diff --git "a/ir_dump/debug-phase4.py" "b/ir_dump/debug-phase4.py" new file mode 100644--- /dev/null +++ "b/ir_dump/debug-phase4.py" @@ -0,0 +1,6639 @@ +# 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, 8, 256), "float16") + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(pages_handle, (max_num_pages, 2, 8, 16, 256), "float16", offset_factor=1) + page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (B,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) + output = T.match_buffer(output_handle, (B, 8, 256), "float16") + lse = T.match_buffer(lse_handle, (B, 8)) + # with T.block("root"): + sm_scale: T.float32 = T.float32(0.090168440055560212) + for bx in T.thread_binding(B, thread="blockIdx.x"): + for fused_by_bz in T.thread_binding(8, thread="blockIdx.y"): + for ty in T.thread_binding(1, thread="threadIdx.y"): + for tx in T.thread_binding(64, thread="threadIdx.x"): + for tz in T.thread_binding(4, thread="threadIdx.z"): + with T.block("attn"): + T.reads(page_table_indptr[bx:bx + 2], length_info[bx], q_rope_position[bx], Q[bx, fused_by_bz // 8 + ty + fused_by_bz % 8, tx * 4 - 128:tx * 4 - 128 + 260]) + T.writes(output[bx, fused_by_bz % 8 + fused_by_bz // 8 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 8 + fused_by_bz // 8 + ty]) + Q_local = T.alloc_buffer((4,), "float16", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + K_smem = T.alloc_buffer((4, 256), "float16", scope="shared") + V_smem = T.alloc_buffer((4, 256), "float16", scope="shared") + O_allreduce = T.alloc_buffer((4, 1, 256), scope="shared") + md_allreduce = T.alloc_buffer((4, 1, 2), scope="shared") + S_reduce_local = T.alloc_buffer((1,), scope="local") + t0 = T.alloc_buffer((1,), scope="local") + S_local = T.alloc_buffer((1,), scope="local") + QK_local = T.alloc_buffer((4,), scope="local") + V_local = T.alloc_buffer((4,), "float16", scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_prev = T.alloc_buffer((1,), scope="local") + other_m = T.alloc_buffer((1,), scope="local") + other_d = T.alloc_buffer((1,), scope="local") + exp_mprev = T.alloc_buffer((1,), scope="local") + exp_otherm = T.alloc_buffer((1,), scope="local") + other_o = T.alloc_buffer((4,), scope="local") + st_m = T.alloc_buffer((1,), scope="local") + st_d = T.alloc_buffer((1,), scope="local") + O_local = T.alloc_buffer((4,), scope="local") + by: T.int32 = fused_by_bz % 8 + bz: T.int32 = fused_by_bz // 8 + batch_idx: T.int32 = bx + cur_page_indptr_begin: T.int32 = page_table_indptr[batch_idx] + cur_page_indptr_end: T.int32 = page_table_indptr[batch_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[batch_idx], 0) + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for vec in T.vectorized(4): + freq = T.float32() + Q_local[vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", Q[bx, by + bz + ty, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 128, Q[bx, by + bz + ty, tx * 4 + vec + 128] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 128]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 256) / T.float32(256.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) + for iterator in range((kv_chunk_len[0] + 3) // 4): + tile_start_s: T.int32 = tz + ty + tile_start_g: T.int32 = iterator * 4 + tz + ty + for j in range(1): + with T.block("KV_load"): + T.reads() + T.writes() + row_g: T.int32 = tile_start_g + j + if row_g < kv_chunk_len[0]: + seq_offset: T.int32 = row_g + page_no: T.int32 = page_table_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + for vec in T.vectorized(4): + freq = T.float32() + K_smem[tile_start_s + j, tx * 4 + vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 128, pages[page_no, 0, by, page_offset, tx * 4 + vec + 128] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 128]))), where={freq: T.Cast("float32", k_rope_pos_offset[batch_idx] + row_g) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 256) / T.float32(256.0))}), pages[page_no, 0, by, page_offset, tx * 4 + vec]) + V_smem[tile_start_s + j, tx * 4 + vec] = pages[page_no, 1, by, page_offset, tx * 4 + vec] + else: + for vec in T.vectorized(4): + K_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + V_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + T.tvm_storage_sync("shared") + m_prev[0] = st_m[0] + for j in range(1): + for vec in T.vectorized(4): + QK_local[vec] = T.Cast("float32", Q_local[vec]) * T.Cast("float32", K_smem[tz + j, tx * 4 + vec]) * attn_score_scaling_factor * sm_scale + S_reduce_local[0] = T.float32(0.0) + for vec in T.unroll(4): + S_reduce_local[0] = S_reduce_local[0] + QK_local[vec] + with T.block("block_cross_thread"): + T.reads(S_reduce_local[0]) + T.writes(t0[0]) + T.attr(T.comm_reducer(lambda x0, y0: x0 + y0, [T.float32(0.0)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))) + T.tvm_thread_allreduce(T.uint32(1), S_reduce_local[0], T.bool(True), t0[0], tx) + S_local[j] = T.float32(-50000.0) + if iterator * 4 + tz + j < kv_chunk_len[0]: + S_local[j] = t0[0] + st_m[0] = T.max(st_m[0], S_local[j]) + o_scale: T.float32 = T.exp2(m_prev[0] - st_m[0]) + st_d[0] = st_d[0] * o_scale + for j in range(1): + S_local[j] = T.exp2(S_local[j] - st_m[0]) + st_d[0] = st_d[0] + S_local[j] + for j in T.vectorized(4): + O_local[j] = O_local[j] * o_scale + for j in range(1): + for vec in T.vectorized(4): + V_local[vec] = V_smem[tz + j, tx * 4 + vec] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] + T.Cast("float32", V_local[vec]) * S_local[j] + for vec in T.vectorized(4): + O_allreduce[tz, ty, tx * 4 + vec] = O_local[vec] + md_allreduce[tz, ty, 0] = st_m[0] + md_allreduce[tz, ty, 1] = st_d[0] + T.tvm_storage_sync("shared") + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for j in range(4): + m_prev[0] = st_m[0] + d_prev[0] = st_d[0] + other_m[0] = md_allreduce[j, ty, 0] + other_d[0] = md_allreduce[j, ty, 1] + for vec in T.vectorized(4): + other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] + st_m[0] = T.max(st_m[0], other_m[0]) + st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) + exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) + exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] / st_d[0] + for vec in T.vectorized(4): + output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) + lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) + + @T.prim_func + def batch_decode_paged_kv_sliding_window(_0: T.int32, Q_handle: T.handle, pages_handle: T.handle, page_table_indptr_handle: T.handle, page_table_values_handle: T.handle, var_length_info: T.handle, k_rope_pos_offset_handle: T.handle, q_rope_position_handle: T.handle, output_handle: T.handle, lse_handle: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + B = T.int32(is_size_var=True) + Q = T.match_buffer(Q_handle, (B, 8, 256), "float16") + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(pages_handle, (max_num_pages, 2, 8, 16, 256), "float16", offset_factor=1) + page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (3, B), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) + output = T.match_buffer(output_handle, (B, 8, 256), "float16") + lse = T.match_buffer(lse_handle, (B, 8)) + # with T.block("root"): + sm_scale: T.float32 = T.float32(0.090168440055560212) + for bx in T.thread_binding(B, thread="blockIdx.x"): + for fused_by_bz in T.thread_binding(8, thread="blockIdx.y"): + for ty in T.thread_binding(1, thread="threadIdx.y"): + for tx in T.thread_binding(64, thread="threadIdx.x"): + for tz in T.thread_binding(4, thread="threadIdx.z"): + with T.block("attn"): + T.reads(page_table_indptr[bx:bx + 2], length_info[0:3, bx], q_rope_position[bx], Q[bx, fused_by_bz // 8 + ty + fused_by_bz % 8, tx * 4 - 128:tx * 4 - 128 + 260]) + T.writes(output[bx, fused_by_bz % 8 + fused_by_bz // 8 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 8 + fused_by_bz // 8 + ty]) + Q_local = T.alloc_buffer((4,), "float16", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + K_smem = T.alloc_buffer((4, 256), "float16", scope="shared") + V_smem = T.alloc_buffer((4, 256), "float16", scope="shared") + O_allreduce = T.alloc_buffer((4, 1, 256), scope="shared") + md_allreduce = T.alloc_buffer((4, 1, 2), scope="shared") + S_reduce_local = T.alloc_buffer((1,), scope="local") + t0 = T.alloc_buffer((1,), scope="local") + S_local = T.alloc_buffer((1,), scope="local") + QK_local = T.alloc_buffer((4,), scope="local") + V_local = T.alloc_buffer((4,), "float16", scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_prev = T.alloc_buffer((1,), scope="local") + other_m = T.alloc_buffer((1,), scope="local") + other_d = T.alloc_buffer((1,), scope="local") + exp_mprev = T.alloc_buffer((1,), scope="local") + exp_otherm = T.alloc_buffer((1,), scope="local") + other_o = T.alloc_buffer((4,), scope="local") + st_m = T.alloc_buffer((1,), scope="local") + st_d = T.alloc_buffer((1,), scope="local") + O_local = T.alloc_buffer((4,), scope="local") + by: T.int32 = fused_by_bz % 8 + bz: T.int32 = fused_by_bz // 8 + batch_idx: T.int32 = bx + cur_page_indptr_begin: T.int32 = page_table_indptr[batch_idx] + cur_page_indptr_end: T.int32 = page_table_indptr[batch_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[0, batch_idx] - length_info[1, batch_idx] + length_info[2, batch_idx], 0) + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for vec in T.vectorized(4): + freq = T.float32() + Q_local[vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", Q[bx, by + bz + ty, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 128, Q[bx, by + bz + ty, tx * 4 + vec + 128] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 128]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 256) / T.float32(256.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) + for iterator in range((kv_chunk_len[0] + 3) // 4): + tile_start_s: T.int32 = tz + ty + tile_start_g: T.int32 = iterator * 4 + tz + ty + for j in range(1): + with T.block("KV_load"): + T.reads() + T.writes() + row_g: T.int32 = tile_start_g + j + if row_g < kv_chunk_len[0]: + seq_offset: T.int32 = T.if_then_else(row_g < length_info[2, batch_idx], row_g, row_g - length_info[2, batch_idx] + length_info[1, batch_idx]) + page_no: T.int32 = page_table_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + for vec in T.vectorized(4): + freq = T.float32() + K_smem[tile_start_s + j, tx * 4 + vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 128, pages[page_no, 0, by, page_offset, tx * 4 + vec + 128] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 128]))), where={freq: T.Cast("float32", k_rope_pos_offset[batch_idx] + row_g) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 256) / T.float32(256.0))}), pages[page_no, 0, by, page_offset, tx * 4 + vec]) + V_smem[tile_start_s + j, tx * 4 + vec] = pages[page_no, 1, by, page_offset, tx * 4 + vec] + else: + for vec in T.vectorized(4): + K_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + V_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + T.tvm_storage_sync("shared") + m_prev[0] = st_m[0] + for j in range(1): + for vec in T.vectorized(4): + QK_local[vec] = T.Cast("float32", Q_local[vec]) * T.Cast("float32", K_smem[tz + j, tx * 4 + vec]) * attn_score_scaling_factor * sm_scale + S_reduce_local[0] = T.float32(0.0) + for vec in T.unroll(4): + S_reduce_local[0] = S_reduce_local[0] + QK_local[vec] + with T.block("block_cross_thread"): + T.reads(S_reduce_local[0]) + T.writes(t0[0]) + T.attr(T.comm_reducer(lambda x0, y0: x0 + y0, [T.float32(0.0)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))) + T.tvm_thread_allreduce(T.uint32(1), S_reduce_local[0], T.bool(True), t0[0], tx) + S_local[j] = T.float32(-50000.0) + if iterator * 4 + tz + j < kv_chunk_len[0]: + S_local[j] = t0[0] + st_m[0] = T.max(st_m[0], S_local[j]) + o_scale: T.float32 = T.exp2(m_prev[0] - st_m[0]) + st_d[0] = st_d[0] * o_scale + for j in range(1): + S_local[j] = T.exp2(S_local[j] - st_m[0]) + st_d[0] = st_d[0] + S_local[j] + for j in T.vectorized(4): + O_local[j] = O_local[j] * o_scale + for j in range(1): + for vec in T.vectorized(4): + V_local[vec] = V_smem[tz + j, tx * 4 + vec] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] + T.Cast("float32", V_local[vec]) * S_local[j] + for vec in T.vectorized(4): + O_allreduce[tz, ty, tx * 4 + vec] = O_local[vec] + md_allreduce[tz, ty, 0] = st_m[0] + md_allreduce[tz, ty, 1] = st_d[0] + T.tvm_storage_sync("shared") + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for j in range(4): + m_prev[0] = st_m[0] + d_prev[0] = st_d[0] + other_m[0] = md_allreduce[j, ty, 0] + other_d[0] = md_allreduce[j, ty, 1] + for vec in T.vectorized(4): + other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] + st_m[0] = T.max(st_m[0], other_m[0]) + st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) + exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) + exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] / st_d[0] + for vec in T.vectorized(4): + output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) + lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) + + @T.prim_func + def batch_prefill_paged_kv(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + total_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (total_len, 8, 256), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 8, 16, 256), "float16", offset_factor=1) + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 8, 256), "float16") + lse = T.match_buffer(var_lse, (total_len, 8)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(8, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + K_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + S_smem = T.alloc_buffer((16, 16), scope="shared") + S_local = T.alloc_buffer((16, 16), scope="local") + O_local = T.alloc_buffer((16, 256), scope="local") + m_smem = T.alloc_buffer((16,), scope="shared") + m_prev_smem = T.alloc_buffer((16,), scope="shared") + d_smem = T.alloc_buffer((16,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 16 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 8): + with T.block("O_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(8): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(16, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 256) + j = T.axis.spatial(256, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 256) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 15) // 16): + L_kv_start: T.int32 = iterator_1 * 16 + for lz_ly_fused_0 in range(8): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + freq = T.float32() + K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, pages[page_no, 0, by, page_offset, j + 128] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 128]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), pages[page_no, 0, by, page_offset, j]) + else: + K_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_ly_fused_0 in range(8): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:16, 0:256], K_smem[0:16, 0:256]) + T.writes(S_local[0:16, 0:16]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(1, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 + li_1_init) + j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(32, 1, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + k = T.axis.reduce(256, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.090168440055560212) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(1, 2): + with T.block("S_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(16): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) + T.writes(S_smem[row, 0:16]) + for j in range(16): + if row < 16: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(16): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:16], m_smem[0:16], S_smem[0:16, 0:16], V_smem[0:16, 0:256]) + T.writes(O_local[0:16, 0:256]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 8): + with T.block("O_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) + j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(2, 8, 4, 8): + with T.block("O_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) + k = T.axis.reduce(16, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 8): + with T.block("O_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(16, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def batch_prefill_paged_kv_sliding_window(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + total_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (total_len, 8, 256), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 8, 16, 256), "float16", offset_factor=1) + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (3, batch_size), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 8, 256), "float16") + lse = T.match_buffer(var_lse, (total_len, 8)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(8, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + K_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + S_smem = T.alloc_buffer((16, 16), scope="shared") + S_local = T.alloc_buffer((16, 16), scope="local") + O_local = T.alloc_buffer((16, 256), scope="local") + m_smem = T.alloc_buffer((16,), scope="shared") + m_prev_smem = T.alloc_buffer((16,), scope="shared") + d_smem = T.alloc_buffer((16,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 16 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[0, b_idx] - length_info[1, b_idx] + length_info[2, b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 8): + with T.block("O_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(8): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(16, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 256) + j = T.axis.spatial(256, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 256) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 15) // 16): + L_kv_start: T.int32 = iterator_1 * 16 + for lz_ly_fused_0 in range(8): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = T.if_then_else(cur_L < length_info[2, b_idx], cur_L, cur_L - length_info[2, b_idx] + length_info[1, b_idx]) + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + freq = T.float32() + K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, pages[page_no, 0, by, page_offset, j + 128] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 128]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), pages[page_no, 0, by, page_offset, j]) + else: + K_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_ly_fused_0 in range(8): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = T.if_then_else(cur_L < length_info[2, b_idx], cur_L, cur_L - length_info[2, b_idx] + length_info[1, b_idx]) + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:16, 0:256], K_smem[0:16, 0:256]) + T.writes(S_local[0:16, 0:16]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(1, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 + li_1_init) + j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(32, 1, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + k = T.axis.reduce(256, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.090168440055560212) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(1, 2): + with T.block("S_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(16): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) + T.writes(S_smem[row, 0:16]) + for j in range(16): + if row < 16: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(16): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:16], m_smem[0:16], S_smem[0:16, 0:16], V_smem[0:16, 0:256]) + T.writes(O_local[0:16, 0:256]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 8): + with T.block("O_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) + j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(2, 8, 4, 8): + with T.block("O_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) + k = T.axis.reduce(16, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 8): + with T.block("O_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(16, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def batch_prefill_ragged_kv(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_k_rope_pos_offset: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + qo_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (qo_len, 8, 256), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + kv_len = T.int32(is_size_var=True) + k = T.match_buffer(var_k, (kv_len, 8, 256), "float16") + v = T.match_buffer(var_v, (kv_len, 8, 256), "float16") + kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (qo_len, 8, 256), "float16") + lse = T.match_buffer(var_lse, (qo_len, 8)) + # with T.block("root"): + for lbx in T.thread_binding(8, thread="blockIdx.x"): + for lby in T.thread_binding(8, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + K_smem = T.alloc_buffer((256, 16), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + S_smem = T.alloc_buffer((16, 16), scope="shared") + S_local = T.alloc_buffer((16, 16), scope="local") + O_local = T.alloc_buffer((16, 256), scope="local") + m_smem = T.alloc_buffer((16,), scope="shared") + m_prev_smem = T.alloc_buffer((16,), scope="shared") + d_smem = T.alloc_buffer((16,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + q_indptr_val: T.int32 = q_indptr[b_idx] + LH_start: T.int32 = tile_id[0] * 16 + kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(8): + with T.block("O_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1_0 * 8 + lj_1_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_1_lj_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_1_lj_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_0, lj_0_0 in T.grid(2, 2): + for lj_1 in T.vectorized(8): + with T.block("Q_load"): + i = T.axis.spatial(16, li_0 * 8 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) // 16) + j = T.axis.spatial(256, lj_0_0 * 128 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) % 16 * 8 + lj_1) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 15) // 16): + L_kv_start: T.int32 = iterator_1 * 16 + L_kv_base: T.int32 = kv_indptr[b_idx] + for lz_1_ly_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for lz_1_ly_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lz_0, ly_0_0 in T.grid(2, 2): + for ly_1 in T.vectorized(8): + with T.block("K_load"): + i = T.axis.spatial(16, lz_0 * 8 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 16) + j = T.axis.spatial(256, ly_0_0 * 128 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 16 * 8 + ly_1) + T.reads(kv_chunk_len[0], k_rope_pos_offset[b_idx], k[L_kv_base + L_kv_start + i, by, j - 128:j - 128 + 257]) + T.writes(K_smem[j, i]) + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + freq = T.float32() + K_smem[j, i] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", k[L_kv_base + cur_L, by, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, k[L_kv_base + cur_L, by, j + 128] * T.float16(-1.0), k[L_kv_base + cur_L, by, j - 128]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), k[L_kv_base + cur_L, by, j]) + else: + K_smem[j, i] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_1_ly_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for lz_1_ly_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lz_0, ly_0_0 in T.grid(2, 2): + for ly_1 in T.vectorized(8): + with T.block("V_load"): + i = T.axis.spatial(16, lz_0 * 8 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 16) + j = T.axis.spatial(256, ly_0_0 * 128 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 16 * 8 + ly_1) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + V_smem[i, j] = v[L_kv_base + cur_L, by, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:16, 0:256], K_smem[0:256, 0:16]) + T.writes(S_local[0:16, 0:16]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init in T.unroll(1): + for lj_1_0_init in T.unroll(1): + for lj_1_1_init in T.vectorized(2): + with T.block("S_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 + li_1_init) + j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_0_init * 2 + lj_1_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0 in range(16): + for li_1 in T.unroll(1): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(2): + for lk_1 in range(16): + with T.block("S_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1_0 * 2 + lj_1_1) + k_1 = T.axis.reduce(256, lk_0 * 16 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k_1], K_smem[k_1, j]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k_1]) * T.Cast("float32", K_smem[k_1, j]) * attn_score_scaling_factor * T.float32(0.090168440055560212) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(1): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(2): + with T.block("S_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1_0 * 2 + lj_1_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(16): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) + T.writes(S_smem[row, 0:16]) + for j in range(16): + if row < 16: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(16): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:16], m_smem[0:16], S_smem[0:16, 0:16], V_smem[0:16, 0:256]) + T.writes(O_local[0:16, 0:256]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init in T.unroll(4): + for lj_1_0_init in T.unroll(1): + for lj_1_1_init in T.vectorized(8): + with T.block("O_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) + j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + lj_1_0_init * 8 + lj_1_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1 in T.grid(1, 16): + for li_1 in T.unroll(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(8): + with T.block("O_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1_0 * 8 + lj_1_1) + k_1 = T.axis.reduce(16, lk_0 * 16 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k_1], V_smem[k_1, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k_1] * T.Cast("float32", V_smem[k_1, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(8): + with T.block("O_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1_0 * 8 + lj_1_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(16, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 8 + + @T.prim_func + def batch_tree_attn(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_mn_indptr: T.handle, var_mask: T.handle, var_output: T.handle, var_lse: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32, batch_size: T.int32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + qo_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (qo_len, 8, 256), "float16") + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + kv_len = T.int32(is_size_var=True) + k = T.match_buffer(var_k, (kv_len, 8, 256), "float16") + v = T.match_buffer(var_v, (kv_len, 8, 256), "float16") + kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) + mn_indptr = T.match_buffer(var_mn_indptr, (batch_size + 1,), "int32", offset_factor=1) + tree_size = T.int32(is_size_var=True) + mask = T.match_buffer(var_mask, (tree_size, 2), "int32", offset_factor=1) + output = T.match_buffer(var_output, (qo_len, 8, 256), "float16") + lse = T.match_buffer(var_lse, (qo_len, 8)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(8, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + K_smem = T.alloc_buffer((32, 256), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 256), "float16", scope="shared") + S_smem = T.alloc_buffer((16, 32), scope="shared") + S_local = T.alloc_buffer((16, 32), scope="local") + O_local = T.alloc_buffer((16, 256), scope="local") + m_smem = T.alloc_buffer((16,), scope="shared") + m_prev_smem = T.alloc_buffer((16,), scope="shared") + d_smem = T.alloc_buffer((16,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 16 + q_indptr_val: T.int32 = q_indptr[b_idx] + kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 8): + with T.block("O_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(8): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(16, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 256) + j = T.axis.spatial(256, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 256) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 31) // 32): + L_kv_start: T.int32 = iterator_1 * 32 + L_kv_base: T.int32 = kv_indptr[b_idx] + for lz_ly_fused_0 in range(16): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("KV_load"): + i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_base + L_kv_start + i + if L_kv_start + i < kv_chunk_len[0]: + freq = T.float32() + K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", k[cur_L, by, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, k[cur_L, by, j + 128] * T.float16(-1.0), k[cur_L, by, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), k[cur_L, by, j]) + V_smem[i, j] = v[cur_L, by, j] + else: + K_smem[i, j] = T.float16(0.0) + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:16, 0:256], K_smem[0:32, 0:256]) + T.writes(S_local[0:16, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 2 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(32, 2, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 2 + lj_1) + k_1 = T.axis.reduce(256, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k_1], K_smem[j, k_1]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k_1]) * T.Cast("float32", K_smem[j, k_1]) * attn_score_scaling_factor * T.float32(0.090168440055560212) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 2): + with T.block("S_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 2 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 16: + row_: T.int32 = LH_start + row + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:16], m_smem[0:16], S_smem[0:16, 0:32], V_smem[0:32, 0:256]) + T.writes(O_local[0:16, 0:256]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 8): + with T.block("O_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) + j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(4, 8, 4, 8): + with T.block("O_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) + k_1 = T.axis.reduce(32, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k_1], V_smem[k_1, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k_1] * T.Cast("float32", V_smem[k_1, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 8): + with T.block("O_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(16, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def 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, 8, 16, 256), "float16", offset_factor=1) + copy_length_indptr = T.match_buffer(var_copy_length_indptr, (batch_size + 1,), "int32", offset_factor=1) + total_copy_length = T.int32() + copy_src_dst_pos = T.match_buffer(var_copy_src_dst_pos, (2, total_copy_length), "int32", offset_factor=1) + with T.block("root"): + T.reads() + T.writes() + for bhd_o in T.thread_binding(batch_size * 8, thread="blockIdx.x"): + for bhd_i in T.thread_binding(256, thread="threadIdx.x"): + b: T.int32 = (bhd_o * 256 + bhd_i) // 2048 + h: T.int32 = (bhd_o * 256 + bhd_i) // 256 % 8 + d: T.int32 = (bhd_o * 256 + bhd_i) % 256 + if bhd_o * 256 + bhd_i < batch_size * 8 * 256: + for i in range(copy_length_indptr[b + 1] - copy_length_indptr[b]): + src_pos: T.int32 = copy_src_dst_pos[0, copy_length_indptr[b] + i] + dst_pos: T.int32 = copy_src_dst_pos[1, copy_length_indptr[b] + i] + pages[dst_pos // 16, 0, h, dst_pos % 16, d] = pages[src_pos // 16, 0, h, src_pos % 16, d] + pages[dst_pos // 16, 1, h, dst_pos % 16, d] = pages[src_pos // 16, 1, h, src_pos % 16, d] + + @T.prim_func + def 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, 8, page_size, 256), "float16", offset_factor=1) + # with T.block("root"): + for b in T.thread_binding(copy_length * T.int64(8), thread="blockIdx.x"): + for t in T.thread_binding(256, thread="threadIdx.x"): + with T.block("copy"): + vh = T.axis.spatial(8, T.Cast("int32", (b * T.int64(256) + T.Cast("int64", t)) // (copy_length * T.int64(256)))) + vp = T.axis.spatial(copy_length, (b * T.int64(256) + T.Cast("int64", t)) % (copy_length * T.int64(256)) // T.int64(256)) + vd = T.axis.spatial(256, T.Cast("int32", (b * T.int64(256) + T.Cast("int64", t)) % T.int64(256))) + T.where(b * T.int64(256) + T.Cast("int64", t) < copy_length * T.int64(8) * T.int64(256)) + T.reads(pages[src_page_id, 0:2, vh, vp, vd]) + T.writes(pages[tgt_page_id, 0:2, vh, vp, vd]) + pages[tgt_page_id, 0, vh, vp, vd] = pages[src_page_id, 0, vh, vp, vd] + pages[tgt_page_id, 1, vh, vp, vd] = pages[src_page_id, 1, vh, vp, vd] + + @T.prim_func(private=True) + def fused_dequantize1_fused_NT_matmul10_add10(gpt_neox_layers_0_attention_query_key_value_q_weight2: T.Buffer((T.int64(6144), T.int64(256)), "uint32"), gpt_neox_layers_0_attention_query_key_value_q_scale2: T.Buffer((T.int64(6144), T.int64(64)), "float16"), layer_norm33: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16"), gpt_neox_layers_0_attention_query_key_value_bias2: T.Buffer((T.int64(6144),), "float16"), T_add_intermediate_intermediate: T.Buffer((T.int64(1), T.int64(1), T.int64(6144)), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + NT_matmul_intermediate_local = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(6144)), "float16", scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), T.int64(1), T.int64(6144)), "float16", scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), T.int64(1), T.int64(6144)), "float16", scope="local") + gpt_neox_layers_0_attention_query_key_value_q_weight2_local = T.alloc_buffer((T.int64(6144), T.int64(256)), "uint32", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(T.int64(3072), 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(6144), 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(8), 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(6144), u_fused_ax0_fused_fused_0 * T.int64(2) + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(T.int64(256), 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(6144), 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_norm33[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_norm33[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(6144), 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(6144), 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(6144), 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(6144), 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(6144), T.int64(256)), "uint32"), gpt_neox_layers_0_attention_query_key_value_q_scale3: T.Buffer((T.int64(6144), T.int64(64)), "float16"), p_layer_norm66: T.handle, gpt_neox_layers_0_attention_query_key_value_bias3: T.Buffer((T.int64(6144),), "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_norm66 = T.match_buffer(p_layer_norm66, (T.int64(1), seq_len, T.int64(2048)), "float16") + T_add_intermediate_intermediate = T.match_buffer(p_output0, (T.int64(1), seq_len, T.int64(6144)), "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(6144), T.int64(2048)), "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(6144)), "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(6144)), "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(6144)), "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(1536), 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(6144), 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(8), 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(6144), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(2048), 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(6144), 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_norm66[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_norm66[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(6144), 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(6144), 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(6144), 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(6144), 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(6144), T.int64(2048)), "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(6144)), "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(6144)), "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(6144)), "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(1536), 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(6144), 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(8), 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(6144), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(2048), 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(6144), 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_norm66[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_norm66[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(6144), 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(6144), 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(6144), 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(6144), 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(6144)), "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(24), 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(6144), 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(128), 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(6144), 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(2048), ax3_0 * T.int64(16) + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], layer_norm66[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_norm66[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(6144), 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(6144), T.int64(256)), "uint32"), gpt_neox_layers_0_attention_query_key_value_q_scale4: T.Buffer((T.int64(6144), T.int64(64)), "float16"), p_layer_norm99: T.handle, gpt_neox_layers_0_attention_query_key_value_bias4: T.Buffer((T.int64(6144),), "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_norm99 = T.match_buffer(p_layer_norm99, (batch_size, T.int64(1), T.int64(2048)), "float16") + T_add_intermediate_intermediate = T.match_buffer(p_output0, (batch_size, T.int64(1), T.int64(6144)), "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(6144), T.int64(2048)), "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(6144)), "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(6144)), "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(6144)), "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(1536), 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(6144), 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(8), 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(6144), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(2048), 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(6144), 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_norm99[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_norm99[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(6144), 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(6144), 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(6144), 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(6144), 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(6144), T.int64(2048)), "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(6144)), "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(6144)), "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(6144)), "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(1536), 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(6144), 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(8), 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(6144), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(2048), 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(6144), 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_norm99[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_norm99[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(6144), 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(6144), 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(6144), 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(6144), 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(6144)), "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(24), 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(6144), 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(128), 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(6144), 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(2048), ax3_0 * T.int64(16) + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], layer_norm99[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_norm99[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(6144), 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(2048), T.int64(256)), "uint32"), gpt_neox_layers_0_attention_dense_q_scale2: T.Buffer((T.int64(2048), T.int64(64)), "float16"), lv197: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16"), gpt_neox_layers_0_attention_dense_bias2: T.Buffer((T.int64(2048),), "float16"), lv146: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16"), input_embed: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16"), T_add_intermediate_1_2_intermediate: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + NT_matmul_intermediate_local = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16", scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), T.int64(1), T.int64(2048)), "float16", scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), T.int64(1), T.int64(2048)), "float16", scope="local") + gpt_neox_layers_0_attention_dense_q_weight2_local = T.alloc_buffer((T.int64(2048), T.int64(256)), "uint32", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(T.int64(1024), 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(2048), 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(8), 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(2048), u_fused_ax0_fused_fused_0 * T.int64(2) + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(T.int64(256), 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(2048), 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], lv197[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] + lv197[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(2048), 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(2048), 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(2048), 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(2048), u_fused_ax0_fused_fused_0 * T.int64(2) + ax0_fused_0_ax0_fused_1_fused + ax0_fused_2) + T.reads(lv146[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] = lv146[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(2048), T.int64(256)), "uint32"), gpt_neox_layers_0_attention_dense_q_scale4: T.Buffer((T.int64(2048), T.int64(64)), "float16"), p_reshape195: T.handle, gpt_neox_layers_0_attention_dense_bias4: T.Buffer((T.int64(2048),), "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() + reshape195 = T.match_buffer(p_reshape195, (batch_size, T.int64(1), T.int64(2048)), "float16") + lv2 = T.match_buffer(p_lv2, (batch_size, T.int64(1), T.int64(2048)), "float16") + input_embeds = T.match_buffer(p_input_embeds, (batch_size, T.int64(1), T.int64(2048)), "float16") + T_add_intermediate_1_2_intermediate = T.match_buffer(p_output0, (batch_size, T.int64(1), T.int64(2048)), "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(2048), T.int64(2048)), "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(2048)), "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(2048)), "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(2048)), "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(512), 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(2048), 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(8), 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(2048), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(2048), 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(2048), 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], reshape195[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, reshape195[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(2048), 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(2048), 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(2048), 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(2048), 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(2048), T.int64(2048)), "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(2048)), "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(2048)), "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(2048)), "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(512), 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(2048), 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(8), 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(2048), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(2048), 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(2048), 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], reshape195[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, reshape195[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(2048), 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(2048), 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(2048), 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(2048), 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(2048)), "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(8), 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(2048), 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(128), 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(2048), 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(2048), ax3_0 * T.int64(16) + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], reshape195[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, reshape195[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(2048), 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(2048), T.int64(256)), "uint32"), gpt_neox_layers_0_attention_dense_q_scale3: T.Buffer((T.int64(2048), T.int64(64)), "float16"), p_reshape131: T.handle, gpt_neox_layers_0_attention_dense_bias3: T.Buffer((T.int64(2048),), "float16"), p_lv50: 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() + reshape131 = T.match_buffer(p_reshape131, (T.int64(1), seq_len, T.int64(2048)), "float16") + lv50 = T.match_buffer(p_lv50, (T.int64(1), seq_len, T.int64(2048)), "float16") + input_embeds = T.match_buffer(p_input_embeds, (T.int64(1), seq_len, T.int64(2048)), "float16") + T_add_intermediate_1_2_intermediate = T.match_buffer(p_output0, (T.int64(1), seq_len, T.int64(2048)), "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(2048), T.int64(2048)), "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(2048)), "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(2048)), "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(2048)), "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(512), 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(2048), 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(8), 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(2048), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(2048), 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(2048), 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], reshape131[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, reshape131[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(2048), 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(2048), 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(2048), 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(2048), 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(lv50[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] = lv50[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(2048), T.int64(2048)), "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(2048)), "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(2048)), "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(2048)), "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(512), 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(2048), 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(8), 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(2048), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(2048), 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(2048), 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], reshape131[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, reshape131[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(2048), 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(2048), 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(2048), 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(2048), 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(lv50[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] = lv50[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(2048)), "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(8), 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(2048), 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(128), 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(2048), 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(2048), ax3_0 * T.int64(16) + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], reshape131[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, reshape131[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(2048), 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(lv50[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] = lv50[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(8192), T.int64(256)), "uint32"), gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale2: T.Buffer((T.int64(8192), T.int64(64)), "float16"), layer_norm34: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: T.Buffer((T.int64(8192),), "float32"), compute_intermediate_intermediate: T.Buffer((T.int64(1), T.int64(1), T.int64(8192)), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + NT_matmul_intermediate_local = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(8192)), scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), T.int64(1), T.int64(8192)), scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), T.int64(1), T.int64(8192)), scope="local") + gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2_local = T.alloc_buffer((T.int64(8192), T.int64(256)), "uint32", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(T.int64(4096), 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(8192), 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(8), 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(8192), u_fused_ax0_fused_fused_0 * T.int64(2) + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(T.int64(256), 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(8192), 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_norm34[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_norm34[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(8192), 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(8192), 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(8192), 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(8192), 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(8192), T.int64(256)), "uint32"), gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4: T.Buffer((T.int64(8192), T.int64(64)), "float16"), p_layer_norm100: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: T.Buffer((T.int64(8192),), "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_norm100 = T.match_buffer(p_layer_norm100, (batch_size, T.int64(1), T.int64(2048)), "float16") + compute_intermediate_intermediate = T.match_buffer(p_output0, (batch_size, T.int64(1), T.int64(8192)), "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(8192), T.int64(2048)), "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(8192)), 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(8192)), 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(8192)), 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(2048), 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(8192), 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(8), 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(8192), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(2048), 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(8192), 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_norm100[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_norm100[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(8192), 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(8192), 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(8192), 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(8192), 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(8192), T.int64(2048)), "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(8192)), 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(8192)), 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(8192)), 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(2048), 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(8192), 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(8), 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(8192), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(2048), 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(8192), 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_norm100[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_norm100[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(8192), 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(8192), 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(8192), 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(8192), 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(8192)), 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(32), 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(8192), 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(128), 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(8192), 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(2048), ax3_0 * T.int64(16) + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], layer_norm100[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_norm100[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(8192), 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(8192), T.int64(256)), "uint32"), gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3: T.Buffer((T.int64(8192), T.int64(64)), "float16"), p_layer_norm67: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: T.Buffer((T.int64(8192),), "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_norm67 = T.match_buffer(p_layer_norm67, (T.int64(1), seq_len, T.int64(2048)), "float16") + compute_intermediate_intermediate = T.match_buffer(p_output0, (T.int64(1), seq_len, T.int64(8192)), "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(8192), T.int64(2048)), "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(8192)), 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(8192)), 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(8192)), 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(2048), 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(8192), 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(8), 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(8192), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(2048), 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(8192), 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_norm67[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_norm67[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(8192), 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(8192), 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(8192), 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(8192), 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(8192), T.int64(2048)), "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(8192)), 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(8192)), 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(8192)), 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(2048), 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(8192), 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(8), 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(8192), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(2048), 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(8192), 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_norm67[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_norm67[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(8192), 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(8192), 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(8192), 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(8192), 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(8192)), 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(32), 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(8192), 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(128), 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(8192), 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(2048), ax3_0 * T.int64(16) + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], layer_norm67[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_norm67[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(8192), 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(2048), T.int64(1024)), "uint32"), gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale2: T.Buffer((T.int64(2048), T.int64(256)), "float16"), lv198: T.Buffer((T.int64(1), T.int64(1), T.int64(8192)), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: T.Buffer((T.int64(2048),), "float32"), compute_intermediate_intermediate: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + NT_matmul_intermediate_local = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(2048)), scope="local") + NT_matmul_intermediate_rf_local = T.alloc_buffer((T.int64(256), T.int64(1), T.int64(1), T.int64(2048)), scope="local") + NT_matmul_intermediate_rf_local_1 = T.alloc_buffer((T.int64(32), T.int64(1), T.int64(1), T.int64(2048)), scope="local") + gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2_local = T.alloc_buffer((T.int64(2048), T.int64(1024)), "uint32", scope="local") + for u_fused_ax0_fused_fused_0 in T.thread_binding(T.int64(1024), 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(2048), 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(32), 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(2048), u_fused_ax0_fused_fused_0 * T.int64(2) + u_fused_ax0_fused_fused_1) + v1 = T.axis.spatial(T.int64(1024), 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(2048), 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], lv198[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", lv198[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(2048), 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(2048), 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(2048), 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(2048), 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(2048), T.int64(1024)), "uint32"), gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4: T.Buffer((T.int64(2048), T.int64(256)), "float16"), p_lv1: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: T.Buffer((T.int64(2048),), "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(8192)), "float16") + compute_intermediate_intermediate = T.match_buffer(p_output0, (batch_size, T.int64(1), T.int64(2048)), "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(2048), T.int64(8192)), "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(2048)), 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(2048)), 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(2048)), 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(512), 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(2048), 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(32), 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(2048), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(8192), 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(2048), 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(2048), 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(2048), 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(2048), 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(2048), 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(2048), T.int64(8192)), "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(2048)), 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(2048)), 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(2048)), 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(512), 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(2048), 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(32), 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(2048), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(8192), 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(2048), 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(2048), 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(2048), 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(2048), 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(2048), 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(2048)), 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(8), 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(2048), 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(512), 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(2048), 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(8192), 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(2048), 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(2048), T.int64(1024)), "uint32"), gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3: T.Buffer((T.int64(2048), T.int64(256)), "float16"), p_lv66: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: T.Buffer((T.int64(2048),), "float32"), p_output0: T.handle): + T.func_attr({"tir.HoistIfThenElseExprWithBlock": 1, "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int64() + lv66 = T.match_buffer(p_lv66, (T.int64(1), seq_len, T.int64(8192)), "float16") + compute_intermediate_intermediate = T.match_buffer(p_output0, (T.int64(1), seq_len, T.int64(2048)), "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(2048), T.int64(8192)), "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(2048)), 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(2048)), 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(2048)), 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(512), 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(2048), 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(32), 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(2048), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(8192), 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(2048), 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], lv66[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, lv66[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(2048), 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(2048), 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(2048), 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(2048), 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(2048), T.int64(8192)), "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(2048)), 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(2048)), 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(2048)), 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(512), 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(2048), 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(32), 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(2048), ax1_fused_0 * T.int64(4) + ax1_fused_1 * T.int64(2) + ax0_0_1 + ax0_1) + v1 = T.axis.spatial(T.int64(8192), 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(2048), 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], lv66[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, lv66[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(2048), 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(2048), 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(2048), 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(2048), 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(2048)), 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(8), 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(2048), 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(512), 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(2048), 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(8192), ax3_0 * T.int64(16) + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], lv66[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, lv66[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(2048), 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_norm65: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "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(256)), "uint32") + embed_out_q_scale2 = T.match_buffer(p_embed_out_q_scale2, (vocab_size, T.int64(64)), "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(256)), "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(8), 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(256), 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_norm65[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_norm65[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_norm131: 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(256)), "uint32") + embed_out_q_scale4 = T.match_buffer(p_embed_out_q_scale4, (vocab_size, T.int64(64)), "float16") + batch_size = T.int64() + layer_norm131 = T.match_buffer(p_layer_norm131, (batch_size, T.int64(1), T.int64(2048)), "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(128), 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(2048), ax3_0 * T.int64(16) + ax3_1) + T.reads(NT_matmul_intermediate_reindex_pad_local[T.int64(0), v1, v2], layer_norm131[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_norm131[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(256)), "uint32") + embed_out_q_scale3 = T.match_buffer(p_embed_out_q_scale3, (vocab_size, T.int64(64)), "float16") + batch_size = T.int64() + take1 = T.match_buffer(p_take1, (T.int64(1), batch_size, T.int64(2048)), "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(128), 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(2048), 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, 256), "uint32") + gpt_neox_embed_in_q_scale = T.match_buffer(p_gpt_neox_embed_in_q_scale, (vocab_size, 64), "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, 2048), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(seq_len * 8, thread="blockIdx.x"): + for ax0_ax1_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("T_take"): + v0 = T.axis.spatial(seq_len, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) // 2048) + v1 = T.axis.spatial(2048, (ax0_ax1_fused_0 * 256 + ax0_ax1_fused_1) % 2048) + T.reads(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(lv84: T.Buffer((T.int64(1), T.int64(8), T.int64(256)), "float16"), T_reshape_intermediate_1: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + for ax0_fused_0 in T.thread_binding(T.int64(8), 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(2048), ax0_fused_0 * T.int64(256) + ax0_fused_1) + T.reads(lv84[T.int64(0), v0 // T.int64(256), v0 % T.int64(256)]) + T.writes(T_reshape_intermediate_1[T.int64(0), T.int64(0), v0]) + T_reshape_intermediate_1[T.int64(0), T.int64(0), v0] = lv84[T.int64(0), v0 // T.int64(256), v0 % T.int64(256)] + + @T.prim_func(private=True) + def fused_reshape8_reshape9(add96: T.Buffer((T.int64(1), T.int64(1), T.int64(6144)), "float16"), T_reshape_intermediate_1: T.Buffer((T.int64(1), T.int64(24), T.int64(256)), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(T.int64(24), 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(24), (ax0_ax1_fused_0 * T.int64(256) + ax0_ax1_fused_1) // T.int64(256)) + v1 = T.axis.spatial(T.int64(256), (ax0_ax1_fused_0 * T.int64(256) + ax0_ax1_fused_1) % T.int64(256)) + T.reads(add96[T.int64(0), T.int64(0), v0 * T.int64(256) + v1]) + T.writes(T_reshape_intermediate_1[T.int64(0), v0, v1]) + T_reshape_intermediate_1[T.int64(0), v0, v1] = add96[T.int64(0), T.int64(0), v0 * T.int64(256) + 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, 24, 256), "float16") + position_map = T.match_buffer(var_position_map, (seq_len,), "int32", offset_factor=1) + q = T.match_buffer(var_q, (seq_len, 8, 256), "float16") + k = T.match_buffer(var_k, (seq_len, 8, 256), "float16") + v = T.match_buffer(var_v, (seq_len, 8, 256), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * 24, thread="blockIdx.x"): + for ax0_ax1_ax2_fused_1 in T.thread_binding(256, thread="threadIdx.x"): + with T.block("llama_fused_rope"): + v0 = T.axis.spatial(seq_len, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) // 6144) + v1 = T.axis.spatial(24, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 6144 // 256) + v2 = T.axis.spatial(256, (ax0_ax1_ax2_fused_0 * 256 + ax0_ax1_ax2_fused_1) % 256) + T.reads(position_map[v0], qkv[v0, v1, v2 + -32:v2 + -32 + 65]) + T.writes(q[v0, v1, v2], k[v0, v1 + -8, v2], v[v0, v1 + -16, v2]) + if v1 < 8: + freq = T.float32() + q[v0, v1, v2] = T.if_then_else(apply_rope > 0 and v2 < 64, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", qkv[v0, v1, v2]) + T.sin(freq) * T.Cast("float32", T.if_then_else(v2 < 32, qkv[v0, v1, v2 + 32] * T.float16(-1.0), qkv[v0, v1, v2 + -32]))), where={freq: T.Cast("float32", position_map[v0]) / T.pow(T.float32(10000.0), T.Cast("float32", v2 * 2 % 64) / T.float32(64.0))}), qkv[v0, v1, v2]) + else: + if v1 < 16: + freq = T.float32() + k[v0, v1 + -8, v2] = T.if_then_else(apply_rope > 0 and v2 < 64, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", qkv[v0, v1, v2]) + T.sin(freq) * T.Cast("float32", T.if_then_else(v2 < 32, qkv[v0, v1, v2 + 32] * T.float16(-1.0), qkv[v0, v1, v2 + -32]))), where={freq: T.Cast("float32", position_map[v0]) / T.pow(T.float32(10000.0), T.Cast("float32", v2 * 2 % 64) / T.float32(64.0))}), qkv[v0, v1, v2]) + else: + v[v0, v1 + -16, v2] = qkv[v0, v1, v2] + + @T.prim_func + def 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_norm32: T.handle, index: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16")): + T.func_attr({"target": T.target({"keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int64() + layer_norm32 = T.match_buffer(var_layer_norm32, (T.int64(1), seq_len, T.int64(2048)), "float16") + # with T.block("root"): + for ax0_fused_0 in T.thread_binding(T.int64(8), 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(2048), ax0_fused_0 * T.int64(256) + ax0_fused_1) + T.reads(layer_norm32[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_norm32[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(2048),), "float16"), gpt_neox_layers_0_input_layernorm_bias4: T.Buffer((T.int64(2048),), "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(2048)), "float16") + T_layer_norm = T.match_buffer(var_T_layer_norm, (batch_size, T.int64(1), T.int64(2048)), "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(8), annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("input_embeds_red_temp"): + v0 = T.axis.spatial(batch_size, ax0_fused + ax0) + v1 = T.axis.reduce(T.int64(2048), 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(8), annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("T_layer_norm"): + v0 = T.axis.spatial(batch_size, ax0_fused) + v1 = T.axis.spatial(T.int64(2048), 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(0.00048828125)) * T.rsqrt(input_embeds_red_temp_v1_shared[v0, T.int64(0)] * T.float32(0.00048828125) - input_embeds_red_temp_v0_shared[v0, T.int64(0)] * T.float32(0.00048828125) * (input_embeds_red_temp_v0_shared[v0, T.int64(0)] * T.float32(0.00048828125)) + T.float32(1.0000000000000001e-05))) * gpt_neox_layers_0_input_layernorm_weight4[v1] + gpt_neox_layers_0_input_layernorm_bias4[v1] + + @T.prim_func(private=True) + def layer_norm1(var_input_embeds: T.handle, gpt_neox_layers_0_input_layernorm_weight3: T.Buffer((T.int64(2048),), "float16"), gpt_neox_layers_0_input_layernorm_bias3: T.Buffer((T.int64(2048),), "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(2048)), "float16") + T_layer_norm = T.match_buffer(var_T_layer_norm, (T.int64(1), seq_len, T.int64(2048)), "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(8), annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("input_embeds_red_temp"): + v0 = T.axis.spatial(seq_len, ax0_fused + ax0) + v1 = T.axis.reduce(T.int64(2048), 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(8), annotations={"pragma_auto_unroll_max_step": 64, "pragma_unroll_explicit": 1}): + with T.block("T_layer_norm"): + v0 = T.axis.spatial(seq_len, ax0_fused) + v1 = T.axis.spatial(T.int64(2048), 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(0.00048828125)) * T.rsqrt(input_embeds_red_temp_v1_shared[T.int64(0), v0] * T.float32(0.00048828125) - input_embeds_red_temp_v0_shared[T.int64(0), v0] * T.float32(0.00048828125) * (input_embeds_red_temp_v0_shared[T.int64(0), v0] * T.float32(0.00048828125)) + T.float32(1.0000000000000001e-05))) * gpt_neox_layers_0_input_layernorm_weight3[v1] + gpt_neox_layers_0_input_layernorm_bias3[v1] + + @T.prim_func(private=True) + def layer_norm2(input_embed: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16"), gpt_neox_layers_0_input_layernorm_weight2: T.Buffer((T.int64(2048),), "float16"), gpt_neox_layers_0_input_layernorm_bias2: T.Buffer((T.int64(2048),), "float16"), T_layer_norm: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16")): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + # with T.block("root"): + input_embed_red_temp_v0_shared = T.alloc_buffer((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(8), 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(2048), 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(8), 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(2048), 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(0.00048828125)) * T.rsqrt(input_embed_red_temp_v1_shared[T.int64(0), T.int64(0)] * T.float32(0.00048828125) - input_embed_red_temp_v0_shared[T.int64(0), T.int64(0)] * T.float32(0.00048828125) * (input_embed_red_temp_v0_shared[T.int64(0), T.int64(0)] * T.float32(0.00048828125)) + T.float32(1.0000000000000001e-05))) * gpt_neox_layers_0_input_layernorm_weight2[v1] + gpt_neox_layers_0_input_layernorm_bias2[v1] + + @T.prim_func + def merge_state_inplace(v: T.handle, s: T.handle, v_other: T.handle, s_other: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + N, H, D = T.int32(is_size_var=True), T.int32(is_size_var=True), T.int32(is_size_var=True) + V = T.match_buffer(v, (N, H, D), "float16") + S = T.match_buffer(s, (N, H)) + V_other = T.match_buffer(v_other, (N, H, D), "float16") + S_other = T.match_buffer(s_other, (N, H)) + # with T.block("root"): + for bx in T.thread_binding(N, thread="blockIdx.x"): + for by in T.thread_binding(2, thread="blockIdx.y"): + for ty in T.thread_binding(4, thread="threadIdx.y"): + for tx in T.thread_binding(64, thread="threadIdx.x"): + with T.block("merge"): + T.reads(S[bx, ty + by * 4], S_other[bx, ty + by * 4], V[bx, ty + by * 4, tx * 4:tx * 4 + 4], V_other[bx, ty + by * 4, tx * 4:tx * 4 + 4]) + T.writes(V[bx, ty + by * 4, tx * 4:tx * 4 + 4], S[bx, ty + by * 4]) + s_val = T.alloc_buffer((1,), scope="local") + s_other_val = T.alloc_buffer((1,), scope="local") + s_max = T.alloc_buffer((1,), scope="local") + scale = T.alloc_buffer((1,), scope="local") + other_scale = T.alloc_buffer((1,), scope="local") + v_vec = T.alloc_buffer((4,), "float16", scope="local") + v_other_vec = T.alloc_buffer((4,), "float16", scope="local") + s_val[0] = S[bx, ty + by * 4] + s_other_val[0] = S_other[bx, ty + by * 4] + s_max[0] = T.max(s_val[0], s_other_val[0]) + s_val[0] = T.exp2(s_val[0] - s_max[0]) + s_other_val[0] = T.exp2(s_other_val[0] - s_max[0]) + scale[0] = s_val[0] / (s_val[0] + s_other_val[0]) + other_scale[0] = s_other_val[0] / (s_val[0] + s_other_val[0]) + for vec in T.vectorized(4): + v_vec[vec] = V[bx, ty + by * 4, tx * 4 + vec] + for vec in T.vectorized(4): + v_other_vec[vec] = V_other[bx, ty + by * 4, tx * 4 + vec] + for vec in range(4): + v_vec[vec] = T.Cast("float16", T.Cast("float32", v_vec[vec]) * scale[0] + T.Cast("float32", v_other_vec[vec]) * other_scale[0]) + for vec in T.vectorized(4): + V[bx, ty + by * 4, tx * 4 + vec] = v_vec[vec] + S[bx, ty + by * 4] = T.log2(s_val[0] + s_other_val[0]) + s_max[0] + + @T.prim_func(private=True) + def reshape(var_add288: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int64() + add288 = T.match_buffer(var_add288, (batch_size, T.int64(1), T.int64(6144)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, T.int64(1), T.int64(24), T.int64(256)), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * T.int64(24), 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(6144)) + v1 = T.axis.spatial(T.int64(24), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(6144) // T.int64(256)) + v2 = T.axis.spatial(T.int64(256), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(256)) + T.reads(add288[v0, T.int64(0), v1 * T.int64(256) + v2]) + T.writes(T_reshape[v0, T.int64(0), v1, v2]) + T_reshape[v0, T.int64(0), v1, v2] = add288[v0, T.int64(0), v1 * T.int64(256) + v2] + + @T.prim_func(private=True) + def reshape1(var_reshape192: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int64() + reshape192 = T.match_buffer(var_reshape192, (batch_size, T.int64(1), T.int64(24), T.int64(256)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, T.int64(24), T.int64(256)), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * T.int64(24), 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(6144)) + v1 = T.axis.spatial(T.int64(24), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(6144) // T.int64(256)) + v2 = T.axis.spatial(T.int64(256), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(256)) + T.reads(reshape192[v0, T.int64(0), v1, v2]) + T.writes(T_reshape[v0, v1, v2]) + T_reshape[v0, v1, v2] = reshape192[v0, T.int64(0), v1, v2] + + @T.prim_func(private=True) + def reshape2(var_lv246: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int64() + lv246 = T.match_buffer(var_lv246, (batch_size, T.int64(8), T.int64(256)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, T.int64(1), T.int64(8), T.int64(256)), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(batch_size * T.int64(8), 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(2048)) + v1 = T.axis.spatial(T.int64(8), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(2048) // T.int64(256)) + v2 = T.axis.spatial(T.int64(256), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(256)) + T.reads(lv246[v0, v1, v2]) + T.writes(T_reshape[v0, T.int64(0), v1, v2]) + T_reshape[v0, T.int64(0), v1, v2] = lv246[v0, v1, v2] + + @T.prim_func(private=True) + def reshape3(var_reshape194: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + batch_size = T.int64() + reshape194 = T.match_buffer(var_reshape194, (batch_size, T.int64(1), T.int64(8), T.int64(256)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, T.int64(1), T.int64(2048)), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(batch_size * T.int64(8), 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(2048)) + v1 = T.axis.spatial(T.int64(2048), (ax0_ax1_fused_0 * T.int64(256) + ax0_ax1_fused_1) % T.int64(2048)) + T.reads(reshape194[v0, T.int64(0), v1 // T.int64(256), v1 % T.int64(256)]) + T.writes(T_reshape[v0, T.int64(0), v1]) + T_reshape[v0, T.int64(0), v1] = reshape194[v0, T.int64(0), v1 // T.int64(256), v1 % T.int64(256)] + + @T.prim_func(private=True) + def reshape4(var_add192: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int64() + add192 = T.match_buffer(var_add192, (T.int64(1), seq_len, T.int64(6144)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (T.int64(1), seq_len, T.int64(24), T.int64(256)), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * T.int64(24), 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(6144)) + v1 = T.axis.spatial(T.int64(24), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(6144) // T.int64(256)) + v2 = T.axis.spatial(T.int64(256), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(256)) + T.reads(add192[T.int64(0), v0, v1 * T.int64(256) + v2]) + T.writes(T_reshape[T.int64(0), v0, v1, v2]) + T_reshape[T.int64(0), v0, v1, v2] = add192[T.int64(0), v0, v1 * T.int64(256) + v2] + + @T.prim_func(private=True) + def reshape5(var_reshape128: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int64() + reshape128 = T.match_buffer(var_reshape128, (T.int64(1), seq_len, T.int64(24), T.int64(256)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (seq_len, T.int64(24), T.int64(256)), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * T.int64(24), 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(6144)) + v1 = T.axis.spatial(T.int64(24), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(6144) // T.int64(256)) + v2 = T.axis.spatial(T.int64(256), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(256)) + T.reads(reshape128[T.int64(0), v0, v1, v2]) + T.writes(T_reshape[v0, v1, v2]) + T_reshape[v0, v1, v2] = reshape128[T.int64(0), v0, v1, v2] + + @T.prim_func(private=True) + def reshape6(var_lv165: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int64() + lv165 = T.match_buffer(var_lv165, (seq_len, T.int64(8), T.int64(256)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (T.int64(1), seq_len, T.int64(8), T.int64(256)), "float16") + # with T.block("root"): + for ax0_ax1_ax2_fused_0 in T.thread_binding(seq_len * T.int64(8), 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(2048)) + v1 = T.axis.spatial(T.int64(8), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(2048) // T.int64(256)) + v2 = T.axis.spatial(T.int64(256), (ax0_ax1_ax2_fused_0 * T.int64(256) + ax0_ax1_ax2_fused_1) % T.int64(256)) + T.reads(lv165[v0, v1, v2]) + T.writes(T_reshape[T.int64(0), v0, v1, v2]) + T_reshape[T.int64(0), v0, v1, v2] = lv165[v0, v1, v2] + + @T.prim_func(private=True) + def reshape7(var_reshape130: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int64() + reshape130 = T.match_buffer(var_reshape130, (T.int64(1), seq_len, T.int64(8), T.int64(256)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (T.int64(1), seq_len, T.int64(2048)), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(seq_len * T.int64(8), 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(2048)) + v1 = T.axis.spatial(T.int64(2048), (ax0_ax1_fused_0 * T.int64(256) + ax0_ax1_fused_1) % T.int64(2048)) + T.reads(reshape130[T.int64(0), v0, v1 // T.int64(256), v1 % T.int64(256)]) + T.writes(T_reshape[T.int64(0), v0, v1]) + T_reshape[T.int64(0), v0, v1] = reshape130[T.int64(0), v0, v1 // T.int64(256), v1 % T.int64(256)] + + @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_norm98: T.handle, var_logit_positions: T.handle, var_T_take: T.handle): + T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)}) + seq_len = T.int64() + layer_norm98 = T.match_buffer(var_layer_norm98, (T.int64(1), seq_len, T.int64(2048)), "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(2048)), "float16") + # with T.block("root"): + for ax0_ax1_fused_0 in T.thread_binding(batch_size * T.int64(8), 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(2048)) + v1 = T.axis.spatial(T.int64(2048), (ax0_ax1_fused_0 * T.int64(256) + ax0_ax1_fused_1) % T.int64(2048)) + T.reads(layer_norm98[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_norm98[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, 8, page_size, 256), "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, (16, seqlen, 8, 256), "float16") + v_data = T.match_buffer(var_v_data, (16, seqlen, 8, 256), "float16") + # with T.block("root"): + for p_h_d_fused_0 in T.thread_binding(seqlen * T.int64(8), 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(2048)) + vh = T.axis.spatial(8, T.Cast("int32", (p_h_d_fused_0 * T.int64(256) + p_h_d_fused_1) % T.int64(2048) // T.int64(256))) + vd = T.axis.spatial(256, T.Cast("int32", (p_h_d_fused_0 * T.int64(256) + p_h_d_fused_1) % T.int64(256))) + 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, 8, 16, 256), "float16", offset_factor=1) + ntoken = T.int64(is_size_var=True) + k_data = T.match_buffer(var_k_data, (ntoken, 8, 256), "float16") + v_data = T.match_buffer(var_v_data, (ntoken, 8, 256), "float16") + position_map = T.match_buffer(var_position_map, (ntoken,), "int32", offset_factor=1) + # with T.block("root"): + for global_pos_h_f_fused_0 in T.thread_binding(ntoken * T.int64(8), 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(2048)] != -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(2048)) + vh = T.axis.spatial(8, T.Cast("int32", (global_pos_h_f_fused_0 * T.int64(256) + global_pos_h_f_fused_1) % T.int64(2048) // T.int64(256))) + vf = T.axis.spatial(256, T.Cast("int32", (global_pos_h_f_fused_0 * T.int64(256) + global_pos_h_f_fused_1) % T.int64(256))) + T.reads(position_map[vgpos], k_data[vgpos, vh, vf]) + T.writes(pages[position_map[vgpos] // 16, 0, vh, position_map[vgpos] % 16, vf]) + position: T.int32 = position_map[vgpos] + pages[position // 16, 0, vh, position % 16, vf] = k_data[vgpos, vh, vf] + with T.block("v_transpose_append"): + vgpos = T.axis.spatial(ntoken, (global_pos_h_f_fused_0 * T.int64(256) + global_pos_h_f_fused_1) // T.int64(2048)) + vh = T.axis.spatial(8, T.Cast("int32", (global_pos_h_f_fused_0 * T.int64(256) + global_pos_h_f_fused_1) % T.int64(2048) // T.int64(256))) + vf = T.axis.spatial(256, T.Cast("int32", (global_pos_h_f_fused_0 * T.int64(256) + global_pos_h_f_fused_1) % T.int64(256))) + T.reads(position_map[vgpos], v_data[vgpos, vh, vf]) + T.writes(pages[position_map[vgpos] // 16, 1, vh, position_map[vgpos] % 16, vf]) + position: T.int32 = position_map[vgpos] + pages[position // 16, 1, vh, position % 16, vf] = v_data[vgpos, vh, vf] + + @T.prim_func + def tree_attn_paged_kv(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32, tree_order_indptr_handle: T.handle, tree_order_handle: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + total_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (total_len, 8, 256), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 8, 32, 256), "float16") + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 8, 256), "float16") + lse = T.match_buffer(var_lse, (total_len, 8)) + tree_order_indptr = T.match_buffer(tree_order_indptr_handle, (batch_size + 1,), "int32", offset_factor=1) + total_tree_order_len = T.int32(is_size_var=True) + tree_order = T.match_buffer(tree_order_handle, (total_tree_order_len, 2), "int32", offset_factor=1) + # with T.block("root"): + assert rotary_mode == 0, "Inline rotary mode is not supported in tree attention." + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(8, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((16, 256), "float16", scope="shared") + K_smem = T.alloc_buffer((32, 256), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 256), "float16", scope="shared") + S_smem = T.alloc_buffer((16, 32), scope="shared") + S_local = T.alloc_buffer((16, 32), scope="local") + O_local = T.alloc_buffer((16, 256), scope="local") + m_smem = T.alloc_buffer((16,), scope="shared") + m_prev_smem = T.alloc_buffer((16,), scope="shared") + d_smem = T.alloc_buffer((16,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 16 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 32 + length_info[b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 8): + with T.block("O_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(8): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(16, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 256) + j = T.axis.spatial(256, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 256) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 31) // 32): + L_kv_start: T.int32 = iterator_1 * 32 + for lz_ly_fused_0 in range(16): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 32] + page_offset: T.int32 = seq_offset % 32 + K_smem[i, j] = pages[page_no, 0, by, page_offset, j] + else: + K_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_ly_fused_0 in range(16): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) + j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 32] + page_offset: T.int32 = seq_offset % 32 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:16, 0:256], K_smem[0:32, 0:256]) + T.writes(S_local[0:16, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 2 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(32, 2, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 2 + lj_1) + k = T.axis.reduce(256, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.090168440055560212) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 2): + with T.block("S_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 2 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 16: + row_: T.int32 = LH_start + row + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 16: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:16], m_smem[0:16], S_smem[0:16, 0:32], V_smem[0:32, 0:256]) + T.writes(O_local[0:16, 0:256]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 8): + with T.block("O_gemm_init"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) + j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(4, 8, 4, 8): + with T.block("O_gemm_update"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) + k = T.axis.reduce(32, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 8): + with T.block("O_store"): + i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) + j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(16, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @R.function + def alloc_embedding_tensor() -> R.Tensor((2048, 2048), dtype="float16"): + R.func_attr({"relax.memory_plan_dynamic_func_output": True}) + gv: R.Tensor((2048, 2048), dtype="float16") = R.builtin.alloc_tensor(R.shape([2048, 2048]), R.dtype("float16"), R.prim_value(0), R.str("global")) + return gv + + @R.function + def batch_decode(input_embeds: R.Tensor(("batch_size", 1, 2048), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 256), dtype="uint32"), R.Tensor(("vocab_size", 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 256), dtype="uint32"), R.Tensor(("vocab_size", 64), 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((2048,), dtype="float16") = packed_params[2] + gpt_neox_layers_0_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[4] + gpt_neox_layers_0_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[6] + gpt_neox_layers_0_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[8] + gpt_neox_layers_0_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[9] + gpt_neox_layers_0_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[10] + gpt_neox_layers_0_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[12] + gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[13] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[14] + gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[15] + gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[16] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[17] + gpt_neox_layers_1_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[18] + gpt_neox_layers_1_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[19] + gpt_neox_layers_1_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[20] + gpt_neox_layers_1_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[21] + gpt_neox_layers_1_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[22] + gpt_neox_layers_1_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[23] + gpt_neox_layers_1_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[24] + gpt_neox_layers_1_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[25] + gpt_neox_layers_1_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[26] + gpt_neox_layers_1_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[27] + gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[28] + gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[29] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[30] + gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[31] + gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[32] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[33] + gpt_neox_layers_2_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[34] + gpt_neox_layers_2_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[35] + gpt_neox_layers_2_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[36] + gpt_neox_layers_2_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[37] + gpt_neox_layers_2_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[38] + gpt_neox_layers_2_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[39] + gpt_neox_layers_2_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[40] + gpt_neox_layers_2_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[41] + gpt_neox_layers_2_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[42] + gpt_neox_layers_2_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[43] + gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[44] + gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[45] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[46] + gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[47] + gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[48] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[49] + gpt_neox_layers_3_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[50] + gpt_neox_layers_3_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[51] + gpt_neox_layers_3_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[52] + gpt_neox_layers_3_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[53] + gpt_neox_layers_3_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[54] + gpt_neox_layers_3_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[55] + gpt_neox_layers_3_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[56] + gpt_neox_layers_3_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[57] + gpt_neox_layers_3_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[58] + gpt_neox_layers_3_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[59] + gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[60] + gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[61] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[62] + gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[63] + gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[64] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[65] + gpt_neox_layers_4_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[66] + gpt_neox_layers_4_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[67] + gpt_neox_layers_4_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[68] + gpt_neox_layers_4_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[69] + gpt_neox_layers_4_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[70] + gpt_neox_layers_4_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[71] + gpt_neox_layers_4_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[72] + gpt_neox_layers_4_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[73] + gpt_neox_layers_4_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[74] + gpt_neox_layers_4_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[75] + gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[76] + gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[77] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[78] + gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[79] + gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[80] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[81] + gpt_neox_layers_5_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[82] + gpt_neox_layers_5_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[83] + gpt_neox_layers_5_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[84] + gpt_neox_layers_5_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[85] + gpt_neox_layers_5_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[86] + gpt_neox_layers_5_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[87] + gpt_neox_layers_5_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[88] + gpt_neox_layers_5_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[89] + gpt_neox_layers_5_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[90] + gpt_neox_layers_5_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[91] + gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[92] + gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[93] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[94] + gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[95] + gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[96] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[97] + gpt_neox_layers_6_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[98] + gpt_neox_layers_6_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[99] + gpt_neox_layers_6_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[100] + gpt_neox_layers_6_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[101] + gpt_neox_layers_6_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[102] + gpt_neox_layers_6_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[103] + gpt_neox_layers_6_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[104] + gpt_neox_layers_6_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[105] + gpt_neox_layers_6_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[106] + gpt_neox_layers_6_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[107] + gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[108] + gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[109] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[110] + gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[111] + gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[112] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[113] + gpt_neox_layers_7_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[114] + gpt_neox_layers_7_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[115] + gpt_neox_layers_7_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[116] + gpt_neox_layers_7_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[117] + gpt_neox_layers_7_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[118] + gpt_neox_layers_7_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[119] + gpt_neox_layers_7_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[120] + gpt_neox_layers_7_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[121] + gpt_neox_layers_7_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[122] + gpt_neox_layers_7_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[123] + gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[124] + gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[125] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[126] + gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[127] + gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[128] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[129] + gpt_neox_layers_8_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[130] + gpt_neox_layers_8_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[131] + gpt_neox_layers_8_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[132] + gpt_neox_layers_8_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[133] + gpt_neox_layers_8_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[134] + gpt_neox_layers_8_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[135] + gpt_neox_layers_8_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[136] + gpt_neox_layers_8_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[137] + gpt_neox_layers_8_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[138] + gpt_neox_layers_8_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[139] + gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[140] + gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[141] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[142] + gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[143] + gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[144] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[145] + gpt_neox_layers_9_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[146] + gpt_neox_layers_9_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[147] + gpt_neox_layers_9_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[148] + gpt_neox_layers_9_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[149] + gpt_neox_layers_9_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[150] + gpt_neox_layers_9_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[151] + gpt_neox_layers_9_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[152] + gpt_neox_layers_9_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[153] + gpt_neox_layers_9_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[154] + gpt_neox_layers_9_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[155] + gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[156] + gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[157] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[158] + gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[159] + gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[160] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[161] + gpt_neox_layers_10_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[162] + gpt_neox_layers_10_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[163] + gpt_neox_layers_10_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[164] + gpt_neox_layers_10_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[165] + gpt_neox_layers_10_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[166] + gpt_neox_layers_10_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[167] + gpt_neox_layers_10_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[168] + gpt_neox_layers_10_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[169] + gpt_neox_layers_10_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[170] + gpt_neox_layers_10_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[171] + gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[172] + gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[173] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[174] + gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[175] + gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[176] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[177] + gpt_neox_layers_11_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[178] + gpt_neox_layers_11_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[179] + gpt_neox_layers_11_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[180] + gpt_neox_layers_11_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[181] + gpt_neox_layers_11_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[182] + gpt_neox_layers_11_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[183] + gpt_neox_layers_11_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[184] + gpt_neox_layers_11_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[185] + gpt_neox_layers_11_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[186] + gpt_neox_layers_11_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[187] + gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[188] + gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[189] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[190] + gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[191] + gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[192] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[193] + gpt_neox_layers_12_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[194] + gpt_neox_layers_12_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[195] + gpt_neox_layers_12_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[196] + gpt_neox_layers_12_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[197] + gpt_neox_layers_12_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[198] + gpt_neox_layers_12_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[199] + gpt_neox_layers_12_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[200] + gpt_neox_layers_12_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[201] + gpt_neox_layers_12_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[202] + gpt_neox_layers_12_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[203] + gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[204] + gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[205] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[206] + gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[207] + gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[208] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[209] + gpt_neox_layers_13_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[210] + gpt_neox_layers_13_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[211] + gpt_neox_layers_13_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[212] + gpt_neox_layers_13_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[213] + gpt_neox_layers_13_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[214] + gpt_neox_layers_13_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[215] + gpt_neox_layers_13_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[216] + gpt_neox_layers_13_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[217] + gpt_neox_layers_13_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[218] + gpt_neox_layers_13_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[219] + gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[220] + gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[221] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[222] + gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[223] + gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[224] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[225] + gpt_neox_layers_14_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[226] + gpt_neox_layers_14_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[227] + gpt_neox_layers_14_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[228] + gpt_neox_layers_14_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[229] + gpt_neox_layers_14_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[230] + gpt_neox_layers_14_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[231] + gpt_neox_layers_14_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[232] + gpt_neox_layers_14_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[233] + gpt_neox_layers_14_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[234] + gpt_neox_layers_14_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[235] + gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[236] + gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[237] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[238] + gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[239] + gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[240] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[241] + gpt_neox_layers_15_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[242] + gpt_neox_layers_15_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[243] + gpt_neox_layers_15_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[244] + gpt_neox_layers_15_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[245] + gpt_neox_layers_15_attention_query_key_value_q_weight4: R.Tensor((6144, 256), dtype="uint32") = packed_params[246] + gpt_neox_layers_15_attention_query_key_value_q_scale4: R.Tensor((6144, 64), dtype="float16") = packed_params[247] + gpt_neox_layers_15_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[248] + gpt_neox_layers_15_attention_dense_q_weight4: R.Tensor((2048, 256), dtype="uint32") = packed_params[249] + gpt_neox_layers_15_attention_dense_q_scale4: R.Tensor((2048, 64), dtype="float16") = packed_params[250] + gpt_neox_layers_15_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[251] + gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight4: R.Tensor((8192, 256), dtype="uint32") = packed_params[252] + gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale4: R.Tensor((8192, 64), dtype="float16") = packed_params[253] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[254] + gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight4: R.Tensor((2048, 1024), dtype="uint32") = packed_params[255] + gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale4: R.Tensor((2048, 256), dtype="float16") = packed_params[256] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[257] + gpt_neox_final_layer_norm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[258] + gpt_neox_final_layer_norm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[259] + embed_out_q_weight4: R.Tensor((vocab_size, 256), dtype="uint32") = packed_params[260] + embed_out_q_scale4: R.Tensor((vocab_size, 64), dtype="float16") = packed_params[261] + layer_norm99 = 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, 2048), 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_norm99, gpt_neox_layers_0_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape192 = R.call_tir(cls.reshape, (lv,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape193 = R.call_tir(cls.reshape1, (reshape192,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv246 = 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((batch_size, 8, 256), dtype="float16")) + reshape194 = R.call_tir(cls.reshape2, (lv246,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape195 = R.call_tir(cls.reshape3, (reshape194,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm100 = 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, 2048), 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_norm100, gpt_neox_layers_0_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape195, gpt_neox_layers_0_attention_dense_bias4, lv2, input_embeds), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm101 = 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, 2048), 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_norm101, gpt_neox_layers_1_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape196 = R.call_tir(cls.reshape, (lv3,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape197 = R.call_tir(cls.reshape1, (reshape196,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv251 = 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((batch_size, 8, 256), dtype="float16")) + reshape198 = R.call_tir(cls.reshape2, (lv251,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape199 = R.call_tir(cls.reshape3, (reshape198,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm102 = 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, 2048), 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_norm102, gpt_neox_layers_1_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape199, gpt_neox_layers_1_attention_dense_bias4, lv5, lv_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm103 = 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, 2048), 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_norm103, gpt_neox_layers_2_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape200 = R.call_tir(cls.reshape, (lv6,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape201 = R.call_tir(cls.reshape1, (reshape200,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv256 = 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((batch_size, 8, 256), dtype="float16")) + reshape202 = R.call_tir(cls.reshape2, (lv256,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape203 = R.call_tir(cls.reshape3, (reshape202,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm104 = 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, 2048), 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_norm104, gpt_neox_layers_2_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape203, gpt_neox_layers_2_attention_dense_bias4, lv8, lv1_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm105 = 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, 2048), 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_norm105, gpt_neox_layers_3_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape204 = R.call_tir(cls.reshape, (lv9,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape205 = R.call_tir(cls.reshape1, (reshape204,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv261 = 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((batch_size, 8, 256), dtype="float16")) + reshape206 = R.call_tir(cls.reshape2, (lv261,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape207 = R.call_tir(cls.reshape3, (reshape206,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm106 = 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, 2048), 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_norm106, gpt_neox_layers_3_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape207, gpt_neox_layers_3_attention_dense_bias4, lv11, lv2_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm107 = 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, 2048), 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_norm107, gpt_neox_layers_4_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape208 = R.call_tir(cls.reshape, (lv12,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape209 = R.call_tir(cls.reshape1, (reshape208,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv266 = 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((batch_size, 8, 256), dtype="float16")) + reshape210 = R.call_tir(cls.reshape2, (lv266,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape211 = R.call_tir(cls.reshape3, (reshape210,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm108 = 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, 2048), 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_norm108, gpt_neox_layers_4_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape211, gpt_neox_layers_4_attention_dense_bias4, lv14, lv3_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm109 = 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, 2048), 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_norm109, gpt_neox_layers_5_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape212 = R.call_tir(cls.reshape, (lv15,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape213 = R.call_tir(cls.reshape1, (reshape212,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv271 = 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((batch_size, 8, 256), dtype="float16")) + reshape214 = R.call_tir(cls.reshape2, (lv271,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape215 = R.call_tir(cls.reshape3, (reshape214,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm110 = 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, 2048), 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_norm110, gpt_neox_layers_5_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape215, gpt_neox_layers_5_attention_dense_bias4, lv17, lv4_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm111 = 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, 2048), 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_norm111, gpt_neox_layers_6_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape216 = R.call_tir(cls.reshape, (lv18,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape217 = R.call_tir(cls.reshape1, (reshape216,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv276 = 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((batch_size, 8, 256), dtype="float16")) + reshape218 = R.call_tir(cls.reshape2, (lv276,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape219 = R.call_tir(cls.reshape3, (reshape218,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm112 = 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, 2048), 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_norm112, gpt_neox_layers_6_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape219, gpt_neox_layers_6_attention_dense_bias4, lv20, lv5_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm113 = 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, 2048), 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_norm113, gpt_neox_layers_7_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape220 = R.call_tir(cls.reshape, (lv21,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape221 = R.call_tir(cls.reshape1, (reshape220,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv281 = 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((batch_size, 8, 256), dtype="float16")) + reshape222 = R.call_tir(cls.reshape2, (lv281,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape223 = R.call_tir(cls.reshape3, (reshape222,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm114 = 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, 2048), 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_norm114, gpt_neox_layers_7_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape223, gpt_neox_layers_7_attention_dense_bias4, lv23, lv6_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm115 = 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, 2048), 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_norm115, gpt_neox_layers_8_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape224 = R.call_tir(cls.reshape, (lv24,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape225 = R.call_tir(cls.reshape1, (reshape224,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv286 = 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((batch_size, 8, 256), dtype="float16")) + reshape226 = R.call_tir(cls.reshape2, (lv286,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape227 = R.call_tir(cls.reshape3, (reshape226,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm116 = 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, 2048), 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_norm116, gpt_neox_layers_8_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape227, gpt_neox_layers_8_attention_dense_bias4, lv26, lv7_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm117 = 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, 2048), 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_norm117, gpt_neox_layers_9_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape228 = R.call_tir(cls.reshape, (lv27,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape229 = R.call_tir(cls.reshape1, (reshape228,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv291 = 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((batch_size, 8, 256), dtype="float16")) + reshape230 = R.call_tir(cls.reshape2, (lv291,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape231 = R.call_tir(cls.reshape3, (reshape230,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm118 = 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, 2048), 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_norm118, gpt_neox_layers_9_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape231, gpt_neox_layers_9_attention_dense_bias4, lv29, lv8_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm119 = 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, 2048), 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_norm119, gpt_neox_layers_10_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape232 = R.call_tir(cls.reshape, (lv30,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape233 = R.call_tir(cls.reshape1, (reshape232,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv296 = 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((batch_size, 8, 256), dtype="float16")) + reshape234 = R.call_tir(cls.reshape2, (lv296,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape235 = R.call_tir(cls.reshape3, (reshape234,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm120 = 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, 2048), 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_norm120, gpt_neox_layers_10_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape235, gpt_neox_layers_10_attention_dense_bias4, lv32, lv9_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm121 = 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, 2048), 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_norm121, gpt_neox_layers_11_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape236 = R.call_tir(cls.reshape, (lv33,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape237 = R.call_tir(cls.reshape1, (reshape236,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv301 = 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((batch_size, 8, 256), dtype="float16")) + reshape238 = R.call_tir(cls.reshape2, (lv301,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape239 = R.call_tir(cls.reshape3, (reshape238,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm122 = 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, 2048), 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_norm122, gpt_neox_layers_11_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape239, gpt_neox_layers_11_attention_dense_bias4, lv35, lv10_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm123 = 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, 2048), 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_norm123, gpt_neox_layers_12_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape240 = R.call_tir(cls.reshape, (lv36,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape241 = R.call_tir(cls.reshape1, (reshape240,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv306 = 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((batch_size, 8, 256), dtype="float16")) + reshape242 = R.call_tir(cls.reshape2, (lv306,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape243 = R.call_tir(cls.reshape3, (reshape242,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm124 = 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, 2048), 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_norm124, gpt_neox_layers_12_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape243, gpt_neox_layers_12_attention_dense_bias4, lv38, lv11_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm125 = 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, 2048), 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_norm125, gpt_neox_layers_13_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape244 = R.call_tir(cls.reshape, (lv39,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape245 = R.call_tir(cls.reshape1, (reshape244,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv311 = 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((batch_size, 8, 256), dtype="float16")) + reshape246 = R.call_tir(cls.reshape2, (lv311,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape247 = R.call_tir(cls.reshape3, (reshape246,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm126 = 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, 2048), 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_norm126, gpt_neox_layers_13_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape247, gpt_neox_layers_13_attention_dense_bias4, lv41, lv12_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm127 = 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, 2048), 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_norm127, gpt_neox_layers_14_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape248 = R.call_tir(cls.reshape, (lv42,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape249 = R.call_tir(cls.reshape1, (reshape248,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv316 = 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((batch_size, 8, 256), dtype="float16")) + reshape250 = R.call_tir(cls.reshape2, (lv316,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape251 = R.call_tir(cls.reshape3, (reshape250,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm128 = 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, 2048), 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_norm128, gpt_neox_layers_14_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape251, gpt_neox_layers_14_attention_dense_bias4, lv44, lv13_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm129 = 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, 2048), 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_norm129, gpt_neox_layers_15_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + reshape252 = R.call_tir(cls.reshape, (lv45,), out_sinfo=R.Tensor((batch_size, 1, 24, 256), dtype="float16")) + reshape253 = R.call_tir(cls.reshape1, (reshape252,), out_sinfo=R.Tensor((batch_size, 24, 256), dtype="float16")) + lv321 = 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((batch_size, 8, 256), dtype="float16")) + reshape254 = R.call_tir(cls.reshape2, (lv321,), out_sinfo=R.Tensor((batch_size, 1, 8, 256), dtype="float16")) + reshape255 = R.call_tir(cls.reshape3, (reshape254,), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm130 = 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, 2048), 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_norm130, gpt_neox_layers_15_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 8192), 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, 2048), 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, reshape255, gpt_neox_layers_15_attention_dense_bias4, lv47, lv14_1), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + layer_norm131 = R.call_tir(cls.layer_norm, (lv15_1, gpt_neox_final_layer_norm_weight4, gpt_neox_final_layer_norm_bias4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + lv_2 = R.call_tir(cls.fused_dequantize_fused_NT_matmul4_cast2, (embed_out_q_weight4, embed_out_q_scale4, layer_norm131), 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", 2048), dtype="float16"), logit_positions: R.Tensor(("batch_size",), dtype="int32"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 256), dtype="uint32"), R.Tensor(("vocab_size", 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 256), dtype="uint32"), R.Tensor(("vocab_size", 64), 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((2048,), dtype="float16") = packed_params[2] + gpt_neox_layers_0_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[4] + gpt_neox_layers_0_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[6] + gpt_neox_layers_0_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[8] + gpt_neox_layers_0_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[9] + gpt_neox_layers_0_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[10] + gpt_neox_layers_0_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[12] + gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[13] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[14] + gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[15] + gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[16] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[17] + gpt_neox_layers_1_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[18] + gpt_neox_layers_1_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[19] + gpt_neox_layers_1_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[20] + gpt_neox_layers_1_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[21] + gpt_neox_layers_1_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[22] + gpt_neox_layers_1_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[23] + gpt_neox_layers_1_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[24] + gpt_neox_layers_1_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[25] + gpt_neox_layers_1_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[26] + gpt_neox_layers_1_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[27] + gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[28] + gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[29] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[30] + gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[31] + gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[32] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[33] + gpt_neox_layers_2_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[34] + gpt_neox_layers_2_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[35] + gpt_neox_layers_2_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[36] + gpt_neox_layers_2_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[37] + gpt_neox_layers_2_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[38] + gpt_neox_layers_2_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[39] + gpt_neox_layers_2_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[40] + gpt_neox_layers_2_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[41] + gpt_neox_layers_2_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[42] + gpt_neox_layers_2_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[43] + gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[44] + gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[45] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[46] + gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[47] + gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[48] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[49] + gpt_neox_layers_3_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[50] + gpt_neox_layers_3_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[51] + gpt_neox_layers_3_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[52] + gpt_neox_layers_3_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[53] + gpt_neox_layers_3_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[54] + gpt_neox_layers_3_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[55] + gpt_neox_layers_3_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[56] + gpt_neox_layers_3_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[57] + gpt_neox_layers_3_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[58] + gpt_neox_layers_3_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[59] + gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[60] + gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[61] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[62] + gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[63] + gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[64] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[65] + gpt_neox_layers_4_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[66] + gpt_neox_layers_4_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[67] + gpt_neox_layers_4_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[68] + gpt_neox_layers_4_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[69] + gpt_neox_layers_4_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[70] + gpt_neox_layers_4_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[71] + gpt_neox_layers_4_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[72] + gpt_neox_layers_4_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[73] + gpt_neox_layers_4_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[74] + gpt_neox_layers_4_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[75] + gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[76] + gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[77] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[78] + gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[79] + gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[80] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[81] + gpt_neox_layers_5_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[82] + gpt_neox_layers_5_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[83] + gpt_neox_layers_5_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[84] + gpt_neox_layers_5_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[85] + gpt_neox_layers_5_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[86] + gpt_neox_layers_5_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[87] + gpt_neox_layers_5_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[88] + gpt_neox_layers_5_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[89] + gpt_neox_layers_5_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[90] + gpt_neox_layers_5_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[91] + gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[92] + gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[93] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[94] + gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[95] + gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[96] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[97] + gpt_neox_layers_6_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[98] + gpt_neox_layers_6_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[99] + gpt_neox_layers_6_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[100] + gpt_neox_layers_6_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[101] + gpt_neox_layers_6_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[102] + gpt_neox_layers_6_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[103] + gpt_neox_layers_6_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[104] + gpt_neox_layers_6_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[105] + gpt_neox_layers_6_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[106] + gpt_neox_layers_6_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[107] + gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[108] + gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[109] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[110] + gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[111] + gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[112] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[113] + gpt_neox_layers_7_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[114] + gpt_neox_layers_7_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[115] + gpt_neox_layers_7_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[116] + gpt_neox_layers_7_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[117] + gpt_neox_layers_7_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[118] + gpt_neox_layers_7_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[119] + gpt_neox_layers_7_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[120] + gpt_neox_layers_7_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[121] + gpt_neox_layers_7_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[122] + gpt_neox_layers_7_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[123] + gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[124] + gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[125] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[126] + gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[127] + gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[128] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[129] + gpt_neox_layers_8_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[130] + gpt_neox_layers_8_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[131] + gpt_neox_layers_8_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[132] + gpt_neox_layers_8_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[133] + gpt_neox_layers_8_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[134] + gpt_neox_layers_8_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[135] + gpt_neox_layers_8_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[136] + gpt_neox_layers_8_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[137] + gpt_neox_layers_8_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[138] + gpt_neox_layers_8_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[139] + gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[140] + gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[141] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[142] + gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[143] + gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[144] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[145] + gpt_neox_layers_9_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[146] + gpt_neox_layers_9_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[147] + gpt_neox_layers_9_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[148] + gpt_neox_layers_9_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[149] + gpt_neox_layers_9_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[150] + gpt_neox_layers_9_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[151] + gpt_neox_layers_9_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[152] + gpt_neox_layers_9_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[153] + gpt_neox_layers_9_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[154] + gpt_neox_layers_9_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[155] + gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[156] + gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[157] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[158] + gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[159] + gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[160] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[161] + gpt_neox_layers_10_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[162] + gpt_neox_layers_10_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[163] + gpt_neox_layers_10_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[164] + gpt_neox_layers_10_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[165] + gpt_neox_layers_10_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[166] + gpt_neox_layers_10_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[167] + gpt_neox_layers_10_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[168] + gpt_neox_layers_10_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[169] + gpt_neox_layers_10_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[170] + gpt_neox_layers_10_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[171] + gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[172] + gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[173] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[174] + gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[175] + gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[176] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[177] + gpt_neox_layers_11_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[178] + gpt_neox_layers_11_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[179] + gpt_neox_layers_11_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[180] + gpt_neox_layers_11_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[181] + gpt_neox_layers_11_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[182] + gpt_neox_layers_11_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[183] + gpt_neox_layers_11_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[184] + gpt_neox_layers_11_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[185] + gpt_neox_layers_11_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[186] + gpt_neox_layers_11_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[187] + gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[188] + gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[189] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[190] + gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[191] + gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[192] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[193] + gpt_neox_layers_12_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[194] + gpt_neox_layers_12_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[195] + gpt_neox_layers_12_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[196] + gpt_neox_layers_12_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[197] + gpt_neox_layers_12_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[198] + gpt_neox_layers_12_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[199] + gpt_neox_layers_12_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[200] + gpt_neox_layers_12_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[201] + gpt_neox_layers_12_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[202] + gpt_neox_layers_12_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[203] + gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[204] + gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[205] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[206] + gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[207] + gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[208] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[209] + gpt_neox_layers_13_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[210] + gpt_neox_layers_13_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[211] + gpt_neox_layers_13_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[212] + gpt_neox_layers_13_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[213] + gpt_neox_layers_13_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[214] + gpt_neox_layers_13_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[215] + gpt_neox_layers_13_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[216] + gpt_neox_layers_13_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[217] + gpt_neox_layers_13_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[218] + gpt_neox_layers_13_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[219] + gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[220] + gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[221] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[222] + gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[223] + gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[224] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[225] + gpt_neox_layers_14_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[226] + gpt_neox_layers_14_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[227] + gpt_neox_layers_14_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[228] + gpt_neox_layers_14_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[229] + gpt_neox_layers_14_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[230] + gpt_neox_layers_14_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[231] + gpt_neox_layers_14_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[232] + gpt_neox_layers_14_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[233] + gpt_neox_layers_14_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[234] + gpt_neox_layers_14_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[235] + gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[236] + gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[237] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[238] + gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[239] + gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[240] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[241] + gpt_neox_layers_15_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[242] + gpt_neox_layers_15_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[243] + gpt_neox_layers_15_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[244] + gpt_neox_layers_15_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[245] + gpt_neox_layers_15_attention_query_key_value_q_weight3: R.Tensor((6144, 256), dtype="uint32") = packed_params[246] + gpt_neox_layers_15_attention_query_key_value_q_scale3: R.Tensor((6144, 64), dtype="float16") = packed_params[247] + gpt_neox_layers_15_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[248] + gpt_neox_layers_15_attention_dense_q_weight3: R.Tensor((2048, 256), dtype="uint32") = packed_params[249] + gpt_neox_layers_15_attention_dense_q_scale3: R.Tensor((2048, 64), dtype="float16") = packed_params[250] + gpt_neox_layers_15_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[251] + gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight3: R.Tensor((8192, 256), dtype="uint32") = packed_params[252] + gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale3: R.Tensor((8192, 64), dtype="float16") = packed_params[253] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[254] + gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight3: R.Tensor((2048, 1024), dtype="uint32") = packed_params[255] + gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale3: R.Tensor((2048, 256), dtype="float16") = packed_params[256] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[257] + gpt_neox_final_layer_norm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[258] + gpt_neox_final_layer_norm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[259] + embed_out_q_weight3: R.Tensor((vocab_size, 256), dtype="uint32") = packed_params[260] + embed_out_q_scale3: R.Tensor((vocab_size, 64), dtype="float16") = packed_params[261] + layer_norm66 = 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, 2048), dtype="float16")) + lv48 = 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_norm66, gpt_neox_layers_0_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape128 = R.call_tir(cls.reshape4, (lv48,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape129 = R.call_tir(cls.reshape5, (reshape128,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv165 = 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)), reshape129), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape130 = R.call_tir(cls.reshape6, (lv165,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape131 = R.call_tir(cls.reshape7, (reshape130,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm67 = 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, 2048), dtype="float16")) + lv49 = 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_norm67, gpt_neox_layers_0_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv50 = 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, lv49, gpt_neox_layers_0_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv16 = 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, reshape131, gpt_neox_layers_0_attention_dense_bias3, lv50, input_embeds), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm68 = R.call_tir(cls.layer_norm1, (lv16, gpt_neox_layers_1_input_layernorm_weight3, gpt_neox_layers_1_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv51 = 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_norm68, gpt_neox_layers_1_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape132 = R.call_tir(cls.reshape4, (lv51,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape133 = R.call_tir(cls.reshape5, (reshape132,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv170 = 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)), reshape133), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape134 = R.call_tir(cls.reshape6, (lv170,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape135 = R.call_tir(cls.reshape7, (reshape134,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm69 = R.call_tir(cls.layer_norm1, (lv16, gpt_neox_layers_1_post_attention_layernorm_weight3, gpt_neox_layers_1_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv52 = 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_norm69, gpt_neox_layers_1_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv53 = 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, lv52, gpt_neox_layers_1_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv17 = 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, reshape135, gpt_neox_layers_1_attention_dense_bias3, lv53, lv16), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm70 = R.call_tir(cls.layer_norm1, (lv17, gpt_neox_layers_2_input_layernorm_weight3, gpt_neox_layers_2_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv54 = 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_norm70, gpt_neox_layers_2_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape136 = R.call_tir(cls.reshape4, (lv54,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape137 = R.call_tir(cls.reshape5, (reshape136,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv175 = 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)), reshape137), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape138 = R.call_tir(cls.reshape6, (lv175,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape139 = R.call_tir(cls.reshape7, (reshape138,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm71 = R.call_tir(cls.layer_norm1, (lv17, gpt_neox_layers_2_post_attention_layernorm_weight3, gpt_neox_layers_2_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv55 = 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_norm71, gpt_neox_layers_2_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv56 = 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, lv55, gpt_neox_layers_2_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv18 = 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, reshape139, gpt_neox_layers_2_attention_dense_bias3, lv56, lv17), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm72 = R.call_tir(cls.layer_norm1, (lv18, gpt_neox_layers_3_input_layernorm_weight3, gpt_neox_layers_3_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv57 = 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_norm72, gpt_neox_layers_3_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape140 = R.call_tir(cls.reshape4, (lv57,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape141 = R.call_tir(cls.reshape5, (reshape140,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv180 = 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)), reshape141), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape142 = R.call_tir(cls.reshape6, (lv180,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape143 = R.call_tir(cls.reshape7, (reshape142,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm73 = R.call_tir(cls.layer_norm1, (lv18, gpt_neox_layers_3_post_attention_layernorm_weight3, gpt_neox_layers_3_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv58 = 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_norm73, gpt_neox_layers_3_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv59 = 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, lv58, gpt_neox_layers_3_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv19 = 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, reshape143, gpt_neox_layers_3_attention_dense_bias3, lv59, lv18), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm74 = R.call_tir(cls.layer_norm1, (lv19, gpt_neox_layers_4_input_layernorm_weight3, gpt_neox_layers_4_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv60 = 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_norm74, gpt_neox_layers_4_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape144 = R.call_tir(cls.reshape4, (lv60,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape145 = R.call_tir(cls.reshape5, (reshape144,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv185 = 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)), reshape145), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape146 = R.call_tir(cls.reshape6, (lv185,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape147 = R.call_tir(cls.reshape7, (reshape146,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm75 = R.call_tir(cls.layer_norm1, (lv19, gpt_neox_layers_4_post_attention_layernorm_weight3, gpt_neox_layers_4_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv61 = 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_norm75, gpt_neox_layers_4_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv62 = 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, lv61, gpt_neox_layers_4_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv20 = 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, reshape147, gpt_neox_layers_4_attention_dense_bias3, lv62, lv19), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm76 = R.call_tir(cls.layer_norm1, (lv20, gpt_neox_layers_5_input_layernorm_weight3, gpt_neox_layers_5_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv63 = 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_norm76, gpt_neox_layers_5_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape148 = R.call_tir(cls.reshape4, (lv63,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape149 = R.call_tir(cls.reshape5, (reshape148,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv190 = 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)), reshape149), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape150 = R.call_tir(cls.reshape6, (lv190,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape151 = R.call_tir(cls.reshape7, (reshape150,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm77 = R.call_tir(cls.layer_norm1, (lv20, gpt_neox_layers_5_post_attention_layernorm_weight3, gpt_neox_layers_5_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv64 = 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_norm77, gpt_neox_layers_5_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv65 = 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, lv64, gpt_neox_layers_5_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv21 = 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, reshape151, gpt_neox_layers_5_attention_dense_bias3, lv65, lv20), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm78 = R.call_tir(cls.layer_norm1, (lv21, gpt_neox_layers_6_input_layernorm_weight3, gpt_neox_layers_6_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv66 = 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_norm78, gpt_neox_layers_6_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape152 = R.call_tir(cls.reshape4, (lv66,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape153 = R.call_tir(cls.reshape5, (reshape152,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv195 = 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)), reshape153), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape154 = R.call_tir(cls.reshape6, (lv195,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape155 = R.call_tir(cls.reshape7, (reshape154,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm79 = R.call_tir(cls.layer_norm1, (lv21, gpt_neox_layers_6_post_attention_layernorm_weight3, gpt_neox_layers_6_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv67 = 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_norm79, gpt_neox_layers_6_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv68 = 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, lv67, gpt_neox_layers_6_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv22 = 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, reshape155, gpt_neox_layers_6_attention_dense_bias3, lv68, lv21), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm80 = R.call_tir(cls.layer_norm1, (lv22, gpt_neox_layers_7_input_layernorm_weight3, gpt_neox_layers_7_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv69 = 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_norm80, gpt_neox_layers_7_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape156 = R.call_tir(cls.reshape4, (lv69,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape157 = R.call_tir(cls.reshape5, (reshape156,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv200 = 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)), reshape157), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape158 = R.call_tir(cls.reshape6, (lv200,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape159 = R.call_tir(cls.reshape7, (reshape158,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm81 = R.call_tir(cls.layer_norm1, (lv22, gpt_neox_layers_7_post_attention_layernorm_weight3, gpt_neox_layers_7_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv70 = 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_norm81, gpt_neox_layers_7_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv71 = 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, lv70, gpt_neox_layers_7_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv23 = 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, reshape159, gpt_neox_layers_7_attention_dense_bias3, lv71, lv22), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm82 = R.call_tir(cls.layer_norm1, (lv23, gpt_neox_layers_8_input_layernorm_weight3, gpt_neox_layers_8_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv72 = 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_norm82, gpt_neox_layers_8_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape160 = R.call_tir(cls.reshape4, (lv72,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape161 = R.call_tir(cls.reshape5, (reshape160,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv205 = 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)), reshape161), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape162 = R.call_tir(cls.reshape6, (lv205,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape163 = R.call_tir(cls.reshape7, (reshape162,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm83 = R.call_tir(cls.layer_norm1, (lv23, gpt_neox_layers_8_post_attention_layernorm_weight3, gpt_neox_layers_8_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv73 = 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_norm83, gpt_neox_layers_8_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv74 = 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, lv73, gpt_neox_layers_8_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv24 = 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, reshape163, gpt_neox_layers_8_attention_dense_bias3, lv74, lv23), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm84 = R.call_tir(cls.layer_norm1, (lv24, gpt_neox_layers_9_input_layernorm_weight3, gpt_neox_layers_9_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv75 = 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_norm84, gpt_neox_layers_9_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape164 = R.call_tir(cls.reshape4, (lv75,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape165 = R.call_tir(cls.reshape5, (reshape164,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv210 = 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)), reshape165), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape166 = R.call_tir(cls.reshape6, (lv210,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape167 = R.call_tir(cls.reshape7, (reshape166,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm85 = R.call_tir(cls.layer_norm1, (lv24, gpt_neox_layers_9_post_attention_layernorm_weight3, gpt_neox_layers_9_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv76 = 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_norm85, gpt_neox_layers_9_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv77 = 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, lv76, gpt_neox_layers_9_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv25 = 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, reshape167, gpt_neox_layers_9_attention_dense_bias3, lv77, lv24), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm86 = R.call_tir(cls.layer_norm1, (lv25, gpt_neox_layers_10_input_layernorm_weight3, gpt_neox_layers_10_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv78 = 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_norm86, gpt_neox_layers_10_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape168 = R.call_tir(cls.reshape4, (lv78,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape169 = R.call_tir(cls.reshape5, (reshape168,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv215 = 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)), reshape169), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape170 = R.call_tir(cls.reshape6, (lv215,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape171 = R.call_tir(cls.reshape7, (reshape170,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm87 = R.call_tir(cls.layer_norm1, (lv25, gpt_neox_layers_10_post_attention_layernorm_weight3, gpt_neox_layers_10_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv79 = 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_norm87, gpt_neox_layers_10_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv80 = 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, lv79, gpt_neox_layers_10_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv26 = 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, reshape171, gpt_neox_layers_10_attention_dense_bias3, lv80, lv25), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm88 = R.call_tir(cls.layer_norm1, (lv26, gpt_neox_layers_11_input_layernorm_weight3, gpt_neox_layers_11_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv81 = 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_norm88, gpt_neox_layers_11_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape172 = R.call_tir(cls.reshape4, (lv81,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape173 = R.call_tir(cls.reshape5, (reshape172,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv220 = 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)), reshape173), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape174 = R.call_tir(cls.reshape6, (lv220,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape175 = R.call_tir(cls.reshape7, (reshape174,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm89 = R.call_tir(cls.layer_norm1, (lv26, gpt_neox_layers_11_post_attention_layernorm_weight3, gpt_neox_layers_11_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv82 = 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_norm89, gpt_neox_layers_11_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv83 = 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, lv82, gpt_neox_layers_11_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv27 = 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, reshape175, gpt_neox_layers_11_attention_dense_bias3, lv83, lv26), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm90 = R.call_tir(cls.layer_norm1, (lv27, gpt_neox_layers_12_input_layernorm_weight3, gpt_neox_layers_12_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv84 = 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_norm90, gpt_neox_layers_12_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape176 = R.call_tir(cls.reshape4, (lv84,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape177 = R.call_tir(cls.reshape5, (reshape176,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv225 = 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)), reshape177), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape178 = R.call_tir(cls.reshape6, (lv225,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape179 = R.call_tir(cls.reshape7, (reshape178,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm91 = R.call_tir(cls.layer_norm1, (lv27, gpt_neox_layers_12_post_attention_layernorm_weight3, gpt_neox_layers_12_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv85 = 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_norm91, gpt_neox_layers_12_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv86 = 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, lv85, gpt_neox_layers_12_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv28 = 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, reshape179, gpt_neox_layers_12_attention_dense_bias3, lv86, lv27), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm92 = R.call_tir(cls.layer_norm1, (lv28, gpt_neox_layers_13_input_layernorm_weight3, gpt_neox_layers_13_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv87 = 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_norm92, gpt_neox_layers_13_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape180 = R.call_tir(cls.reshape4, (lv87,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape181 = R.call_tir(cls.reshape5, (reshape180,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv230 = 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)), reshape181), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape182 = R.call_tir(cls.reshape6, (lv230,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape183 = R.call_tir(cls.reshape7, (reshape182,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm93 = R.call_tir(cls.layer_norm1, (lv28, gpt_neox_layers_13_post_attention_layernorm_weight3, gpt_neox_layers_13_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv88 = 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_norm93, gpt_neox_layers_13_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv89 = 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, lv88, gpt_neox_layers_13_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv29 = 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, reshape183, gpt_neox_layers_13_attention_dense_bias3, lv89, lv28), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm94 = R.call_tir(cls.layer_norm1, (lv29, gpt_neox_layers_14_input_layernorm_weight3, gpt_neox_layers_14_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv90 = 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_norm94, gpt_neox_layers_14_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape184 = R.call_tir(cls.reshape4, (lv90,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape185 = R.call_tir(cls.reshape5, (reshape184,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv235 = 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)), reshape185), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape186 = R.call_tir(cls.reshape6, (lv235,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape187 = R.call_tir(cls.reshape7, (reshape186,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm95 = R.call_tir(cls.layer_norm1, (lv29, gpt_neox_layers_14_post_attention_layernorm_weight3, gpt_neox_layers_14_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv91 = 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_norm95, gpt_neox_layers_14_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv92 = 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, lv91, gpt_neox_layers_14_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv30 = 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, reshape187, gpt_neox_layers_14_attention_dense_bias3, lv92, lv29), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm96 = R.call_tir(cls.layer_norm1, (lv30, gpt_neox_layers_15_input_layernorm_weight3, gpt_neox_layers_15_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv93 = 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_norm96, gpt_neox_layers_15_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape188 = R.call_tir(cls.reshape4, (lv93,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape189 = R.call_tir(cls.reshape5, (reshape188,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv240 = 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)), reshape189), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape190 = R.call_tir(cls.reshape6, (lv240,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape191 = R.call_tir(cls.reshape7, (reshape190,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm97 = R.call_tir(cls.layer_norm1, (lv30, gpt_neox_layers_15_post_attention_layernorm_weight3, gpt_neox_layers_15_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv94 = 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_norm97, gpt_neox_layers_15_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv95 = 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, lv94, gpt_neox_layers_15_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv31 = 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, reshape191, gpt_neox_layers_15_attention_dense_bias3, lv95, lv30), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm98 = R.call_tir(cls.layer_norm1, (lv31, gpt_neox_final_layer_norm_weight3, gpt_neox_final_layer_norm_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + take1 = R.call_tir(cls.take, (layer_norm98, logit_positions), out_sinfo=R.Tensor((1, batch_size, 2048), 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", 2048), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 256), dtype="uint32"), R.Tensor(("vocab_size", 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 256), dtype="uint32"), R.Tensor(("vocab_size", 64), 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((2048,), dtype="float16") = packed_params[2] + gpt_neox_layers_0_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[4] + gpt_neox_layers_0_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[6] + gpt_neox_layers_0_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[8] + gpt_neox_layers_0_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[9] + gpt_neox_layers_0_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[10] + gpt_neox_layers_0_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[12] + gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[13] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[14] + gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[15] + gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[16] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[17] + gpt_neox_layers_1_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[18] + gpt_neox_layers_1_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[19] + gpt_neox_layers_1_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[20] + gpt_neox_layers_1_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[21] + gpt_neox_layers_1_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[22] + gpt_neox_layers_1_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[23] + gpt_neox_layers_1_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[24] + gpt_neox_layers_1_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[25] + gpt_neox_layers_1_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[26] + gpt_neox_layers_1_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[27] + gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[28] + gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[29] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[30] + gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[31] + gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[32] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[33] + gpt_neox_layers_2_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[34] + gpt_neox_layers_2_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[35] + gpt_neox_layers_2_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[36] + gpt_neox_layers_2_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[37] + gpt_neox_layers_2_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[38] + gpt_neox_layers_2_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[39] + gpt_neox_layers_2_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[40] + gpt_neox_layers_2_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[41] + gpt_neox_layers_2_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[42] + gpt_neox_layers_2_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[43] + gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[44] + gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[45] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[46] + gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[47] + gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[48] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[49] + gpt_neox_layers_3_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[50] + gpt_neox_layers_3_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[51] + gpt_neox_layers_3_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[52] + gpt_neox_layers_3_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[53] + gpt_neox_layers_3_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[54] + gpt_neox_layers_3_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[55] + gpt_neox_layers_3_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[56] + gpt_neox_layers_3_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[57] + gpt_neox_layers_3_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[58] + gpt_neox_layers_3_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[59] + gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[60] + gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[61] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[62] + gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[63] + gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[64] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[65] + gpt_neox_layers_4_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[66] + gpt_neox_layers_4_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[67] + gpt_neox_layers_4_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[68] + gpt_neox_layers_4_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[69] + gpt_neox_layers_4_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[70] + gpt_neox_layers_4_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[71] + gpt_neox_layers_4_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[72] + gpt_neox_layers_4_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[73] + gpt_neox_layers_4_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[74] + gpt_neox_layers_4_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[75] + gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[76] + gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[77] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[78] + gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[79] + gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[80] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[81] + gpt_neox_layers_5_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[82] + gpt_neox_layers_5_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[83] + gpt_neox_layers_5_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[84] + gpt_neox_layers_5_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[85] + gpt_neox_layers_5_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[86] + gpt_neox_layers_5_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[87] + gpt_neox_layers_5_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[88] + gpt_neox_layers_5_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[89] + gpt_neox_layers_5_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[90] + gpt_neox_layers_5_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[91] + gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[92] + gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[93] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[94] + gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[95] + gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[96] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[97] + gpt_neox_layers_6_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[98] + gpt_neox_layers_6_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[99] + gpt_neox_layers_6_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[100] + gpt_neox_layers_6_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[101] + gpt_neox_layers_6_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[102] + gpt_neox_layers_6_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[103] + gpt_neox_layers_6_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[104] + gpt_neox_layers_6_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[105] + gpt_neox_layers_6_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[106] + gpt_neox_layers_6_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[107] + gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[108] + gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[109] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[110] + gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[111] + gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[112] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[113] + gpt_neox_layers_7_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[114] + gpt_neox_layers_7_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[115] + gpt_neox_layers_7_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[116] + gpt_neox_layers_7_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[117] + gpt_neox_layers_7_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[118] + gpt_neox_layers_7_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[119] + gpt_neox_layers_7_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[120] + gpt_neox_layers_7_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[121] + gpt_neox_layers_7_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[122] + gpt_neox_layers_7_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[123] + gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[124] + gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[125] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[126] + gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[127] + gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[128] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[129] + gpt_neox_layers_8_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[130] + gpt_neox_layers_8_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[131] + gpt_neox_layers_8_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[132] + gpt_neox_layers_8_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[133] + gpt_neox_layers_8_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[134] + gpt_neox_layers_8_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[135] + gpt_neox_layers_8_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[136] + gpt_neox_layers_8_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[137] + gpt_neox_layers_8_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[138] + gpt_neox_layers_8_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[139] + gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[140] + gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[141] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[142] + gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[143] + gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[144] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[145] + gpt_neox_layers_9_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[146] + gpt_neox_layers_9_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[147] + gpt_neox_layers_9_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[148] + gpt_neox_layers_9_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[149] + gpt_neox_layers_9_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[150] + gpt_neox_layers_9_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[151] + gpt_neox_layers_9_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[152] + gpt_neox_layers_9_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[153] + gpt_neox_layers_9_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[154] + gpt_neox_layers_9_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[155] + gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[156] + gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[157] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[158] + gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[159] + gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[160] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[161] + gpt_neox_layers_10_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[162] + gpt_neox_layers_10_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[163] + gpt_neox_layers_10_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[164] + gpt_neox_layers_10_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[165] + gpt_neox_layers_10_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[166] + gpt_neox_layers_10_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[167] + gpt_neox_layers_10_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[168] + gpt_neox_layers_10_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[169] + gpt_neox_layers_10_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[170] + gpt_neox_layers_10_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[171] + gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[172] + gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[173] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[174] + gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[175] + gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[176] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[177] + gpt_neox_layers_11_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[178] + gpt_neox_layers_11_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[179] + gpt_neox_layers_11_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[180] + gpt_neox_layers_11_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[181] + gpt_neox_layers_11_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[182] + gpt_neox_layers_11_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[183] + gpt_neox_layers_11_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[184] + gpt_neox_layers_11_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[185] + gpt_neox_layers_11_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[186] + gpt_neox_layers_11_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[187] + gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[188] + gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[189] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[190] + gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[191] + gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[192] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[193] + gpt_neox_layers_12_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[194] + gpt_neox_layers_12_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[195] + gpt_neox_layers_12_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[196] + gpt_neox_layers_12_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[197] + gpt_neox_layers_12_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[198] + gpt_neox_layers_12_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[199] + gpt_neox_layers_12_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[200] + gpt_neox_layers_12_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[201] + gpt_neox_layers_12_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[202] + gpt_neox_layers_12_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[203] + gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[204] + gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[205] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[206] + gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[207] + gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[208] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[209] + gpt_neox_layers_13_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[210] + gpt_neox_layers_13_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[211] + gpt_neox_layers_13_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[212] + gpt_neox_layers_13_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[213] + gpt_neox_layers_13_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[214] + gpt_neox_layers_13_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[215] + gpt_neox_layers_13_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[216] + gpt_neox_layers_13_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[217] + gpt_neox_layers_13_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[218] + gpt_neox_layers_13_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[219] + gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[220] + gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[221] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[222] + gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[223] + gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[224] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[225] + gpt_neox_layers_14_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[226] + gpt_neox_layers_14_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[227] + gpt_neox_layers_14_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[228] + gpt_neox_layers_14_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[229] + gpt_neox_layers_14_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[230] + gpt_neox_layers_14_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[231] + gpt_neox_layers_14_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[232] + gpt_neox_layers_14_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[233] + gpt_neox_layers_14_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[234] + gpt_neox_layers_14_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[235] + gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[236] + gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[237] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[238] + gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[239] + gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[240] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[241] + gpt_neox_layers_15_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[242] + gpt_neox_layers_15_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[243] + gpt_neox_layers_15_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[244] + gpt_neox_layers_15_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[245] + gpt_neox_layers_15_attention_query_key_value_q_weight5: R.Tensor((6144, 256), dtype="uint32") = packed_params[246] + gpt_neox_layers_15_attention_query_key_value_q_scale5: R.Tensor((6144, 64), dtype="float16") = packed_params[247] + gpt_neox_layers_15_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[248] + gpt_neox_layers_15_attention_dense_q_weight5: R.Tensor((2048, 256), dtype="uint32") = packed_params[249] + gpt_neox_layers_15_attention_dense_q_scale5: R.Tensor((2048, 64), dtype="float16") = packed_params[250] + gpt_neox_layers_15_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[251] + gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight5: R.Tensor((8192, 256), dtype="uint32") = packed_params[252] + gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale5: R.Tensor((8192, 64), dtype="float16") = packed_params[253] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[254] + gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight5: R.Tensor((2048, 1024), dtype="uint32") = packed_params[255] + gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale5: R.Tensor((2048, 256), dtype="float16") = packed_params[256] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[257] + gpt_neox_final_layer_norm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[258] + gpt_neox_final_layer_norm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[259] + embed_out_q_weight5: R.Tensor((vocab_size, 256), dtype="uint32") = packed_params[260] + embed_out_q_scale5: R.Tensor((vocab_size, 64), dtype="float16") = packed_params[261] + layer_norm132 = 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, 2048), dtype="float16")) + lv96 = 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_norm132, gpt_neox_layers_0_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape256 = R.call_tir(cls.reshape4, (lv96,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape257 = R.call_tir(cls.reshape5, (reshape256,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv327 = 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)), reshape257), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape258 = R.call_tir(cls.reshape6, (lv327,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape259 = R.call_tir(cls.reshape7, (reshape258,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm133 = 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, 2048), dtype="float16")) + lv97 = 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_norm133, gpt_neox_layers_0_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv98 = 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, lv97, gpt_neox_layers_0_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv32 = 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, reshape259, gpt_neox_layers_0_attention_dense_bias5, lv98, input_embeds), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm134 = R.call_tir(cls.layer_norm1, (lv32, gpt_neox_layers_1_input_layernorm_weight5, gpt_neox_layers_1_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv99 = 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_norm134, gpt_neox_layers_1_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape260 = R.call_tir(cls.reshape4, (lv99,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape261 = R.call_tir(cls.reshape5, (reshape260,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv332 = 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)), reshape261), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape262 = R.call_tir(cls.reshape6, (lv332,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape263 = R.call_tir(cls.reshape7, (reshape262,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm135 = R.call_tir(cls.layer_norm1, (lv32, gpt_neox_layers_1_post_attention_layernorm_weight5, gpt_neox_layers_1_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv100 = 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_norm135, gpt_neox_layers_1_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv101 = 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, lv100, gpt_neox_layers_1_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv33 = 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, reshape263, gpt_neox_layers_1_attention_dense_bias5, lv101, lv32), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm136 = R.call_tir(cls.layer_norm1, (lv33, gpt_neox_layers_2_input_layernorm_weight5, gpt_neox_layers_2_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv102 = 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_norm136, gpt_neox_layers_2_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape264 = R.call_tir(cls.reshape4, (lv102,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape265 = R.call_tir(cls.reshape5, (reshape264,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv337 = 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)), reshape265), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape266 = R.call_tir(cls.reshape6, (lv337,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape267 = R.call_tir(cls.reshape7, (reshape266,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm137 = R.call_tir(cls.layer_norm1, (lv33, gpt_neox_layers_2_post_attention_layernorm_weight5, gpt_neox_layers_2_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv103 = 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_norm137, gpt_neox_layers_2_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv104 = 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, lv103, gpt_neox_layers_2_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv34 = 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, reshape267, gpt_neox_layers_2_attention_dense_bias5, lv104, lv33), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm138 = R.call_tir(cls.layer_norm1, (lv34, gpt_neox_layers_3_input_layernorm_weight5, gpt_neox_layers_3_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv105 = 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_norm138, gpt_neox_layers_3_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape268 = R.call_tir(cls.reshape4, (lv105,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape269 = R.call_tir(cls.reshape5, (reshape268,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv342 = 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)), reshape269), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape270 = R.call_tir(cls.reshape6, (lv342,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape271 = R.call_tir(cls.reshape7, (reshape270,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm139 = R.call_tir(cls.layer_norm1, (lv34, gpt_neox_layers_3_post_attention_layernorm_weight5, gpt_neox_layers_3_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv106 = 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_norm139, gpt_neox_layers_3_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv107 = 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, lv106, gpt_neox_layers_3_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv35 = 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, reshape271, gpt_neox_layers_3_attention_dense_bias5, lv107, lv34), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm140 = R.call_tir(cls.layer_norm1, (lv35, gpt_neox_layers_4_input_layernorm_weight5, gpt_neox_layers_4_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv108 = 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_norm140, gpt_neox_layers_4_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape272 = R.call_tir(cls.reshape4, (lv108,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape273 = R.call_tir(cls.reshape5, (reshape272,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv347 = 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)), reshape273), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape274 = R.call_tir(cls.reshape6, (lv347,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape275 = R.call_tir(cls.reshape7, (reshape274,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm141 = R.call_tir(cls.layer_norm1, (lv35, gpt_neox_layers_4_post_attention_layernorm_weight5, gpt_neox_layers_4_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv109 = 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_norm141, gpt_neox_layers_4_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv110 = 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, lv109, gpt_neox_layers_4_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv36 = 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, reshape275, gpt_neox_layers_4_attention_dense_bias5, lv110, lv35), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm142 = R.call_tir(cls.layer_norm1, (lv36, gpt_neox_layers_5_input_layernorm_weight5, gpt_neox_layers_5_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv111 = 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_norm142, gpt_neox_layers_5_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape276 = R.call_tir(cls.reshape4, (lv111,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape277 = R.call_tir(cls.reshape5, (reshape276,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv352 = 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)), reshape277), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape278 = R.call_tir(cls.reshape6, (lv352,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape279 = R.call_tir(cls.reshape7, (reshape278,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm143 = R.call_tir(cls.layer_norm1, (lv36, gpt_neox_layers_5_post_attention_layernorm_weight5, gpt_neox_layers_5_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv112 = 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_norm143, gpt_neox_layers_5_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv113 = 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, lv112, gpt_neox_layers_5_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv37 = 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, reshape279, gpt_neox_layers_5_attention_dense_bias5, lv113, lv36), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm144 = R.call_tir(cls.layer_norm1, (lv37, gpt_neox_layers_6_input_layernorm_weight5, gpt_neox_layers_6_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv114 = 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_norm144, gpt_neox_layers_6_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape280 = R.call_tir(cls.reshape4, (lv114,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape281 = R.call_tir(cls.reshape5, (reshape280,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv357 = 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)), reshape281), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape282 = R.call_tir(cls.reshape6, (lv357,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape283 = R.call_tir(cls.reshape7, (reshape282,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm145 = R.call_tir(cls.layer_norm1, (lv37, gpt_neox_layers_6_post_attention_layernorm_weight5, gpt_neox_layers_6_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv115 = 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_norm145, gpt_neox_layers_6_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv116 = 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, lv115, gpt_neox_layers_6_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv38 = 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, reshape283, gpt_neox_layers_6_attention_dense_bias5, lv116, lv37), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm146 = R.call_tir(cls.layer_norm1, (lv38, gpt_neox_layers_7_input_layernorm_weight5, gpt_neox_layers_7_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv117 = 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_norm146, gpt_neox_layers_7_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape284 = R.call_tir(cls.reshape4, (lv117,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape285 = R.call_tir(cls.reshape5, (reshape284,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv362 = 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)), reshape285), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape286 = R.call_tir(cls.reshape6, (lv362,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape287 = R.call_tir(cls.reshape7, (reshape286,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm147 = R.call_tir(cls.layer_norm1, (lv38, gpt_neox_layers_7_post_attention_layernorm_weight5, gpt_neox_layers_7_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv118 = 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_norm147, gpt_neox_layers_7_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv119 = 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, lv118, gpt_neox_layers_7_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv39 = 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, reshape287, gpt_neox_layers_7_attention_dense_bias5, lv119, lv38), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm148 = R.call_tir(cls.layer_norm1, (lv39, gpt_neox_layers_8_input_layernorm_weight5, gpt_neox_layers_8_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv120 = 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_norm148, gpt_neox_layers_8_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape288 = R.call_tir(cls.reshape4, (lv120,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape289 = R.call_tir(cls.reshape5, (reshape288,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv367 = 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)), reshape289), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape290 = R.call_tir(cls.reshape6, (lv367,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape291 = R.call_tir(cls.reshape7, (reshape290,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm149 = R.call_tir(cls.layer_norm1, (lv39, gpt_neox_layers_8_post_attention_layernorm_weight5, gpt_neox_layers_8_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv121 = 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_norm149, gpt_neox_layers_8_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv122 = 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, lv121, gpt_neox_layers_8_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv40 = 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, reshape291, gpt_neox_layers_8_attention_dense_bias5, lv122, lv39), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm150 = R.call_tir(cls.layer_norm1, (lv40, gpt_neox_layers_9_input_layernorm_weight5, gpt_neox_layers_9_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv123 = 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_norm150, gpt_neox_layers_9_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape292 = R.call_tir(cls.reshape4, (lv123,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape293 = R.call_tir(cls.reshape5, (reshape292,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv372 = 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)), reshape293), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape294 = R.call_tir(cls.reshape6, (lv372,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape295 = R.call_tir(cls.reshape7, (reshape294,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm151 = R.call_tir(cls.layer_norm1, (lv40, gpt_neox_layers_9_post_attention_layernorm_weight5, gpt_neox_layers_9_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv124 = 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_norm151, gpt_neox_layers_9_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv125 = 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, lv124, gpt_neox_layers_9_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv41 = 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, reshape295, gpt_neox_layers_9_attention_dense_bias5, lv125, lv40), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm152 = R.call_tir(cls.layer_norm1, (lv41, gpt_neox_layers_10_input_layernorm_weight5, gpt_neox_layers_10_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv126 = 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_norm152, gpt_neox_layers_10_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape296 = R.call_tir(cls.reshape4, (lv126,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape297 = R.call_tir(cls.reshape5, (reshape296,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv377 = 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)), reshape297), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape298 = R.call_tir(cls.reshape6, (lv377,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape299 = R.call_tir(cls.reshape7, (reshape298,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm153 = R.call_tir(cls.layer_norm1, (lv41, gpt_neox_layers_10_post_attention_layernorm_weight5, gpt_neox_layers_10_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv127 = 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_norm153, gpt_neox_layers_10_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv128 = 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, lv127, gpt_neox_layers_10_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv42 = 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, reshape299, gpt_neox_layers_10_attention_dense_bias5, lv128, lv41), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm154 = R.call_tir(cls.layer_norm1, (lv42, gpt_neox_layers_11_input_layernorm_weight5, gpt_neox_layers_11_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv129 = 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_norm154, gpt_neox_layers_11_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape300 = R.call_tir(cls.reshape4, (lv129,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape301 = R.call_tir(cls.reshape5, (reshape300,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv382 = 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)), reshape301), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape302 = R.call_tir(cls.reshape6, (lv382,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape303 = R.call_tir(cls.reshape7, (reshape302,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm155 = R.call_tir(cls.layer_norm1, (lv42, gpt_neox_layers_11_post_attention_layernorm_weight5, gpt_neox_layers_11_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv130 = 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_norm155, gpt_neox_layers_11_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv131 = 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, lv130, gpt_neox_layers_11_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv43 = 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, reshape303, gpt_neox_layers_11_attention_dense_bias5, lv131, lv42), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm156 = R.call_tir(cls.layer_norm1, (lv43, gpt_neox_layers_12_input_layernorm_weight5, gpt_neox_layers_12_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv132 = 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_norm156, gpt_neox_layers_12_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape304 = R.call_tir(cls.reshape4, (lv132,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape305 = R.call_tir(cls.reshape5, (reshape304,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv387 = 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)), reshape305), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape306 = R.call_tir(cls.reshape6, (lv387,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape307 = R.call_tir(cls.reshape7, (reshape306,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm157 = R.call_tir(cls.layer_norm1, (lv43, gpt_neox_layers_12_post_attention_layernorm_weight5, gpt_neox_layers_12_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv133 = 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_norm157, gpt_neox_layers_12_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv134 = 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, lv133, gpt_neox_layers_12_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv44 = 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, reshape307, gpt_neox_layers_12_attention_dense_bias5, lv134, lv43), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm158 = R.call_tir(cls.layer_norm1, (lv44, gpt_neox_layers_13_input_layernorm_weight5, gpt_neox_layers_13_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv135 = 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_norm158, gpt_neox_layers_13_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape308 = R.call_tir(cls.reshape4, (lv135,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape309 = R.call_tir(cls.reshape5, (reshape308,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv392 = 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)), reshape309), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape310 = R.call_tir(cls.reshape6, (lv392,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape311 = R.call_tir(cls.reshape7, (reshape310,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm159 = R.call_tir(cls.layer_norm1, (lv44, gpt_neox_layers_13_post_attention_layernorm_weight5, gpt_neox_layers_13_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv136 = 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_norm159, gpt_neox_layers_13_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv137 = 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, lv136, gpt_neox_layers_13_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv45 = 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, reshape311, gpt_neox_layers_13_attention_dense_bias5, lv137, lv44), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm160 = R.call_tir(cls.layer_norm1, (lv45, gpt_neox_layers_14_input_layernorm_weight5, gpt_neox_layers_14_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv138 = 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_norm160, gpt_neox_layers_14_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape312 = R.call_tir(cls.reshape4, (lv138,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape313 = R.call_tir(cls.reshape5, (reshape312,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv397 = 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)), reshape313), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape314 = R.call_tir(cls.reshape6, (lv397,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape315 = R.call_tir(cls.reshape7, (reshape314,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm161 = R.call_tir(cls.layer_norm1, (lv45, gpt_neox_layers_14_post_attention_layernorm_weight5, gpt_neox_layers_14_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv139 = 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_norm161, gpt_neox_layers_14_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv140 = 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, lv139, gpt_neox_layers_14_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv46 = 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, reshape315, gpt_neox_layers_14_attention_dense_bias5, lv140, lv45), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm162 = R.call_tir(cls.layer_norm1, (lv46, gpt_neox_layers_15_input_layernorm_weight5, gpt_neox_layers_15_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv141 = 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_norm162, gpt_neox_layers_15_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + reshape316 = R.call_tir(cls.reshape4, (lv141,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape317 = R.call_tir(cls.reshape5, (reshape316,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv402 = 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)), reshape317), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) + reshape318 = R.call_tir(cls.reshape6, (lv402,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape319 = R.call_tir(cls.reshape7, (reshape318,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm163 = R.call_tir(cls.layer_norm1, (lv46, gpt_neox_layers_15_post_attention_layernorm_weight5, gpt_neox_layers_15_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv142 = 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_norm163, gpt_neox_layers_15_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float16")) + lv143 = 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, lv142, gpt_neox_layers_15_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv47 = 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, reshape319, gpt_neox_layers_15_attention_dense_bias5, lv143, lv46), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm164 = R.call_tir(cls.layer_norm1, (lv47, gpt_neox_final_layer_norm_weight5, gpt_neox_final_layer_norm_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv2 = R.call_tir(cls.fused_dequantize_fused_NT_matmul9_cast5, (embed_out_q_weight5, embed_out_q_scale5, layer_norm164), 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, 16]), R.prim_value(8), R.prim_value(8), R.prim_value(256), R.prim_value(1), R.prim_value(1), R.prim_value(10000), R.const(0.0, "float16"), cls.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, 2048), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 256), dtype="uint32"), R.Tensor(("vocab_size", 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 256), dtype="uint32"), R.Tensor(("vocab_size", 64), 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((2048,), dtype="float16") = packed_params[2] + gpt_neox_layers_0_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[4] + gpt_neox_layers_0_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[6] + gpt_neox_layers_0_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[8] + gpt_neox_layers_0_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[9] + gpt_neox_layers_0_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[10] + gpt_neox_layers_0_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[12] + gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[13] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[14] + gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[15] + gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[16] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[17] + gpt_neox_layers_1_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[18] + gpt_neox_layers_1_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[19] + gpt_neox_layers_1_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[20] + gpt_neox_layers_1_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[21] + gpt_neox_layers_1_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[22] + gpt_neox_layers_1_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[23] + gpt_neox_layers_1_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[24] + gpt_neox_layers_1_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[25] + gpt_neox_layers_1_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[26] + gpt_neox_layers_1_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[27] + gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[28] + gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[29] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[30] + gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[31] + gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[32] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[33] + gpt_neox_layers_2_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[34] + gpt_neox_layers_2_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[35] + gpt_neox_layers_2_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[36] + gpt_neox_layers_2_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[37] + gpt_neox_layers_2_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[38] + gpt_neox_layers_2_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[39] + gpt_neox_layers_2_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[40] + gpt_neox_layers_2_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[41] + gpt_neox_layers_2_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[42] + gpt_neox_layers_2_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[43] + gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[44] + gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[45] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[46] + gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[47] + gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[48] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[49] + gpt_neox_layers_3_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[50] + gpt_neox_layers_3_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[51] + gpt_neox_layers_3_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[52] + gpt_neox_layers_3_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[53] + gpt_neox_layers_3_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[54] + gpt_neox_layers_3_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[55] + gpt_neox_layers_3_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[56] + gpt_neox_layers_3_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[57] + gpt_neox_layers_3_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[58] + gpt_neox_layers_3_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[59] + gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[60] + gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[61] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[62] + gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[63] + gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[64] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[65] + gpt_neox_layers_4_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[66] + gpt_neox_layers_4_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[67] + gpt_neox_layers_4_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[68] + gpt_neox_layers_4_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[69] + gpt_neox_layers_4_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[70] + gpt_neox_layers_4_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[71] + gpt_neox_layers_4_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[72] + gpt_neox_layers_4_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[73] + gpt_neox_layers_4_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[74] + gpt_neox_layers_4_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[75] + gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[76] + gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[77] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[78] + gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[79] + gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[80] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[81] + gpt_neox_layers_5_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[82] + gpt_neox_layers_5_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[83] + gpt_neox_layers_5_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[84] + gpt_neox_layers_5_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[85] + gpt_neox_layers_5_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[86] + gpt_neox_layers_5_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[87] + gpt_neox_layers_5_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[88] + gpt_neox_layers_5_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[89] + gpt_neox_layers_5_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[90] + gpt_neox_layers_5_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[91] + gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[92] + gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[93] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[94] + gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[95] + gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[96] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[97] + gpt_neox_layers_6_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[98] + gpt_neox_layers_6_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[99] + gpt_neox_layers_6_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[100] + gpt_neox_layers_6_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[101] + gpt_neox_layers_6_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[102] + gpt_neox_layers_6_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[103] + gpt_neox_layers_6_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[104] + gpt_neox_layers_6_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[105] + gpt_neox_layers_6_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[106] + gpt_neox_layers_6_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[107] + gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[108] + gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[109] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[110] + gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[111] + gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[112] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[113] + gpt_neox_layers_7_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[114] + gpt_neox_layers_7_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[115] + gpt_neox_layers_7_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[116] + gpt_neox_layers_7_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[117] + gpt_neox_layers_7_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[118] + gpt_neox_layers_7_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[119] + gpt_neox_layers_7_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[120] + gpt_neox_layers_7_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[121] + gpt_neox_layers_7_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[122] + gpt_neox_layers_7_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[123] + gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[124] + gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[125] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[126] + gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[127] + gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[128] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[129] + gpt_neox_layers_8_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[130] + gpt_neox_layers_8_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[131] + gpt_neox_layers_8_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[132] + gpt_neox_layers_8_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[133] + gpt_neox_layers_8_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[134] + gpt_neox_layers_8_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[135] + gpt_neox_layers_8_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[136] + gpt_neox_layers_8_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[137] + gpt_neox_layers_8_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[138] + gpt_neox_layers_8_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[139] + gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[140] + gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[141] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[142] + gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[143] + gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[144] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[145] + gpt_neox_layers_9_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[146] + gpt_neox_layers_9_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[147] + gpt_neox_layers_9_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[148] + gpt_neox_layers_9_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[149] + gpt_neox_layers_9_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[150] + gpt_neox_layers_9_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[151] + gpt_neox_layers_9_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[152] + gpt_neox_layers_9_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[153] + gpt_neox_layers_9_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[154] + gpt_neox_layers_9_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[155] + gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[156] + gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[157] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[158] + gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[159] + gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[160] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[161] + gpt_neox_layers_10_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[162] + gpt_neox_layers_10_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[163] + gpt_neox_layers_10_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[164] + gpt_neox_layers_10_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[165] + gpt_neox_layers_10_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[166] + gpt_neox_layers_10_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[167] + gpt_neox_layers_10_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[168] + gpt_neox_layers_10_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[169] + gpt_neox_layers_10_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[170] + gpt_neox_layers_10_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[171] + gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[172] + gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[173] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[174] + gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[175] + gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[176] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[177] + gpt_neox_layers_11_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[178] + gpt_neox_layers_11_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[179] + gpt_neox_layers_11_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[180] + gpt_neox_layers_11_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[181] + gpt_neox_layers_11_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[182] + gpt_neox_layers_11_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[183] + gpt_neox_layers_11_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[184] + gpt_neox_layers_11_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[185] + gpt_neox_layers_11_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[186] + gpt_neox_layers_11_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[187] + gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[188] + gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[189] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[190] + gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[191] + gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[192] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[193] + gpt_neox_layers_12_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[194] + gpt_neox_layers_12_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[195] + gpt_neox_layers_12_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[196] + gpt_neox_layers_12_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[197] + gpt_neox_layers_12_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[198] + gpt_neox_layers_12_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[199] + gpt_neox_layers_12_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[200] + gpt_neox_layers_12_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[201] + gpt_neox_layers_12_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[202] + gpt_neox_layers_12_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[203] + gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[204] + gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[205] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[206] + gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[207] + gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[208] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[209] + gpt_neox_layers_13_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[210] + gpt_neox_layers_13_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[211] + gpt_neox_layers_13_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[212] + gpt_neox_layers_13_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[213] + gpt_neox_layers_13_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[214] + gpt_neox_layers_13_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[215] + gpt_neox_layers_13_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[216] + gpt_neox_layers_13_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[217] + gpt_neox_layers_13_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[218] + gpt_neox_layers_13_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[219] + gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[220] + gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[221] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[222] + gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[223] + gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[224] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[225] + gpt_neox_layers_14_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[226] + gpt_neox_layers_14_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[227] + gpt_neox_layers_14_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[228] + gpt_neox_layers_14_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[229] + gpt_neox_layers_14_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[230] + gpt_neox_layers_14_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[231] + gpt_neox_layers_14_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[232] + gpt_neox_layers_14_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[233] + gpt_neox_layers_14_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[234] + gpt_neox_layers_14_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[235] + gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[236] + gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[237] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[238] + gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[239] + gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[240] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[241] + gpt_neox_layers_15_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[242] + gpt_neox_layers_15_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[243] + gpt_neox_layers_15_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[244] + gpt_neox_layers_15_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[245] + gpt_neox_layers_15_attention_query_key_value_q_weight2: R.Tensor((6144, 256), dtype="uint32") = packed_params[246] + gpt_neox_layers_15_attention_query_key_value_q_scale2: R.Tensor((6144, 64), dtype="float16") = packed_params[247] + gpt_neox_layers_15_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[248] + gpt_neox_layers_15_attention_dense_q_weight2: R.Tensor((2048, 256), dtype="uint32") = packed_params[249] + gpt_neox_layers_15_attention_dense_q_scale2: R.Tensor((2048, 64), dtype="float16") = packed_params[250] + gpt_neox_layers_15_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[251] + gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight2: R.Tensor((8192, 256), dtype="uint32") = packed_params[252] + gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale2: R.Tensor((8192, 64), dtype="float16") = packed_params[253] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[254] + gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight2: R.Tensor((2048, 1024), dtype="uint32") = packed_params[255] + gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale2: R.Tensor((2048, 256), dtype="float16") = packed_params[256] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[257] + gpt_neox_final_layer_norm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[258] + gpt_neox_final_layer_norm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[259] + embed_out_q_weight2: R.Tensor((vocab_size, 256), dtype="uint32") = packed_params[260] + embed_out_q_scale2: R.Tensor((vocab_size, 64), dtype="float16") = packed_params[261] + layer_norm33 = 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, 2048), dtype="float16")) + lv144 = 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_norm33, gpt_neox_layers_0_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv196 = R.call_tir(cls.fused_reshape8_reshape9, (lv144,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv84 = 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)), lv196), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv197 = R.call_tir(cls.fused_reshape10_reshape11, (lv84,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm34 = 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, 2048), dtype="float16")) + lv145 = 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_norm34, gpt_neox_layers_0_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv146 = 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, lv145, gpt_neox_layers_0_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv48 = 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, lv197, gpt_neox_layers_0_attention_dense_bias2, lv146, input_embed), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm35 = R.call_tir(cls.layer_norm2, (lv48, gpt_neox_layers_1_input_layernorm_weight2, gpt_neox_layers_1_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv147 = 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_norm35, gpt_neox_layers_1_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv202 = R.call_tir(cls.fused_reshape8_reshape9, (lv147,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv89 = 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)), lv202), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv203 = R.call_tir(cls.fused_reshape10_reshape11, (lv89,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm36 = R.call_tir(cls.layer_norm2, (lv48, gpt_neox_layers_1_post_attention_layernorm_weight2, gpt_neox_layers_1_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv148 = 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_norm36, gpt_neox_layers_1_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv149 = 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, lv148, gpt_neox_layers_1_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv49 = 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, lv203, gpt_neox_layers_1_attention_dense_bias2, lv149, lv48), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm37 = R.call_tir(cls.layer_norm2, (lv49, gpt_neox_layers_2_input_layernorm_weight2, gpt_neox_layers_2_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv150 = 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_norm37, gpt_neox_layers_2_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv208 = R.call_tir(cls.fused_reshape8_reshape9, (lv150,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv94 = 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)), lv208), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv209 = R.call_tir(cls.fused_reshape10_reshape11, (lv94,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm38 = R.call_tir(cls.layer_norm2, (lv49, gpt_neox_layers_2_post_attention_layernorm_weight2, gpt_neox_layers_2_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv151 = 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_norm38, gpt_neox_layers_2_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv152 = 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, lv151, gpt_neox_layers_2_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv50 = 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, lv209, gpt_neox_layers_2_attention_dense_bias2, lv152, lv49), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm39 = R.call_tir(cls.layer_norm2, (lv50, gpt_neox_layers_3_input_layernorm_weight2, gpt_neox_layers_3_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv153 = 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_norm39, gpt_neox_layers_3_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv214 = R.call_tir(cls.fused_reshape8_reshape9, (lv153,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv99 = 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)), lv214), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv215 = R.call_tir(cls.fused_reshape10_reshape11, (lv99,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm40 = R.call_tir(cls.layer_norm2, (lv50, gpt_neox_layers_3_post_attention_layernorm_weight2, gpt_neox_layers_3_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv154 = 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_norm40, gpt_neox_layers_3_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv155 = 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, lv154, gpt_neox_layers_3_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv51 = 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, lv215, gpt_neox_layers_3_attention_dense_bias2, lv155, lv50), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm41 = R.call_tir(cls.layer_norm2, (lv51, gpt_neox_layers_4_input_layernorm_weight2, gpt_neox_layers_4_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv156 = 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_norm41, gpt_neox_layers_4_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv220 = R.call_tir(cls.fused_reshape8_reshape9, (lv156,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv104 = 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)), lv220), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv221 = R.call_tir(cls.fused_reshape10_reshape11, (lv104,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm42 = R.call_tir(cls.layer_norm2, (lv51, gpt_neox_layers_4_post_attention_layernorm_weight2, gpt_neox_layers_4_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv157 = 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_norm42, gpt_neox_layers_4_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv158 = 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, lv157, gpt_neox_layers_4_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv52 = 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, lv221, gpt_neox_layers_4_attention_dense_bias2, lv158, lv51), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm43 = R.call_tir(cls.layer_norm2, (lv52, gpt_neox_layers_5_input_layernorm_weight2, gpt_neox_layers_5_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv159 = 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_norm43, gpt_neox_layers_5_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv226 = R.call_tir(cls.fused_reshape8_reshape9, (lv159,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv109 = 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)), lv226), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv227 = R.call_tir(cls.fused_reshape10_reshape11, (lv109,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm44 = R.call_tir(cls.layer_norm2, (lv52, gpt_neox_layers_5_post_attention_layernorm_weight2, gpt_neox_layers_5_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv160 = 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_norm44, gpt_neox_layers_5_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv161 = 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, lv160, gpt_neox_layers_5_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv53 = 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, lv227, gpt_neox_layers_5_attention_dense_bias2, lv161, lv52), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm45 = R.call_tir(cls.layer_norm2, (lv53, gpt_neox_layers_6_input_layernorm_weight2, gpt_neox_layers_6_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv162 = 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_norm45, gpt_neox_layers_6_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv232 = R.call_tir(cls.fused_reshape8_reshape9, (lv162,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv114 = 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)), lv232), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv233 = R.call_tir(cls.fused_reshape10_reshape11, (lv114,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm46 = R.call_tir(cls.layer_norm2, (lv53, gpt_neox_layers_6_post_attention_layernorm_weight2, gpt_neox_layers_6_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv163 = 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_norm46, gpt_neox_layers_6_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv164 = 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, lv163, gpt_neox_layers_6_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv54 = 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, lv233, gpt_neox_layers_6_attention_dense_bias2, lv164, lv53), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm47 = R.call_tir(cls.layer_norm2, (lv54, gpt_neox_layers_7_input_layernorm_weight2, gpt_neox_layers_7_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv165 = 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_norm47, gpt_neox_layers_7_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv238 = R.call_tir(cls.fused_reshape8_reshape9, (lv165,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv119 = 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)), lv238), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv239 = R.call_tir(cls.fused_reshape10_reshape11, (lv119,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm48 = R.call_tir(cls.layer_norm2, (lv54, gpt_neox_layers_7_post_attention_layernorm_weight2, gpt_neox_layers_7_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv166 = 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_norm48, gpt_neox_layers_7_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv167 = 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, lv166, gpt_neox_layers_7_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv55 = 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, lv239, gpt_neox_layers_7_attention_dense_bias2, lv167, lv54), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm49 = R.call_tir(cls.layer_norm2, (lv55, gpt_neox_layers_8_input_layernorm_weight2, gpt_neox_layers_8_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv168 = 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_norm49, gpt_neox_layers_8_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv244 = R.call_tir(cls.fused_reshape8_reshape9, (lv168,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv124 = 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)), lv244), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv245 = R.call_tir(cls.fused_reshape10_reshape11, (lv124,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm50 = R.call_tir(cls.layer_norm2, (lv55, gpt_neox_layers_8_post_attention_layernorm_weight2, gpt_neox_layers_8_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv169 = 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_norm50, gpt_neox_layers_8_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv170 = 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, lv169, gpt_neox_layers_8_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv56 = 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, lv245, gpt_neox_layers_8_attention_dense_bias2, lv170, lv55), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm51 = R.call_tir(cls.layer_norm2, (lv56, gpt_neox_layers_9_input_layernorm_weight2, gpt_neox_layers_9_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv171 = 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_norm51, gpt_neox_layers_9_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv250 = R.call_tir(cls.fused_reshape8_reshape9, (lv171,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv129 = 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)), lv250), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv251 = R.call_tir(cls.fused_reshape10_reshape11, (lv129,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm52 = R.call_tir(cls.layer_norm2, (lv56, gpt_neox_layers_9_post_attention_layernorm_weight2, gpt_neox_layers_9_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv172 = 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_norm52, gpt_neox_layers_9_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv173 = 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, lv172, gpt_neox_layers_9_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv57 = 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, lv251, gpt_neox_layers_9_attention_dense_bias2, lv173, lv56), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm53 = R.call_tir(cls.layer_norm2, (lv57, gpt_neox_layers_10_input_layernorm_weight2, gpt_neox_layers_10_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv174 = 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_norm53, gpt_neox_layers_10_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv256 = R.call_tir(cls.fused_reshape8_reshape9, (lv174,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv134 = 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)), lv256), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv257 = R.call_tir(cls.fused_reshape10_reshape11, (lv134,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm54 = R.call_tir(cls.layer_norm2, (lv57, gpt_neox_layers_10_post_attention_layernorm_weight2, gpt_neox_layers_10_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv175 = 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_norm54, gpt_neox_layers_10_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv176 = 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, lv175, gpt_neox_layers_10_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv58 = 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, lv257, gpt_neox_layers_10_attention_dense_bias2, lv176, lv57), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm55 = R.call_tir(cls.layer_norm2, (lv58, gpt_neox_layers_11_input_layernorm_weight2, gpt_neox_layers_11_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv177 = 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_norm55, gpt_neox_layers_11_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv262 = R.call_tir(cls.fused_reshape8_reshape9, (lv177,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv139 = 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)), lv262), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv263 = R.call_tir(cls.fused_reshape10_reshape11, (lv139,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm56 = R.call_tir(cls.layer_norm2, (lv58, gpt_neox_layers_11_post_attention_layernorm_weight2, gpt_neox_layers_11_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv178 = 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_norm56, gpt_neox_layers_11_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv179 = 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, lv178, gpt_neox_layers_11_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv59 = 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, lv263, gpt_neox_layers_11_attention_dense_bias2, lv179, lv58), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm57 = R.call_tir(cls.layer_norm2, (lv59, gpt_neox_layers_12_input_layernorm_weight2, gpt_neox_layers_12_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv180 = 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_norm57, gpt_neox_layers_12_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv268 = R.call_tir(cls.fused_reshape8_reshape9, (lv180,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv144_1 = 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)), lv268), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv269 = R.call_tir(cls.fused_reshape10_reshape11, (lv144_1,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm58 = R.call_tir(cls.layer_norm2, (lv59, gpt_neox_layers_12_post_attention_layernorm_weight2, gpt_neox_layers_12_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv181 = 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_norm58, gpt_neox_layers_12_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv182 = 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, lv181, gpt_neox_layers_12_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv60 = 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, lv269, gpt_neox_layers_12_attention_dense_bias2, lv182, lv59), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm59 = R.call_tir(cls.layer_norm2, (lv60, gpt_neox_layers_13_input_layernorm_weight2, gpt_neox_layers_13_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv183 = 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_norm59, gpt_neox_layers_13_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv274 = R.call_tir(cls.fused_reshape8_reshape9, (lv183,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv149_1 = 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)), lv274), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv275 = R.call_tir(cls.fused_reshape10_reshape11, (lv149_1,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm60 = R.call_tir(cls.layer_norm2, (lv60, gpt_neox_layers_13_post_attention_layernorm_weight2, gpt_neox_layers_13_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv184 = 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_norm60, gpt_neox_layers_13_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv185 = 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, lv184, gpt_neox_layers_13_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv61 = 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, lv275, gpt_neox_layers_13_attention_dense_bias2, lv185, lv60), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm61 = R.call_tir(cls.layer_norm2, (lv61, gpt_neox_layers_14_input_layernorm_weight2, gpt_neox_layers_14_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv186 = 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_norm61, gpt_neox_layers_14_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv280 = R.call_tir(cls.fused_reshape8_reshape9, (lv186,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv154_1 = 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)), lv280), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv281 = R.call_tir(cls.fused_reshape10_reshape11, (lv154_1,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm62 = R.call_tir(cls.layer_norm2, (lv61, gpt_neox_layers_14_post_attention_layernorm_weight2, gpt_neox_layers_14_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv187 = 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_norm62, gpt_neox_layers_14_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv188 = 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, lv187, gpt_neox_layers_14_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv62 = 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, lv281, gpt_neox_layers_14_attention_dense_bias2, lv188, lv61), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm63 = R.call_tir(cls.layer_norm2, (lv62, gpt_neox_layers_15_input_layernorm_weight2, gpt_neox_layers_15_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv189 = 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_norm63, gpt_neox_layers_15_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + lv286 = R.call_tir(cls.fused_reshape8_reshape9, (lv189,), out_sinfo=R.Tensor((1, 24, 256), dtype="float16")) + lv159_1 = 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)), lv286), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) + lv287 = R.call_tir(cls.fused_reshape10_reshape11, (lv159_1,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm64 = R.call_tir(cls.layer_norm2, (lv62, gpt_neox_layers_15_post_attention_layernorm_weight2, gpt_neox_layers_15_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv190 = 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_norm64, gpt_neox_layers_15_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float16")) + lv191 = 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, lv190, gpt_neox_layers_15_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv63 = 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, lv287, gpt_neox_layers_15_attention_dense_bias2, lv191, lv62), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + layer_norm65 = R.call_tir(cls.layer_norm2, (lv63, gpt_neox_final_layer_norm_weight2, gpt_neox_final_layer_norm_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv3 = R.call_tir(cls.fused_dequantize_fused_NT_matmul14_cast8, (embed_out_q_weight2, embed_out_q_scale2, layer_norm65), 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", 256), dtype="uint32"), R.Tensor(("vocab_size", 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 256), dtype="uint32"), R.Tensor(("vocab_size", 64), dtype="float16"))) -> R.Tensor(("seq_len", 2048), 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, 256), dtype="uint32") = packed_params[0] + gpt_neox_embed_in_q_scale: R.Tensor((vocab_size, 64), 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, 2048), dtype="float16")) + R.output(gv) + return gv + + @R.function + def prefill(input_embed: R.Tensor((1, "seq_len", 2048), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 256), dtype="uint32"), R.Tensor(("vocab_size", 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 256), dtype="uint32"), R.Tensor((6144, 64), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 256), dtype="uint32"), R.Tensor((2048, 64), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 256), dtype="uint32"), R.Tensor((8192, 64), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 1024), dtype="uint32"), R.Tensor((2048, 256), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 256), dtype="uint32"), R.Tensor(("vocab_size", 64), 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((2048,), dtype="float16") = packed_params[2] + gpt_neox_layers_0_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[4] + gpt_neox_layers_0_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[6] + gpt_neox_layers_0_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[8] + gpt_neox_layers_0_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[9] + gpt_neox_layers_0_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[10] + gpt_neox_layers_0_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[12] + gpt_neox_layers_0_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[13] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[14] + gpt_neox_layers_0_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[15] + gpt_neox_layers_0_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[16] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[17] + gpt_neox_layers_1_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[18] + gpt_neox_layers_1_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[19] + gpt_neox_layers_1_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[20] + gpt_neox_layers_1_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[21] + gpt_neox_layers_1_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[22] + gpt_neox_layers_1_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[23] + gpt_neox_layers_1_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[24] + gpt_neox_layers_1_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[25] + gpt_neox_layers_1_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[26] + gpt_neox_layers_1_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[27] + gpt_neox_layers_1_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[28] + gpt_neox_layers_1_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[29] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[30] + gpt_neox_layers_1_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[31] + gpt_neox_layers_1_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[32] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[33] + gpt_neox_layers_2_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[34] + gpt_neox_layers_2_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[35] + gpt_neox_layers_2_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[36] + gpt_neox_layers_2_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[37] + gpt_neox_layers_2_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[38] + gpt_neox_layers_2_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[39] + gpt_neox_layers_2_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[40] + gpt_neox_layers_2_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[41] + gpt_neox_layers_2_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[42] + gpt_neox_layers_2_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[43] + gpt_neox_layers_2_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[44] + gpt_neox_layers_2_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[45] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[46] + gpt_neox_layers_2_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[47] + gpt_neox_layers_2_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[48] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[49] + gpt_neox_layers_3_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[50] + gpt_neox_layers_3_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[51] + gpt_neox_layers_3_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[52] + gpt_neox_layers_3_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[53] + gpt_neox_layers_3_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[54] + gpt_neox_layers_3_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[55] + gpt_neox_layers_3_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[56] + gpt_neox_layers_3_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[57] + gpt_neox_layers_3_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[58] + gpt_neox_layers_3_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[59] + gpt_neox_layers_3_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[60] + gpt_neox_layers_3_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[61] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[62] + gpt_neox_layers_3_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[63] + gpt_neox_layers_3_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[64] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[65] + gpt_neox_layers_4_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[66] + gpt_neox_layers_4_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[67] + gpt_neox_layers_4_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[68] + gpt_neox_layers_4_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[69] + gpt_neox_layers_4_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[70] + gpt_neox_layers_4_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[71] + gpt_neox_layers_4_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[72] + gpt_neox_layers_4_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[73] + gpt_neox_layers_4_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[74] + gpt_neox_layers_4_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[75] + gpt_neox_layers_4_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[76] + gpt_neox_layers_4_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[77] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[78] + gpt_neox_layers_4_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[79] + gpt_neox_layers_4_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[80] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[81] + gpt_neox_layers_5_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[82] + gpt_neox_layers_5_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[83] + gpt_neox_layers_5_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[84] + gpt_neox_layers_5_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[85] + gpt_neox_layers_5_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[86] + gpt_neox_layers_5_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[87] + gpt_neox_layers_5_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[88] + gpt_neox_layers_5_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[89] + gpt_neox_layers_5_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[90] + gpt_neox_layers_5_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[91] + gpt_neox_layers_5_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[92] + gpt_neox_layers_5_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[93] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[94] + gpt_neox_layers_5_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[95] + gpt_neox_layers_5_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[96] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[97] + gpt_neox_layers_6_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[98] + gpt_neox_layers_6_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[99] + gpt_neox_layers_6_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[100] + gpt_neox_layers_6_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[101] + gpt_neox_layers_6_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[102] + gpt_neox_layers_6_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[103] + gpt_neox_layers_6_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[104] + gpt_neox_layers_6_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[105] + gpt_neox_layers_6_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[106] + gpt_neox_layers_6_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[107] + gpt_neox_layers_6_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[108] + gpt_neox_layers_6_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[109] + gpt_neox_layers_6_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[110] + gpt_neox_layers_6_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[111] + gpt_neox_layers_6_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[112] + gpt_neox_layers_6_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[113] + gpt_neox_layers_7_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[114] + gpt_neox_layers_7_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[115] + gpt_neox_layers_7_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[116] + gpt_neox_layers_7_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[117] + gpt_neox_layers_7_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[118] + gpt_neox_layers_7_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[119] + gpt_neox_layers_7_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[120] + gpt_neox_layers_7_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[121] + gpt_neox_layers_7_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[122] + gpt_neox_layers_7_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[123] + gpt_neox_layers_7_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[124] + gpt_neox_layers_7_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[125] + gpt_neox_layers_7_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[126] + gpt_neox_layers_7_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[127] + gpt_neox_layers_7_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[128] + gpt_neox_layers_7_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[129] + gpt_neox_layers_8_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[130] + gpt_neox_layers_8_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[131] + gpt_neox_layers_8_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[132] + gpt_neox_layers_8_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[133] + gpt_neox_layers_8_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[134] + gpt_neox_layers_8_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[135] + gpt_neox_layers_8_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[136] + gpt_neox_layers_8_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[137] + gpt_neox_layers_8_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[138] + gpt_neox_layers_8_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[139] + gpt_neox_layers_8_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[140] + gpt_neox_layers_8_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[141] + gpt_neox_layers_8_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[142] + gpt_neox_layers_8_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[143] + gpt_neox_layers_8_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[144] + gpt_neox_layers_8_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[145] + gpt_neox_layers_9_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[146] + gpt_neox_layers_9_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[147] + gpt_neox_layers_9_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[148] + gpt_neox_layers_9_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[149] + gpt_neox_layers_9_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[150] + gpt_neox_layers_9_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[151] + gpt_neox_layers_9_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[152] + gpt_neox_layers_9_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[153] + gpt_neox_layers_9_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[154] + gpt_neox_layers_9_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[155] + gpt_neox_layers_9_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[156] + gpt_neox_layers_9_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[157] + gpt_neox_layers_9_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[158] + gpt_neox_layers_9_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[159] + gpt_neox_layers_9_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[160] + gpt_neox_layers_9_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[161] + gpt_neox_layers_10_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[162] + gpt_neox_layers_10_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[163] + gpt_neox_layers_10_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[164] + gpt_neox_layers_10_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[165] + gpt_neox_layers_10_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[166] + gpt_neox_layers_10_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[167] + gpt_neox_layers_10_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[168] + gpt_neox_layers_10_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[169] + gpt_neox_layers_10_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[170] + gpt_neox_layers_10_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[171] + gpt_neox_layers_10_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[172] + gpt_neox_layers_10_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[173] + gpt_neox_layers_10_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[174] + gpt_neox_layers_10_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[175] + gpt_neox_layers_10_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[176] + gpt_neox_layers_10_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[177] + gpt_neox_layers_11_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[178] + gpt_neox_layers_11_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[179] + gpt_neox_layers_11_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[180] + gpt_neox_layers_11_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[181] + gpt_neox_layers_11_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[182] + gpt_neox_layers_11_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[183] + gpt_neox_layers_11_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[184] + gpt_neox_layers_11_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[185] + gpt_neox_layers_11_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[186] + gpt_neox_layers_11_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[187] + gpt_neox_layers_11_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[188] + gpt_neox_layers_11_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[189] + gpt_neox_layers_11_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[190] + gpt_neox_layers_11_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[191] + gpt_neox_layers_11_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[192] + gpt_neox_layers_11_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[193] + gpt_neox_layers_12_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[194] + gpt_neox_layers_12_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[195] + gpt_neox_layers_12_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[196] + gpt_neox_layers_12_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[197] + gpt_neox_layers_12_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[198] + gpt_neox_layers_12_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[199] + gpt_neox_layers_12_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[200] + gpt_neox_layers_12_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[201] + gpt_neox_layers_12_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[202] + gpt_neox_layers_12_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[203] + gpt_neox_layers_12_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[204] + gpt_neox_layers_12_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[205] + gpt_neox_layers_12_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[206] + gpt_neox_layers_12_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[207] + gpt_neox_layers_12_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[208] + gpt_neox_layers_12_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[209] + gpt_neox_layers_13_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[210] + gpt_neox_layers_13_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[211] + gpt_neox_layers_13_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[212] + gpt_neox_layers_13_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[213] + gpt_neox_layers_13_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[214] + gpt_neox_layers_13_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[215] + gpt_neox_layers_13_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[216] + gpt_neox_layers_13_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[217] + gpt_neox_layers_13_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[218] + gpt_neox_layers_13_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[219] + gpt_neox_layers_13_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[220] + gpt_neox_layers_13_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[221] + gpt_neox_layers_13_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[222] + gpt_neox_layers_13_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[223] + gpt_neox_layers_13_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[224] + gpt_neox_layers_13_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[225] + gpt_neox_layers_14_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[226] + gpt_neox_layers_14_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[227] + gpt_neox_layers_14_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[228] + gpt_neox_layers_14_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[229] + gpt_neox_layers_14_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[230] + gpt_neox_layers_14_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[231] + gpt_neox_layers_14_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[232] + gpt_neox_layers_14_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[233] + gpt_neox_layers_14_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[234] + gpt_neox_layers_14_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[235] + gpt_neox_layers_14_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[236] + gpt_neox_layers_14_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[237] + gpt_neox_layers_14_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[238] + gpt_neox_layers_14_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[239] + gpt_neox_layers_14_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[240] + gpt_neox_layers_14_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[241] + gpt_neox_layers_15_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[242] + gpt_neox_layers_15_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[243] + gpt_neox_layers_15_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[244] + gpt_neox_layers_15_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[245] + gpt_neox_layers_15_attention_query_key_value_q_weight1: R.Tensor((6144, 256), dtype="uint32") = packed_params[246] + gpt_neox_layers_15_attention_query_key_value_q_scale1: R.Tensor((6144, 64), dtype="float16") = packed_params[247] + gpt_neox_layers_15_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[248] + gpt_neox_layers_15_attention_dense_q_weight1: R.Tensor((2048, 256), dtype="uint32") = packed_params[249] + gpt_neox_layers_15_attention_dense_q_scale1: R.Tensor((2048, 64), dtype="float16") = packed_params[250] + gpt_neox_layers_15_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[251] + gpt_neox_layers_15_mlp_dense_h_to_4h_q_weight1: R.Tensor((8192, 256), dtype="uint32") = packed_params[252] + gpt_neox_layers_15_mlp_dense_h_to_4h_q_scale1: R.Tensor((8192, 64), dtype="float16") = packed_params[253] + gpt_neox_layers_15_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[254] + gpt_neox_layers_15_mlp_dense_4h_to_h_q_weight1: R.Tensor((2048, 1024), dtype="uint32") = packed_params[255] + gpt_neox_layers_15_mlp_dense_4h_to_h_q_scale1: R.Tensor((2048, 256), dtype="float16") = packed_params[256] + gpt_neox_layers_15_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[257] + gpt_neox_final_layer_norm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[258] + gpt_neox_final_layer_norm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[259] + embed_out_q_weight1: R.Tensor((vocab_size, 256), dtype="uint32") = packed_params[260] + embed_out_q_scale1: R.Tensor((vocab_size, 64), dtype="float16") = packed_params[261] + 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, 2048), dtype="float16")) + lv192 = 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, 6144), dtype="float16")) + reshape = R.call_tir(cls.reshape4, (lv192,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape1 = R.call_tir(cls.reshape5, (reshape,), out_sinfo=R.Tensor((seq_len, 24, 256), 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, 8, 256), dtype="float16")) + reshape2 = R.call_tir(cls.reshape6, (lv2,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape3 = R.call_tir(cls.reshape7, (reshape2,), out_sinfo=R.Tensor((1, seq_len, 2048), 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, 2048), dtype="float16")) + lv193 = 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, 8192), dtype="float16")) + lv194 = 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, lv193, gpt_neox_layers_0_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv64 = 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, lv194, input_embed), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm2 = R.call_tir(cls.layer_norm1, (lv64, gpt_neox_layers_1_input_layernorm_weight1, gpt_neox_layers_1_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv195 = 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, 6144), dtype="float16")) + reshape4 = R.call_tir(cls.reshape4, (lv195,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape5 = R.call_tir(cls.reshape5, (reshape4,), out_sinfo=R.Tensor((seq_len, 24, 256), 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, 8, 256), dtype="float16")) + reshape6 = R.call_tir(cls.reshape6, (lv7,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape7 = R.call_tir(cls.reshape7, (reshape6,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm3 = R.call_tir(cls.layer_norm1, (lv64, gpt_neox_layers_1_post_attention_layernorm_weight1, gpt_neox_layers_1_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv196 = 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, 8192), dtype="float16")) + lv197 = 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, lv196, gpt_neox_layers_1_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv65 = 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, lv197, lv64), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm4 = R.call_tir(cls.layer_norm1, (lv65, gpt_neox_layers_2_input_layernorm_weight1, gpt_neox_layers_2_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv198 = 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, 6144), dtype="float16")) + reshape8 = R.call_tir(cls.reshape4, (lv198,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape9 = R.call_tir(cls.reshape5, (reshape8,), out_sinfo=R.Tensor((seq_len, 24, 256), 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, 8, 256), dtype="float16")) + reshape10 = R.call_tir(cls.reshape6, (lv12,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape11 = R.call_tir(cls.reshape7, (reshape10,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm5 = R.call_tir(cls.layer_norm1, (lv65, gpt_neox_layers_2_post_attention_layernorm_weight1, gpt_neox_layers_2_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv199 = 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, 8192), dtype="float16")) + lv200 = 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, lv199, gpt_neox_layers_2_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv66 = 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, lv200, lv65), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm6 = R.call_tir(cls.layer_norm1, (lv66, gpt_neox_layers_3_input_layernorm_weight1, gpt_neox_layers_3_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv201 = 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, 6144), dtype="float16")) + reshape12 = R.call_tir(cls.reshape4, (lv201,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape13 = R.call_tir(cls.reshape5, (reshape12,), out_sinfo=R.Tensor((seq_len, 24, 256), 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, 8, 256), dtype="float16")) + reshape14 = R.call_tir(cls.reshape6, (lv17,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape15 = R.call_tir(cls.reshape7, (reshape14,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm7 = R.call_tir(cls.layer_norm1, (lv66, gpt_neox_layers_3_post_attention_layernorm_weight1, gpt_neox_layers_3_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv202 = 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, 8192), dtype="float16")) + lv203 = 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, lv202, gpt_neox_layers_3_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv67 = 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, lv203, lv66), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm8 = R.call_tir(cls.layer_norm1, (lv67, gpt_neox_layers_4_input_layernorm_weight1, gpt_neox_layers_4_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv204 = 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, 6144), dtype="float16")) + reshape16 = R.call_tir(cls.reshape4, (lv204,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape17 = R.call_tir(cls.reshape5, (reshape16,), out_sinfo=R.Tensor((seq_len, 24, 256), 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, 8, 256), dtype="float16")) + reshape18 = R.call_tir(cls.reshape6, (lv22,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape19 = R.call_tir(cls.reshape7, (reshape18,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm9 = R.call_tir(cls.layer_norm1, (lv67, gpt_neox_layers_4_post_attention_layernorm_weight1, gpt_neox_layers_4_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv205 = 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, 8192), dtype="float16")) + lv206 = 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, lv205, gpt_neox_layers_4_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv68 = 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, lv206, lv67), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm10 = R.call_tir(cls.layer_norm1, (lv68, gpt_neox_layers_5_input_layernorm_weight1, gpt_neox_layers_5_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv207 = 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, 6144), dtype="float16")) + reshape20 = R.call_tir(cls.reshape4, (lv207,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape21 = R.call_tir(cls.reshape5, (reshape20,), out_sinfo=R.Tensor((seq_len, 24, 256), 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, 8, 256), dtype="float16")) + reshape22 = R.call_tir(cls.reshape6, (lv27,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape23 = R.call_tir(cls.reshape7, (reshape22,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm11 = R.call_tir(cls.layer_norm1, (lv68, gpt_neox_layers_5_post_attention_layernorm_weight1, gpt_neox_layers_5_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv208 = 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, 8192), dtype="float16")) + lv209 = 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, lv208, gpt_neox_layers_5_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv69 = 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, lv209, lv68), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm12 = R.call_tir(cls.layer_norm1, (lv69, gpt_neox_layers_6_input_layernorm_weight1, gpt_neox_layers_6_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv210 = 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, 6144), dtype="float16")) + reshape24 = R.call_tir(cls.reshape4, (lv210,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape25 = R.call_tir(cls.reshape5, (reshape24,), out_sinfo=R.Tensor((seq_len, 24, 256), 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, 8, 256), dtype="float16")) + reshape26 = R.call_tir(cls.reshape6, (lv32,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape27 = R.call_tir(cls.reshape7, (reshape26,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm13 = R.call_tir(cls.layer_norm1, (lv69, gpt_neox_layers_6_post_attention_layernorm_weight1, gpt_neox_layers_6_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv211 = 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, 8192), dtype="float16")) + lv212 = 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, lv211, gpt_neox_layers_6_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv70 = 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, lv212, lv69), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm14 = R.call_tir(cls.layer_norm1, (lv70, gpt_neox_layers_7_input_layernorm_weight1, gpt_neox_layers_7_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv213 = 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, 6144), dtype="float16")) + reshape28 = R.call_tir(cls.reshape4, (lv213,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape29 = R.call_tir(cls.reshape5, (reshape28,), out_sinfo=R.Tensor((seq_len, 24, 256), 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, 8, 256), dtype="float16")) + reshape30 = R.call_tir(cls.reshape6, (lv37,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape31 = R.call_tir(cls.reshape7, (reshape30,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm15 = R.call_tir(cls.layer_norm1, (lv70, gpt_neox_layers_7_post_attention_layernorm_weight1, gpt_neox_layers_7_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv214 = 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, 8192), dtype="float16")) + lv215 = 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, lv214, gpt_neox_layers_7_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv71 = 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, lv215, lv70), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm16 = R.call_tir(cls.layer_norm1, (lv71, gpt_neox_layers_8_input_layernorm_weight1, gpt_neox_layers_8_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv216 = 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, 6144), dtype="float16")) + reshape32 = R.call_tir(cls.reshape4, (lv216,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape33 = R.call_tir(cls.reshape5, (reshape32,), out_sinfo=R.Tensor((seq_len, 24, 256), 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, 8, 256), dtype="float16")) + reshape34 = R.call_tir(cls.reshape6, (lv42,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape35 = R.call_tir(cls.reshape7, (reshape34,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm17 = R.call_tir(cls.layer_norm1, (lv71, gpt_neox_layers_8_post_attention_layernorm_weight1, gpt_neox_layers_8_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv217 = 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, 8192), dtype="float16")) + lv218 = 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, lv217, gpt_neox_layers_8_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv72 = 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, lv218, lv71), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm18 = R.call_tir(cls.layer_norm1, (lv72, gpt_neox_layers_9_input_layernorm_weight1, gpt_neox_layers_9_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv219 = 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, 6144), dtype="float16")) + reshape36 = R.call_tir(cls.reshape4, (lv219,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape37 = R.call_tir(cls.reshape5, (reshape36,), out_sinfo=R.Tensor((seq_len, 24, 256), 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, 8, 256), dtype="float16")) + reshape38 = R.call_tir(cls.reshape6, (lv47,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape39 = R.call_tir(cls.reshape7, (reshape38,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm19 = R.call_tir(cls.layer_norm1, (lv72, gpt_neox_layers_9_post_attention_layernorm_weight1, gpt_neox_layers_9_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv220 = 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, 8192), dtype="float16")) + lv221 = 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, lv220, gpt_neox_layers_9_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv73 = 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, lv221, lv72), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm20 = R.call_tir(cls.layer_norm1, (lv73, gpt_neox_layers_10_input_layernorm_weight1, gpt_neox_layers_10_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv222 = 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, 6144), dtype="float16")) + reshape40 = R.call_tir(cls.reshape4, (lv222,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape41 = R.call_tir(cls.reshape5, (reshape40,), out_sinfo=R.Tensor((seq_len, 24, 256), 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, 8, 256), dtype="float16")) + reshape42 = R.call_tir(cls.reshape6, (lv52,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape43 = R.call_tir(cls.reshape7, (reshape42,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm21 = R.call_tir(cls.layer_norm1, (lv73, gpt_neox_layers_10_post_attention_layernorm_weight1, gpt_neox_layers_10_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv223 = 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, 8192), dtype="float16")) + lv224 = 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, lv223, gpt_neox_layers_10_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv74 = 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, lv224, lv73), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm22 = R.call_tir(cls.layer_norm1, (lv74, gpt_neox_layers_11_input_layernorm_weight1, gpt_neox_layers_11_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv225 = 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, 6144), dtype="float16")) + reshape44 = R.call_tir(cls.reshape4, (lv225,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape45 = R.call_tir(cls.reshape5, (reshape44,), out_sinfo=R.Tensor((seq_len, 24, 256), 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, 8, 256), dtype="float16")) + reshape46 = R.call_tir(cls.reshape6, (lv57,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape47 = R.call_tir(cls.reshape7, (reshape46,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm23 = R.call_tir(cls.layer_norm1, (lv74, gpt_neox_layers_11_post_attention_layernorm_weight1, gpt_neox_layers_11_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv226 = 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, 8192), dtype="float16")) + lv227 = 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, lv226, gpt_neox_layers_11_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv75 = 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, lv227, lv74), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm24 = R.call_tir(cls.layer_norm1, (lv75, gpt_neox_layers_12_input_layernorm_weight1, gpt_neox_layers_12_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv228 = 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, 6144), dtype="float16")) + reshape48 = R.call_tir(cls.reshape4, (lv228,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape49 = R.call_tir(cls.reshape5, (reshape48,), out_sinfo=R.Tensor((seq_len, 24, 256), 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, 8, 256), dtype="float16")) + reshape50 = R.call_tir(cls.reshape6, (lv62,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape51 = R.call_tir(cls.reshape7, (reshape50,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm25 = R.call_tir(cls.layer_norm1, (lv75, gpt_neox_layers_12_post_attention_layernorm_weight1, gpt_neox_layers_12_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv229 = 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, 8192), dtype="float16")) + lv230 = 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, lv229, gpt_neox_layers_12_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv76 = 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, lv230, lv75), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm26 = R.call_tir(cls.layer_norm1, (lv76, gpt_neox_layers_13_input_layernorm_weight1, gpt_neox_layers_13_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv231 = 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, 6144), dtype="float16")) + reshape52 = R.call_tir(cls.reshape4, (lv231,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape53 = R.call_tir(cls.reshape5, (reshape52,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv67_1 = 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, 8, 256), dtype="float16")) + reshape54 = R.call_tir(cls.reshape6, (lv67_1,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape55 = R.call_tir(cls.reshape7, (reshape54,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm27 = R.call_tir(cls.layer_norm1, (lv76, gpt_neox_layers_13_post_attention_layernorm_weight1, gpt_neox_layers_13_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv232 = 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, 8192), dtype="float16")) + lv233 = 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, lv232, gpt_neox_layers_13_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv77 = 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, lv233, lv76), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm28 = R.call_tir(cls.layer_norm1, (lv77, gpt_neox_layers_14_input_layernorm_weight1, gpt_neox_layers_14_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv234 = 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, 6144), dtype="float16")) + reshape56 = R.call_tir(cls.reshape4, (lv234,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape57 = R.call_tir(cls.reshape5, (reshape56,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv72_1 = 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, 8, 256), dtype="float16")) + reshape58 = R.call_tir(cls.reshape6, (lv72_1,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape59 = R.call_tir(cls.reshape7, (reshape58,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm29 = R.call_tir(cls.layer_norm1, (lv77, gpt_neox_layers_14_post_attention_layernorm_weight1, gpt_neox_layers_14_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv235 = 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, 8192), dtype="float16")) + lv236 = 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, lv235, gpt_neox_layers_14_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv78 = 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, lv236, lv77), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm30 = R.call_tir(cls.layer_norm1, (lv78, gpt_neox_layers_15_input_layernorm_weight1, gpt_neox_layers_15_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv237 = 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, 6144), dtype="float16")) + reshape60 = R.call_tir(cls.reshape4, (lv237,), out_sinfo=R.Tensor((1, seq_len, 24, 256), dtype="float16")) + reshape61 = R.call_tir(cls.reshape5, (reshape60,), out_sinfo=R.Tensor((seq_len, 24, 256), dtype="float16")) + lv77_1 = 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, 8, 256), dtype="float16")) + reshape62 = R.call_tir(cls.reshape6, (lv77_1,), out_sinfo=R.Tensor((1, seq_len, 8, 256), dtype="float16")) + reshape63 = R.call_tir(cls.reshape7, (reshape62,), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm31 = R.call_tir(cls.layer_norm1, (lv78, gpt_neox_layers_15_post_attention_layernorm_weight1, gpt_neox_layers_15_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv238 = 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, 8192), dtype="float16")) + lv239 = 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, lv238, gpt_neox_layers_15_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv79 = 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, lv239, lv78), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + layer_norm32 = R.call_tir(cls.layer_norm1, (lv79, gpt_neox_final_layer_norm_weight1, gpt_neox_final_layer_norm_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv81 = R.call_tir(cls.index, (layer_norm32,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv4 = R.call_tir(cls.fused_dequantize_fused_NT_matmul14_cast8, (embed_out_q_weight1, embed_out_q_scale1, lv81), 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 \ No newline at end of file