072, 128), dtype="uint32") = packed_params[374] gpt_neox_layers_23_attention_query_key_value_q_scale3: R.Tensor((3072, 32), dtype="float16") = packed_params[375] gpt_neox_layers_23_attention_query_key_value_bias3: R.Tensor((3072,), dtype="float16") = packed_params[376] gpt_neox_layers_23_attention_dense_q_weight3: R.Tensor((1024, 128), dtype="uint32") = packed_params[377] gpt_neox_layers_23_attention_dense_q_scale3: R.Tensor((1024, 32), dtype="float16") = packed_params[378] gpt_neox_layers_23_attention_dense_bias3: R.Tensor((1024,), dtype="float16") = packed_params[379] gpt_neox_layers_23_mlp_dense_h_to_4h_q_weight3: R.Tensor((4096, 128), dtype="uint32") = packed_params[380] gpt_neox_layers_23_mlp_dense_h_to_4h_q_scale3: R.Tensor((4096, 32), dtype="float16") = packed_params[381] gpt_neox_layers_23_mlp_dense_h_to_4h_bias3: R.Tensor((4096,), dtype="float32") = packed_params[382] gpt_neox_layers_23_mlp_dense_4h_to_h_q_weight3: R.Tensor((1024, 512), dtype="uint32") = packed_params[383] gpt_neox_layers_23_mlp_dense_4h_to_h_q_scale3: R.Tensor((1024, 128), dtype="float16") = packed_params[384] gpt_neox_layers_23_mlp_dense_4h_to_h_bias3: R.Tensor((1024,), dtype="float32") = packed_params[385] gpt_neox_final_layer_norm_weight3: R.Tensor((1024,), dtype="float16") = packed_params[386] gpt_neox_final_layer_norm_bias3: R.Tensor((1024,), dtype="float16") = packed_params[387] embed_out_q_weight3: R.Tensor((vocab_size, 128), dtype="uint32") = packed_params[388] embed_out_q_scale3: R.Tensor((vocab_size, 32), dtype="float16") = packed_params[389] layer_norm98 = R.call_tir(cls.layer_norm1, (input_embeds, gpt_neox_layers_0_input_layernorm_weight3, gpt_neox_layers_0_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv72 = R.call_tir(cls.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, layer_norm98, gpt_neox_layers_0_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape192 = R.call_tir(cls.reshape4, (lv72,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape193 = R.call_tir(cls.reshape5, (reshape192,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv245 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape194 = R.call_tir(cls.reshape6, (lv245,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape195 = R.call_tir(cls.reshape7, (reshape194,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm99 = R.call_tir(cls.layer_norm1, (input_embeds, gpt_neox_layers_0_post_attention_layernorm_weight3, gpt_neox_layers_0_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv73 = R.call_tir(cls.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, layer_norm99, gpt_neox_layers_0_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv74 = R.call_tir(cls.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, lv73, gpt_neox_layers_0_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv24 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_0_attention_dense_q_weight3, gpt_neox_layers_0_attention_dense_q_scale3, reshape195, gpt_neox_layers_0_attention_dense_bias3, lv74, input_embeds), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm100 = R.call_tir(cls.layer_norm1, (lv24, gpt_neox_layers_1_input_layernorm_weight3, gpt_neox_layers_1_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv75 = R.call_tir(cls.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, layer_norm100, gpt_neox_layers_1_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape196 = R.call_tir(cls.reshape4, (lv75,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape197 = R.call_tir(cls.reshape5, (reshape196,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv250 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape198 = R.call_tir(cls.reshape6, (lv250,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape199 = R.call_tir(cls.reshape7, (reshape198,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm101 = R.call_tir(cls.layer_norm1, (lv24, gpt_neox_layers_1_post_attention_layernorm_weight3, gpt_neox_layers_1_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv76 = R.call_tir(cls.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, layer_norm101, gpt_neox_layers_1_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv77 = R.call_tir(cls.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, lv76, gpt_neox_layers_1_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv25 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_1_attention_dense_q_weight3, gpt_neox_layers_1_attention_dense_q_scale3, reshape199, gpt_neox_layers_1_attention_dense_bias3, lv77, lv24), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm102 = R.call_tir(cls.layer_norm1, (lv25, gpt_neox_layers_2_input_layernorm_weight3, gpt_neox_layers_2_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv78 = R.call_tir(cls.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, layer_norm102, gpt_neox_layers_2_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape200 = R.call_tir(cls.reshape4, (lv78,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape201 = R.call_tir(cls.reshape5, (reshape200,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv255 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape202 = R.call_tir(cls.reshape6, (lv255,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape203 = R.call_tir(cls.reshape7, (reshape202,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm103 = R.call_tir(cls.layer_norm1, (lv25, gpt_neox_layers_2_post_attention_layernorm_weight3, gpt_neox_layers_2_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv79 = R.call_tir(cls.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, layer_norm103, gpt_neox_layers_2_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv80 = R.call_tir(cls.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, lv79, gpt_neox_layers_2_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv26 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_2_attention_dense_q_weight3, gpt_neox_layers_2_attention_dense_q_scale3, reshape203, gpt_neox_layers_2_attention_dense_bias3, lv80, lv25), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm104 = R.call_tir(cls.layer_norm1, (lv26, gpt_neox_layers_3_input_layernorm_weight3, gpt_neox_layers_3_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv81 = R.call_tir(cls.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, layer_norm104, gpt_neox_layers_3_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape204 = R.call_tir(cls.reshape4, (lv81,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape205 = R.call_tir(cls.reshape5, (reshape204,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv260 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape206 = R.call_tir(cls.reshape6, (lv260,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape207 = R.call_tir(cls.reshape7, (reshape206,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm105 = R.call_tir(cls.layer_norm1, (lv26, gpt_neox_layers_3_post_attention_layernorm_weight3, gpt_neox_layers_3_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv82 = R.call_tir(cls.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, layer_norm105, gpt_neox_layers_3_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv83 = R.call_tir(cls.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, lv82, gpt_neox_layers_3_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv27 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_3_attention_dense_q_weight3, gpt_neox_layers_3_attention_dense_q_scale3, reshape207, gpt_neox_layers_3_attention_dense_bias3, lv83, lv26), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm106 = R.call_tir(cls.layer_norm1, (lv27, gpt_neox_layers_4_input_layernorm_weight3, gpt_neox_layers_4_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv84 = R.call_tir(cls.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, layer_norm106, gpt_neox_layers_4_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape208 = R.call_tir(cls.reshape4, (lv84,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape209 = R.call_tir(cls.reshape5, (reshape208,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv265 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape210 = R.call_tir(cls.reshape6, (lv265,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape211 = R.call_tir(cls.reshape7, (reshape210,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm107 = R.call_tir(cls.layer_norm1, (lv27, gpt_neox_layers_4_post_attention_layernorm_weight3, gpt_neox_layers_4_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv85 = R.call_tir(cls.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, layer_norm107, gpt_neox_layers_4_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv86 = R.call_tir(cls.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, lv85, gpt_neox_layers_4_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv28 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_4_attention_dense_q_weight3, gpt_neox_layers_4_attention_dense_q_scale3, reshape211, gpt_neox_layers_4_attention_dense_bias3, lv86, lv27), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm108 = R.call_tir(cls.layer_norm1, (lv28, gpt_neox_layers_5_input_layernorm_weight3, gpt_neox_layers_5_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv87 = R.call_tir(cls.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, layer_norm108, gpt_neox_layers_5_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape212 = R.call_tir(cls.reshape4, (lv87,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape213 = R.call_tir(cls.reshape5, (reshape212,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv270 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape214 = R.call_tir(cls.reshape6, (lv270,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape215 = R.call_tir(cls.reshape7, (reshape214,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm109 = R.call_tir(cls.layer_norm1, (lv28, gpt_neox_layers_5_post_attention_layernorm_weight3, gpt_neox_layers_5_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv88 = R.call_tir(cls.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, layer_norm109, gpt_neox_layers_5_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv89 = R.call_tir(cls.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, lv88, gpt_neox_layers_5_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv29 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_5_attention_dense_q_weight3, gpt_neox_layers_5_attention_dense_q_scale3, reshape215, gpt_neox_layers_5_attention_dense_bias3, lv89, lv28), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm110 = R.call_tir(cls.layer_norm1, (lv29, gpt_neox_layers_6_input_layernorm_weight3, gpt_neox_layers_6_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv90 = R.call_tir(cls.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, layer_norm110, gpt_neox_layers_6_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape216 = R.call_tir(cls.reshape4, (lv90,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape217 = R.call_tir(cls.reshape5, (reshape216,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv275 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape218 = R.call_tir(cls.reshape6, (lv275,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape219 = R.call_tir(cls.reshape7, (reshape218,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm111 = R.call_tir(cls.layer_norm1, (lv29, gpt_neox_layers_6_post_attention_layernorm_weight3, gpt_neox_layers_6_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv91 = R.call_tir(cls.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, layer_norm111, gpt_neox_layers_6_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv92 = R.call_tir(cls.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, lv91, gpt_neox_layers_6_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv30 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_6_attention_dense_q_weight3, gpt_neox_layers_6_attention_dense_q_scale3, reshape219, gpt_neox_layers_6_attention_dense_bias3, lv92, lv29), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm112 = R.call_tir(cls.layer_norm1, (lv30, gpt_neox_layers_7_input_layernorm_weight3, gpt_neox_layers_7_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv93 = R.call_tir(cls.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, layer_norm112, gpt_neox_layers_7_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape220 = R.call_tir(cls.reshape4, (lv93,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape221 = R.call_tir(cls.reshape5, (reshape220,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv280 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape222 = R.call_tir(cls.reshape6, (lv280,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape223 = R.call_tir(cls.reshape7, (reshape222,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm113 = R.call_tir(cls.layer_norm1, (lv30, gpt_neox_layers_7_post_attention_layernorm_weight3, gpt_neox_layers_7_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv94 = R.call_tir(cls.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, layer_norm113, gpt_neox_layers_7_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv95 = R.call_tir(cls.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, lv94, gpt_neox_layers_7_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv31 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_7_attention_dense_q_weight3, gpt_neox_layers_7_attention_dense_q_scale3, reshape223, gpt_neox_layers_7_attention_dense_bias3, lv95, lv30), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm114 = R.call_tir(cls.layer_norm1, (lv31, gpt_neox_layers_8_input_layernorm_weight3, gpt_neox_layers_8_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv96 = R.call_tir(cls.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, layer_norm114, gpt_neox_layers_8_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape224 = R.call_tir(cls.reshape4, (lv96,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape225 = R.call_tir(cls.reshape5, (reshape224,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv285 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape226 = R.call_tir(cls.reshape6, (lv285,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape227 = R.call_tir(cls.reshape7, (reshape226,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm115 = R.call_tir(cls.layer_norm1, (lv31, gpt_neox_layers_8_post_attention_layernorm_weight3, gpt_neox_layers_8_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv97 = R.call_tir(cls.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, layer_norm115, gpt_neox_layers_8_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv98 = R.call_tir(cls.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, lv97, gpt_neox_layers_8_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv32 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_8_attention_dense_q_weight3, gpt_neox_layers_8_attention_dense_q_scale3, reshape227, gpt_neox_layers_8_attention_dense_bias3, lv98, lv31), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm116 = R.call_tir(cls.layer_norm1, (lv32, gpt_neox_layers_9_input_layernorm_weight3, gpt_neox_layers_9_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv99 = R.call_tir(cls.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, layer_norm116, gpt_neox_layers_9_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape228 = R.call_tir(cls.reshape4, (lv99,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape229 = R.call_tir(cls.reshape5, (reshape228,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv290 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape230 = R.call_tir(cls.reshape6, (lv290,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape231 = R.call_tir(cls.reshape7, (reshape230,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm117 = R.call_tir(cls.layer_norm1, (lv32, gpt_neox_layers_9_post_attention_layernorm_weight3, gpt_neox_layers_9_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv100 = R.call_tir(cls.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, layer_norm117, gpt_neox_layers_9_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv101 = R.call_tir(cls.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, lv100, gpt_neox_layers_9_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv33 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_9_attention_dense_q_weight3, gpt_neox_layers_9_attention_dense_q_scale3, reshape231, gpt_neox_layers_9_attention_dense_bias3, lv101, lv32), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm118 = R.call_tir(cls.layer_norm1, (lv33, gpt_neox_layers_10_input_layernorm_weight3, gpt_neox_layers_10_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv102 = R.call_tir(cls.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, layer_norm118, gpt_neox_layers_10_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape232 = R.call_tir(cls.reshape4, (lv102,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape233 = R.call_tir(cls.reshape5, (reshape232,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv295 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape234 = R.call_tir(cls.reshape6, (lv295,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape235 = R.call_tir(cls.reshape7, (reshape234,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm119 = R.call_tir(cls.layer_norm1, (lv33, gpt_neox_layers_10_post_attention_layernorm_weight3, gpt_neox_layers_10_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv103 = R.call_tir(cls.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, layer_norm119, gpt_neox_layers_10_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv104 = R.call_tir(cls.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, lv103, gpt_neox_layers_10_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv34 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_10_attention_dense_q_weight3, gpt_neox_layers_10_attention_dense_q_scale3, reshape235, gpt_neox_layers_10_attention_dense_bias3, lv104, lv33), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm120 = R.call_tir(cls.layer_norm1, (lv34, gpt_neox_layers_11_input_layernorm_weight3, gpt_neox_layers_11_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv105 = R.call_tir(cls.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, layer_norm120, gpt_neox_layers_11_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape236 = R.call_tir(cls.reshape4, (lv105,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape237 = R.call_tir(cls.reshape5, (reshape236,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv300 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape238 = R.call_tir(cls.reshape6, (lv300,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape239 = R.call_tir(cls.reshape7, (reshape238,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm121 = R.call_tir(cls.layer_norm1, (lv34, gpt_neox_layers_11_post_attention_layernorm_weight3, gpt_neox_layers_11_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv106 = R.call_tir(cls.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, layer_norm121, gpt_neox_layers_11_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv107 = R.call_tir(cls.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, lv106, gpt_neox_layers_11_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv35 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_11_attention_dense_q_weight3, gpt_neox_layers_11_attention_dense_q_scale3, reshape239, gpt_neox_layers_11_attention_dense_bias3, lv107, lv34), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm122 = R.call_tir(cls.layer_norm1, (lv35, gpt_neox_layers_12_input_layernorm_weight3, gpt_neox_layers_12_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv108 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_12_attention_query_key_value_q_weight3, gpt_neox_layers_12_attention_query_key_value_q_scale3, layer_norm122, gpt_neox_layers_12_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape240 = R.call_tir(cls.reshape4, (lv108,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape241 = R.call_tir(cls.reshape5, (reshape240,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv305 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape241), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape242 = R.call_tir(cls.reshape6, (lv305,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape243 = R.call_tir(cls.reshape7, (reshape242,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm123 = R.call_tir(cls.layer_norm1, (lv35, gpt_neox_layers_12_post_attention_layernorm_weight3, gpt_neox_layers_12_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv109 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale3, layer_norm123, gpt_neox_layers_12_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv110 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale3, lv109, gpt_neox_layers_12_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv36 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_12_attention_dense_q_weight3, gpt_neox_layers_12_attention_dense_q_scale3, reshape243, gpt_neox_layers_12_attention_dense_bias3, lv110, lv35), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm124 = R.call_tir(cls.layer_norm1, (lv36, gpt_neox_layers_13_input_layernorm_weight3, gpt_neox_layers_13_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv111 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_13_attention_query_key_value_q_weight3, gpt_neox_layers_13_attention_query_key_value_q_scale3, layer_norm124, gpt_neox_layers_13_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape244 = R.call_tir(cls.reshape4, (lv111,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape245 = R.call_tir(cls.reshape5, (reshape244,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv310 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape245), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape246 = R.call_tir(cls.reshape6, (lv310,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape247 = R.call_tir(cls.reshape7, (reshape246,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm125 = R.call_tir(cls.layer_norm1, (lv36, gpt_neox_layers_13_post_attention_layernorm_weight3, gpt_neox_layers_13_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv112 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale3, layer_norm125, gpt_neox_layers_13_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv113 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale3, lv112, gpt_neox_layers_13_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv37 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_13_attention_dense_q_weight3, gpt_neox_layers_13_attention_dense_q_scale3, reshape247, gpt_neox_layers_13_attention_dense_bias3, lv113, lv36), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm126 = R.call_tir(cls.layer_norm1, (lv37, gpt_neox_layers_14_input_layernorm_weight3, gpt_neox_layers_14_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv114 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_14_attention_query_key_value_q_weight3, gpt_neox_layers_14_attention_query_key_value_q_scale3, layer_norm126, gpt_neox_layers_14_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape248 = R.call_tir(cls.reshape4, (lv114,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape249 = R.call_tir(cls.reshape5, (reshape248,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv315 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape249), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape250 = R.call_tir(cls.reshape6, (lv315,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape251 = R.call_tir(cls.reshape7, (reshape250,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm127 = R.call_tir(cls.layer_norm1, (lv37, gpt_neox_layers_14_post_attention_layernorm_weight3, gpt_neox_layers_14_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv115 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale3, layer_norm127, gpt_neox_layers_14_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv116 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale3, lv115, gpt_neox_layers_14_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv38 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_14_attention_dense_q_weight3, gpt_neox_layers_14_attention_dense_q_scale3, reshape251, gpt_neox_layers_14_attention_dense_bias3, lv116, lv37), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm128 = R.call_tir(cls.layer_norm1, (lv38, gpt_neox_layers_15_input_layernorm_weight3, gpt_neox_layers_15_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv117 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_15_attention_query_key_value_q_weight3, gpt_neox_layers_15_attention_query_key_value_q_scale3, layer_norm128, gpt_neox_layers_15_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape252 = R.call_tir(cls.reshape4, (lv117,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape253 = R.call_tir(cls.reshape5, (reshape252,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv320 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape253), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape254 = R.call_tir(cls.reshape6, (lv320,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape255 = R.call_tir(cls.reshape7, (reshape254,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm129 = R.call_tir(cls.layer_norm1, (lv38, gpt_neox_layers_15_post_attention_layernorm_weight3, gpt_neox_layers_15_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv118 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale3, layer_norm129, gpt_neox_layers_15_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv119 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale3, lv118, gpt_neox_layers_15_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv39 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_15_attention_dense_q_weight3, gpt_neox_layers_15_attention_dense_q_scale3, reshape255, gpt_neox_layers_15_attention_dense_bias3, lv119, lv38), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm130 = R.call_tir(cls.layer_norm1, (lv39, gpt_neox_layers_16_input_layernorm_weight3, gpt_neox_layers_16_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv120 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_16_attention_query_key_value_q_weight3, gpt_neox_layers_16_attention_query_key_value_q_scale3, layer_norm130, gpt_neox_layers_16_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape256 = R.call_tir(cls.reshape4, (lv120,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape257 = R.call_tir(cls.reshape5, (reshape256,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv325 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), reshape257), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape258 = R.call_tir(cls.reshape6, (lv325,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape259 = R.call_tir(cls.reshape7, (reshape258,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm131 = R.call_tir(cls.layer_norm1, (lv39, gpt_neox_layers_16_post_attention_layernorm_weight3, gpt_neox_layers_16_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv121 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_16_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_16_mlp_dense_h_to_4h_q_scale3, layer_norm131, gpt_neox_layers_16_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv122 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_16_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_16_mlp_dense_4h_to_h_q_scale3, lv121, gpt_neox_layers_16_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv40 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_16_attention_dense_q_weight3, gpt_neox_layers_16_attention_dense_q_scale3, reshape259, gpt_neox_layers_16_attention_dense_bias3, lv122, lv39), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm132 = R.call_tir(cls.layer_norm1, (lv40, gpt_neox_layers_17_input_layernorm_weight3, gpt_neox_layers_17_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv123 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_17_attention_query_key_value_q_weight3, gpt_neox_layers_17_attention_query_key_value_q_scale3, layer_norm132, gpt_neox_layers_17_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape260 = R.call_tir(cls.reshape4, (lv123,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape261 = R.call_tir(cls.reshape5, (reshape260,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv330 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), reshape261), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape262 = R.call_tir(cls.reshape6, (lv330,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape263 = R.call_tir(cls.reshape7, (reshape262,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm133 = R.call_tir(cls.layer_norm1, (lv40, gpt_neox_layers_17_post_attention_layernorm_weight3, gpt_neox_layers_17_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv124 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_17_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_17_mlp_dense_h_to_4h_q_scale3, layer_norm133, gpt_neox_layers_17_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv125 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_17_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_17_mlp_dense_4h_to_h_q_scale3, lv124, gpt_neox_layers_17_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv41 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_17_attention_dense_q_weight3, gpt_neox_layers_17_attention_dense_q_scale3, reshape263, gpt_neox_layers_17_attention_dense_bias3, lv125, lv40), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm134 = R.call_tir(cls.layer_norm1, (lv41, gpt_neox_layers_18_input_layernorm_weight3, gpt_neox_layers_18_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv126 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_18_attention_query_key_value_q_weight3, gpt_neox_layers_18_attention_query_key_value_q_scale3, layer_norm134, gpt_neox_layers_18_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape264 = R.call_tir(cls.reshape4, (lv126,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape265 = R.call_tir(cls.reshape5, (reshape264,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv335 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), reshape265), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape266 = R.call_tir(cls.reshape6, (lv335,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape267 = R.call_tir(cls.reshape7, (reshape266,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm135 = R.call_tir(cls.layer_norm1, (lv41, gpt_neox_layers_18_post_attention_layernorm_weight3, gpt_neox_layers_18_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv127 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_18_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_18_mlp_dense_h_to_4h_q_scale3, layer_norm135, gpt_neox_layers_18_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv128 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_18_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_18_mlp_dense_4h_to_h_q_scale3, lv127, gpt_neox_layers_18_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv42 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_18_attention_dense_q_weight3, gpt_neox_layers_18_attention_dense_q_scale3, reshape267, gpt_neox_layers_18_attention_dense_bias3, lv128, lv41), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm136 = R.call_tir(cls.layer_norm1, (lv42, gpt_neox_layers_19_input_layernorm_weight3, gpt_neox_layers_19_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv129 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_19_attention_query_key_value_q_weight3, gpt_neox_layers_19_attention_query_key_value_q_scale3, layer_norm136, gpt_neox_layers_19_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape268 = R.call_tir(cls.reshape4, (lv129,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape269 = R.call_tir(cls.reshape5, (reshape268,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv340 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), reshape269), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape270 = R.call_tir(cls.reshape6, (lv340,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape271 = R.call_tir(cls.reshape7, (reshape270,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm137 = R.call_tir(cls.layer_norm1, (lv42, gpt_neox_layers_19_post_attention_layernorm_weight3, gpt_neox_layers_19_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv130 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_19_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_19_mlp_dense_h_to_4h_q_scale3, layer_norm137, gpt_neox_layers_19_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv131 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_19_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_19_mlp_dense_4h_to_h_q_scale3, lv130, gpt_neox_layers_19_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv43 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_19_attention_dense_q_weight3, gpt_neox_layers_19_attention_dense_q_scale3, reshape271, gpt_neox_layers_19_attention_dense_bias3, lv131, lv42), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm138 = R.call_tir(cls.layer_norm1, (lv43, gpt_neox_layers_20_input_layernorm_weight3, gpt_neox_layers_20_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv132 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_20_attention_query_key_value_q_weight3, gpt_neox_layers_20_attention_query_key_value_q_scale3, layer_norm138, gpt_neox_layers_20_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape272 = R.call_tir(cls.reshape4, (lv132,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape273 = R.call_tir(cls.reshape5, (reshape272,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv345 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), reshape273), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape274 = R.call_tir(cls.reshape6, (lv345,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape275 = R.call_tir(cls.reshape7, (reshape274,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm139 = R.call_tir(cls.layer_norm1, (lv43, gpt_neox_layers_20_post_attention_layernorm_weight3, gpt_neox_layers_20_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv133 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_20_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_20_mlp_dense_h_to_4h_q_scale3, layer_norm139, gpt_neox_layers_20_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv134 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_20_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_20_mlp_dense_4h_to_h_q_scale3, lv133, gpt_neox_layers_20_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv44 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_20_attention_dense_q_weight3, gpt_neox_layers_20_attention_dense_q_scale3, reshape275, gpt_neox_layers_20_attention_dense_bias3, lv134, lv43), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm140 = R.call_tir(cls.layer_norm1, (lv44, gpt_neox_layers_21_input_layernorm_weight3, gpt_neox_layers_21_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv135 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_21_attention_query_key_value_q_weight3, gpt_neox_layers_21_attention_query_key_value_q_scale3, layer_norm140, gpt_neox_layers_21_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape276 = R.call_tir(cls.reshape4, (lv135,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape277 = R.call_tir(cls.reshape5, (reshape276,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv350 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), reshape277), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape278 = R.call_tir(cls.reshape6, (lv350,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape279 = R.call_tir(cls.reshape7, (reshape278,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm141 = R.call_tir(cls.layer_norm1, (lv44, gpt_neox_layers_21_post_attention_layernorm_weight3, gpt_neox_layers_21_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv136 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_21_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_21_mlp_dense_h_to_4h_q_scale3, layer_norm141, gpt_neox_layers_21_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv137 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_21_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_21_mlp_dense_4h_to_h_q_scale3, lv136, gpt_neox_layers_21_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv45 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_21_attention_dense_q_weight3, gpt_neox_layers_21_attention_dense_q_scale3, reshape279, gpt_neox_layers_21_attention_dense_bias3, lv137, lv44), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm142 = R.call_tir(cls.layer_norm1, (lv45, gpt_neox_layers_22_input_layernorm_weight3, gpt_neox_layers_22_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv138 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_22_attention_query_key_value_q_weight3, gpt_neox_layers_22_attention_query_key_value_q_scale3, layer_norm142, gpt_neox_layers_22_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape280 = R.call_tir(cls.reshape4, (lv138,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape281 = R.call_tir(cls.reshape5, (reshape280,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv355 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), reshape281), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape282 = R.call_tir(cls.reshape6, (lv355,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape283 = R.call_tir(cls.reshape7, (reshape282,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm143 = R.call_tir(cls.layer_norm1, (lv45, gpt_neox_layers_22_post_attention_layernorm_weight3, gpt_neox_layers_22_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv139 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_22_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_22_mlp_dense_h_to_4h_q_scale3, layer_norm143, gpt_neox_layers_22_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv140 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_22_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_22_mlp_dense_4h_to_h_q_scale3, lv139, gpt_neox_layers_22_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv46 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_22_attention_dense_q_weight3, gpt_neox_layers_22_attention_dense_q_scale3, reshape283, gpt_neox_layers_22_attention_dense_bias3, lv140, lv45), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm144 = R.call_tir(cls.layer_norm1, (lv46, gpt_neox_layers_23_input_layernorm_weight3, gpt_neox_layers_23_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv141 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_23_attention_query_key_value_q_weight3, gpt_neox_layers_23_attention_query_key_value_q_scale3, layer_norm144, gpt_neox_layers_23_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape284 = R.call_tir(cls.reshape4, (lv141,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape285 = R.call_tir(cls.reshape5, (reshape284,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv360 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), reshape285), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape286 = R.call_tir(cls.reshape6, (lv360,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape287 = R.call_tir(cls.reshape7, (reshape286,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm145 = R.call_tir(cls.layer_norm1, (lv46, gpt_neox_layers_23_post_attention_layernorm_weight3, gpt_neox_layers_23_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv142 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_23_mlp_dense_h_to_4h_q_weight3, gpt_neox_layers_23_mlp_dense_h_to_4h_q_scale3, layer_norm145, gpt_neox_layers_23_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv143 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_23_mlp_dense_4h_to_h_q_weight3, gpt_neox_layers_23_mlp_dense_4h_to_h_q_scale3, lv142, gpt_neox_layers_23_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv47 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_23_attention_dense_q_weight3, gpt_neox_layers_23_attention_dense_q_scale3, reshape287, gpt_neox_layers_23_attention_dense_bias3, lv143, lv46), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm146 = R.call_tir(cls.layer_norm1, (lv47, gpt_neox_final_layer_norm_weight3, gpt_neox_final_layer_norm_bias3), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) take1 = R.call_tir(cls.take, (layer_norm146, logit_positions), out_sinfo=R.Tensor((1, batch_size, 1024), dtype="float16")) lv1 = R.call_tir(cls.fused_dequantize_fused_NT_matmul9_cast5, (embed_out_q_weight3, embed_out_q_scale3, take1), out_sinfo=R.Tensor((1, batch_size, vocab_size), dtype="float32")) gv3: R.Tuple(R.Tensor((1, batch_size, vocab_size), dtype="float32"), R.Object) = lv1, paged_kv_cache R.output(gv3) return gv3 @R.function def batch_verify(input_embeds: R.Tensor((1, "seq_len", 1024), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 128), dtype="uint32"), R.Tensor(("vocab_size", 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor(("vocab_size", 128), dtype="uint32"), R.Tensor(("vocab_size", 32), 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.memory_plan_dynamic_func_output": 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 with R.dataflow(): gpt_neox_layers_0_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[2] gpt_neox_layers_0_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[3] gpt_neox_layers_0_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[4] gpt_neox_layers_0_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[5] gpt_neox_layers_0_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[6] gpt_neox_layers_0_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[8] gpt_neox_layers_0_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[9] gpt_neox_layers_0_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[10] gpt_neox_layers_0_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[11] gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[12] gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[13] gpt_neox_layers_0_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[14] gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[15] gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[16] gpt_neox_layers_0_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[17] gpt_neox_layers_1_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[18] gpt_neox_layers_1_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[19] gpt_neox_layers_1_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[20] gpt_neox_layers_1_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[21] gpt_neox_layers_1_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[22] gpt_neox_layers_1_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[23] gpt_neox_layers_1_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[24] gpt_neox_layers_1_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[25] gpt_neox_layers_1_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[26] gpt_neox_layers_1_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[27] gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[28] gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[29] gpt_neox_layers_1_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[30] gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[31] gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[32] gpt_neox_layers_1_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[33] gpt_neox_layers_2_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[34] gpt_neox_layers_2_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[35] gpt_neox_layers_2_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[36] gpt_neox_layers_2_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[37] gpt_neox_layers_2_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[38] gpt_neox_layers_2_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[39] gpt_neox_layers_2_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[40] gpt_neox_layers_2_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[41] gpt_neox_layers_2_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[42] gpt_neox_layers_2_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[43] gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[44] gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[45] gpt_neox_layers_2_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[46] gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[47] gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[48] gpt_neox_layers_2_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[49] gpt_neox_layers_3_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[50] gpt_neox_layers_3_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[51] gpt_neox_layers_3_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[52] gpt_neox_layers_3_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[53] gpt_neox_layers_3_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[54] gpt_neox_layers_3_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[55] gpt_neox_layers_3_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[56] gpt_neox_layers_3_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[57] gpt_neox_layers_3_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[58] gpt_neox_layers_3_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[59] gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[60] gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[61] gpt_neox_layers_3_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[62] gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[63] gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[64] gpt_neox_layers_3_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[65] gpt_neox_layers_4_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[66] gpt_neox_layers_4_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[67] gpt_neox_layers_4_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[68] gpt_neox_layers_4_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[69] gpt_neox_layers_4_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[70] gpt_neox_layers_4_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[71] gpt_neox_layers_4_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[72] gpt_neox_layers_4_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[73] gpt_neox_layers_4_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[74] gpt_neox_layers_4_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[75] gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[76] gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[77] gpt_neox_layers_4_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[78] gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[79] gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[80] gpt_neox_layers_4_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[81] gpt_neox_layers_5_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[82] gpt_neox_layers_5_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[83] gpt_neox_layers_5_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[84] gpt_neox_layers_5_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[85] gpt_neox_layers_5_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[86] gpt_neox_layers_5_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[87] gpt_neox_layers_5_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[88] gpt_neox_layers_5_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[89] gpt_neox_layers_5_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[90] gpt_neox_layers_5_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[91] gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[92] gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[93] gpt_neox_layers_5_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[94] gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[95] gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[96] gpt_neox_layers_5_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[97] gpt_neox_layers_6_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[98] gpt_neox_layers_6_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[99] gpt_neox_layers_6_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[100] gpt_neox_layers_6_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[101] gpt_neox_layers_6_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[102] gpt_neox_layers_6_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[103] gpt_neox_layers_6_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[104] gpt_neox_layers_6_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[105] gpt_neox_layers_6_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[106] gpt_neox_layers_6_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[107] gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[108] gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[109] gpt_neox_layers_6_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[110] gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[111] gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[112] gpt_neox_layers_6_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[113] gpt_neox_layers_7_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[114] gpt_neox_layers_7_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[115] gpt_neox_layers_7_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[116] gpt_neox_layers_7_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[117] gpt_neox_layers_7_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[118] gpt_neox_layers_7_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[119] gpt_neox_layers_7_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[120] gpt_neox_layers_7_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[121] gpt_neox_layers_7_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[122] gpt_neox_layers_7_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[123] gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[124] gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[125] gpt_neox_layers_7_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[126] gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[127] gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[128] gpt_neox_layers_7_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[129] gpt_neox_layers_8_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[130] gpt_neox_layers_8_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[131] gpt_neox_layers_8_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[132] gpt_neox_layers_8_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[133] gpt_neox_layers_8_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[134] gpt_neox_layers_8_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[135] gpt_neox_layers_8_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[136] gpt_neox_layers_8_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[137] gpt_neox_layers_8_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[138] gpt_neox_layers_8_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[139] gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[140] gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[141] gpt_neox_layers_8_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[142] gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[143] gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[144] gpt_neox_layers_8_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[145] gpt_neox_layers_9_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[146] gpt_neox_layers_9_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[147] gpt_neox_layers_9_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[148] gpt_neox_layers_9_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[149] gpt_neox_layers_9_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[150] gpt_neox_layers_9_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[151] gpt_neox_layers_9_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[152] gpt_neox_layers_9_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[153] gpt_neox_layers_9_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[154] gpt_neox_layers_9_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[155] gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[156] gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[157] gpt_neox_layers_9_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[158] gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[159] gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[160] gpt_neox_layers_9_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[161] gpt_neox_layers_10_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[162] gpt_neox_layers_10_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[163] gpt_neox_layers_10_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[164] gpt_neox_layers_10_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[165] gpt_neox_layers_10_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[166] gpt_neox_layers_10_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[167] gpt_neox_layers_10_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[168] gpt_neox_layers_10_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[169] gpt_neox_layers_10_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[170] gpt_neox_layers_10_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[171] gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[172] gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[173] gpt_neox_layers_10_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[174] gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[175] gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[176] gpt_neox_layers_10_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[177] gpt_neox_layers_11_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[178] gpt_neox_layers_11_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[179] gpt_neox_layers_11_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[180] gpt_neox_layers_11_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[181] gpt_neox_layers_11_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[182] gpt_neox_layers_11_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[183] gpt_neox_layers_11_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[184] gpt_neox_layers_11_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[185] gpt_neox_layers_11_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[186] gpt_neox_layers_11_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[187] gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[188] gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[189] gpt_neox_layers_11_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[190] gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[191] gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[192] gpt_neox_layers_11_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[193] gpt_neox_layers_12_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[194] gpt_neox_layers_12_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[195] gpt_neox_layers_12_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[196] gpt_neox_layers_12_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[197] gpt_neox_layers_12_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[198] gpt_neox_layers_12_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[199] gpt_neox_layers_12_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[200] gpt_neox_layers_12_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[201] gpt_neox_layers_12_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[202] gpt_neox_layers_12_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[203] gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[204] gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[205] gpt_neox_layers_12_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[206] gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[207] gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[208] gpt_neox_layers_12_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[209] gpt_neox_layers_13_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[210] gpt_neox_layers_13_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[211] gpt_neox_layers_13_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[212] gpt_neox_layers_13_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[213] gpt_neox_layers_13_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[214] gpt_neox_layers_13_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[215] gpt_neox_layers_13_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[216] gpt_neox_layers_13_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[217] gpt_neox_layers_13_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[218] gpt_neox_layers_13_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[219] gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[220] gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[221] gpt_neox_layers_13_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[222] gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[223] gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[224] gpt_neox_layers_13_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[225] gpt_neox_layers_14_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[226] gpt_neox_layers_14_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[227] gpt_neox_layers_14_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[228] gpt_neox_layers_14_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[229] gpt_neox_layers_14_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[230] gpt_neox_layers_14_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[231] gpt_neox_layers_14_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[232] gpt_neox_layers_14_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[233] gpt_neox_layers_14_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[234] gpt_neox_layers_14_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[235] gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[236] gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[237] gpt_neox_layers_14_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[238] gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[239] gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[240] gpt_neox_layers_14_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[241] gpt_neox_layers_15_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[242] gpt_neox_layers_15_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[243] gpt_neox_layers_15_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[244] gpt_neox_layers_15_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[245] gpt_neox_layers_15_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[246] gpt_neox_layers_15_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[247] gpt_neox_layers_15_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[248] gpt_neox_layers_15_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[249] gpt_neox_layers_15_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[250] gpt_neox_layers_15_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[251] gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[252] gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[253] gpt_neox_layers_15_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[254] gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[255] gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[256] gpt_neox_layers_15_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[257] gpt_neox_layers_16_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[258] gpt_neox_layers_16_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[259] gpt_neox_layers_16_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[260] gpt_neox_layers_16_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[261] gpt_neox_layers_16_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[262] gpt_neox_layers_16_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[263] gpt_neox_layers_16_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[264] gpt_neox_layers_16_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[265] gpt_neox_layers_16_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[266] gpt_neox_layers_16_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[267] gpt_neox_layers_16_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[268] gpt_neox_layers_16_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[269] gpt_neox_layers_16_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[270] gpt_neox_layers_16_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[271] gpt_neox_layers_16_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[272] gpt_neox_layers_16_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[273] gpt_neox_layers_17_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[274] gpt_neox_layers_17_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[275] gpt_neox_layers_17_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[276] gpt_neox_layers_17_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[277] gpt_neox_layers_17_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[278] gpt_neox_layers_17_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[279] gpt_neox_layers_17_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[280] gpt_neox_layers_17_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[281] gpt_neox_layers_17_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[282] gpt_neox_layers_17_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[283] gpt_neox_layers_17_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[284] gpt_neox_layers_17_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[285] gpt_neox_layers_17_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[286] gpt_neox_layers_17_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[287] gpt_neox_layers_17_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[288] gpt_neox_layers_17_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[289] gpt_neox_layers_18_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[290] gpt_neox_layers_18_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[291] gpt_neox_layers_18_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[292] gpt_neox_layers_18_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[293] gpt_neox_layers_18_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[294] gpt_neox_layers_18_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[295] gpt_neox_layers_18_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[296] gpt_neox_layers_18_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[297] gpt_neox_layers_18_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[298] gpt_neox_layers_18_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[299] gpt_neox_layers_18_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[300] gpt_neox_layers_18_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[301] gpt_neox_layers_18_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[302] gpt_neox_layers_18_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[303] gpt_neox_layers_18_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[304] gpt_neox_layers_18_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[305] gpt_neox_layers_19_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[306] gpt_neox_layers_19_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[307] gpt_neox_layers_19_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[308] gpt_neox_layers_19_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[309] gpt_neox_layers_19_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[310] gpt_neox_layers_19_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[311] gpt_neox_layers_19_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[312] gpt_neox_layers_19_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[313] gpt_neox_layers_19_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[314] gpt_neox_layers_19_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[315] gpt_neox_layers_19_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[316] gpt_neox_layers_19_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[317] gpt_neox_layers_19_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[318] gpt_neox_layers_19_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[319] gpt_neox_layers_19_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[320] gpt_neox_layers_19_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[321] gpt_neox_layers_20_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[322] gpt_neox_layers_20_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[323] gpt_neox_layers_20_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[324] gpt_neox_layers_20_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[325] gpt_neox_layers_20_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[326] gpt_neox_layers_20_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[327] gpt_neox_layers_20_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[328] gpt_neox_layers_20_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[329] gpt_neox_layers_20_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[330] gpt_neox_layers_20_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[331] gpt_neox_layers_20_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[332] gpt_neox_layers_20_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[333] gpt_neox_layers_20_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[334] gpt_neox_layers_20_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[335] gpt_neox_layers_20_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[336] gpt_neox_layers_20_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[337] gpt_neox_layers_21_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[338] gpt_neox_layers_21_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[339] gpt_neox_layers_21_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[340] gpt_neox_layers_21_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[341] gpt_neox_layers_21_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[342] gpt_neox_layers_21_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[343] gpt_neox_layers_21_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[344] gpt_neox_layers_21_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[345] gpt_neox_layers_21_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[346] gpt_neox_layers_21_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[347] gpt_neox_layers_21_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[348] gpt_neox_layers_21_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[349] gpt_neox_layers_21_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[350] gpt_neox_layers_21_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[351] gpt_neox_layers_21_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[352] gpt_neox_layers_21_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[353] gpt_neox_layers_22_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[354] gpt_neox_layers_22_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[355] gpt_neox_layers_22_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[356] gpt_neox_layers_22_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[357] gpt_neox_layers_22_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[358] gpt_neox_layers_22_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[359] gpt_neox_layers_22_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[360] gpt_neox_layers_22_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[361] gpt_neox_layers_22_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[362] gpt_neox_layers_22_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[363] gpt_neox_layers_22_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[364] gpt_neox_layers_22_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[365] gpt_neox_layers_22_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[366] gpt_neox_layers_22_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[367] gpt_neox_layers_22_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[368] gpt_neox_layers_22_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[369] gpt_neox_layers_23_input_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[370] gpt_neox_layers_23_input_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[371] gpt_neox_layers_23_post_attention_layernorm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[372] gpt_neox_layers_23_post_attention_layernorm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[373] gpt_neox_layers_23_attention_query_key_value_q_weight5: R.Tensor((3072, 128), dtype="uint32") = packed_params[374] gpt_neox_layers_23_attention_query_key_value_q_scale5: R.Tensor((3072, 32), dtype="float16") = packed_params[375] gpt_neox_layers_23_attention_query_key_value_bias5: R.Tensor((3072,), dtype="float16") = packed_params[376] gpt_neox_layers_23_attention_dense_q_weight5: R.Tensor((1024, 128), dtype="uint32") = packed_params[377] gpt_neox_layers_23_attention_dense_q_scale5: R.Tensor((1024, 32), dtype="float16") = packed_params[378] gpt_neox_layers_23_attention_dense_bias5: R.Tensor((1024,), dtype="float16") = packed_params[379] gpt_neox_layers_23_mlp_dense_h_to_4h_q_weight5: R.Tensor((4096, 128), dtype="uint32") = packed_params[380] gpt_neox_layers_23_mlp_dense_h_to_4h_q_scale5: R.Tensor((4096, 32), dtype="float16") = packed_params[381] gpt_neox_layers_23_mlp_dense_h_to_4h_bias5: R.Tensor((4096,), dtype="float32") = packed_params[382] gpt_neox_layers_23_mlp_dense_4h_to_h_q_weight5: R.Tensor((1024, 512), dtype="uint32") = packed_params[383] gpt_neox_layers_23_mlp_dense_4h_to_h_q_scale5: R.Tensor((1024, 128), dtype="float16") = packed_params[384] gpt_neox_layers_23_mlp_dense_4h_to_h_bias5: R.Tensor((1024,), dtype="float32") = packed_params[385] gpt_neox_final_layer_norm_weight5: R.Tensor((1024,), dtype="float16") = packed_params[386] gpt_neox_final_layer_norm_bias5: R.Tensor((1024,), dtype="float16") = packed_params[387] embed_out_q_weight5: R.Tensor((vocab_size, 128), dtype="uint32") = packed_params[388] embed_out_q_scale5: R.Tensor((vocab_size, 32), dtype="float16") = packed_params[389] layer_norm196 = R.call_tir(cls.layer_norm1, (input_embeds, gpt_neox_layers_0_input_layernorm_weight5, gpt_neox_layers_0_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv144 = R.call_tir(cls.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, layer_norm196, gpt_neox_layers_0_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape384 = R.call_tir(cls.reshape4, (lv144,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape385 = R.call_tir(cls.reshape5, (reshape384,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv487 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), reshape385), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape386 = R.call_tir(cls.reshape6, (lv487,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape387 = R.call_tir(cls.reshape7, (reshape386,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm197 = R.call_tir(cls.layer_norm1, (input_embeds, gpt_neox_layers_0_post_attention_layernorm_weight5, gpt_neox_layers_0_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv145 = R.call_tir(cls.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, layer_norm197, gpt_neox_layers_0_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv146 = R.call_tir(cls.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, lv145, gpt_neox_layers_0_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv48 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_0_attention_dense_q_weight5, gpt_neox_layers_0_attention_dense_q_scale5, reshape387, gpt_neox_layers_0_attention_dense_bias5, lv146, input_embeds), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm198 = R.call_tir(cls.layer_norm1, (lv48, gpt_neox_layers_1_input_layernorm_weight5, gpt_neox_layers_1_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv147 = R.call_tir(cls.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, layer_norm198, gpt_neox_layers_1_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape388 = R.call_tir(cls.reshape4, (lv147,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape389 = R.call_tir(cls.reshape5, (reshape388,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv492 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), reshape389), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape390 = R.call_tir(cls.reshape6, (lv492,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape391 = R.call_tir(cls.reshape7, (reshape390,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm199 = R.call_tir(cls.layer_norm1, (lv48, gpt_neox_layers_1_post_attention_layernorm_weight5, gpt_neox_layers_1_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv148 = R.call_tir(cls.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, layer_norm199, gpt_neox_layers_1_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv149 = R.call_tir(cls.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, lv148, gpt_neox_layers_1_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv49 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_1_attention_dense_q_weight5, gpt_neox_layers_1_attention_dense_q_scale5, reshape391, gpt_neox_layers_1_attention_dense_bias5, lv149, lv48), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm200 = R.call_tir(cls.layer_norm1, (lv49, gpt_neox_layers_2_input_layernorm_weight5, gpt_neox_layers_2_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv150 = R.call_tir(cls.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, layer_norm200, gpt_neox_layers_2_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape392 = R.call_tir(cls.reshape4, (lv150,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape393 = R.call_tir(cls.reshape5, (reshape392,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv497 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), reshape393), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape394 = R.call_tir(cls.reshape6, (lv497,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape395 = R.call_tir(cls.reshape7, (reshape394,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm201 = R.call_tir(cls.layer_norm1, (lv49, gpt_neox_layers_2_post_attention_layernorm_weight5, gpt_neox_layers_2_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv151 = R.call_tir(cls.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, layer_norm201, gpt_neox_layers_2_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv152 = R.call_tir(cls.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, lv151, gpt_neox_layers_2_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv50 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_2_attention_dense_q_weight5, gpt_neox_layers_2_attention_dense_q_scale5, reshape395, gpt_neox_layers_2_attention_dense_bias5, lv152, lv49), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm202 = R.call_tir(cls.layer_norm1, (lv50, gpt_neox_layers_3_input_layernorm_weight5, gpt_neox_layers_3_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv153 = R.call_tir(cls.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, layer_norm202, gpt_neox_layers_3_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape396 = R.call_tir(cls.reshape4, (lv153,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape397 = R.call_tir(cls.reshape5, (reshape396,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv502 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), reshape397), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape398 = R.call_tir(cls.reshape6, (lv502,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape399 = R.call_tir(cls.reshape7, (reshape398,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm203 = R.call_tir(cls.layer_norm1, (lv50, gpt_neox_layers_3_post_attention_layernorm_weight5, gpt_neox_layers_3_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv154 = R.call_tir(cls.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, layer_norm203, gpt_neox_layers_3_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv155 = R.call_tir(cls.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, lv154, gpt_neox_layers_3_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv51 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_3_attention_dense_q_weight5, gpt_neox_layers_3_attention_dense_q_scale5, reshape399, gpt_neox_layers_3_attention_dense_bias5, lv155, lv50), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm204 = R.call_tir(cls.layer_norm1, (lv51, gpt_neox_layers_4_input_layernorm_weight5, gpt_neox_layers_4_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv156 = R.call_tir(cls.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, layer_norm204, gpt_neox_layers_4_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape400 = R.call_tir(cls.reshape4, (lv156,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape401 = R.call_tir(cls.reshape5, (reshape400,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv507 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), reshape401), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape402 = R.call_tir(cls.reshape6, (lv507,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape403 = R.call_tir(cls.reshape7, (reshape402,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm205 = R.call_tir(cls.layer_norm1, (lv51, gpt_neox_layers_4_post_attention_layernorm_weight5, gpt_neox_layers_4_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv157 = R.call_tir(cls.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, layer_norm205, gpt_neox_layers_4_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv158 = R.call_tir(cls.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, lv157, gpt_neox_layers_4_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv52 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_4_attention_dense_q_weight5, gpt_neox_layers_4_attention_dense_q_scale5, reshape403, gpt_neox_layers_4_attention_dense_bias5, lv158, lv51), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm206 = R.call_tir(cls.layer_norm1, (lv52, gpt_neox_layers_5_input_layernorm_weight5, gpt_neox_layers_5_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv159 = R.call_tir(cls.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, layer_norm206, gpt_neox_layers_5_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape404 = R.call_tir(cls.reshape4, (lv159,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape405 = R.call_tir(cls.reshape5, (reshape404,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv512 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), reshape405), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape406 = R.call_tir(cls.reshape6, (lv512,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape407 = R.call_tir(cls.reshape7, (reshape406,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm207 = R.call_tir(cls.layer_norm1, (lv52, gpt_neox_layers_5_post_attention_layernorm_weight5, gpt_neox_layers_5_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv160 = R.call_tir(cls.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, layer_norm207, gpt_neox_layers_5_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv161 = R.call_tir(cls.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, lv160, gpt_neox_layers_5_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv53 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_5_attention_dense_q_weight5, gpt_neox_layers_5_attention_dense_q_scale5, reshape407, gpt_neox_layers_5_attention_dense_bias5, lv161, lv52), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm208 = R.call_tir(cls.layer_norm1, (lv53, gpt_neox_layers_6_input_layernorm_weight5, gpt_neox_layers_6_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv162 = R.call_tir(cls.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, layer_norm208, gpt_neox_layers_6_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape408 = R.call_tir(cls.reshape4, (lv162,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape409 = R.call_tir(cls.reshape5, (reshape408,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv517 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), reshape409), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape410 = R.call_tir(cls.reshape6, (lv517,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape411 = R.call_tir(cls.reshape7, (reshape410,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm209 = R.call_tir(cls.layer_norm1, (lv53, gpt_neox_layers_6_post_attention_layernorm_weight5, gpt_neox_layers_6_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv163 = R.call_tir(cls.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, layer_norm209, gpt_neox_layers_6_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv164 = R.call_tir(cls.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, lv163, gpt_neox_layers_6_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv54 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_6_attention_dense_q_weight5, gpt_neox_layers_6_attention_dense_q_scale5, reshape411, gpt_neox_layers_6_attention_dense_bias5, lv164, lv53), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm210 = R.call_tir(cls.layer_norm1, (lv54, gpt_neox_layers_7_input_layernorm_weight5, gpt_neox_layers_7_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv165 = R.call_tir(cls.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, layer_norm210, gpt_neox_layers_7_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape412 = R.call_tir(cls.reshape4, (lv165,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape413 = R.call_tir(cls.reshape5, (reshape412,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv522 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), reshape413), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape414 = R.call_tir(cls.reshape6, (lv522,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape415 = R.call_tir(cls.reshape7, (reshape414,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm211 = R.call_tir(cls.layer_norm1, (lv54, gpt_neox_layers_7_post_attention_layernorm_weight5, gpt_neox_layers_7_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv166 = R.call_tir(cls.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, layer_norm211, gpt_neox_layers_7_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv167 = R.call_tir(cls.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, lv166, gpt_neox_layers_7_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv55 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_7_attention_dense_q_weight5, gpt_neox_layers_7_attention_dense_q_scale5, reshape415, gpt_neox_layers_7_attention_dense_bias5, lv167, lv54), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm212 = R.call_tir(cls.layer_norm1, (lv55, gpt_neox_layers_8_input_layernorm_weight5, gpt_neox_layers_8_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv168 = R.call_tir(cls.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, layer_norm212, gpt_neox_layers_8_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape416 = R.call_tir(cls.reshape4, (lv168,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape417 = R.call_tir(cls.reshape5, (reshape416,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv527 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), reshape417), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape418 = R.call_tir(cls.reshape6, (lv527,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape419 = R.call_tir(cls.reshape7, (reshape418,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm213 = R.call_tir(cls.layer_norm1, (lv55, gpt_neox_layers_8_post_attention_layernorm_weight5, gpt_neox_layers_8_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv169 = R.call_tir(cls.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, layer_norm213, gpt_neox_layers_8_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv170 = R.call_tir(cls.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, lv169, gpt_neox_layers_8_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv56 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_8_attention_dense_q_weight5, gpt_neox_layers_8_attention_dense_q_scale5, reshape419, gpt_neox_layers_8_attention_dense_bias5, lv170, lv55), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm214 = R.call_tir(cls.layer_norm1, (lv56, gpt_neox_layers_9_input_layernorm_weight5, gpt_neox_layers_9_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv171 = R.call_tir(cls.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, layer_norm214, gpt_neox_layers_9_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape420 = R.call_tir(cls.reshape4, (lv171,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape421 = R.call_tir(cls.reshape5, (reshape420,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv532 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), reshape421), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape422 = R.call_tir(cls.reshape6, (lv532,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape423 = R.call_tir(cls.reshape7, (reshape422,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm215 = R.call_tir(cls.layer_norm1, (lv56, gpt_neox_layers_9_post_attention_layernorm_weight5, gpt_neox_layers_9_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv172 = R.call_tir(cls.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, layer_norm215, gpt_neox_layers_9_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv173 = R.call_tir(cls.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, lv172, gpt_neox_layers_9_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv57 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_9_attention_dense_q_weight5, gpt_neox_layers_9_attention_dense_q_scale5, reshape423, gpt_neox_layers_9_attention_dense_bias5, lv173, lv56), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm216 = R.call_tir(cls.layer_norm1, (lv57, gpt_neox_layers_10_input_layernorm_weight5, gpt_neox_layers_10_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv174 = R.call_tir(cls.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, layer_norm216, gpt_neox_layers_10_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape424 = R.call_tir(cls.reshape4, (lv174,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape425 = R.call_tir(cls.reshape5, (reshape424,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv537 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), reshape425), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape426 = R.call_tir(cls.reshape6, (lv537,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape427 = R.call_tir(cls.reshape7, (reshape426,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm217 = R.call_tir(cls.layer_norm1, (lv57, gpt_neox_layers_10_post_attention_layernorm_weight5, gpt_neox_layers_10_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv175 = R.call_tir(cls.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, layer_norm217, gpt_neox_layers_10_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv176 = R.call_tir(cls.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, lv175, gpt_neox_layers_10_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv58 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_10_attention_dense_q_weight5, gpt_neox_layers_10_attention_dense_q_scale5, reshape427, gpt_neox_layers_10_attention_dense_bias5, lv176, lv57), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm218 = R.call_tir(cls.layer_norm1, (lv58, gpt_neox_layers_11_input_layernorm_weight5, gpt_neox_layers_11_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv177 = R.call_tir(cls.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, layer_norm218, gpt_neox_layers_11_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape428 = R.call_tir(cls.reshape4, (lv177,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape429 = R.call_tir(cls.reshape5, (reshape428,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv542 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), reshape429), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape430 = R.call_tir(cls.reshape6, (lv542,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape431 = R.call_tir(cls.reshape7, (reshape430,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm219 = R.call_tir(cls.layer_norm1, (lv58, gpt_neox_layers_11_post_attention_layernorm_weight5, gpt_neox_layers_11_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv178 = R.call_tir(cls.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, layer_norm219, gpt_neox_layers_11_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv179 = R.call_tir(cls.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, lv178, gpt_neox_layers_11_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv59 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_11_attention_dense_q_weight5, gpt_neox_layers_11_attention_dense_q_scale5, reshape431, gpt_neox_layers_11_attention_dense_bias5, lv179, lv58), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm220 = R.call_tir(cls.layer_norm1, (lv59, gpt_neox_layers_12_input_layernorm_weight5, gpt_neox_layers_12_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv180 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_12_attention_query_key_value_q_weight5, gpt_neox_layers_12_attention_query_key_value_q_scale5, layer_norm220, gpt_neox_layers_12_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape432 = R.call_tir(cls.reshape4, (lv180,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape433 = R.call_tir(cls.reshape5, (reshape432,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv547 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape433), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape434 = R.call_tir(cls.reshape6, (lv547,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape435 = R.call_tir(cls.reshape7, (reshape434,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm221 = R.call_tir(cls.layer_norm1, (lv59, gpt_neox_layers_12_post_attention_layernorm_weight5, gpt_neox_layers_12_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv181 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale5, layer_norm221, gpt_neox_layers_12_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv182 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale5, lv181, gpt_neox_layers_12_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv60 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_12_attention_dense_q_weight5, gpt_neox_layers_12_attention_dense_q_scale5, reshape435, gpt_neox_layers_12_attention_dense_bias5, lv182, lv59), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm222 = R.call_tir(cls.layer_norm1, (lv60, gpt_neox_layers_13_input_layernorm_weight5, gpt_neox_layers_13_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv183 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_13_attention_query_key_value_q_weight5, gpt_neox_layers_13_attention_query_key_value_q_scale5, layer_norm222, gpt_neox_layers_13_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape436 = R.call_tir(cls.reshape4, (lv183,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape437 = R.call_tir(cls.reshape5, (reshape436,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv552 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape437), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape438 = R.call_tir(cls.reshape6, (lv552,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape439 = R.call_tir(cls.reshape7, (reshape438,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm223 = R.call_tir(cls.layer_norm1, (lv60, gpt_neox_layers_13_post_attention_layernorm_weight5, gpt_neox_layers_13_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv184 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale5, layer_norm223, gpt_neox_layers_13_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv185 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale5, lv184, gpt_neox_layers_13_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv61 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_13_attention_dense_q_weight5, gpt_neox_layers_13_attention_dense_q_scale5, reshape439, gpt_neox_layers_13_attention_dense_bias5, lv185, lv60), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm224 = R.call_tir(cls.layer_norm1, (lv61, gpt_neox_layers_14_input_layernorm_weight5, gpt_neox_layers_14_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv186 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_14_attention_query_key_value_q_weight5, gpt_neox_layers_14_attention_query_key_value_q_scale5, layer_norm224, gpt_neox_layers_14_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape440 = R.call_tir(cls.reshape4, (lv186,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape441 = R.call_tir(cls.reshape5, (reshape440,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv557 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape441), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape442 = R.call_tir(cls.reshape6, (lv557,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape443 = R.call_tir(cls.reshape7, (reshape442,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm225 = R.call_tir(cls.layer_norm1, (lv61, gpt_neox_layers_14_post_attention_layernorm_weight5, gpt_neox_layers_14_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv187 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale5, layer_norm225, gpt_neox_layers_14_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv188 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale5, lv187, gpt_neox_layers_14_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv62 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_14_attention_dense_q_weight5, gpt_neox_layers_14_attention_dense_q_scale5, reshape443, gpt_neox_layers_14_attention_dense_bias5, lv188, lv61), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm226 = R.call_tir(cls.layer_norm1, (lv62, gpt_neox_layers_15_input_layernorm_weight5, gpt_neox_layers_15_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv189 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_15_attention_query_key_value_q_weight5, gpt_neox_layers_15_attention_query_key_value_q_scale5, layer_norm226, gpt_neox_layers_15_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape444 = R.call_tir(cls.reshape4, (lv189,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape445 = R.call_tir(cls.reshape5, (reshape444,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv562 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape445), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape446 = R.call_tir(cls.reshape6, (lv562,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape447 = R.call_tir(cls.reshape7, (reshape446,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm227 = R.call_tir(cls.layer_norm1, (lv62, gpt_neox_layers_15_post_attention_layernorm_weight5, gpt_neox_layers_15_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv190 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale5, layer_norm227, gpt_neox_layers_15_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv191 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale5, lv190, gpt_neox_layers_15_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv63 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_15_attention_dense_q_weight5, gpt_neox_layers_15_attention_dense_q_scale5, reshape447, gpt_neox_layers_15_attention_dense_bias5, lv191, lv62), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm228 = R.call_tir(cls.layer_norm1, (lv63, gpt_neox_layers_16_input_layernorm_weight5, gpt_neox_layers_16_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv192 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_16_attention_query_key_value_q_weight5, gpt_neox_layers_16_attention_query_key_value_q_scale5, layer_norm228, gpt_neox_layers_16_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape448 = R.call_tir(cls.reshape4, (lv192,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape449 = R.call_tir(cls.reshape5, (reshape448,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv567 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), reshape449), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape450 = R.call_tir(cls.reshape6, (lv567,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape451 = R.call_tir(cls.reshape7, (reshape450,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm229 = R.call_tir(cls.layer_norm1, (lv63, gpt_neox_layers_16_post_attention_layernorm_weight5, gpt_neox_layers_16_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv193 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_16_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_16_mlp_dense_h_to_4h_q_scale5, layer_norm229, gpt_neox_layers_16_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv194 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_16_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_16_mlp_dense_4h_to_h_q_scale5, lv193, gpt_neox_layers_16_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv64 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_16_attention_dense_q_weight5, gpt_neox_layers_16_attention_dense_q_scale5, reshape451, gpt_neox_layers_16_attention_dense_bias5, lv194, lv63), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm230 = R.call_tir(cls.layer_norm1, (lv64, gpt_neox_layers_17_input_layernorm_weight5, gpt_neox_layers_17_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv195 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_17_attention_query_key_value_q_weight5, gpt_neox_layers_17_attention_query_key_value_q_scale5, layer_norm230, gpt_neox_layers_17_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape452 = R.call_tir(cls.reshape4, (lv195,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape453 = R.call_tir(cls.reshape5, (reshape452,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv572 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), reshape453), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape454 = R.call_tir(cls.reshape6, (lv572,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape455 = R.call_tir(cls.reshape7, (reshape454,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm231 = R.call_tir(cls.layer_norm1, (lv64, gpt_neox_layers_17_post_attention_layernorm_weight5, gpt_neox_layers_17_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv196 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_17_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_17_mlp_dense_h_to_4h_q_scale5, layer_norm231, gpt_neox_layers_17_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv197 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_17_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_17_mlp_dense_4h_to_h_q_scale5, lv196, gpt_neox_layers_17_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv65 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_17_attention_dense_q_weight5, gpt_neox_layers_17_attention_dense_q_scale5, reshape455, gpt_neox_layers_17_attention_dense_bias5, lv197, lv64), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm232 = R.call_tir(cls.layer_norm1, (lv65, gpt_neox_layers_18_input_layernorm_weight5, gpt_neox_layers_18_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv198 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_18_attention_query_key_value_q_weight5, gpt_neox_layers_18_attention_query_key_value_q_scale5, layer_norm232, gpt_neox_layers_18_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape456 = R.call_tir(cls.reshape4, (lv198,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape457 = R.call_tir(cls.reshape5, (reshape456,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv577 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), reshape457), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape458 = R.call_tir(cls.reshape6, (lv577,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape459 = R.call_tir(cls.reshape7, (reshape458,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm233 = R.call_tir(cls.layer_norm1, (lv65, gpt_neox_layers_18_post_attention_layernorm_weight5, gpt_neox_layers_18_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv199 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_18_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_18_mlp_dense_h_to_4h_q_scale5, layer_norm233, gpt_neox_layers_18_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv200 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_18_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_18_mlp_dense_4h_to_h_q_scale5, lv199, gpt_neox_layers_18_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv66 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_18_attention_dense_q_weight5, gpt_neox_layers_18_attention_dense_q_scale5, reshape459, gpt_neox_layers_18_attention_dense_bias5, lv200, lv65), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm234 = R.call_tir(cls.layer_norm1, (lv66, gpt_neox_layers_19_input_layernorm_weight5, gpt_neox_layers_19_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv201 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_19_attention_query_key_value_q_weight5, gpt_neox_layers_19_attention_query_key_value_q_scale5, layer_norm234, gpt_neox_layers_19_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape460 = R.call_tir(cls.reshape4, (lv201,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape461 = R.call_tir(cls.reshape5, (reshape460,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv582 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), reshape461), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape462 = R.call_tir(cls.reshape6, (lv582,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape463 = R.call_tir(cls.reshape7, (reshape462,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm235 = R.call_tir(cls.layer_norm1, (lv66, gpt_neox_layers_19_post_attention_layernorm_weight5, gpt_neox_layers_19_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv202 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_19_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_19_mlp_dense_h_to_4h_q_scale5, layer_norm235, gpt_neox_layers_19_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv203 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_19_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_19_mlp_dense_4h_to_h_q_scale5, lv202, gpt_neox_layers_19_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv67 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_19_attention_dense_q_weight5, gpt_neox_layers_19_attention_dense_q_scale5, reshape463, gpt_neox_layers_19_attention_dense_bias5, lv203, lv66), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm236 = R.call_tir(cls.layer_norm1, (lv67, gpt_neox_layers_20_input_layernorm_weight5, gpt_neox_layers_20_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv204 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_20_attention_query_key_value_q_weight5, gpt_neox_layers_20_attention_query_key_value_q_scale5, layer_norm236, gpt_neox_layers_20_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape464 = R.call_tir(cls.reshape4, (lv204,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape465 = R.call_tir(cls.reshape5, (reshape464,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv587 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), reshape465), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape466 = R.call_tir(cls.reshape6, (lv587,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape467 = R.call_tir(cls.reshape7, (reshape466,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm237 = R.call_tir(cls.layer_norm1, (lv67, gpt_neox_layers_20_post_attention_layernorm_weight5, gpt_neox_layers_20_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv205 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_20_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_20_mlp_dense_h_to_4h_q_scale5, layer_norm237, gpt_neox_layers_20_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv206 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_20_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_20_mlp_dense_4h_to_h_q_scale5, lv205, gpt_neox_layers_20_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv68 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_20_attention_dense_q_weight5, gpt_neox_layers_20_attention_dense_q_scale5, reshape467, gpt_neox_layers_20_attention_dense_bias5, lv206, lv67), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm238 = R.call_tir(cls.layer_norm1, (lv68, gpt_neox_layers_21_input_layernorm_weight5, gpt_neox_layers_21_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv207 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_21_attention_query_key_value_q_weight5, gpt_neox_layers_21_attention_query_key_value_q_scale5, layer_norm238, gpt_neox_layers_21_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape468 = R.call_tir(cls.reshape4, (lv207,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape469 = R.call_tir(cls.reshape5, (reshape468,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv592 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), reshape469), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape470 = R.call_tir(cls.reshape6, (lv592,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape471 = R.call_tir(cls.reshape7, (reshape470,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm239 = R.call_tir(cls.layer_norm1, (lv68, gpt_neox_layers_21_post_attention_layernorm_weight5, gpt_neox_layers_21_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv208 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_21_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_21_mlp_dense_h_to_4h_q_scale5, layer_norm239, gpt_neox_layers_21_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv209 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_21_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_21_mlp_dense_4h_to_h_q_scale5, lv208, gpt_neox_layers_21_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv69 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_21_attention_dense_q_weight5, gpt_neox_layers_21_attention_dense_q_scale5, reshape471, gpt_neox_layers_21_attention_dense_bias5, lv209, lv68), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm240 = R.call_tir(cls.layer_norm1, (lv69, gpt_neox_layers_22_input_layernorm_weight5, gpt_neox_layers_22_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv210 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_22_attention_query_key_value_q_weight5, gpt_neox_layers_22_attention_query_key_value_q_scale5, layer_norm240, gpt_neox_layers_22_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape472 = R.call_tir(cls.reshape4, (lv210,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape473 = R.call_tir(cls.reshape5, (reshape472,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv597 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), reshape473), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape474 = R.call_tir(cls.reshape6, (lv597,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape475 = R.call_tir(cls.reshape7, (reshape474,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm241 = R.call_tir(cls.layer_norm1, (lv69, gpt_neox_layers_22_post_attention_layernorm_weight5, gpt_neox_layers_22_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv211 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_22_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_22_mlp_dense_h_to_4h_q_scale5, layer_norm241, gpt_neox_layers_22_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv212 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_22_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_22_mlp_dense_4h_to_h_q_scale5, lv211, gpt_neox_layers_22_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv70 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_22_attention_dense_q_weight5, gpt_neox_layers_22_attention_dense_q_scale5, reshape475, gpt_neox_layers_22_attention_dense_bias5, lv212, lv69), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm242 = R.call_tir(cls.layer_norm1, (lv70, gpt_neox_layers_23_input_layernorm_weight5, gpt_neox_layers_23_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv213 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_23_attention_query_key_value_q_weight5, gpt_neox_layers_23_attention_query_key_value_q_scale5, layer_norm242, gpt_neox_layers_23_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape476 = R.call_tir(cls.reshape4, (lv213,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape477 = R.call_tir(cls.reshape5, (reshape476,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv602 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), reshape477), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape478 = R.call_tir(cls.reshape6, (lv602,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape479 = R.call_tir(cls.reshape7, (reshape478,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm243 = R.call_tir(cls.layer_norm1, (lv70, gpt_neox_layers_23_post_attention_layernorm_weight5, gpt_neox_layers_23_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv214 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_23_mlp_dense_h_to_4h_q_weight5, gpt_neox_layers_23_mlp_dense_h_to_4h_q_scale5, layer_norm243, gpt_neox_layers_23_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv215 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_23_mlp_dense_4h_to_h_q_weight5, gpt_neox_layers_23_mlp_dense_4h_to_h_q_scale5, lv214, gpt_neox_layers_23_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv71 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_23_attention_dense_q_weight5, gpt_neox_layers_23_attention_dense_q_scale5, reshape479, gpt_neox_layers_23_attention_dense_bias5, lv215, lv70), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm244 = R.call_tir(cls.layer_norm1, (lv71, gpt_neox_final_layer_norm_weight5, gpt_neox_final_layer_norm_bias5), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv2 = R.call_tir(cls.fused_dequantize_fused_NT_matmul9_cast5, (embed_out_q_weight5, embed_out_q_scale5, layer_norm244), out_sinfo=R.Tensor((1, seq_len, vocab_size), dtype="float32")) gv5: R.Tuple(R.Tensor((1, seq_len, vocab_size), dtype="float32"), R.Object) = lv2, paged_kv_cache R.output(gv5) 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.memory_plan_dynamic_func_output": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) cls = Module paged_kv_cache: R.Object = R.call_pure_packed("vm.builtin.paged_attention_kv_cache_create_reduced", R.shape([max_batch_size, max_total_seq_len, prefill_chunk_size, page_size, support_sliding_window]), R.shape([0, 24]), R.prim_value(16), R.prim_value(16), R.prim_value(64), R.prim_value(1), R.prim_value(1), R.prim_value(10000), R.const(0.0, "float16"), cls.tir_kv_cache_transpose_append, cls.batch_prefill_paged_kv, cls.batch_decode_paged_kv, cls.batch_prefill_paged_kv_sliding_window, cls.batch_decode_paged_kv_sliding_window, cls.batch_prefill_ragged_kv, cls.merge_state_inplace, cls.fused_rope, cls.copy_single_page, cls.tir_kv_cache_debug_get_kv, cls.compact_kv_copy, cls.batch_tree_attn, cls.tree_attn_paged_kv, R.prim_value(0), R.prim_value(0), sinfo_args=(R.Object,)) return paged_kv_cache @R.function def decode(input_embed: R.Tensor((1, 1, 1024), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 128), dtype="uint32"), R.Tensor(("vocab_size", 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor(("vocab_size", 128), dtype="uint32"), R.Tensor(("vocab_size", 32), 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.memory_plan_dynamic_func_output": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) cls = Module with R.dataflow(): gpt_neox_layers_0_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[2] gpt_neox_layers_0_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[3] gpt_neox_layers_0_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[4] gpt_neox_layers_0_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[5] gpt_neox_layers_0_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[6] gpt_neox_layers_0_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[8] gpt_neox_layers_0_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[9] gpt_neox_layers_0_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[10] gpt_neox_layers_0_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[11] gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[12] gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[13] gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[14] gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[15] gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[16] gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[17] gpt_neox_layers_1_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[18] gpt_neox_layers_1_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[19] gpt_neox_layers_1_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[20] gpt_neox_layers_1_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[21] gpt_neox_layers_1_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[22] gpt_neox_layers_1_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[23] gpt_neox_layers_1_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[24] gpt_neox_layers_1_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[25] gpt_neox_layers_1_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[26] gpt_neox_layers_1_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[27] gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[28] gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[29] gpt_neox_layers_1_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[30] gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[31] gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[32] gpt_neox_layers_1_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[33] gpt_neox_layers_2_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[34] gpt_neox_layers_2_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[35] gpt_neox_layers_2_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[36] gpt_neox_layers_2_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[37] gpt_neox_layers_2_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[38] gpt_neox_layers_2_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[39] gpt_neox_layers_2_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[40] gpt_neox_layers_2_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[41] gpt_neox_layers_2_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[42] gpt_neox_layers_2_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[43] gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[44] gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[45] gpt_neox_layers_2_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[46] gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[47] gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[48] gpt_neox_layers_2_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[49] gpt_neox_layers_3_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[50] gpt_neox_layers_3_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[51] gpt_neox_layers_3_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[52] gpt_neox_layers_3_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[53] gpt_neox_layers_3_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[54] gpt_neox_layers_3_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[55] gpt_neox_layers_3_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[56] gpt_neox_layers_3_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[57] gpt_neox_layers_3_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[58] gpt_neox_layers_3_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[59] gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[60] gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[61] gpt_neox_layers_3_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[62] gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[63] gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[64] gpt_neox_layers_3_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[65] gpt_neox_layers_4_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[66] gpt_neox_layers_4_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[67] gpt_neox_layers_4_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[68] gpt_neox_layers_4_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[69] gpt_neox_layers_4_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[70] gpt_neox_layers_4_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[71] gpt_neox_layers_4_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[72] gpt_neox_layers_4_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[73] gpt_neox_layers_4_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[74] gpt_neox_layers_4_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[75] gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[76] gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[77] gpt_neox_layers_4_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[78] gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[79] gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[80] gpt_neox_layers_4_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[81] gpt_neox_layers_5_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[82] gpt_neox_layers_5_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[83] gpt_neox_layers_5_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[84] gpt_neox_layers_5_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[85] gpt_neox_layers_5_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[86] gpt_neox_layers_5_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[87] gpt_neox_layers_5_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[88] gpt_neox_layers_5_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[89] gpt_neox_layers_5_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[90] gpt_neox_layers_5_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[91] gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[92] gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[93] gpt_neox_layers_5_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[94] gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[95] gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[96] gpt_neox_layers_5_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[97] gpt_neox_layers_6_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[98] gpt_neox_layers_6_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[99] gpt_neox_layers_6_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[100] gpt_neox_layers_6_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[101] gpt_neox_layers_6_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[102] gpt_neox_layers_6_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[103] gpt_neox_layers_6_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[104] gpt_neox_layers_6_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[105] gpt_neox_layers_6_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[106] gpt_neox_layers_6_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[107] gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[108] gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[109] gpt_neox_layers_6_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[110] gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[111] gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[112] gpt_neox_layers_6_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[113] gpt_neox_layers_7_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[114] gpt_neox_layers_7_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[115] gpt_neox_layers_7_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[116] gpt_neox_layers_7_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[117] gpt_neox_layers_7_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[118] gpt_neox_layers_7_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[119] gpt_neox_layers_7_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[120] gpt_neox_layers_7_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[121] gpt_neox_layers_7_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[122] gpt_neox_layers_7_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[123] gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[124] gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[125] gpt_neox_layers_7_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[126] gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[127] gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[128] gpt_neox_layers_7_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[129] gpt_neox_layers_8_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[130] gpt_neox_layers_8_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[131] gpt_neox_layers_8_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[132] gpt_neox_layers_8_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[133] gpt_neox_layers_8_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[134] gpt_neox_layers_8_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[135] gpt_neox_layers_8_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[136] gpt_neox_layers_8_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[137] gpt_neox_layers_8_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[138] gpt_neox_layers_8_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[139] gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[140] gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[141] gpt_neox_layers_8_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[142] gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[143] gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[144] gpt_neox_layers_8_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[145] gpt_neox_layers_9_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[146] gpt_neox_layers_9_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[147] gpt_neox_layers_9_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[148] gpt_neox_layers_9_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[149] gpt_neox_layers_9_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[150] gpt_neox_layers_9_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[151] gpt_neox_layers_9_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[152] gpt_neox_layers_9_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[153] gpt_neox_layers_9_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[154] gpt_neox_layers_9_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[155] gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[156] gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[157] gpt_neox_layers_9_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[158] gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[159] gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[160] gpt_neox_layers_9_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[161] gpt_neox_layers_10_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[162] gpt_neox_layers_10_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[163] gpt_neox_layers_10_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[164] gpt_neox_layers_10_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[165] gpt_neox_layers_10_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[166] gpt_neox_layers_10_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[167] gpt_neox_layers_10_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[168] gpt_neox_layers_10_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[169] gpt_neox_layers_10_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[170] gpt_neox_layers_10_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[171] gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[172] gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[173] gpt_neox_layers_10_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[174] gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[175] gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[176] gpt_neox_layers_10_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[177] gpt_neox_layers_11_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[178] gpt_neox_layers_11_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[179] gpt_neox_layers_11_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[180] gpt_neox_layers_11_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[181] gpt_neox_layers_11_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[182] gpt_neox_layers_11_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[183] gpt_neox_layers_11_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[184] gpt_neox_layers_11_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[185] gpt_neox_layers_11_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[186] gpt_neox_layers_11_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[187] gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[188] gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[189] gpt_neox_layers_11_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[190] gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[191] gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[192] gpt_neox_layers_11_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[193] gpt_neox_layers_12_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[194] gpt_neox_layers_12_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[195] gpt_neox_layers_12_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[196] gpt_neox_layers_12_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[197] gpt_neox_layers_12_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[198] gpt_neox_layers_12_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[199] gpt_neox_layers_12_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[200] gpt_neox_layers_12_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[201] gpt_neox_layers_12_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[202] gpt_neox_layers_12_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[203] gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[204] gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[205] gpt_neox_layers_12_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[206] gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[207] gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[208] gpt_neox_layers_12_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[209] gpt_neox_layers_13_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[210] gpt_neox_layers_13_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[211] gpt_neox_layers_13_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[212] gpt_neox_layers_13_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[213] gpt_neox_layers_13_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[214] gpt_neox_layers_13_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[215] gpt_neox_layers_13_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[216] gpt_neox_layers_13_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[217] gpt_neox_layers_13_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[218] gpt_neox_layers_13_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[219] gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[220] gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[221] gpt_neox_layers_13_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[222] gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[223] gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[224] gpt_neox_layers_13_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[225] gpt_neox_layers_14_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[226] gpt_neox_layers_14_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[227] gpt_neox_layers_14_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[228] gpt_neox_layers_14_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[229] gpt_neox_layers_14_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[230] gpt_neox_layers_14_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[231] gpt_neox_layers_14_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[232] gpt_neox_layers_14_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[233] gpt_neox_layers_14_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[234] gpt_neox_layers_14_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[235] gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[236] gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[237] gpt_neox_layers_14_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[238] gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[239] gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[240] gpt_neox_layers_14_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[241] gpt_neox_layers_15_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[242] gpt_neox_layers_15_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[243] gpt_neox_layers_15_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[244] gpt_neox_layers_15_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[245] gpt_neox_layers_15_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[246] gpt_neox_layers_15_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[247] gpt_neox_layers_15_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[248] gpt_neox_layers_15_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[249] gpt_neox_layers_15_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[250] gpt_neox_layers_15_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[251] gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[252] gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[253] gpt_neox_layers_15_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[254] gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[255] gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[256] gpt_neox_layers_15_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[257] gpt_neox_layers_16_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[258] gpt_neox_layers_16_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[259] gpt_neox_layers_16_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[260] gpt_neox_layers_16_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[261] gpt_neox_layers_16_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[262] gpt_neox_layers_16_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[263] gpt_neox_layers_16_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[264] gpt_neox_layers_16_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[265] gpt_neox_layers_16_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[266] gpt_neox_layers_16_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[267] gpt_neox_layers_16_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[268] gpt_neox_layers_16_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[269] gpt_neox_layers_16_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[270] gpt_neox_layers_16_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[271] gpt_neox_layers_16_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[272] gpt_neox_layers_16_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[273] gpt_neox_layers_17_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[274] gpt_neox_layers_17_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[275] gpt_neox_layers_17_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[276] gpt_neox_layers_17_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[277] gpt_neox_layers_17_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[278] gpt_neox_layers_17_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[279] gpt_neox_layers_17_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[280] gpt_neox_layers_17_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[281] gpt_neox_layers_17_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[282] gpt_neox_layers_17_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[283] gpt_neox_layers_17_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[284] gpt_neox_layers_17_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[285] gpt_neox_layers_17_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[286] gpt_neox_layers_17_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[287] gpt_neox_layers_17_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[288] gpt_neox_layers_17_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[289] gpt_neox_layers_18_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[290] gpt_neox_layers_18_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[291] gpt_neox_layers_18_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[292] gpt_neox_layers_18_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[293] gpt_neox_layers_18_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[294] gpt_neox_layers_18_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[295] gpt_neox_layers_18_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[296] gpt_neox_layers_18_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[297] gpt_neox_layers_18_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[298] gpt_neox_layers_18_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[299] gpt_neox_layers_18_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[300] gpt_neox_layers_18_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[301] gpt_neox_layers_18_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[302] gpt_neox_layers_18_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[303] gpt_neox_layers_18_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[304] gpt_neox_layers_18_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[305] gpt_neox_layers_19_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[306] gpt_neox_layers_19_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[307] gpt_neox_layers_19_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[308] gpt_neox_layers_19_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[309] gpt_neox_layers_19_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[310] gpt_neox_layers_19_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[311] gpt_neox_layers_19_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[312] gpt_neox_layers_19_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[313] gpt_neox_layers_19_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[314] gpt_neox_layers_19_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[315] gpt_neox_layers_19_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[316] gpt_neox_layers_19_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[317] gpt_neox_layers_19_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[318] gpt_neox_layers_19_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[319] gpt_neox_layers_19_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[320] gpt_neox_layers_19_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[321] gpt_neox_layers_20_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[322] gpt_neox_layers_20_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[323] gpt_neox_layers_20_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[324] gpt_neox_layers_20_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[325] gpt_neox_layers_20_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[326] gpt_neox_layers_20_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[327] gpt_neox_layers_20_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[328] gpt_neox_layers_20_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[329] gpt_neox_layers_20_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[330] gpt_neox_layers_20_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[331] gpt_neox_layers_20_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[332] gpt_neox_layers_20_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[333] gpt_neox_layers_20_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[334] gpt_neox_layers_20_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[335] gpt_neox_layers_20_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[336] gpt_neox_layers_20_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[337] gpt_neox_layers_21_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[338] gpt_neox_layers_21_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[339] gpt_neox_layers_21_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[340] gpt_neox_layers_21_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[341] gpt_neox_layers_21_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[342] gpt_neox_layers_21_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[343] gpt_neox_layers_21_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[344] gpt_neox_layers_21_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[345] gpt_neox_layers_21_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[346] gpt_neox_layers_21_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[347] gpt_neox_layers_21_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[348] gpt_neox_layers_21_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[349] gpt_neox_layers_21_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[350] gpt_neox_layers_21_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[351] gpt_neox_layers_21_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[352] gpt_neox_layers_21_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[353] gpt_neox_layers_22_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[354] gpt_neox_layers_22_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[355] gpt_neox_layers_22_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[356] gpt_neox_layers_22_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[357] gpt_neox_layers_22_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[358] gpt_neox_layers_22_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[359] gpt_neox_layers_22_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[360] gpt_neox_layers_22_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[361] gpt_neox_layers_22_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[362] gpt_neox_layers_22_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[363] gpt_neox_layers_22_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[364] gpt_neox_layers_22_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[365] gpt_neox_layers_22_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[366] gpt_neox_layers_22_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[367] gpt_neox_layers_22_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[368] gpt_neox_layers_22_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[369] gpt_neox_layers_23_input_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[370] gpt_neox_layers_23_input_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[371] gpt_neox_layers_23_post_attention_layernorm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[372] gpt_neox_layers_23_post_attention_layernorm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[373] gpt_neox_layers_23_attention_query_key_value_q_weight2: R.Tensor((3072, 128), dtype="uint32") = packed_params[374] gpt_neox_layers_23_attention_query_key_value_q_scale2: R.Tensor((3072, 32), dtype="float16") = packed_params[375] gpt_neox_layers_23_attention_query_key_value_bias2: R.Tensor((3072,), dtype="float16") = packed_params[376] gpt_neox_layers_23_attention_dense_q_weight2: R.Tensor((1024, 128), dtype="uint32") = packed_params[377] gpt_neox_layers_23_attention_dense_q_scale2: R.Tensor((1024, 32), dtype="float16") = packed_params[378] gpt_neox_layers_23_attention_dense_bias2: R.Tensor((1024,), dtype="float16") = packed_params[379] gpt_neox_layers_23_mlp_dense_h_to_4h_q_weight2: R.Tensor((4096, 128), dtype="uint32") = packed_params[380] gpt_neox_layers_23_mlp_dense_h_to_4h_q_scale2: R.Tensor((4096, 32), dtype="float16") = packed_params[381] gpt_neox_layers_23_mlp_dense_h_to_4h_bias2: R.Tensor((4096,), dtype="float32") = packed_params[382] gpt_neox_layers_23_mlp_dense_4h_to_h_q_weight2: R.Tensor((1024, 512), dtype="uint32") = packed_params[383] gpt_neox_layers_23_mlp_dense_4h_to_h_q_scale2: R.Tensor((1024, 128), dtype="float16") = packed_params[384] gpt_neox_layers_23_mlp_dense_4h_to_h_bias2: R.Tensor((1024,), dtype="float32") = packed_params[385] gpt_neox_final_layer_norm_weight2: R.Tensor((1024,), dtype="float16") = packed_params[386] gpt_neox_final_layer_norm_bias2: R.Tensor((1024,), dtype="float16") = packed_params[387] embed_out_q_weight2: R.Tensor((vocab_size, 128), dtype="uint32") = packed_params[388] embed_out_q_scale2: R.Tensor((vocab_size, 32), dtype="float16") = packed_params[389] layer_norm49 = R.call_tir(cls.layer_norm2, (input_embed, gpt_neox_layers_0_input_layernorm_weight2, gpt_neox_layers_0_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv216 = R.call_tir(cls.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, layer_norm49, gpt_neox_layers_0_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv292 = R.call_tir(cls.fused_reshape8_reshape9, (lv216,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv124 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), lv292), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv293 = R.call_tir(cls.fused_reshape10_reshape11, (lv124,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm50 = R.call_tir(cls.layer_norm2, (input_embed, gpt_neox_layers_0_post_attention_layernorm_weight2, gpt_neox_layers_0_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv217 = R.call_tir(cls.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, layer_norm50, gpt_neox_layers_0_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv218 = R.call_tir(cls.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, lv217, gpt_neox_layers_0_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv72 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_0_attention_dense_q_weight2, gpt_neox_layers_0_attention_dense_q_scale2, lv293, gpt_neox_layers_0_attention_dense_bias2, lv218, input_embed), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm51 = R.call_tir(cls.layer_norm2, (lv72, gpt_neox_layers_1_input_layernorm_weight2, gpt_neox_layers_1_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv219 = R.call_tir(cls.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, layer_norm51, gpt_neox_layers_1_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv298 = R.call_tir(cls.fused_reshape8_reshape9, (lv219,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv129 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), lv298), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv299 = R.call_tir(cls.fused_reshape10_reshape11, (lv129,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm52 = R.call_tir(cls.layer_norm2, (lv72, gpt_neox_layers_1_post_attention_layernorm_weight2, gpt_neox_layers_1_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv220 = R.call_tir(cls.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, layer_norm52, gpt_neox_layers_1_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv221 = R.call_tir(cls.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, lv220, gpt_neox_layers_1_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv73 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_1_attention_dense_q_weight2, gpt_neox_layers_1_attention_dense_q_scale2, lv299, gpt_neox_layers_1_attention_dense_bias2, lv221, lv72), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm53 = R.call_tir(cls.layer_norm2, (lv73, gpt_neox_layers_2_input_layernorm_weight2, gpt_neox_layers_2_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv222 = R.call_tir(cls.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, layer_norm53, gpt_neox_layers_2_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv304 = R.call_tir(cls.fused_reshape8_reshape9, (lv222,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv134 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), lv304), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv305 = R.call_tir(cls.fused_reshape10_reshape11, (lv134,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm54 = R.call_tir(cls.layer_norm2, (lv73, gpt_neox_layers_2_post_attention_layernorm_weight2, gpt_neox_layers_2_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv223 = R.call_tir(cls.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, layer_norm54, gpt_neox_layers_2_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv224 = R.call_tir(cls.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, lv223, gpt_neox_layers_2_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv74 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_2_attention_dense_q_weight2, gpt_neox_layers_2_attention_dense_q_scale2, lv305, gpt_neox_layers_2_attention_dense_bias2, lv224, lv73), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm55 = R.call_tir(cls.layer_norm2, (lv74, gpt_neox_layers_3_input_layernorm_weight2, gpt_neox_layers_3_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv225 = R.call_tir(cls.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, layer_norm55, gpt_neox_layers_3_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv310 = R.call_tir(cls.fused_reshape8_reshape9, (lv225,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv139 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), lv310), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv311 = R.call_tir(cls.fused_reshape10_reshape11, (lv139,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm56 = R.call_tir(cls.layer_norm2, (lv74, gpt_neox_layers_3_post_attention_layernorm_weight2, gpt_neox_layers_3_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv226 = R.call_tir(cls.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, layer_norm56, gpt_neox_layers_3_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv227 = R.call_tir(cls.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, lv226, gpt_neox_layers_3_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv75 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_3_attention_dense_q_weight2, gpt_neox_layers_3_attention_dense_q_scale2, lv311, gpt_neox_layers_3_attention_dense_bias2, lv227, lv74), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm57 = R.call_tir(cls.layer_norm2, (lv75, gpt_neox_layers_4_input_layernorm_weight2, gpt_neox_layers_4_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv228 = R.call_tir(cls.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, layer_norm57, gpt_neox_layers_4_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv316 = R.call_tir(cls.fused_reshape8_reshape9, (lv228,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv144 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), lv316), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv317 = R.call_tir(cls.fused_reshape10_reshape11, (lv144,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm58 = R.call_tir(cls.layer_norm2, (lv75, gpt_neox_layers_4_post_attention_layernorm_weight2, gpt_neox_layers_4_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv229 = R.call_tir(cls.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, layer_norm58, gpt_neox_layers_4_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv230 = R.call_tir(cls.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, lv229, gpt_neox_layers_4_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv76 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_4_attention_dense_q_weight2, gpt_neox_layers_4_attention_dense_q_scale2, lv317, gpt_neox_layers_4_attention_dense_bias2, lv230, lv75), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm59 = R.call_tir(cls.layer_norm2, (lv76, gpt_neox_layers_5_input_layernorm_weight2, gpt_neox_layers_5_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv231 = R.call_tir(cls.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, layer_norm59, gpt_neox_layers_5_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv322 = R.call_tir(cls.fused_reshape8_reshape9, (lv231,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv149 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), lv322), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv323 = R.call_tir(cls.fused_reshape10_reshape11, (lv149,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm60 = R.call_tir(cls.layer_norm2, (lv76, gpt_neox_layers_5_post_attention_layernorm_weight2, gpt_neox_layers_5_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv232 = R.call_tir(cls.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, layer_norm60, gpt_neox_layers_5_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv233 = R.call_tir(cls.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, lv232, gpt_neox_layers_5_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv77 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_5_attention_dense_q_weight2, gpt_neox_layers_5_attention_dense_q_scale2, lv323, gpt_neox_layers_5_attention_dense_bias2, lv233, lv76), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm61 = R.call_tir(cls.layer_norm2, (lv77, gpt_neox_layers_6_input_layernorm_weight2, gpt_neox_layers_6_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv234 = R.call_tir(cls.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, layer_norm61, gpt_neox_layers_6_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv328 = R.call_tir(cls.fused_reshape8_reshape9, (lv234,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv154 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), lv328), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv329 = R.call_tir(cls.fused_reshape10_reshape11, (lv154,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm62 = R.call_tir(cls.layer_norm2, (lv77, gpt_neox_layers_6_post_attention_layernorm_weight2, gpt_neox_layers_6_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv235 = R.call_tir(cls.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, layer_norm62, gpt_neox_layers_6_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv236 = R.call_tir(cls.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, lv235, gpt_neox_layers_6_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv78 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_6_attention_dense_q_weight2, gpt_neox_layers_6_attention_dense_q_scale2, lv329, gpt_neox_layers_6_attention_dense_bias2, lv236, lv77), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm63 = R.call_tir(cls.layer_norm2, (lv78, gpt_neox_layers_7_input_layernorm_weight2, gpt_neox_layers_7_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv237 = R.call_tir(cls.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, layer_norm63, gpt_neox_layers_7_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv334 = R.call_tir(cls.fused_reshape8_reshape9, (lv237,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv159 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), lv334), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv335 = R.call_tir(cls.fused_reshape10_reshape11, (lv159,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm64 = R.call_tir(cls.layer_norm2, (lv78, gpt_neox_layers_7_post_attention_layernorm_weight2, gpt_neox_layers_7_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv238 = R.call_tir(cls.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, layer_norm64, gpt_neox_layers_7_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv239 = R.call_tir(cls.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, lv238, gpt_neox_layers_7_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv79 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_7_attention_dense_q_weight2, gpt_neox_layers_7_attention_dense_q_scale2, lv335, gpt_neox_layers_7_attention_dense_bias2, lv239, lv78), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm65 = R.call_tir(cls.layer_norm2, (lv79, gpt_neox_layers_8_input_layernorm_weight2, gpt_neox_layers_8_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv240 = R.call_tir(cls.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, layer_norm65, gpt_neox_layers_8_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv340 = R.call_tir(cls.fused_reshape8_reshape9, (lv240,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv164 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), lv340), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv341 = R.call_tir(cls.fused_reshape10_reshape11, (lv164,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm66 = R.call_tir(cls.layer_norm2, (lv79, gpt_neox_layers_8_post_attention_layernorm_weight2, gpt_neox_layers_8_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv241 = R.call_tir(cls.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, layer_norm66, gpt_neox_layers_8_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv242 = R.call_tir(cls.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, lv241, gpt_neox_layers_8_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv80 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_8_attention_dense_q_weight2, gpt_neox_layers_8_attention_dense_q_scale2, lv341, gpt_neox_layers_8_attention_dense_bias2, lv242, lv79), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm67 = R.call_tir(cls.layer_norm2, (lv80, gpt_neox_layers_9_input_layernorm_weight2, gpt_neox_layers_9_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv243 = R.call_tir(cls.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, layer_norm67, gpt_neox_layers_9_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv346 = R.call_tir(cls.fused_reshape8_reshape9, (lv243,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv169 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), lv346), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv347 = R.call_tir(cls.fused_reshape10_reshape11, (lv169,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm68 = R.call_tir(cls.layer_norm2, (lv80, gpt_neox_layers_9_post_attention_layernorm_weight2, gpt_neox_layers_9_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv244 = R.call_tir(cls.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, layer_norm68, gpt_neox_layers_9_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv245 = R.call_tir(cls.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, lv244, gpt_neox_layers_9_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv81 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_9_attention_dense_q_weight2, gpt_neox_layers_9_attention_dense_q_scale2, lv347, gpt_neox_layers_9_attention_dense_bias2, lv245, lv80), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm69 = R.call_tir(cls.layer_norm2, (lv81, gpt_neox_layers_10_input_layernorm_weight2, gpt_neox_layers_10_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv246 = R.call_tir(cls.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, layer_norm69, gpt_neox_layers_10_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv352 = R.call_tir(cls.fused_reshape8_reshape9, (lv246,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv174 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), lv352), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv353 = R.call_tir(cls.fused_reshape10_reshape11, (lv174,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm70 = R.call_tir(cls.layer_norm2, (lv81, gpt_neox_layers_10_post_attention_layernorm_weight2, gpt_neox_layers_10_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv247 = R.call_tir(cls.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, layer_norm70, gpt_neox_layers_10_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv248 = R.call_tir(cls.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, lv247, gpt_neox_layers_10_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv82 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_10_attention_dense_q_weight2, gpt_neox_layers_10_attention_dense_q_scale2, lv353, gpt_neox_layers_10_attention_dense_bias2, lv248, lv81), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm71 = R.call_tir(cls.layer_norm2, (lv82, gpt_neox_layers_11_input_layernorm_weight2, gpt_neox_layers_11_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv249 = R.call_tir(cls.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, layer_norm71, gpt_neox_layers_11_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv358 = R.call_tir(cls.fused_reshape8_reshape9, (lv249,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv179 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), lv358), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv359 = R.call_tir(cls.fused_reshape10_reshape11, (lv179,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm72 = R.call_tir(cls.layer_norm2, (lv82, gpt_neox_layers_11_post_attention_layernorm_weight2, gpt_neox_layers_11_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv250 = R.call_tir(cls.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, layer_norm72, gpt_neox_layers_11_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv251 = R.call_tir(cls.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, lv250, gpt_neox_layers_11_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv83 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_11_attention_dense_q_weight2, gpt_neox_layers_11_attention_dense_q_scale2, lv359, gpt_neox_layers_11_attention_dense_bias2, lv251, lv82), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm73 = R.call_tir(cls.layer_norm2, (lv83, gpt_neox_layers_12_input_layernorm_weight2, gpt_neox_layers_12_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv252 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul10_add10, (gpt_neox_layers_12_attention_query_key_value_q_weight2, gpt_neox_layers_12_attention_query_key_value_q_scale2, layer_norm73, gpt_neox_layers_12_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv364 = R.call_tir(cls.fused_reshape8_reshape9, (lv252,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv184 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), lv364), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv365 = R.call_tir(cls.fused_reshape10_reshape11, (lv184,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm74 = R.call_tir(cls.layer_norm2, (lv83, gpt_neox_layers_12_post_attention_layernorm_weight2, gpt_neox_layers_12_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv253 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6, (gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale2, layer_norm74, gpt_neox_layers_12_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv254 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul13_add13_cast7, (gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale2, lv253, gpt_neox_layers_12_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv84 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_12_attention_dense_q_weight2, gpt_neox_layers_12_attention_dense_q_scale2, lv365, gpt_neox_layers_12_attention_dense_bias2, lv254, lv83), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm75 = R.call_tir(cls.layer_norm2, (lv84, gpt_neox_layers_13_input_layernorm_weight2, gpt_neox_layers_13_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv255 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul10_add10, (gpt_neox_layers_13_attention_query_key_value_q_weight2, gpt_neox_layers_13_attention_query_key_value_q_scale2, layer_norm75, gpt_neox_layers_13_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv370 = R.call_tir(cls.fused_reshape8_reshape9, (lv255,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv189 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), lv370), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv371 = R.call_tir(cls.fused_reshape10_reshape11, (lv189,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm76 = R.call_tir(cls.layer_norm2, (lv84, gpt_neox_layers_13_post_attention_layernorm_weight2, gpt_neox_layers_13_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv256 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6, (gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale2, layer_norm76, gpt_neox_layers_13_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv257 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul13_add13_cast7, (gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale2, lv256, gpt_neox_layers_13_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv85 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_13_attention_dense_q_weight2, gpt_neox_layers_13_attention_dense_q_scale2, lv371, gpt_neox_layers_13_attention_dense_bias2, lv257, lv84), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm77 = R.call_tir(cls.layer_norm2, (lv85, gpt_neox_layers_14_input_layernorm_weight2, gpt_neox_layers_14_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv258 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul10_add10, (gpt_neox_layers_14_attention_query_key_value_q_weight2, gpt_neox_layers_14_attention_query_key_value_q_scale2, layer_norm77, gpt_neox_layers_14_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv376 = R.call_tir(cls.fused_reshape8_reshape9, (lv258,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv194 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), lv376), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv377 = R.call_tir(cls.fused_reshape10_reshape11, (lv194,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm78 = R.call_tir(cls.layer_norm2, (lv85, gpt_neox_layers_14_post_attention_layernorm_weight2, gpt_neox_layers_14_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv259 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6, (gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale2, layer_norm78, gpt_neox_layers_14_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv260 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul13_add13_cast7, (gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale2, lv259, gpt_neox_layers_14_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv86 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_14_attention_dense_q_weight2, gpt_neox_layers_14_attention_dense_q_scale2, lv377, gpt_neox_layers_14_attention_dense_bias2, lv260, lv85), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm79 = R.call_tir(cls.layer_norm2, (lv86, gpt_neox_layers_15_input_layernorm_weight2, gpt_neox_layers_15_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv261 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul10_add10, (gpt_neox_layers_15_attention_query_key_value_q_weight2, gpt_neox_layers_15_attention_query_key_value_q_scale2, layer_norm79, gpt_neox_layers_15_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv382 = R.call_tir(cls.fused_reshape8_reshape9, (lv261,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv199 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), lv382), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv383 = R.call_tir(cls.fused_reshape10_reshape11, (lv199,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm80 = R.call_tir(cls.layer_norm2, (lv86, gpt_neox_layers_15_post_attention_layernorm_weight2, gpt_neox_layers_15_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv262 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6, (gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale2, layer_norm80, gpt_neox_layers_15_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv263 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul13_add13_cast7, (gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale2, lv262, gpt_neox_layers_15_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv87 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_15_attention_dense_q_weight2, gpt_neox_layers_15_attention_dense_q_scale2, lv383, gpt_neox_layers_15_attention_dense_bias2, lv263, lv86), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm81 = R.call_tir(cls.layer_norm2, (lv87, gpt_neox_layers_16_input_layernorm_weight2, gpt_neox_layers_16_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv264 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul10_add10, (gpt_neox_layers_16_attention_query_key_value_q_weight2, gpt_neox_layers_16_attention_query_key_value_q_scale2, layer_norm81, gpt_neox_layers_16_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv388 = R.call_tir(cls.fused_reshape8_reshape9, (lv264,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv204 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), lv388), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv389 = R.call_tir(cls.fused_reshape10_reshape11, (lv204,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm82 = R.call_tir(cls.layer_norm2, (lv87, gpt_neox_layers_16_post_attention_layernorm_weight2, gpt_neox_layers_16_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv265 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6, (gpt_neox_layers_16_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_16_mlp_dense_h_to_4h_q_scale2, layer_norm82, gpt_neox_layers_16_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv266 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul13_add13_cast7, (gpt_neox_layers_16_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_16_mlp_dense_4h_to_h_q_scale2, lv265, gpt_neox_layers_16_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv88 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_16_attention_dense_q_weight2, gpt_neox_layers_16_attention_dense_q_scale2, lv389, gpt_neox_layers_16_attention_dense_bias2, lv266, lv87), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm83 = R.call_tir(cls.layer_norm2, (lv88, gpt_neox_layers_17_input_layernorm_weight2, gpt_neox_layers_17_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv267 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul10_add10, (gpt_neox_layers_17_attention_query_key_value_q_weight2, gpt_neox_layers_17_attention_query_key_value_q_scale2, layer_norm83, gpt_neox_layers_17_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv394 = R.call_tir(cls.fused_reshape8_reshape9, (lv267,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv209 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), lv394), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv395 = R.call_tir(cls.fused_reshape10_reshape11, (lv209,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm84 = R.call_tir(cls.layer_norm2, (lv88, gpt_neox_layers_17_post_attention_layernorm_weight2, gpt_neox_layers_17_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv268 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6, (gpt_neox_layers_17_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_17_mlp_dense_h_to_4h_q_scale2, layer_norm84, gpt_neox_layers_17_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv269 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul13_add13_cast7, (gpt_neox_layers_17_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_17_mlp_dense_4h_to_h_q_scale2, lv268, gpt_neox_layers_17_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv89 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_17_attention_dense_q_weight2, gpt_neox_layers_17_attention_dense_q_scale2, lv395, gpt_neox_layers_17_attention_dense_bias2, lv269, lv88), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm85 = R.call_tir(cls.layer_norm2, (lv89, gpt_neox_layers_18_input_layernorm_weight2, gpt_neox_layers_18_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv270 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul10_add10, (gpt_neox_layers_18_attention_query_key_value_q_weight2, gpt_neox_layers_18_attention_query_key_value_q_scale2, layer_norm85, gpt_neox_layers_18_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv400 = R.call_tir(cls.fused_reshape8_reshape9, (lv270,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv214 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), lv400), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv401 = R.call_tir(cls.fused_reshape10_reshape11, (lv214,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm86 = R.call_tir(cls.layer_norm2, (lv89, gpt_neox_layers_18_post_attention_layernorm_weight2, gpt_neox_layers_18_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv271 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6, (gpt_neox_layers_18_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_18_mlp_dense_h_to_4h_q_scale2, layer_norm86, gpt_neox_layers_18_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv272 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul13_add13_cast7, (gpt_neox_layers_18_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_18_mlp_dense_4h_to_h_q_scale2, lv271, gpt_neox_layers_18_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv90 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_18_attention_dense_q_weight2, gpt_neox_layers_18_attention_dense_q_scale2, lv401, gpt_neox_layers_18_attention_dense_bias2, lv272, lv89), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm87 = R.call_tir(cls.layer_norm2, (lv90, gpt_neox_layers_19_input_layernorm_weight2, gpt_neox_layers_19_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv273 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul10_add10, (gpt_neox_layers_19_attention_query_key_value_q_weight2, gpt_neox_layers_19_attention_query_key_value_q_scale2, layer_norm87, gpt_neox_layers_19_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv406 = R.call_tir(cls.fused_reshape8_reshape9, (lv273,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv219_1 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), lv406), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv407 = R.call_tir(cls.fused_reshape10_reshape11, (lv219_1,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm88 = R.call_tir(cls.layer_norm2, (lv90, gpt_neox_layers_19_post_attention_layernorm_weight2, gpt_neox_layers_19_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv274 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6, (gpt_neox_layers_19_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_19_mlp_dense_h_to_4h_q_scale2, layer_norm88, gpt_neox_layers_19_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv275 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul13_add13_cast7, (gpt_neox_layers_19_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_19_mlp_dense_4h_to_h_q_scale2, lv274, gpt_neox_layers_19_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv91 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_19_attention_dense_q_weight2, gpt_neox_layers_19_attention_dense_q_scale2, lv407, gpt_neox_layers_19_attention_dense_bias2, lv275, lv90), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm89 = R.call_tir(cls.layer_norm2, (lv91, gpt_neox_layers_20_input_layernorm_weight2, gpt_neox_layers_20_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv276 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul10_add10, (gpt_neox_layers_20_attention_query_key_value_q_weight2, gpt_neox_layers_20_attention_query_key_value_q_scale2, layer_norm89, gpt_neox_layers_20_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv412 = R.call_tir(cls.fused_reshape8_reshape9, (lv276,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv224_1 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), lv412), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv413 = R.call_tir(cls.fused_reshape10_reshape11, (lv224_1,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm90 = R.call_tir(cls.layer_norm2, (lv91, gpt_neox_layers_20_post_attention_layernorm_weight2, gpt_neox_layers_20_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv277 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6, (gpt_neox_layers_20_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_20_mlp_dense_h_to_4h_q_scale2, layer_norm90, gpt_neox_layers_20_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv278 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul13_add13_cast7, (gpt_neox_layers_20_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_20_mlp_dense_4h_to_h_q_scale2, lv277, gpt_neox_layers_20_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv92 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_20_attention_dense_q_weight2, gpt_neox_layers_20_attention_dense_q_scale2, lv413, gpt_neox_layers_20_attention_dense_bias2, lv278, lv91), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm91 = R.call_tir(cls.layer_norm2, (lv92, gpt_neox_layers_21_input_layernorm_weight2, gpt_neox_layers_21_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv279 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul10_add10, (gpt_neox_layers_21_attention_query_key_value_q_weight2, gpt_neox_layers_21_attention_query_key_value_q_scale2, layer_norm91, gpt_neox_layers_21_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv418 = R.call_tir(cls.fused_reshape8_reshape9, (lv279,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv229_1 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), lv418), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv419 = R.call_tir(cls.fused_reshape10_reshape11, (lv229_1,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm92 = R.call_tir(cls.layer_norm2, (lv92, gpt_neox_layers_21_post_attention_layernorm_weight2, gpt_neox_layers_21_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv280 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6, (gpt_neox_layers_21_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_21_mlp_dense_h_to_4h_q_scale2, layer_norm92, gpt_neox_layers_21_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv281 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul13_add13_cast7, (gpt_neox_layers_21_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_21_mlp_dense_4h_to_h_q_scale2, lv280, gpt_neox_layers_21_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv93 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_21_attention_dense_q_weight2, gpt_neox_layers_21_attention_dense_q_scale2, lv419, gpt_neox_layers_21_attention_dense_bias2, lv281, lv92), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm93 = R.call_tir(cls.layer_norm2, (lv93, gpt_neox_layers_22_input_layernorm_weight2, gpt_neox_layers_22_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv282 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul10_add10, (gpt_neox_layers_22_attention_query_key_value_q_weight2, gpt_neox_layers_22_attention_query_key_value_q_scale2, layer_norm93, gpt_neox_layers_22_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv424 = R.call_tir(cls.fused_reshape8_reshape9, (lv282,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv234_1 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), lv424), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv425 = R.call_tir(cls.fused_reshape10_reshape11, (lv234_1,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm94 = R.call_tir(cls.layer_norm2, (lv93, gpt_neox_layers_22_post_attention_layernorm_weight2, gpt_neox_layers_22_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv283 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6, (gpt_neox_layers_22_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_22_mlp_dense_h_to_4h_q_scale2, layer_norm94, gpt_neox_layers_22_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv284 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul13_add13_cast7, (gpt_neox_layers_22_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_22_mlp_dense_4h_to_h_q_scale2, lv283, gpt_neox_layers_22_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv94 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_22_attention_dense_q_weight2, gpt_neox_layers_22_attention_dense_q_scale2, lv425, gpt_neox_layers_22_attention_dense_bias2, lv284, lv93), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm95 = R.call_tir(cls.layer_norm2, (lv94, gpt_neox_layers_23_input_layernorm_weight2, gpt_neox_layers_23_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv285 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul10_add10, (gpt_neox_layers_23_attention_query_key_value_q_weight2, gpt_neox_layers_23_attention_query_key_value_q_scale2, layer_norm95, gpt_neox_layers_23_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 3072), dtype="float16")) lv430 = R.call_tir(cls.fused_reshape8_reshape9, (lv285,), out_sinfo=R.Tensor((1, 48, 64), dtype="float16")) lv239_1 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), lv430), out_sinfo=R.Tensor((1, 16, 64), dtype="float16")) lv431 = R.call_tir(cls.fused_reshape10_reshape11, (lv239_1,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm96 = R.call_tir(cls.layer_norm2, (lv94, gpt_neox_layers_23_post_attention_layernorm_weight2, gpt_neox_layers_23_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv286 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul12_add12_gelu2_cast6, (gpt_neox_layers_23_mlp_dense_h_to_4h_q_weight2, gpt_neox_layers_23_mlp_dense_h_to_4h_q_scale2, layer_norm96, gpt_neox_layers_23_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 4096), dtype="float16")) lv287 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul13_add13_cast7, (gpt_neox_layers_23_mlp_dense_4h_to_h_q_weight2, gpt_neox_layers_23_mlp_dense_4h_to_h_q_scale2, lv286, gpt_neox_layers_23_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv95 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul11_add11_add14_add14, (gpt_neox_layers_23_attention_dense_q_weight2, gpt_neox_layers_23_attention_dense_q_scale2, lv431, gpt_neox_layers_23_attention_dense_bias2, lv287, lv94), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) layer_norm97 = R.call_tir(cls.layer_norm2, (lv95, gpt_neox_final_layer_norm_weight2, gpt_neox_final_layer_norm_bias2), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv3 = R.call_tir(cls.fused_dequantize_fused_NT_matmul14_cast8, (embed_out_q_weight2, embed_out_q_scale2, layer_norm97), out_sinfo=R.Tensor((1, 1, vocab_size), dtype="float32")) gv2: R.Tuple(R.Tensor((1, 1, vocab_size), dtype="float32"), R.Object) = lv3, paged_kv_cache R.output(gv2) return gv2 @R.function def embed(input_ids: R.Tensor(("seq_len",), dtype="int32"), packed_params: R.Tuple(R.Tensor(("vocab_size", 128), dtype="uint32"), R.Tensor(("vocab_size", 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor(("vocab_size", 128), dtype="uint32"), R.Tensor(("vocab_size", 32), dtype="float16"))) -> R.Tensor(("seq_len", 1024), dtype="float16"): seq_len = T.int64() vocab_size = T.int64() R.func_attr({"num_input": 1, "relax.memory_plan_dynamic_func_output": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) cls = Module with R.dataflow(): gpt_neox_embed_in_q_weight: R.Tensor((vocab_size, 128), dtype="uint32") = packed_params[0] gpt_neox_embed_in_q_scale: R.Tensor((vocab_size, 32), dtype="float16") = packed_params[1] gv = R.call_tir(cls.fused_dequantize_take1, (gpt_neox_embed_in_q_weight, gpt_neox_embed_in_q_scale, input_ids), out_sinfo=R.Tensor((seq_len, 1024), dtype="float16")) R.output(gv) return gv @R.function def prefill(input_embed: R.Tensor((1, "seq_len", 1024), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 128), dtype="uint32"), R.Tensor(("vocab_size", 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((3072, 128), dtype="uint32"), R.Tensor((3072, 32), dtype="float16"), R.Tensor((3072,), dtype="float16"), R.Tensor((1024, 128), dtype="uint32"), R.Tensor((1024, 32), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor((4096, 128), dtype="uint32"), R.Tensor((4096, 32), dtype="float16"), R.Tensor((4096,), dtype="float32"), R.Tensor((1024, 512), dtype="uint32"), R.Tensor((1024, 128), dtype="float16"), R.Tensor((1024,), dtype="float32"), R.Tensor((1024,), dtype="float16"), R.Tensor((1024,), dtype="float16"), R.Tensor(("vocab_size", 128), dtype="uint32"), R.Tensor(("vocab_size", 32), 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.memory_plan_dynamic_func_output": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) cls = Module with R.dataflow(): gpt_neox_layers_0_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[2] gpt_neox_layers_0_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[3] gpt_neox_layers_0_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[4] gpt_neox_layers_0_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[5] gpt_neox_layers_0_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[6] gpt_neox_layers_0_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[8] gpt_neox_layers_0_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[9] gpt_neox_layers_0_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[10] gpt_neox_layers_0_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[11] gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[12] gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[13] gpt_neox_layers_0_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[14] gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[15] gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[16] gpt_neox_layers_0_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[17] gpt_neox_layers_1_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[18] gpt_neox_layers_1_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[19] gpt_neox_layers_1_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[20] gpt_neox_layers_1_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[21] gpt_neox_layers_1_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[22] gpt_neox_layers_1_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[23] gpt_neox_layers_1_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[24] gpt_neox_layers_1_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[25] gpt_neox_layers_1_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[26] gpt_neox_layers_1_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[27] gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[28] gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[29] gpt_neox_layers_1_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[30] gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[31] gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[32] gpt_neox_layers_1_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[33] gpt_neox_layers_2_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[34] gpt_neox_layers_2_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[35] gpt_neox_layers_2_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[36] gpt_neox_layers_2_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[37] gpt_neox_layers_2_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[38] gpt_neox_layers_2_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[39] gpt_neox_layers_2_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[40] gpt_neox_layers_2_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[41] gpt_neox_layers_2_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[42] gpt_neox_layers_2_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[43] gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[44] gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[45] gpt_neox_layers_2_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[46] gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[47] gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[48] gpt_neox_layers_2_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[49] gpt_neox_layers_3_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[50] gpt_neox_layers_3_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[51] gpt_neox_layers_3_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[52] gpt_neox_layers_3_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[53] gpt_neox_layers_3_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[54] gpt_neox_layers_3_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[55] gpt_neox_layers_3_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[56] gpt_neox_layers_3_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[57] gpt_neox_layers_3_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[58] gpt_neox_layers_3_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[59] gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[60] gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[61] gpt_neox_layers_3_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[62] gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[63] gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[64] gpt_neox_layers_3_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[65] gpt_neox_layers_4_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[66] gpt_neox_layers_4_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[67] gpt_neox_layers_4_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[68] gpt_neox_layers_4_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[69] gpt_neox_layers_4_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[70] gpt_neox_layers_4_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[71] gpt_neox_layers_4_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[72] gpt_neox_layers_4_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[73] gpt_neox_layers_4_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[74] gpt_neox_layers_4_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[75] gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[76] gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[77] gpt_neox_layers_4_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[78] gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[79] gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[80] gpt_neox_layers_4_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[81] gpt_neox_layers_5_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[82] gpt_neox_layers_5_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[83] gpt_neox_layers_5_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[84] gpt_neox_layers_5_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[85] gpt_neox_layers_5_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[86] gpt_neox_layers_5_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[87] gpt_neox_layers_5_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[88] gpt_neox_layers_5_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[89] gpt_neox_layers_5_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[90] gpt_neox_layers_5_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[91] gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[92] gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[93] gpt_neox_layers_5_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[94] gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[95] gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[96] gpt_neox_layers_5_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[97] gpt_neox_layers_6_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[98] gpt_neox_layers_6_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[99] gpt_neox_layers_6_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[100] gpt_neox_layers_6_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[101] gpt_neox_layers_6_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[102] gpt_neox_layers_6_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[103] gpt_neox_layers_6_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[104] gpt_neox_layers_6_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[105] gpt_neox_layers_6_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[106] gpt_neox_layers_6_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[107] gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[108] gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[109] gpt_neox_layers_6_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[110] gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[111] gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[112] gpt_neox_layers_6_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[113] gpt_neox_layers_7_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[114] gpt_neox_layers_7_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[115] gpt_neox_layers_7_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[116] gpt_neox_layers_7_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[117] gpt_neox_layers_7_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[118] gpt_neox_layers_7_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[119] gpt_neox_layers_7_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[120] gpt_neox_layers_7_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[121] gpt_neox_layers_7_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[122] gpt_neox_layers_7_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[123] gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[124] gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[125] gpt_neox_layers_7_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[126] gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[127] gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[128] gpt_neox_layers_7_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[129] gpt_neox_layers_8_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[130] gpt_neox_layers_8_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[131] gpt_neox_layers_8_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[132] gpt_neox_layers_8_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[133] gpt_neox_layers_8_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[134] gpt_neox_layers_8_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[135] gpt_neox_layers_8_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[136] gpt_neox_layers_8_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[137] gpt_neox_layers_8_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[138] gpt_neox_layers_8_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[139] gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[140] gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[141] gpt_neox_layers_8_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[142] gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[143] gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[144] gpt_neox_layers_8_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[145] gpt_neox_layers_9_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[146] gpt_neox_layers_9_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[147] gpt_neox_layers_9_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[148] gpt_neox_layers_9_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[149] gpt_neox_layers_9_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[150] gpt_neox_layers_9_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[151] gpt_neox_layers_9_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[152] gpt_neox_layers_9_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[153] gpt_neox_layers_9_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[154] gpt_neox_layers_9_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[155] gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[156] gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[157] gpt_neox_layers_9_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[158] gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[159] gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[160] gpt_neox_layers_9_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[161] gpt_neox_layers_10_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[162] gpt_neox_layers_10_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[163] gpt_neox_layers_10_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[164] gpt_neox_layers_10_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[165] gpt_neox_layers_10_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[166] gpt_neox_layers_10_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[167] gpt_neox_layers_10_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[168] gpt_neox_layers_10_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[169] gpt_neox_layers_10_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[170] gpt_neox_layers_10_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[171] gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[172] gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[173] gpt_neox_layers_10_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[174] gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[175] gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[176] gpt_neox_layers_10_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[177] gpt_neox_layers_11_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[178] gpt_neox_layers_11_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[179] gpt_neox_layers_11_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[180] gpt_neox_layers_11_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[181] gpt_neox_layers_11_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[182] gpt_neox_layers_11_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[183] gpt_neox_layers_11_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[184] gpt_neox_layers_11_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[185] gpt_neox_layers_11_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[186] gpt_neox_layers_11_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[187] gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[188] gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[189] gpt_neox_layers_11_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[190] gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[191] gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[192] gpt_neox_layers_11_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[193] gpt_neox_layers_12_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[194] gpt_neox_layers_12_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[195] gpt_neox_layers_12_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[196] gpt_neox_layers_12_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[197] gpt_neox_layers_12_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[198] gpt_neox_layers_12_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[199] gpt_neox_layers_12_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[200] gpt_neox_layers_12_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[201] gpt_neox_layers_12_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[202] gpt_neox_layers_12_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[203] gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[204] gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[205] gpt_neox_layers_12_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[206] gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[207] gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[208] gpt_neox_layers_12_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[209] gpt_neox_layers_13_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[210] gpt_neox_layers_13_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[211] gpt_neox_layers_13_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[212] gpt_neox_layers_13_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[213] gpt_neox_layers_13_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[214] gpt_neox_layers_13_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[215] gpt_neox_layers_13_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[216] gpt_neox_layers_13_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[217] gpt_neox_layers_13_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[218] gpt_neox_layers_13_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[219] gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[220] gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[221] gpt_neox_layers_13_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[222] gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[223] gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[224] gpt_neox_layers_13_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[225] gpt_neox_layers_14_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[226] gpt_neox_layers_14_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[227] gpt_neox_layers_14_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[228] gpt_neox_layers_14_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[229] gpt_neox_layers_14_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[230] gpt_neox_layers_14_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[231] gpt_neox_layers_14_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[232] gpt_neox_layers_14_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[233] gpt_neox_layers_14_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[234] gpt_neox_layers_14_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[235] gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[236] gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[237] gpt_neox_layers_14_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[238] gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[239] gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[240] gpt_neox_layers_14_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[241] gpt_neox_layers_15_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[242] gpt_neox_layers_15_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[243] gpt_neox_layers_15_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[244] gpt_neox_layers_15_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[245] gpt_neox_layers_15_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[246] gpt_neox_layers_15_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[247] gpt_neox_layers_15_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[248] gpt_neox_layers_15_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[249] gpt_neox_layers_15_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[250] gpt_neox_layers_15_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[251] gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[252] gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[253] gpt_neox_layers_15_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[254] gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[255] gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[256] gpt_neox_layers_15_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[257] gpt_neox_layers_16_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[258] gpt_neox_layers_16_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[259] gpt_neox_layers_16_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[260] gpt_neox_layers_16_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[261] gpt_neox_layers_16_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[262] gpt_neox_layers_16_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[263] gpt_neox_layers_16_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[264] gpt_neox_layers_16_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[265] gpt_neox_layers_16_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[266] gpt_neox_layers_16_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[267] gpt_neox_layers_16_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[268] gpt_neox_layers_16_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[269] gpt_neox_layers_16_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[270] gpt_neox_layers_16_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[271] gpt_neox_layers_16_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[272] gpt_neox_layers_16_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[273] gpt_neox_layers_17_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[274] gpt_neox_layers_17_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[275] gpt_neox_layers_17_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[276] gpt_neox_layers_17_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[277] gpt_neox_layers_17_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[278] gpt_neox_layers_17_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[279] gpt_neox_layers_17_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[280] gpt_neox_layers_17_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[281] gpt_neox_layers_17_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[282] gpt_neox_layers_17_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[283] gpt_neox_layers_17_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[284] gpt_neox_layers_17_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[285] gpt_neox_layers_17_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[286] gpt_neox_layers_17_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[287] gpt_neox_layers_17_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[288] gpt_neox_layers_17_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[289] gpt_neox_layers_18_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[290] gpt_neox_layers_18_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[291] gpt_neox_layers_18_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[292] gpt_neox_layers_18_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[293] gpt_neox_layers_18_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[294] gpt_neox_layers_18_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[295] gpt_neox_layers_18_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[296] gpt_neox_layers_18_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[297] gpt_neox_layers_18_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[298] gpt_neox_layers_18_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[299] gpt_neox_layers_18_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[300] gpt_neox_layers_18_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[301] gpt_neox_layers_18_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[302] gpt_neox_layers_18_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[303] gpt_neox_layers_18_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[304] gpt_neox_layers_18_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[305] gpt_neox_layers_19_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[306] gpt_neox_layers_19_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[307] gpt_neox_layers_19_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[308] gpt_neox_layers_19_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[309] gpt_neox_layers_19_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[310] gpt_neox_layers_19_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[311] gpt_neox_layers_19_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[312] gpt_neox_layers_19_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[313] gpt_neox_layers_19_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[314] gpt_neox_layers_19_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[315] gpt_neox_layers_19_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[316] gpt_neox_layers_19_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[317] gpt_neox_layers_19_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[318] gpt_neox_layers_19_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[319] gpt_neox_layers_19_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[320] gpt_neox_layers_19_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[321] gpt_neox_layers_20_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[322] gpt_neox_layers_20_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[323] gpt_neox_layers_20_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[324] gpt_neox_layers_20_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[325] gpt_neox_layers_20_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[326] gpt_neox_layers_20_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[327] gpt_neox_layers_20_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[328] gpt_neox_layers_20_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[329] gpt_neox_layers_20_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[330] gpt_neox_layers_20_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[331] gpt_neox_layers_20_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[332] gpt_neox_layers_20_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[333] gpt_neox_layers_20_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[334] gpt_neox_layers_20_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[335] gpt_neox_layers_20_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[336] gpt_neox_layers_20_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[337] gpt_neox_layers_21_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[338] gpt_neox_layers_21_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[339] gpt_neox_layers_21_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[340] gpt_neox_layers_21_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[341] gpt_neox_layers_21_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[342] gpt_neox_layers_21_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[343] gpt_neox_layers_21_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[344] gpt_neox_layers_21_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[345] gpt_neox_layers_21_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[346] gpt_neox_layers_21_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[347] gpt_neox_layers_21_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[348] gpt_neox_layers_21_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[349] gpt_neox_layers_21_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[350] gpt_neox_layers_21_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[351] gpt_neox_layers_21_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[352] gpt_neox_layers_21_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[353] gpt_neox_layers_22_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[354] gpt_neox_layers_22_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[355] gpt_neox_layers_22_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[356] gpt_neox_layers_22_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[357] gpt_neox_layers_22_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[358] gpt_neox_layers_22_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[359] gpt_neox_layers_22_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[360] gpt_neox_layers_22_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[361] gpt_neox_layers_22_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[362] gpt_neox_layers_22_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[363] gpt_neox_layers_22_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[364] gpt_neox_layers_22_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[365] gpt_neox_layers_22_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[366] gpt_neox_layers_22_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[367] gpt_neox_layers_22_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[368] gpt_neox_layers_22_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[369] gpt_neox_layers_23_input_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[370] gpt_neox_layers_23_input_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[371] gpt_neox_layers_23_post_attention_layernorm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[372] gpt_neox_layers_23_post_attention_layernorm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[373] gpt_neox_layers_23_attention_query_key_value_q_weight1: R.Tensor((3072, 128), dtype="uint32") = packed_params[374] gpt_neox_layers_23_attention_query_key_value_q_scale1: R.Tensor((3072, 32), dtype="float16") = packed_params[375] gpt_neox_layers_23_attention_query_key_value_bias1: R.Tensor((3072,), dtype="float16") = packed_params[376] gpt_neox_layers_23_attention_dense_q_weight1: R.Tensor((1024, 128), dtype="uint32") = packed_params[377] gpt_neox_layers_23_attention_dense_q_scale1: R.Tensor((1024, 32), dtype="float16") = packed_params[378] gpt_neox_layers_23_attention_dense_bias1: R.Tensor((1024,), dtype="float16") = packed_params[379] gpt_neox_layers_23_mlp_dense_h_to_4h_q_weight1: R.Tensor((4096, 128), dtype="uint32") = packed_params[380] gpt_neox_layers_23_mlp_dense_h_to_4h_q_scale1: R.Tensor((4096, 32), dtype="float16") = packed_params[381] gpt_neox_layers_23_mlp_dense_h_to_4h_bias1: R.Tensor((4096,), dtype="float32") = packed_params[382] gpt_neox_layers_23_mlp_dense_4h_to_h_q_weight1: R.Tensor((1024, 512), dtype="uint32") = packed_params[383] gpt_neox_layers_23_mlp_dense_4h_to_h_q_scale1: R.Tensor((1024, 128), dtype="float16") = packed_params[384] gpt_neox_layers_23_mlp_dense_4h_to_h_bias1: R.Tensor((1024,), dtype="float32") = packed_params[385] gpt_neox_final_layer_norm_weight1: R.Tensor((1024,), dtype="float16") = packed_params[386] gpt_neox_final_layer_norm_bias1: R.Tensor((1024,), dtype="float16") = packed_params[387] embed_out_q_weight1: R.Tensor((vocab_size, 128), dtype="uint32") = packed_params[388] embed_out_q_scale1: R.Tensor((vocab_size, 32), dtype="float16") = packed_params[389] layer_norm = R.call_tir(cls.layer_norm1, (input_embed, gpt_neox_layers_0_input_layernorm_weight1, gpt_neox_layers_0_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv288 = R.call_tir(cls.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, layer_norm, gpt_neox_layers_0_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape = R.call_tir(cls.reshape4, (lv288,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape1 = R.call_tir(cls.reshape5, (reshape,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv2 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape2 = R.call_tir(cls.reshape6, (lv2,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape3 = R.call_tir(cls.reshape7, (reshape2,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm1 = R.call_tir(cls.layer_norm1, (input_embed, gpt_neox_layers_0_post_attention_layernorm_weight1, gpt_neox_layers_0_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv289 = R.call_tir(cls.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, layer_norm1, gpt_neox_layers_0_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv290 = R.call_tir(cls.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, lv289, gpt_neox_layers_0_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv96 = R.call_tir(cls.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, lv290, input_embed), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm2 = R.call_tir(cls.layer_norm1, (lv96, gpt_neox_layers_1_input_layernorm_weight1, gpt_neox_layers_1_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv291 = R.call_tir(cls.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, layer_norm2, gpt_neox_layers_1_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape4 = R.call_tir(cls.reshape4, (lv291,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape5 = R.call_tir(cls.reshape5, (reshape4,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv7 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape6 = R.call_tir(cls.reshape6, (lv7,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape7 = R.call_tir(cls.reshape7, (reshape6,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm3 = R.call_tir(cls.layer_norm1, (lv96, gpt_neox_layers_1_post_attention_layernorm_weight1, gpt_neox_layers_1_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv292 = R.call_tir(cls.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, layer_norm3, gpt_neox_layers_1_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv293 = R.call_tir(cls.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, lv292, gpt_neox_layers_1_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv97 = R.call_tir(cls.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, lv293, lv96), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm4 = R.call_tir(cls.layer_norm1, (lv97, gpt_neox_layers_2_input_layernorm_weight1, gpt_neox_layers_2_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv294 = R.call_tir(cls.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, layer_norm4, gpt_neox_layers_2_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape8 = R.call_tir(cls.reshape4, (lv294,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape9 = R.call_tir(cls.reshape5, (reshape8,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv12 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape10 = R.call_tir(cls.reshape6, (lv12,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape11 = R.call_tir(cls.reshape7, (reshape10,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm5 = R.call_tir(cls.layer_norm1, (lv97, gpt_neox_layers_2_post_attention_layernorm_weight1, gpt_neox_layers_2_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv295 = R.call_tir(cls.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, layer_norm5, gpt_neox_layers_2_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv296 = R.call_tir(cls.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, lv295, gpt_neox_layers_2_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv98 = R.call_tir(cls.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, lv296, lv97), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm6 = R.call_tir(cls.layer_norm1, (lv98, gpt_neox_layers_3_input_layernorm_weight1, gpt_neox_layers_3_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv297 = R.call_tir(cls.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, layer_norm6, gpt_neox_layers_3_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape12 = R.call_tir(cls.reshape4, (lv297,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape13 = R.call_tir(cls.reshape5, (reshape12,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv17 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape14 = R.call_tir(cls.reshape6, (lv17,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape15 = R.call_tir(cls.reshape7, (reshape14,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm7 = R.call_tir(cls.layer_norm1, (lv98, gpt_neox_layers_3_post_attention_layernorm_weight1, gpt_neox_layers_3_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv298 = R.call_tir(cls.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, layer_norm7, gpt_neox_layers_3_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv299 = R.call_tir(cls.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, lv298, gpt_neox_layers_3_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv99 = R.call_tir(cls.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, lv299, lv98), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm8 = R.call_tir(cls.layer_norm1, (lv99, gpt_neox_layers_4_input_layernorm_weight1, gpt_neox_layers_4_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv300 = R.call_tir(cls.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, layer_norm8, gpt_neox_layers_4_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape16 = R.call_tir(cls.reshape4, (lv300,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape17 = R.call_tir(cls.reshape5, (reshape16,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv22 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape18 = R.call_tir(cls.reshape6, (lv22,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape19 = R.call_tir(cls.reshape7, (reshape18,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm9 = R.call_tir(cls.layer_norm1, (lv99, gpt_neox_layers_4_post_attention_layernorm_weight1, gpt_neox_layers_4_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv301 = R.call_tir(cls.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, layer_norm9, gpt_neox_layers_4_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv302 = R.call_tir(cls.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, lv301, gpt_neox_layers_4_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv100 = R.call_tir(cls.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, lv302, lv99), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm10 = R.call_tir(cls.layer_norm1, (lv100, gpt_neox_layers_5_input_layernorm_weight1, gpt_neox_layers_5_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv303 = R.call_tir(cls.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, layer_norm10, gpt_neox_layers_5_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape20 = R.call_tir(cls.reshape4, (lv303,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape21 = R.call_tir(cls.reshape5, (reshape20,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv27 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape22 = R.call_tir(cls.reshape6, (lv27,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape23 = R.call_tir(cls.reshape7, (reshape22,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm11 = R.call_tir(cls.layer_norm1, (lv100, gpt_neox_layers_5_post_attention_layernorm_weight1, gpt_neox_layers_5_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv304 = R.call_tir(cls.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, layer_norm11, gpt_neox_layers_5_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv305 = R.call_tir(cls.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, lv304, gpt_neox_layers_5_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv101 = R.call_tir(cls.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, lv305, lv100), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm12 = R.call_tir(cls.layer_norm1, (lv101, gpt_neox_layers_6_input_layernorm_weight1, gpt_neox_layers_6_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv306 = R.call_tir(cls.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, layer_norm12, gpt_neox_layers_6_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape24 = R.call_tir(cls.reshape4, (lv306,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape25 = R.call_tir(cls.reshape5, (reshape24,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv32 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape26 = R.call_tir(cls.reshape6, (lv32,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape27 = R.call_tir(cls.reshape7, (reshape26,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm13 = R.call_tir(cls.layer_norm1, (lv101, gpt_neox_layers_6_post_attention_layernorm_weight1, gpt_neox_layers_6_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv307 = R.call_tir(cls.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, layer_norm13, gpt_neox_layers_6_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv308 = R.call_tir(cls.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, lv307, gpt_neox_layers_6_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv102 = R.call_tir(cls.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, lv308, lv101), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm14 = R.call_tir(cls.layer_norm1, (lv102, gpt_neox_layers_7_input_layernorm_weight1, gpt_neox_layers_7_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv309 = R.call_tir(cls.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, layer_norm14, gpt_neox_layers_7_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape28 = R.call_tir(cls.reshape4, (lv309,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape29 = R.call_tir(cls.reshape5, (reshape28,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv37 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape30 = R.call_tir(cls.reshape6, (lv37,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape31 = R.call_tir(cls.reshape7, (reshape30,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm15 = R.call_tir(cls.layer_norm1, (lv102, gpt_neox_layers_7_post_attention_layernorm_weight1, gpt_neox_layers_7_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv310 = R.call_tir(cls.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, layer_norm15, gpt_neox_layers_7_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv311 = R.call_tir(cls.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, lv310, gpt_neox_layers_7_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv103 = R.call_tir(cls.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, lv311, lv102), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm16 = R.call_tir(cls.layer_norm1, (lv103, gpt_neox_layers_8_input_layernorm_weight1, gpt_neox_layers_8_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv312 = R.call_tir(cls.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, layer_norm16, gpt_neox_layers_8_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape32 = R.call_tir(cls.reshape4, (lv312,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape33 = R.call_tir(cls.reshape5, (reshape32,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv42 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape34 = R.call_tir(cls.reshape6, (lv42,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape35 = R.call_tir(cls.reshape7, (reshape34,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm17 = R.call_tir(cls.layer_norm1, (lv103, gpt_neox_layers_8_post_attention_layernorm_weight1, gpt_neox_layers_8_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv313 = R.call_tir(cls.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, layer_norm17, gpt_neox_layers_8_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv314 = R.call_tir(cls.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, lv313, gpt_neox_layers_8_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv104 = R.call_tir(cls.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, lv314, lv103), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm18 = R.call_tir(cls.layer_norm1, (lv104, gpt_neox_layers_9_input_layernorm_weight1, gpt_neox_layers_9_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv315 = R.call_tir(cls.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, layer_norm18, gpt_neox_layers_9_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape36 = R.call_tir(cls.reshape4, (lv315,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape37 = R.call_tir(cls.reshape5, (reshape36,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv47 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape38 = R.call_tir(cls.reshape6, (lv47,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape39 = R.call_tir(cls.reshape7, (reshape38,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm19 = R.call_tir(cls.layer_norm1, (lv104, gpt_neox_layers_9_post_attention_layernorm_weight1, gpt_neox_layers_9_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv316 = R.call_tir(cls.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, layer_norm19, gpt_neox_layers_9_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv317 = R.call_tir(cls.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, lv316, gpt_neox_layers_9_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv105 = R.call_tir(cls.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, lv317, lv104), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm20 = R.call_tir(cls.layer_norm1, (lv105, gpt_neox_layers_10_input_layernorm_weight1, gpt_neox_layers_10_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv318 = R.call_tir(cls.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, layer_norm20, gpt_neox_layers_10_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape40 = R.call_tir(cls.reshape4, (lv318,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape41 = R.call_tir(cls.reshape5, (reshape40,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv52 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape42 = R.call_tir(cls.reshape6, (lv52,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape43 = R.call_tir(cls.reshape7, (reshape42,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm21 = R.call_tir(cls.layer_norm1, (lv105, gpt_neox_layers_10_post_attention_layernorm_weight1, gpt_neox_layers_10_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv319 = R.call_tir(cls.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, layer_norm21, gpt_neox_layers_10_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv320 = R.call_tir(cls.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, lv319, gpt_neox_layers_10_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv106 = R.call_tir(cls.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, lv320, lv105), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm22 = R.call_tir(cls.layer_norm1, (lv106, gpt_neox_layers_11_input_layernorm_weight1, gpt_neox_layers_11_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv321 = R.call_tir(cls.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, layer_norm22, gpt_neox_layers_11_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape44 = R.call_tir(cls.reshape4, (lv321,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape45 = R.call_tir(cls.reshape5, (reshape44,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv57 = R.call_dps_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), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape46 = R.call_tir(cls.reshape6, (lv57,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape47 = R.call_tir(cls.reshape7, (reshape46,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm23 = R.call_tir(cls.layer_norm1, (lv106, gpt_neox_layers_11_post_attention_layernorm_weight1, gpt_neox_layers_11_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv322 = R.call_tir(cls.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, layer_norm23, gpt_neox_layers_11_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv323 = R.call_tir(cls.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, lv322, gpt_neox_layers_11_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv107 = R.call_tir(cls.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, lv323, lv106), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm24 = R.call_tir(cls.layer_norm1, (lv107, gpt_neox_layers_12_input_layernorm_weight1, gpt_neox_layers_12_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv324 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_12_attention_query_key_value_q_weight1, gpt_neox_layers_12_attention_query_key_value_q_scale1, layer_norm24, gpt_neox_layers_12_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape48 = R.call_tir(cls.reshape4, (lv324,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape49 = R.call_tir(cls.reshape5, (reshape48,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv62 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape49), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape50 = R.call_tir(cls.reshape6, (lv62,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape51 = R.call_tir(cls.reshape7, (reshape50,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm25 = R.call_tir(cls.layer_norm1, (lv107, gpt_neox_layers_12_post_attention_layernorm_weight1, gpt_neox_layers_12_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv325 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale1, layer_norm25, gpt_neox_layers_12_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv326 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale1, lv325, gpt_neox_layers_12_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv108 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_12_attention_dense_q_weight1, gpt_neox_layers_12_attention_dense_q_scale1, reshape51, gpt_neox_layers_12_attention_dense_bias1, lv326, lv107), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm26 = R.call_tir(cls.layer_norm1, (lv108, gpt_neox_layers_13_input_layernorm_weight1, gpt_neox_layers_13_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv327 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_13_attention_query_key_value_q_weight1, gpt_neox_layers_13_attention_query_key_value_q_scale1, layer_norm26, gpt_neox_layers_13_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape52 = R.call_tir(cls.reshape4, (lv327,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape53 = R.call_tir(cls.reshape5, (reshape52,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv67 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape53), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape54 = R.call_tir(cls.reshape6, (lv67,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape55 = R.call_tir(cls.reshape7, (reshape54,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm27 = R.call_tir(cls.layer_norm1, (lv108, gpt_neox_layers_13_post_attention_layernorm_weight1, gpt_neox_layers_13_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv328 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale1, layer_norm27, gpt_neox_layers_13_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv329 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale1, lv328, gpt_neox_layers_13_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv109 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_13_attention_dense_q_weight1, gpt_neox_layers_13_attention_dense_q_scale1, reshape55, gpt_neox_layers_13_attention_dense_bias1, lv329, lv108), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm28 = R.call_tir(cls.layer_norm1, (lv109, gpt_neox_layers_14_input_layernorm_weight1, gpt_neox_layers_14_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv330 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_14_attention_query_key_value_q_weight1, gpt_neox_layers_14_attention_query_key_value_q_scale1, layer_norm28, gpt_neox_layers_14_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape56 = R.call_tir(cls.reshape4, (lv330,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape57 = R.call_tir(cls.reshape5, (reshape56,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv72 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape57), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape58 = R.call_tir(cls.reshape6, (lv72,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape59 = R.call_tir(cls.reshape7, (reshape58,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm29 = R.call_tir(cls.layer_norm1, (lv109, gpt_neox_layers_14_post_attention_layernorm_weight1, gpt_neox_layers_14_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv331 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale1, layer_norm29, gpt_neox_layers_14_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv332 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale1, lv331, gpt_neox_layers_14_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv110 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_14_attention_dense_q_weight1, gpt_neox_layers_14_attention_dense_q_scale1, reshape59, gpt_neox_layers_14_attention_dense_bias1, lv332, lv109), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm30 = R.call_tir(cls.layer_norm1, (lv110, gpt_neox_layers_15_input_layernorm_weight1, gpt_neox_layers_15_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv333 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_15_attention_query_key_value_q_weight1, gpt_neox_layers_15_attention_query_key_value_q_scale1, layer_norm30, gpt_neox_layers_15_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape60 = R.call_tir(cls.reshape4, (lv333,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape61 = R.call_tir(cls.reshape5, (reshape60,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv77 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape61), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape62 = R.call_tir(cls.reshape6, (lv77,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape63 = R.call_tir(cls.reshape7, (reshape62,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm31 = R.call_tir(cls.layer_norm1, (lv110, gpt_neox_layers_15_post_attention_layernorm_weight1, gpt_neox_layers_15_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv334 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale1, layer_norm31, gpt_neox_layers_15_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv335 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale1, lv334, gpt_neox_layers_15_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv111 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_15_attention_dense_q_weight1, gpt_neox_layers_15_attention_dense_q_scale1, reshape63, gpt_neox_layers_15_attention_dense_bias1, lv335, lv110), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm32 = R.call_tir(cls.layer_norm1, (lv111, gpt_neox_layers_16_input_layernorm_weight1, gpt_neox_layers_16_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv336 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_16_attention_query_key_value_q_weight1, gpt_neox_layers_16_attention_query_key_value_q_scale1, layer_norm32, gpt_neox_layers_16_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape64 = R.call_tir(cls.reshape4, (lv336,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape65 = R.call_tir(cls.reshape5, (reshape64,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv82 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), reshape65), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape66 = R.call_tir(cls.reshape6, (lv82,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape67 = R.call_tir(cls.reshape7, (reshape66,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm33 = R.call_tir(cls.layer_norm1, (lv111, gpt_neox_layers_16_post_attention_layernorm_weight1, gpt_neox_layers_16_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv337 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_16_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_16_mlp_dense_h_to_4h_q_scale1, layer_norm33, gpt_neox_layers_16_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv338 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_16_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_16_mlp_dense_4h_to_h_q_scale1, lv337, gpt_neox_layers_16_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv112 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_16_attention_dense_q_weight1, gpt_neox_layers_16_attention_dense_q_scale1, reshape67, gpt_neox_layers_16_attention_dense_bias1, lv338, lv111), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm34 = R.call_tir(cls.layer_norm1, (lv112, gpt_neox_layers_17_input_layernorm_weight1, gpt_neox_layers_17_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv339 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_17_attention_query_key_value_q_weight1, gpt_neox_layers_17_attention_query_key_value_q_scale1, layer_norm34, gpt_neox_layers_17_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape68 = R.call_tir(cls.reshape4, (lv339,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape69 = R.call_tir(cls.reshape5, (reshape68,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv87 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), reshape69), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape70 = R.call_tir(cls.reshape6, (lv87,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape71 = R.call_tir(cls.reshape7, (reshape70,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm35 = R.call_tir(cls.layer_norm1, (lv112, gpt_neox_layers_17_post_attention_layernorm_weight1, gpt_neox_layers_17_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv340 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_17_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_17_mlp_dense_h_to_4h_q_scale1, layer_norm35, gpt_neox_layers_17_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv341 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_17_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_17_mlp_dense_4h_to_h_q_scale1, lv340, gpt_neox_layers_17_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv113 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_17_attention_dense_q_weight1, gpt_neox_layers_17_attention_dense_q_scale1, reshape71, gpt_neox_layers_17_attention_dense_bias1, lv341, lv112), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm36 = R.call_tir(cls.layer_norm1, (lv113, gpt_neox_layers_18_input_layernorm_weight1, gpt_neox_layers_18_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv342 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_18_attention_query_key_value_q_weight1, gpt_neox_layers_18_attention_query_key_value_q_scale1, layer_norm36, gpt_neox_layers_18_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape72 = R.call_tir(cls.reshape4, (lv342,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape73 = R.call_tir(cls.reshape5, (reshape72,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv92 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), reshape73), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape74 = R.call_tir(cls.reshape6, (lv92,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape75 = R.call_tir(cls.reshape7, (reshape74,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm37 = R.call_tir(cls.layer_norm1, (lv113, gpt_neox_layers_18_post_attention_layernorm_weight1, gpt_neox_layers_18_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv343 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_18_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_18_mlp_dense_h_to_4h_q_scale1, layer_norm37, gpt_neox_layers_18_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv344 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_18_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_18_mlp_dense_4h_to_h_q_scale1, lv343, gpt_neox_layers_18_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv114 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_18_attention_dense_q_weight1, gpt_neox_layers_18_attention_dense_q_scale1, reshape75, gpt_neox_layers_18_attention_dense_bias1, lv344, lv113), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm38 = R.call_tir(cls.layer_norm1, (lv114, gpt_neox_layers_19_input_layernorm_weight1, gpt_neox_layers_19_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv345 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_19_attention_query_key_value_q_weight1, gpt_neox_layers_19_attention_query_key_value_q_scale1, layer_norm38, gpt_neox_layers_19_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape76 = R.call_tir(cls.reshape4, (lv345,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape77 = R.call_tir(cls.reshape5, (reshape76,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv97_1 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), reshape77), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape78 = R.call_tir(cls.reshape6, (lv97_1,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape79 = R.call_tir(cls.reshape7, (reshape78,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm39 = R.call_tir(cls.layer_norm1, (lv114, gpt_neox_layers_19_post_attention_layernorm_weight1, gpt_neox_layers_19_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv346 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_19_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_19_mlp_dense_h_to_4h_q_scale1, layer_norm39, gpt_neox_layers_19_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv347 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_19_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_19_mlp_dense_4h_to_h_q_scale1, lv346, gpt_neox_layers_19_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv115 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_19_attention_dense_q_weight1, gpt_neox_layers_19_attention_dense_q_scale1, reshape79, gpt_neox_layers_19_attention_dense_bias1, lv347, lv114), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm40 = R.call_tir(cls.layer_norm1, (lv115, gpt_neox_layers_20_input_layernorm_weight1, gpt_neox_layers_20_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv348 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_20_attention_query_key_value_q_weight1, gpt_neox_layers_20_attention_query_key_value_q_scale1, layer_norm40, gpt_neox_layers_20_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape80 = R.call_tir(cls.reshape4, (lv348,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape81 = R.call_tir(cls.reshape5, (reshape80,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv102_1 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), reshape81), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape82 = R.call_tir(cls.reshape6, (lv102_1,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape83 = R.call_tir(cls.reshape7, (reshape82,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm41 = R.call_tir(cls.layer_norm1, (lv115, gpt_neox_layers_20_post_attention_layernorm_weight1, gpt_neox_layers_20_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv349 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_20_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_20_mlp_dense_h_to_4h_q_scale1, layer_norm41, gpt_neox_layers_20_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv350 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_20_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_20_mlp_dense_4h_to_h_q_scale1, lv349, gpt_neox_layers_20_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv116 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_20_attention_dense_q_weight1, gpt_neox_layers_20_attention_dense_q_scale1, reshape83, gpt_neox_layers_20_attention_dense_bias1, lv350, lv115), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm42 = R.call_tir(cls.layer_norm1, (lv116, gpt_neox_layers_21_input_layernorm_weight1, gpt_neox_layers_21_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv351 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_21_attention_query_key_value_q_weight1, gpt_neox_layers_21_attention_query_key_value_q_scale1, layer_norm42, gpt_neox_layers_21_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape84 = R.call_tir(cls.reshape4, (lv351,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape85 = R.call_tir(cls.reshape5, (reshape84,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv107_1 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), reshape85), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape86 = R.call_tir(cls.reshape6, (lv107_1,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape87 = R.call_tir(cls.reshape7, (reshape86,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm43 = R.call_tir(cls.layer_norm1, (lv116, gpt_neox_layers_21_post_attention_layernorm_weight1, gpt_neox_layers_21_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv352 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_21_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_21_mlp_dense_h_to_4h_q_scale1, layer_norm43, gpt_neox_layers_21_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv353 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_21_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_21_mlp_dense_4h_to_h_q_scale1, lv352, gpt_neox_layers_21_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv117 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_21_attention_dense_q_weight1, gpt_neox_layers_21_attention_dense_q_scale1, reshape87, gpt_neox_layers_21_attention_dense_bias1, lv353, lv116), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm44 = R.call_tir(cls.layer_norm1, (lv117, gpt_neox_layers_22_input_layernorm_weight1, gpt_neox_layers_22_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv354 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_22_attention_query_key_value_q_weight1, gpt_neox_layers_22_attention_query_key_value_q_scale1, layer_norm44, gpt_neox_layers_22_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape88 = R.call_tir(cls.reshape4, (lv354,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape89 = R.call_tir(cls.reshape5, (reshape88,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv112_1 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), reshape89), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape90 = R.call_tir(cls.reshape6, (lv112_1,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape91 = R.call_tir(cls.reshape7, (reshape90,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm45 = R.call_tir(cls.layer_norm1, (lv117, gpt_neox_layers_22_post_attention_layernorm_weight1, gpt_neox_layers_22_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv355 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_22_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_22_mlp_dense_h_to_4h_q_scale1, layer_norm45, gpt_neox_layers_22_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv356 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_22_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_22_mlp_dense_4h_to_h_q_scale1, lv355, gpt_neox_layers_22_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv118 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_22_attention_dense_q_weight1, gpt_neox_layers_22_attention_dense_q_scale1, reshape91, gpt_neox_layers_22_attention_dense_bias1, lv356, lv117), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm46 = R.call_tir(cls.layer_norm1, (lv118, gpt_neox_layers_23_input_layernorm_weight1, gpt_neox_layers_23_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv357 = R.call_tir(cls.fused_dequantize1_fused_NT_matmul5_add5, (gpt_neox_layers_23_attention_query_key_value_q_weight1, gpt_neox_layers_23_attention_query_key_value_q_scale1, layer_norm46, gpt_neox_layers_23_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 3072), dtype="float16")) reshape92 = R.call_tir(cls.reshape4, (lv357,), out_sinfo=R.Tensor((1, seq_len, 48, 64), dtype="float16")) reshape93 = R.call_tir(cls.reshape5, (reshape92,), out_sinfo=R.Tensor((seq_len, 48, 64), dtype="float16")) lv117_1 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), reshape93), out_sinfo=R.Tensor((seq_len, 16, 64), dtype="float16")) reshape94 = R.call_tir(cls.reshape6, (lv117_1,), out_sinfo=R.Tensor((1, seq_len, 16, 64), dtype="float16")) reshape95 = R.call_tir(cls.reshape7, (reshape94,), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm47 = R.call_tir(cls.layer_norm1, (lv118, gpt_neox_layers_23_post_attention_layernorm_weight1, gpt_neox_layers_23_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv358 = R.call_tir(cls.fused_dequantize3_fused_NT_matmul7_add7_gelu1_cast3, (gpt_neox_layers_23_mlp_dense_h_to_4h_q_weight1, gpt_neox_layers_23_mlp_dense_h_to_4h_q_scale1, layer_norm47, gpt_neox_layers_23_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 4096), dtype="float16")) lv359 = R.call_tir(cls.fused_dequantize4_fused_NT_matmul8_add8_cast4, (gpt_neox_layers_23_mlp_dense_4h_to_h_q_weight1, gpt_neox_layers_23_mlp_dense_4h_to_h_q_scale1, lv358, gpt_neox_layers_23_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv119 = R.call_tir(cls.fused_dequantize2_fused_NT_matmul6_add6_add9_add9, (gpt_neox_layers_23_attention_dense_q_weight1, gpt_neox_layers_23_attention_dense_q_scale1, reshape95, gpt_neox_layers_23_attention_dense_bias1, lv359, lv118), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) layer_norm48 = R.call_tir(cls.layer_norm1, (lv119, gpt_neox_final_layer_norm_weight1, gpt_neox_final_layer_norm_bias1), out_sinfo=R.Tensor((1, seq_len, 1024), dtype="float16")) lv121 = R.call_tir(cls.index, (layer_norm48,), out_sinfo=R.Tensor((1, 1, 1024), dtype="float16")) lv4 = R.call_tir(cls.fused_dequantize_fused_NT_matmul14_cast8, (embed_out_q_weight1, embed_out_q_scale1, lv121), out_sinfo=R.Tensor((1, 1, vocab_size), dtype="float32")) gv1: R.Tuple(R.Tensor((1, 1, vocab_size), dtype="float32"), R.Object) = lv4, paged_kv_cache R.output(gv1) 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.memory_plan_dynamic_func_output": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) cls = Module with R.dataflow(): lv: R.Tensor((batch_size, vocab_size), dtype="float32") = R.call_pure_packed("vm.builtin.reshape", logits, R.shape([batch_size, vocab_size]), sinfo_args=(R.Tensor((batch_size, vocab_size), dtype="float32"),)) lv1 = R.call_tir(cls.chunk_lse, (lv, temperature), out_sinfo=[R.Tensor((batch_size, (vocab_size + 4096 - 1) // 4096), dtype="float32"), R.Tensor((batch_size, (vocab_size + 4096 - 1) // 4096), dtype="float32")]) lv2: R.Tensor((batch_size, (vocab_size + 4096 - 1) // 4096), dtype="float32") = lv1[0] lv3: R.Tensor((batch_size, (vocab_size + 4096 - 1) // 4096), dtype="float32") = lv1[1] lv4 = R.call_tir(cls.softmax_with_chunked_sum, (lv, temperature, lv2, lv3), out_sinfo=R.Tensor((batch_size, vocab_size), dtype="float32")) gv: R.Tensor((batch_size, 1, vocab_size), dtype="float32") = R.call_pure_packed("vm.builtin.reshape", lv4, R.shape([batch_size, 1, vocab_size]), sinfo_args=(R.Tensor((batch_size, 1, vocab_size), dtype="float32"),)) R.output(gv) return gv