diff --git "a/ir_dump/debug-phase3.py" "b/ir_dump/debug-phase3.py" new file mode 100644--- /dev/null +++ "b/ir_dump/debug-phase3.py" @@ -0,0 +1,3477 @@ +# from tvm.script import ir as I +# from tvm.script import tir as T +# from tvm.script import relax as R + +@I.ir_module +class Module: + I.module_attrs({"system_lib_prefix": "gpt_neox_q0f16_"}) + @T.prim_func + def 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, 64), "float16") + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(pages_handle, (max_num_pages, 2, 8, 16, 64), "float16", offset_factor=1) + page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (B,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) + output = T.match_buffer(output_handle, (B, 8, 64), "float16") + lse = T.match_buffer(lse_handle, (B, 8)) + # with T.block("root"): + sm_scale: T.float32 = T.float32(0.18033688011112042) + for bx in T.thread_binding(B, thread="blockIdx.x"): + for fused_by_bz in T.thread_binding(8, thread="blockIdx.y"): + for ty in T.thread_binding(1, thread="threadIdx.y"): + for tx in T.thread_binding(16, thread="threadIdx.x"): + for tz in T.thread_binding(16, thread="threadIdx.z"): + with T.block("attn"): + T.reads(page_table_indptr[bx:bx + 2], length_info[bx], q_rope_position[bx], Q[bx, fused_by_bz // 8 + ty + fused_by_bz % 8, tx * 4 - 32:tx * 4 - 32 + 68]) + 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((16, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + O_allreduce = T.alloc_buffer((16, 1, 64), scope="shared") + md_allreduce = T.alloc_buffer((16, 1, 2), scope="shared") + S_reduce_local = T.alloc_buffer((1,), scope="local") + t0 = T.alloc_buffer((1,), scope="local") + S_local = T.alloc_buffer((1,), scope="local") + QK_local = T.alloc_buffer((4,), scope="local") + V_local = T.alloc_buffer((4,), "float16", scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_prev = T.alloc_buffer((1,), scope="local") + other_m = T.alloc_buffer((1,), scope="local") + other_d = T.alloc_buffer((1,), scope="local") + exp_mprev = T.alloc_buffer((1,), scope="local") + exp_otherm = T.alloc_buffer((1,), scope="local") + other_o = T.alloc_buffer((4,), scope="local") + st_m = T.alloc_buffer((1,), scope="local") + st_d = T.alloc_buffer((1,), scope="local") + O_local = T.alloc_buffer((4,), scope="local") + by: T.int32 = fused_by_bz % 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 < 32, Q[bx, by + bz + ty, tx * 4 + vec + 32] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) + for iterator in range((kv_chunk_len[0] + 15) // 16): + tile_start_s: T.int32 = tz + ty + tile_start_g: T.int32 = iterator * 16 + tz + ty + for j in range(1): + with T.block("KV_load"): + T.reads() + T.writes() + row_g: T.int32 = tile_start_g + j + if row_g < kv_chunk_len[0]: + seq_offset: T.int32 = row_g + page_no: T.int32 = page_table_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + for vec in T.vectorized(4): + freq = T.float32() + K_smem[tile_start_s + j, tx * 4 + vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, pages[page_no, 0, by, page_offset, tx * 4 + vec + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[batch_idx] + row_g) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, tx * 4 + vec]) + V_smem[tile_start_s + j, tx * 4 + vec] = pages[page_no, 1, by, page_offset, tx * 4 + vec] + else: + for vec in T.vectorized(4): + K_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + V_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + T.tvm_storage_sync("shared") + m_prev[0] = st_m[0] + for j in range(1): + for vec in T.vectorized(4): + QK_local[vec] = T.Cast("float32", Q_local[vec]) * T.Cast("float32", K_smem[tz + j, tx * 4 + vec]) * attn_score_scaling_factor * sm_scale + S_reduce_local[0] = T.float32(0.0) + for vec in T.unroll(4): + S_reduce_local[0] = S_reduce_local[0] + QK_local[vec] + with T.block("block_cross_thread"): + T.reads(S_reduce_local[0]) + T.writes(t0[0]) + T.attr(T.comm_reducer(lambda x0, y0: x0 + y0, [T.float32(0.0)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))) + T.tvm_thread_allreduce(T.uint32(1), S_reduce_local[0], T.bool(True), t0[0], tx) + S_local[j] = T.float32(-50000.0) + if iterator * 16 + tz + j < kv_chunk_len[0]: + S_local[j] = t0[0] + st_m[0] = T.max(st_m[0], S_local[j]) + o_scale: T.float32 = T.exp2(m_prev[0] - st_m[0]) + st_d[0] = st_d[0] * o_scale + for j in range(1): + S_local[j] = T.exp2(S_local[j] - st_m[0]) + st_d[0] = st_d[0] + S_local[j] + for j in T.vectorized(4): + O_local[j] = O_local[j] * o_scale + for j in range(1): + for vec in T.vectorized(4): + V_local[vec] = V_smem[tz + j, tx * 4 + vec] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] + T.Cast("float32", V_local[vec]) * S_local[j] + for vec in T.vectorized(4): + O_allreduce[tz, ty, tx * 4 + vec] = O_local[vec] + md_allreduce[tz, ty, 0] = st_m[0] + md_allreduce[tz, ty, 1] = st_d[0] + T.tvm_storage_sync("shared") + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for j in range(16): + m_prev[0] = st_m[0] + d_prev[0] = st_d[0] + other_m[0] = md_allreduce[j, ty, 0] + other_d[0] = md_allreduce[j, ty, 1] + for vec in T.vectorized(4): + other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] + st_m[0] = T.max(st_m[0], other_m[0]) + st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) + exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) + exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] / st_d[0] + for vec in T.vectorized(4): + output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) + lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) + + @T.prim_func + def batch_decode_paged_kv_sliding_window(_0: T.int32, Q_handle: T.handle, pages_handle: T.handle, page_table_indptr_handle: T.handle, page_table_values_handle: T.handle, var_length_info: T.handle, k_rope_pos_offset_handle: T.handle, q_rope_position_handle: T.handle, output_handle: T.handle, lse_handle: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + B = T.int32(is_size_var=True) + Q = T.match_buffer(Q_handle, (B, 8, 64), "float16") + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(pages_handle, (max_num_pages, 2, 8, 16, 64), "float16", offset_factor=1) + page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (3, B), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) + output = T.match_buffer(output_handle, (B, 8, 64), "float16") + lse = T.match_buffer(lse_handle, (B, 8)) + # with T.block("root"): + sm_scale: T.float32 = T.float32(0.18033688011112042) + for bx in T.thread_binding(B, thread="blockIdx.x"): + for fused_by_bz in T.thread_binding(8, thread="blockIdx.y"): + for ty in T.thread_binding(1, thread="threadIdx.y"): + for tx in T.thread_binding(16, thread="threadIdx.x"): + for tz in T.thread_binding(16, thread="threadIdx.z"): + with T.block("attn"): + T.reads(page_table_indptr[bx:bx + 2], length_info[0:3, bx], q_rope_position[bx], Q[bx, fused_by_bz // 8 + ty + fused_by_bz % 8, tx * 4 - 32:tx * 4 - 32 + 68]) + 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((16, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + O_allreduce = T.alloc_buffer((16, 1, 64), scope="shared") + md_allreduce = T.alloc_buffer((16, 1, 2), scope="shared") + S_reduce_local = T.alloc_buffer((1,), scope="local") + t0 = T.alloc_buffer((1,), scope="local") + S_local = T.alloc_buffer((1,), scope="local") + QK_local = T.alloc_buffer((4,), scope="local") + V_local = T.alloc_buffer((4,), "float16", scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_prev = T.alloc_buffer((1,), scope="local") + other_m = T.alloc_buffer((1,), scope="local") + other_d = T.alloc_buffer((1,), scope="local") + exp_mprev = T.alloc_buffer((1,), scope="local") + exp_otherm = T.alloc_buffer((1,), scope="local") + other_o = T.alloc_buffer((4,), scope="local") + st_m = T.alloc_buffer((1,), scope="local") + st_d = T.alloc_buffer((1,), scope="local") + O_local = T.alloc_buffer((4,), scope="local") + by: T.int32 = fused_by_bz % 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 < 32, Q[bx, by + bz + ty, tx * 4 + vec + 32] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) + for iterator in range((kv_chunk_len[0] + 15) // 16): + tile_start_s: T.int32 = tz + ty + tile_start_g: T.int32 = iterator * 16 + tz + ty + for j in range(1): + with T.block("KV_load"): + T.reads() + T.writes() + row_g: T.int32 = tile_start_g + j + if row_g < kv_chunk_len[0]: + seq_offset: T.int32 = T.if_then_else(row_g < length_info[2, batch_idx], row_g, row_g - length_info[2, batch_idx] + length_info[1, batch_idx]) + page_no: T.int32 = page_table_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + for vec in T.vectorized(4): + freq = T.float32() + K_smem[tile_start_s + j, tx * 4 + vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 32, pages[page_no, 0, by, page_offset, tx * 4 + vec + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[batch_idx] + row_g) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, tx * 4 + vec]) + V_smem[tile_start_s + j, tx * 4 + vec] = pages[page_no, 1, by, page_offset, tx * 4 + vec] + else: + for vec in T.vectorized(4): + K_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + V_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) + T.tvm_storage_sync("shared") + m_prev[0] = st_m[0] + for j in range(1): + for vec in T.vectorized(4): + QK_local[vec] = T.Cast("float32", Q_local[vec]) * T.Cast("float32", K_smem[tz + j, tx * 4 + vec]) * attn_score_scaling_factor * sm_scale + S_reduce_local[0] = T.float32(0.0) + for vec in T.unroll(4): + S_reduce_local[0] = S_reduce_local[0] + QK_local[vec] + with T.block("block_cross_thread"): + T.reads(S_reduce_local[0]) + T.writes(t0[0]) + T.attr(T.comm_reducer(lambda x0, y0: x0 + y0, [T.float32(0.0)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))) + T.tvm_thread_allreduce(T.uint32(1), S_reduce_local[0], T.bool(True), t0[0], tx) + S_local[j] = T.float32(-50000.0) + if iterator * 16 + tz + j < kv_chunk_len[0]: + S_local[j] = t0[0] + st_m[0] = T.max(st_m[0], S_local[j]) + o_scale: T.float32 = T.exp2(m_prev[0] - st_m[0]) + st_d[0] = st_d[0] * o_scale + for j in range(1): + S_local[j] = T.exp2(S_local[j] - st_m[0]) + st_d[0] = st_d[0] + S_local[j] + for j in T.vectorized(4): + O_local[j] = O_local[j] * o_scale + for j in range(1): + for vec in T.vectorized(4): + V_local[vec] = V_smem[tz + j, tx * 4 + vec] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] + T.Cast("float32", V_local[vec]) * S_local[j] + for vec in T.vectorized(4): + O_allreduce[tz, ty, tx * 4 + vec] = O_local[vec] + md_allreduce[tz, ty, 0] = st_m[0] + md_allreduce[tz, ty, 1] = st_d[0] + T.tvm_storage_sync("shared") + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for j in range(16): + m_prev[0] = st_m[0] + d_prev[0] = st_d[0] + other_m[0] = md_allreduce[j, ty, 0] + other_d[0] = md_allreduce[j, ty, 1] + for vec in T.vectorized(4): + other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] + st_m[0] = T.max(st_m[0], other_m[0]) + st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) + exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) + exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] / st_d[0] + for vec in T.vectorized(4): + output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) + lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) + + @T.prim_func + def batch_prefill_paged_kv(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + total_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (total_len, 8, 64), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 8, 16, 64), "float16", offset_factor=1) + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 8, 64), "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((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 16), scope="shared") + S_local = T.alloc_buffer((32, 16), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(4): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) + j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 15) // 16): + L_kv_start: T.int32 = iterator_1 * 16 + for lz_ly_fused_0 in range(2): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + freq = T.float32() + K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, pages[page_no, 0, by, page_offset, j + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, j]) + else: + K_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_ly_fused_0 in range(2): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:16, 0:64]) + T.writes(S_local[0:32, 0:16]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + k = T.axis.reduce(64, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 2): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(16): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) + T.writes(S_smem[row, 0:16]) + for j in range(16): + if row < 32: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(16): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:16], V_smem[0:16, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(2, 8, 4, 4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + k = T.axis.reduce(16, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def batch_prefill_paged_kv_sliding_window(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + total_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (total_len, 8, 64), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 8, 16, 64), "float16", offset_factor=1) + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (3, batch_size), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 8, 64), "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((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 16), scope="shared") + S_local = T.alloc_buffer((32, 16), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[0, b_idx] - length_info[1, b_idx] + length_info[2, b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(4): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) + j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 15) // 16): + L_kv_start: T.int32 = iterator_1 * 16 + for lz_ly_fused_0 in range(2): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = T.if_then_else(cur_L < length_info[2, b_idx], cur_L, cur_L - length_info[2, b_idx] + length_info[1, b_idx]) + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + freq = T.float32() + K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, pages[page_no, 0, by, page_offset, j + 32] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), pages[page_no, 0, by, page_offset, j]) + else: + K_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_ly_fused_0 in range(2): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = T.if_then_else(cur_L < length_info[2, b_idx], cur_L, cur_L - length_info[2, b_idx] + length_info[1, b_idx]) + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:16, 0:64]) + T.writes(S_local[0:32, 0:16]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + k = T.axis.reduce(64, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 2): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(16): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) + T.writes(S_smem[row, 0:16]) + for j in range(16): + if row < 32: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(16): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:16], V_smem[0:16, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(2, 8, 4, 4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + k = T.axis.reduce(16, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def batch_prefill_ragged_kv(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_k_rope_pos_offset: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + qo_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (qo_len, 8, 64), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + kv_len = T.int32(is_size_var=True) + k = T.match_buffer(var_k, (kv_len, 8, 64), "float16") + v = T.match_buffer(var_v, (kv_len, 8, 64), "float16") + kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (qo_len, 8, 64), "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((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((64, 32), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 32), scope="shared") + S_local = T.alloc_buffer((32, 32), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + q_indptr_val: T.int32 = q_indptr[b_idx] + LH_start: T.int32 = tile_id[0] * 32 + kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1_0 * 4 + lj_1_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_1_lj_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_1_lj_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_0, lj_0_0 in T.grid(1, 2): + for lj_1 in T.vectorized(8): + with T.block("Q_load"): + i = T.axis.spatial(32, li_0 * 32 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) // 4) + j = T.axis.spatial(64, lj_0_0 * 32 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) % 4 * 8 + lj_1) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 31) // 32): + L_kv_start: T.int32 = iterator_1 * 32 + L_kv_base: T.int32 = kv_indptr[b_idx] + for lz_1_ly_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for lz_1_ly_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lz_0, ly_0_0 in T.grid(1, 2): + for ly_1 in T.vectorized(8): + with T.block("K_load"): + i = T.axis.spatial(32, lz_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 4) + j = T.axis.spatial(64, ly_0_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 4 * 8 + ly_1) + T.reads(kv_chunk_len[0], k_rope_pos_offset[b_idx], k[L_kv_base + L_kv_start + i, by, j - 32:j - 32 + 65]) + T.writes(K_smem[j, i]) + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + freq = T.float32() + K_smem[j, i] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", k[L_kv_base + cur_L, by, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, k[L_kv_base + cur_L, by, j + 32] * T.float16(-1.0), k[L_kv_base + cur_L, by, j - 32]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), k[L_kv_base + cur_L, by, j]) + else: + K_smem[j, i] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_1_ly_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for lz_1_ly_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lz_0, ly_0_0 in T.grid(1, 2): + for ly_1 in T.vectorized(8): + with T.block("V_load"): + i = T.axis.spatial(32, lz_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 4) + j = T.axis.spatial(64, ly_0_0 * 32 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 4 * 8 + ly_1) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + V_smem[i, j] = v[L_kv_base + cur_L, by, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:64, 0:32]) + T.writes(S_local[0:32, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init in T.unroll(2): + for lj_1_0_init in T.unroll(1): + for lj_1_1_init in T.vectorized(4): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_0_init * 4 + lj_1_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0 in range(4): + for li_1 in T.unroll(2): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + for lk_1 in range(16): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1_0 * 4 + lj_1_1) + k_1 = T.axis.reduce(64, lk_0 * 16 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k_1], K_smem[k_1, j]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k_1]) * T.Cast("float32", K_smem[k_1, j]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(2): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1_0 * 4 + lj_1_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 32: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init in T.unroll(4): + for lj_1_0_init in T.unroll(1): + for lj_1_1_init in T.vectorized(4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_0_init * 4 + lj_1_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1 in T.grid(2, 16): + for li_1 in T.unroll(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1_0 * 4 + lj_1_1) + k_1 = T.axis.reduce(32, lk_0 * 16 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k_1], V_smem[k_1, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k_1] * T.Cast("float32", V_smem[k_1, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1_0 * 4 + lj_1_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 8 + + @T.prim_func + def batch_tree_attn(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_mn_indptr: T.handle, var_mask: T.handle, var_output: T.handle, var_lse: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32, batch_size: T.int32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + qo_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (qo_len, 8, 64), "float16") + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + kv_len = T.int32(is_size_var=True) + k = T.match_buffer(var_k, (kv_len, 8, 64), "float16") + v = T.match_buffer(var_v, (kv_len, 8, 64), "float16") + kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) + mn_indptr = T.match_buffer(var_mn_indptr, (batch_size + 1,), "int32", offset_factor=1) + tree_size = T.int32(is_size_var=True) + mask = T.match_buffer(var_mask, (tree_size, 2), "int32", offset_factor=1) + output = T.match_buffer(var_output, (qo_len, 8, 64), "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((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 32), scope="shared") + S_local = T.alloc_buffer((32, 32), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(4): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) + j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 31) // 32): + L_kv_start: T.int32 = iterator_1 * 32 + L_kv_base: T.int32 = kv_indptr[b_idx] + for lz_ly_fused_0 in range(4): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("KV_load"): + i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_base + L_kv_start + i + if L_kv_start + i < kv_chunk_len[0]: + freq = T.float32() + K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", k[cur_L, by, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, k[cur_L, by, j + 32] * T.float16(-1.0), k[cur_L, by, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), k[cur_L, by, j]) + V_smem[i, j] = v[cur_L, by, j] + else: + K_smem[i, j] = T.float16(0.0) + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:32, 0:64]) + T.writes(S_local[0:32, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 4): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 4, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + k_1 = T.axis.reduce(64, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k_1], K_smem[j, k_1]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k_1]) * T.Cast("float32", K_smem[j, k_1]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 4): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 32: + row_: T.int32 = LH_start + row + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(4, 8, 4, 4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + k_1 = T.axis.reduce(32, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k_1], V_smem[k_1, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k_1] * T.Cast("float32", V_smem[k_1, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def chunk_lse(var_A: T.handle, var_temperature: T.handle, var_chunked_sum: T.handle, var_chunked_max: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.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"): + A_pad = T.alloc_buffer((batch_size, num_chunks, T.int64(4096))) + temp_max = T.alloc_buffer((batch_size, num_chunks)) + temp_sum = T.alloc_buffer((batch_size, num_chunks)) + for l0, l1, l2 in T.grid(batch_size, num_chunks, T.int64(4096)): + with T.block("pad"): + v0, v1, v2 = T.axis.remap("SSS", [l0, l1, l2]) + T.reads(temperature[v0], A[v0, v1 * T.int64(4096) + v2]) + T.writes(A_pad[v0, v1, v2]) + A_pad[v0, v1, v2] = 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 l0, l1, l2 in T.grid(batch_size, num_chunks, T.int64(4096)): + with T.block("max"): + v0, v1, v2 = T.axis.remap("SSR", [l0, l1, l2]) + T.reads(A_pad[v0, v1, v2]) + T.writes(temp_max[v0, v1]) + with T.init(): + temp_max[v0, v1] = T.float32(-340282346638528859811704183484516925440.0) + temp_max[v0, v1] = T.max(temp_max[v0, v1], A_pad[v0, v1, v2]) + for l0, l1, l2 in T.grid(batch_size, num_chunks, T.int64(4096)): + with T.block("sum_exp"): + v0, v1, v2 = T.axis.remap("SSR", [l0, l1, l2]) + T.reads(temperature[v0], A_pad[v0, v1, v2], temp_max[v0, v1]) + T.writes(temp_sum[v0, v1]) + with T.init(): + temp_sum[v0, v1] = T.float32(0.0) + temp_sum[v0, v1] = temp_sum[v0, v1] + T.if_then_else(v1 * T.int64(4096) + v2 < vocab_size, T.Select(temperature[v0] > T.float32(1.0000000000000001e-05), T.exp(A_pad[v0, v1, v2] - temp_max[v0, v1]), T.Cast("float32", A_pad[v0, v1, v2] == temp_max[v0, v1])), T.float32(0.0)) + for l0, l1, l2 in T.grid(batch_size, num_chunks, T.int64(1)): + with T.block("log"): + v0, v1, v2 = T.axis.remap("SSS", [l0, l1, l2]) + T.reads(temperature[v0], temp_sum[v0, v1], temp_max[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[v0, v1]), temp_sum[v0, v1]) + chunked_max[v0, v1] = temp_max[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, 64), "float16", offset_factor=1) + copy_length_indptr = T.match_buffer(var_copy_length_indptr, (batch_size + 1,), "int32", offset_factor=1) + total_copy_length = T.int32() + copy_src_dst_pos = T.match_buffer(var_copy_src_dst_pos, (2, total_copy_length), "int32", offset_factor=1) + with T.block("root"): + T.reads() + T.writes() + for bhd_o in T.thread_binding(batch_size * 2, thread="blockIdx.x"): + for bhd_i in T.thread_binding(256, thread="threadIdx.x"): + b: T.int32 = (bhd_o * 256 + bhd_i) // 512 + h: T.int32 = (bhd_o * 256 + bhd_i) // 64 % 8 + d: T.int32 = (bhd_o * 256 + bhd_i) % 64 + if bhd_o * 256 + bhd_i < batch_size * 8 * 64: + for i in range(copy_length_indptr[b + 1] - copy_length_indptr[b]): + src_pos: T.int32 = copy_src_dst_pos[0, copy_length_indptr[b] + i] + dst_pos: T.int32 = copy_src_dst_pos[1, copy_length_indptr[b] + i] + pages[dst_pos // 16, 0, h, dst_pos % 16, d] = pages[src_pos // 16, 0, h, src_pos % 16, d] + pages[dst_pos // 16, 1, h, dst_pos % 16, d] = pages[src_pos // 16, 1, h, src_pos % 16, d] + + @T.prim_func + def copy_single_page(var_pages: T.handle, src_page_id: T.int64, tgt_page_id: T.int64, copy_length: T.int64): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + num_pages, page_size = T.int32(), T.int64() + pages = T.match_buffer(var_pages, (num_pages, 2, 8, page_size, 64), "float16", offset_factor=1) + # with T.block("root"): + for b in T.thread_binding(copy_length * T.int64(2), 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(64)))) + vp = T.axis.spatial(copy_length, (b * T.int64(256) + T.Cast("int64", t)) % (copy_length * T.int64(64)) // T.int64(64)) + vd = T.axis.spatial(64, T.Cast("int32", (b * T.int64(256) + T.Cast("int64", t)) % T.int64(64))) + T.where(b * T.int64(256) + T.Cast("int64", t) < copy_length * T.int64(8) * T.int64(64)) + T.reads(pages[src_page_id, 0:2, vh, vp, vd]) + T.writes(pages[tgt_page_id, 0:2, vh, vp, vd]) + pages[tgt_page_id, 0, vh, vp, vd] = pages[src_page_id, 0, vh, vp, vd] + pages[tgt_page_id, 1, vh, vp, vd] = pages[src_page_id, 1, vh, vp, vd] + + @T.prim_func(private=True) + def fused_NT_matmul10_add10(layer_norm13: T.Buffer((T.int64(1), T.int64(1), T.int64(512)), "float16"), gpt_neox_layers_0_attention_query_key_value_weight2: T.Buffer((T.int64(1536), T.int64(512)), "float16"), gpt_neox_layers_0_attention_query_key_value_bias2: T.Buffer((T.int64(1536),), "float16"), T_add_intermediate: T.Buffer((T.int64(1), T.int64(1), T.int64(1536)), "float16")): + T.func_attr({"tir.noalias": T.bool(True)}) + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(1536)), "float16") + for i0, i1, i2, k in T.grid(T.int64(1), T.int64(1), T.int64(1536), T.int64(512)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(layer_norm13[v_i0, v_i1, v_k], gpt_neox_layers_0_attention_query_key_value_weight2[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float16(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + layer_norm13[v_i0, v_i1, v_k] * gpt_neox_layers_0_attention_query_key_value_weight2[v_i2, v_k] + for ax0, ax1, ax2 in T.grid(T.int64(1), T.int64(1), T.int64(1536)): + with T.block("T_add"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(NT_matmul_intermediate[v_ax0, v_ax1, v_ax2], gpt_neox_layers_0_attention_query_key_value_bias2[v_ax2]) + T.writes(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T_add_intermediate[v_ax0, v_ax1, v_ax2] = NT_matmul_intermediate[v_ax0, v_ax1, v_ax2] + gpt_neox_layers_0_attention_query_key_value_bias2[v_ax2] + + @T.prim_func(private=True) + def fused_NT_matmul11_add11_add14_add14(reshape27: T.Buffer((T.int64(1), T.int64(1), T.int64(512)), "float16"), gpt_neox_layers_0_attention_dense_weight2: T.Buffer((T.int64(512), T.int64(512)), "float16"), gpt_neox_layers_0_attention_dense_bias2: T.Buffer((T.int64(512),), "float16"), astype14: T.Buffer((T.int64(1), T.int64(1), T.int64(512)), "float16"), input_embed: T.Buffer((T.int64(1), T.int64(1), T.int64(512)), "float16"), T_add_intermediate_1_2: T.Buffer((T.int64(1), T.int64(1), T.int64(512)), "float16")): + T.func_attr({"tir.noalias": T.bool(True)}) + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(512)), "float16") + T_add_intermediate = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(512)), "float16") + T_add_intermediate_1 = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(512)), "float16") + for i0, i1, i2, k in T.grid(T.int64(1), T.int64(1), T.int64(512), T.int64(512)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(reshape27[v_i0, v_i1, v_k], gpt_neox_layers_0_attention_dense_weight2[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float16(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + reshape27[v_i0, v_i1, v_k] * gpt_neox_layers_0_attention_dense_weight2[v_i2, v_k] + for ax0, ax1, ax2 in T.grid(T.int64(1), T.int64(1), T.int64(512)): + with T.block("T_add"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(NT_matmul_intermediate[v_ax0, v_ax1, v_ax2], gpt_neox_layers_0_attention_dense_bias2[v_ax2]) + T.writes(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T_add_intermediate[v_ax0, v_ax1, v_ax2] = NT_matmul_intermediate[v_ax0, v_ax1, v_ax2] + gpt_neox_layers_0_attention_dense_bias2[v_ax2] + for ax0, ax1, ax2 in T.grid(T.int64(1), T.int64(1), T.int64(512)): + with T.block("T_add_1"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(astype14[v_ax0, v_ax1, v_ax2], T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T.writes(T_add_intermediate_1[v_ax0, v_ax1, v_ax2]) + T_add_intermediate_1[v_ax0, v_ax1, v_ax2] = astype14[v_ax0, v_ax1, v_ax2] + T_add_intermediate[v_ax0, v_ax1, v_ax2] + for ax0, ax1, ax2 in T.grid(T.int64(1), T.int64(1), T.int64(512)): + with T.block("T_add_2"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_add_intermediate_1[v_ax0, v_ax1, v_ax2], input_embed[v_ax0, v_ax1, v_ax2]) + T.writes(T_add_intermediate_1_2[v_ax0, v_ax1, v_ax2]) + T_add_intermediate_1_2[v_ax0, v_ax1, v_ax2] = T_add_intermediate_1[v_ax0, v_ax1, v_ax2] + input_embed[v_ax0, v_ax1, v_ax2] + + @T.prim_func(private=True) + def fused_NT_matmul12_add12_gelu2_cast6(layer_norm14: T.Buffer((T.int64(1), T.int64(1), T.int64(512)), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_weight2: T.Buffer((T.int64(2048), T.int64(512)), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: T.Buffer((T.int64(2048),), "float32"), compute_intermediate: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16")): + T.func_attr({"tir.noalias": T.bool(True)}) + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(2048))) + T_add_intermediate = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(2048))) + T_multiply = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(2048))) + compute = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(2048))) + T_multiply_1 = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(2048))) + T_add = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(2048))) + T_multiply_intermediate = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(2048))) + for i0, i1, i2, k in T.grid(T.int64(1), T.int64(1), T.int64(2048), T.int64(512)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(layer_norm14[v_i0, v_i1, v_k], gpt_neox_layers_0_mlp_dense_h_to_4h_weight2[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float32(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + T.Cast("float32", layer_norm14[v_i0, v_i1, v_k]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_h_to_4h_weight2[v_i2, v_k]) + for ax0, ax1, ax2 in T.grid(T.int64(1), T.int64(1), T.int64(2048)): + with T.block("T_add"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(NT_matmul_intermediate[v_ax0, v_ax1, v_ax2], gpt_neox_layers_0_mlp_dense_h_to_4h_bias2[v_ax2]) + T.writes(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T_add_intermediate[v_ax0, v_ax1, v_ax2] = NT_matmul_intermediate[v_ax0, v_ax1, v_ax2] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias2[v_ax2] + for ax0, ax1, ax2 in T.grid(T.int64(1), T.int64(1), T.int64(2048)): + with T.block("T_multiply"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T.writes(T_multiply[v_ax0, v_ax1, v_ax2]) + T_multiply[v_ax0, v_ax1, v_ax2] = T_add_intermediate[v_ax0, v_ax1, v_ax2] * T.float32(0.70710678118654757) + for i0, i1, i2 in T.grid(T.int64(1), T.int64(1), T.int64(2048)): + with T.block("compute"): + v_i0, v_i1, v_i2 = T.axis.remap("SSS", [i0, i1, i2]) + T.reads(T_multiply[v_i0, v_i1, v_i2]) + T.writes(compute[v_i0, v_i1, v_i2]) + compute[v_i0, v_i1, v_i2] = T.erf(T_multiply[v_i0, v_i1, v_i2]) + for ax0, ax1, ax2 in T.grid(T.int64(1), T.int64(1), T.int64(2048)): + with T.block("T_multiply_1"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(compute[v_ax0, v_ax1, v_ax2]) + T.writes(T_multiply_1[v_ax0, v_ax1, v_ax2]) + T_multiply_1[v_ax0, v_ax1, v_ax2] = compute[v_ax0, v_ax1, v_ax2] * T.float32(0.5) + for ax0, ax1, ax2 in T.grid(T.int64(1), T.int64(1), T.int64(2048)): + with T.block("T_add_1"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_multiply_1[v_ax0, v_ax1, v_ax2]) + T.writes(T_add[v_ax0, v_ax1, v_ax2]) + T_add[v_ax0, v_ax1, v_ax2] = T.float32(0.5) + T_multiply_1[v_ax0, v_ax1, v_ax2] + for ax0, ax1, ax2 in T.grid(T.int64(1), T.int64(1), T.int64(2048)): + with T.block("T_multiply_2"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_add_intermediate[v_ax0, v_ax1, v_ax2], T_add[v_ax0, v_ax1, v_ax2]) + T.writes(T_multiply_intermediate[v_ax0, v_ax1, v_ax2]) + T_multiply_intermediate[v_ax0, v_ax1, v_ax2] = T_add_intermediate[v_ax0, v_ax1, v_ax2] * T_add[v_ax0, v_ax1, v_ax2] + for i0, i1, i2 in T.grid(T.int64(1), T.int64(1), T.int64(2048)): + with T.block("compute_1"): + v_i0, v_i1, v_i2 = T.axis.remap("SSS", [i0, i1, i2]) + T.reads(T_multiply_intermediate[v_i0, v_i1, v_i2]) + T.writes(compute_intermediate[v_i0, v_i1, v_i2]) + compute_intermediate[v_i0, v_i1, v_i2] = T.Cast("float16", T_multiply_intermediate[v_i0, v_i1, v_i2]) + + @T.prim_func(private=True) + def fused_NT_matmul13_add13_cast7(astype13: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_weight2: T.Buffer((T.int64(512), T.int64(2048)), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: T.Buffer((T.int64(512),), "float32"), compute_intermediate: T.Buffer((T.int64(1), T.int64(1), T.int64(512)), "float16")): + T.func_attr({"tir.noalias": T.bool(True)}) + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(512))) + T_add_intermediate = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(512))) + for i0, i1, i2, k in T.grid(T.int64(1), T.int64(1), T.int64(512), T.int64(2048)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(astype13[v_i0, v_i1, v_k], gpt_neox_layers_0_mlp_dense_4h_to_h_weight2[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float32(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + T.Cast("float32", astype13[v_i0, v_i1, v_k]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_4h_to_h_weight2[v_i2, v_k]) + for ax0, ax1, ax2 in T.grid(T.int64(1), T.int64(1), T.int64(512)): + with T.block("T_add"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(NT_matmul_intermediate[v_ax0, v_ax1, v_ax2], gpt_neox_layers_0_mlp_dense_4h_to_h_bias2[v_ax2]) + T.writes(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T_add_intermediate[v_ax0, v_ax1, v_ax2] = NT_matmul_intermediate[v_ax0, v_ax1, v_ax2] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias2[v_ax2] + for i0, i1, i2 in T.grid(T.int64(1), T.int64(1), T.int64(512)): + with T.block("compute"): + v_i0, v_i1, v_i2 = T.axis.remap("SSS", [i0, i1, i2]) + T.reads(T_add_intermediate[v_i0, v_i1, v_i2]) + T.writes(compute_intermediate[v_i0, v_i1, v_i2]) + compute_intermediate[v_i0, v_i1, v_i2] = T.Cast("float16", T_add_intermediate[v_i0, v_i1, v_i2]) + + @T.prim_func(private=True) + def fused_NT_matmul14_cast8(layer_norm25: T.Buffer((T.int64(1), T.int64(1), T.int64(512)), "float16"), p_embed_out_weight2: T.handle, p_output0: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + vocab_size = T.int64() + embed_out_weight2 = T.match_buffer(p_embed_out_weight2, (vocab_size, T.int64(512)), "float16") + compute_intermediate = T.match_buffer(p_output0, (T.int64(1), T.int64(1), vocab_size)) + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((T.int64(1), T.int64(1), vocab_size), "float16") + for i0, i1, i2, k in T.grid(T.int64(1), T.int64(1), vocab_size, T.int64(512)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(layer_norm25[v_i0, v_i1, v_k], embed_out_weight2[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float16(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + layer_norm25[v_i0, v_i1, v_k] * embed_out_weight2[v_i2, v_k] + for i0, i1, i2 in T.grid(T.int64(1), T.int64(1), vocab_size): + with T.block("compute"): + v_i0, v_i1, v_i2 = T.axis.remap("SSS", [i0, i1, i2]) + T.reads(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + T.writes(compute_intermediate[v_i0, v_i1, v_i2]) + compute_intermediate[v_i0, v_i1, v_i2] = T.Cast("float32", NT_matmul_intermediate[v_i0, v_i1, v_i2]) + + @T.prim_func(private=True) + def fused_NT_matmul1_add1_add4_add4(p_reshape75: T.handle, gpt_neox_layers_0_attention_dense_weight4: T.Buffer((T.int64(512), T.int64(512)), "float16"), gpt_neox_layers_0_attention_dense_bias4: T.Buffer((T.int64(512),), "float16"), p_astype40: T.handle, p_input_embeds: T.handle, p_output0: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + batch_size = T.int64() + reshape75 = T.match_buffer(p_reshape75, (batch_size, T.int64(1), T.int64(512)), "float16") + astype40 = T.match_buffer(p_astype40, (batch_size, T.int64(1), T.int64(512)), "float16") + input_embeds = T.match_buffer(p_input_embeds, (batch_size, T.int64(1), T.int64(512)), "float16") + T_add_intermediate_1_2 = T.match_buffer(p_output0, (batch_size, T.int64(1), T.int64(512)), "float16") + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((batch_size, T.int64(1), T.int64(512)), "float16") + T_add_intermediate = T.alloc_buffer((batch_size, T.int64(1), T.int64(512)), "float16") + T_add_intermediate_1 = T.alloc_buffer((batch_size, T.int64(1), T.int64(512)), "float16") + for i0, i1, i2, k in T.grid(batch_size, T.int64(1), T.int64(512), T.int64(512)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(reshape75[v_i0, v_i1, v_k], gpt_neox_layers_0_attention_dense_weight4[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float16(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + reshape75[v_i0, v_i1, v_k] * gpt_neox_layers_0_attention_dense_weight4[v_i2, v_k] + for ax0, ax1, ax2 in T.grid(batch_size, T.int64(1), T.int64(512)): + with T.block("T_add"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(NT_matmul_intermediate[v_ax0, v_ax1, v_ax2], gpt_neox_layers_0_attention_dense_bias4[v_ax2]) + T.writes(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T_add_intermediate[v_ax0, v_ax1, v_ax2] = NT_matmul_intermediate[v_ax0, v_ax1, v_ax2] + gpt_neox_layers_0_attention_dense_bias4[v_ax2] + for ax0, ax1, ax2 in T.grid(batch_size, T.int64(1), T.int64(512)): + with T.block("T_add_1"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(astype40[v_ax0, v_ax1, v_ax2], T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T.writes(T_add_intermediate_1[v_ax0, v_ax1, v_ax2]) + T_add_intermediate_1[v_ax0, v_ax1, v_ax2] = astype40[v_ax0, v_ax1, v_ax2] + T_add_intermediate[v_ax0, v_ax1, v_ax2] + for ax0, ax1, ax2 in T.grid(batch_size, T.int64(1), T.int64(512)): + with T.block("T_add_2"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_add_intermediate_1[v_ax0, v_ax1, v_ax2], input_embeds[v_ax0, v_ax1, v_ax2]) + T.writes(T_add_intermediate_1_2[v_ax0, v_ax1, v_ax2]) + T_add_intermediate_1_2[v_ax0, v_ax1, v_ax2] = T_add_intermediate_1[v_ax0, v_ax1, v_ax2] + input_embeds[v_ax0, v_ax1, v_ax2] + + @T.prim_func(private=True) + def fused_NT_matmul2_add2_gelu_cast(p_layer_norm40: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_weight4: T.Buffer((T.int64(2048), T.int64(512)), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: T.Buffer((T.int64(2048),), "float32"), p_output0: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + batch_size = T.int64() + layer_norm40 = T.match_buffer(p_layer_norm40, (batch_size, T.int64(1), T.int64(512)), "float16") + compute_intermediate = T.match_buffer(p_output0, (batch_size, T.int64(1), T.int64(2048)), "float16") + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((batch_size, T.int64(1), T.int64(2048))) + T_add_intermediate = T.alloc_buffer((batch_size, T.int64(1), T.int64(2048))) + T_multiply = T.alloc_buffer((batch_size, T.int64(1), T.int64(2048))) + compute = T.alloc_buffer((batch_size, T.int64(1), T.int64(2048))) + T_multiply_1 = T.alloc_buffer((batch_size, T.int64(1), T.int64(2048))) + T_add = T.alloc_buffer((batch_size, T.int64(1), T.int64(2048))) + T_multiply_intermediate = T.alloc_buffer((batch_size, T.int64(1), T.int64(2048))) + for i0, i1, i2, k in T.grid(batch_size, T.int64(1), T.int64(2048), T.int64(512)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(layer_norm40[v_i0, v_i1, v_k], gpt_neox_layers_0_mlp_dense_h_to_4h_weight4[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float32(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + T.Cast("float32", layer_norm40[v_i0, v_i1, v_k]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_h_to_4h_weight4[v_i2, v_k]) + for ax0, ax1, ax2 in T.grid(batch_size, T.int64(1), T.int64(2048)): + with T.block("T_add"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(NT_matmul_intermediate[v_ax0, v_ax1, v_ax2], gpt_neox_layers_0_mlp_dense_h_to_4h_bias4[v_ax2]) + T.writes(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T_add_intermediate[v_ax0, v_ax1, v_ax2] = NT_matmul_intermediate[v_ax0, v_ax1, v_ax2] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias4[v_ax2] + for ax0, ax1, ax2 in T.grid(batch_size, T.int64(1), T.int64(2048)): + with T.block("T_multiply"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T.writes(T_multiply[v_ax0, v_ax1, v_ax2]) + T_multiply[v_ax0, v_ax1, v_ax2] = T_add_intermediate[v_ax0, v_ax1, v_ax2] * T.float32(0.70710678118654757) + for i0, i1, i2 in T.grid(batch_size, T.int64(1), T.int64(2048)): + with T.block("compute"): + v_i0, v_i1, v_i2 = T.axis.remap("SSS", [i0, i1, i2]) + T.reads(T_multiply[v_i0, v_i1, v_i2]) + T.writes(compute[v_i0, v_i1, v_i2]) + compute[v_i0, v_i1, v_i2] = T.erf(T_multiply[v_i0, v_i1, v_i2]) + for ax0, ax1, ax2 in T.grid(batch_size, T.int64(1), T.int64(2048)): + with T.block("T_multiply_1"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(compute[v_ax0, v_ax1, v_ax2]) + T.writes(T_multiply_1[v_ax0, v_ax1, v_ax2]) + T_multiply_1[v_ax0, v_ax1, v_ax2] = compute[v_ax0, v_ax1, v_ax2] * T.float32(0.5) + for ax0, ax1, ax2 in T.grid(batch_size, T.int64(1), T.int64(2048)): + with T.block("T_add_1"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_multiply_1[v_ax0, v_ax1, v_ax2]) + T.writes(T_add[v_ax0, v_ax1, v_ax2]) + T_add[v_ax0, v_ax1, v_ax2] = T.float32(0.5) + T_multiply_1[v_ax0, v_ax1, v_ax2] + for ax0, ax1, ax2 in T.grid(batch_size, T.int64(1), T.int64(2048)): + with T.block("T_multiply_2"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_add_intermediate[v_ax0, v_ax1, v_ax2], T_add[v_ax0, v_ax1, v_ax2]) + T.writes(T_multiply_intermediate[v_ax0, v_ax1, v_ax2]) + T_multiply_intermediate[v_ax0, v_ax1, v_ax2] = T_add_intermediate[v_ax0, v_ax1, v_ax2] * T_add[v_ax0, v_ax1, v_ax2] + for i0, i1, i2 in T.grid(batch_size, T.int64(1), T.int64(2048)): + with T.block("compute_1"): + v_i0, v_i1, v_i2 = T.axis.remap("SSS", [i0, i1, i2]) + T.reads(T_multiply_intermediate[v_i0, v_i1, v_i2]) + T.writes(compute_intermediate[v_i0, v_i1, v_i2]) + compute_intermediate[v_i0, v_i1, v_i2] = T.Cast("float16", T_multiply_intermediate[v_i0, v_i1, v_i2]) + + @T.prim_func(private=True) + def fused_NT_matmul3_add3_cast1(p_astype39: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_weight4: T.Buffer((T.int64(512), T.int64(2048)), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: T.Buffer((T.int64(512),), "float32"), p_output0: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + batch_size = T.int64() + astype39 = T.match_buffer(p_astype39, (batch_size, T.int64(1), T.int64(2048)), "float16") + compute_intermediate = T.match_buffer(p_output0, (batch_size, T.int64(1), T.int64(512)), "float16") + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((batch_size, T.int64(1), T.int64(512))) + T_add_intermediate = T.alloc_buffer((batch_size, T.int64(1), T.int64(512))) + for i0, i1, i2, k in T.grid(batch_size, T.int64(1), T.int64(512), T.int64(2048)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(astype39[v_i0, v_i1, v_k], gpt_neox_layers_0_mlp_dense_4h_to_h_weight4[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float32(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + T.Cast("float32", astype39[v_i0, v_i1, v_k]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_4h_to_h_weight4[v_i2, v_k]) + for ax0, ax1, ax2 in T.grid(batch_size, T.int64(1), T.int64(512)): + with T.block("T_add"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(NT_matmul_intermediate[v_ax0, v_ax1, v_ax2], gpt_neox_layers_0_mlp_dense_4h_to_h_bias4[v_ax2]) + T.writes(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T_add_intermediate[v_ax0, v_ax1, v_ax2] = NT_matmul_intermediate[v_ax0, v_ax1, v_ax2] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias4[v_ax2] + for i0, i1, i2 in T.grid(batch_size, T.int64(1), T.int64(512)): + with T.block("compute"): + v_i0, v_i1, v_i2 = T.axis.remap("SSS", [i0, i1, i2]) + T.reads(T_add_intermediate[v_i0, v_i1, v_i2]) + T.writes(compute_intermediate[v_i0, v_i1, v_i2]) + compute_intermediate[v_i0, v_i1, v_i2] = T.Cast("float16", T_add_intermediate[v_i0, v_i1, v_i2]) + + @T.prim_func(private=True) + def fused_NT_matmul4_cast2(p_layer_norm51: T.handle, p_embed_out_weight4: T.handle, p_output0: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + batch_size = T.int64() + layer_norm51 = T.match_buffer(p_layer_norm51, (batch_size, T.int64(1), T.int64(512)), "float16") + vocab_size = T.int64() + embed_out_weight4 = T.match_buffer(p_embed_out_weight4, (vocab_size, T.int64(512)), "float16") + compute_intermediate = T.match_buffer(p_output0, (batch_size, T.int64(1), vocab_size)) + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((batch_size, T.int64(1), vocab_size), "float16") + for i0, i1, i2, k in T.grid(batch_size, T.int64(1), vocab_size, T.int64(512)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(layer_norm51[v_i0, v_i1, v_k], embed_out_weight4[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float16(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + layer_norm51[v_i0, v_i1, v_k] * embed_out_weight4[v_i2, v_k] + for i0, i1, i2 in T.grid(batch_size, T.int64(1), vocab_size): + with T.block("compute"): + v_i0, v_i1, v_i2 = T.axis.remap("SSS", [i0, i1, i2]) + T.reads(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + T.writes(compute_intermediate[v_i0, v_i1, v_i2]) + compute_intermediate[v_i0, v_i1, v_i2] = T.Cast("float32", NT_matmul_intermediate[v_i0, v_i1, v_i2]) + + @T.prim_func(private=True) + def fused_NT_matmul5_add5(p_layer_norm26: T.handle, gpt_neox_layers_0_attention_query_key_value_weight3: T.Buffer((T.int64(1536), T.int64(512)), "float16"), gpt_neox_layers_0_attention_query_key_value_bias3: T.Buffer((T.int64(1536),), "float16"), p_output0: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + seq_len = T.int64() + layer_norm26 = T.match_buffer(p_layer_norm26, (T.int64(1), seq_len, T.int64(512)), "float16") + T_add_intermediate = T.match_buffer(p_output0, (T.int64(1), seq_len, T.int64(1536)), "float16") + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((T.int64(1), seq_len, T.int64(1536)), "float16") + for i0, i1, i2, k in T.grid(T.int64(1), seq_len, T.int64(1536), T.int64(512)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(layer_norm26[v_i0, v_i1, v_k], gpt_neox_layers_0_attention_query_key_value_weight3[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float16(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + layer_norm26[v_i0, v_i1, v_k] * gpt_neox_layers_0_attention_query_key_value_weight3[v_i2, v_k] + for ax0, ax1, ax2 in T.grid(T.int64(1), seq_len, T.int64(1536)): + with T.block("T_add"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(NT_matmul_intermediate[v_ax0, v_ax1, v_ax2], gpt_neox_layers_0_attention_query_key_value_bias3[v_ax2]) + T.writes(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T_add_intermediate[v_ax0, v_ax1, v_ax2] = NT_matmul_intermediate[v_ax0, v_ax1, v_ax2] + gpt_neox_layers_0_attention_query_key_value_bias3[v_ax2] + + @T.prim_func(private=True) + def fused_NT_matmul6_add6_add9_add9(p_reshape51: T.handle, gpt_neox_layers_0_attention_dense_weight3: T.Buffer((T.int64(512), T.int64(512)), "float16"), gpt_neox_layers_0_attention_dense_bias3: T.Buffer((T.int64(512),), "float16"), p_astype27: T.handle, p_input_embeds: T.handle, p_output0: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + seq_len = T.int64() + reshape51 = T.match_buffer(p_reshape51, (T.int64(1), seq_len, T.int64(512)), "float16") + astype27 = T.match_buffer(p_astype27, (T.int64(1), seq_len, T.int64(512)), "float16") + input_embeds = T.match_buffer(p_input_embeds, (T.int64(1), seq_len, T.int64(512)), "float16") + T_add_intermediate_1_2 = T.match_buffer(p_output0, (T.int64(1), seq_len, T.int64(512)), "float16") + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((T.int64(1), seq_len, T.int64(512)), "float16") + T_add_intermediate = T.alloc_buffer((T.int64(1), seq_len, T.int64(512)), "float16") + T_add_intermediate_1 = T.alloc_buffer((T.int64(1), seq_len, T.int64(512)), "float16") + for i0, i1, i2, k in T.grid(T.int64(1), seq_len, T.int64(512), T.int64(512)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(reshape51[v_i0, v_i1, v_k], gpt_neox_layers_0_attention_dense_weight3[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float16(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + reshape51[v_i0, v_i1, v_k] * gpt_neox_layers_0_attention_dense_weight3[v_i2, v_k] + for ax0, ax1, ax2 in T.grid(T.int64(1), seq_len, T.int64(512)): + with T.block("T_add"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(NT_matmul_intermediate[v_ax0, v_ax1, v_ax2], gpt_neox_layers_0_attention_dense_bias3[v_ax2]) + T.writes(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T_add_intermediate[v_ax0, v_ax1, v_ax2] = NT_matmul_intermediate[v_ax0, v_ax1, v_ax2] + gpt_neox_layers_0_attention_dense_bias3[v_ax2] + for ax0, ax1, ax2 in T.grid(T.int64(1), seq_len, T.int64(512)): + with T.block("T_add_1"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(astype27[v_ax0, v_ax1, v_ax2], T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T.writes(T_add_intermediate_1[v_ax0, v_ax1, v_ax2]) + T_add_intermediate_1[v_ax0, v_ax1, v_ax2] = astype27[v_ax0, v_ax1, v_ax2] + T_add_intermediate[v_ax0, v_ax1, v_ax2] + for ax0, ax1, ax2 in T.grid(T.int64(1), seq_len, T.int64(512)): + with T.block("T_add_2"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_add_intermediate_1[v_ax0, v_ax1, v_ax2], input_embeds[v_ax0, v_ax1, v_ax2]) + T.writes(T_add_intermediate_1_2[v_ax0, v_ax1, v_ax2]) + T_add_intermediate_1_2[v_ax0, v_ax1, v_ax2] = T_add_intermediate_1[v_ax0, v_ax1, v_ax2] + input_embeds[v_ax0, v_ax1, v_ax2] + + @T.prim_func(private=True) + def fused_NT_matmul7_add7_gelu1_cast3(p_layer_norm27: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_weight3: T.Buffer((T.int64(2048), T.int64(512)), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: T.Buffer((T.int64(2048),), "float32"), p_output0: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + seq_len = T.int64() + layer_norm27 = T.match_buffer(p_layer_norm27, (T.int64(1), seq_len, T.int64(512)), "float16") + compute_intermediate = T.match_buffer(p_output0, (T.int64(1), seq_len, T.int64(2048)), "float16") + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((T.int64(1), seq_len, T.int64(2048))) + T_add_intermediate = T.alloc_buffer((T.int64(1), seq_len, T.int64(2048))) + T_multiply = T.alloc_buffer((T.int64(1), seq_len, T.int64(2048))) + compute = T.alloc_buffer((T.int64(1), seq_len, T.int64(2048))) + T_multiply_1 = T.alloc_buffer((T.int64(1), seq_len, T.int64(2048))) + T_add = T.alloc_buffer((T.int64(1), seq_len, T.int64(2048))) + T_multiply_intermediate = T.alloc_buffer((T.int64(1), seq_len, T.int64(2048))) + for i0, i1, i2, k in T.grid(T.int64(1), seq_len, T.int64(2048), T.int64(512)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(layer_norm27[v_i0, v_i1, v_k], gpt_neox_layers_0_mlp_dense_h_to_4h_weight3[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float32(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + T.Cast("float32", layer_norm27[v_i0, v_i1, v_k]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_h_to_4h_weight3[v_i2, v_k]) + for ax0, ax1, ax2 in T.grid(T.int64(1), seq_len, T.int64(2048)): + with T.block("T_add"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(NT_matmul_intermediate[v_ax0, v_ax1, v_ax2], gpt_neox_layers_0_mlp_dense_h_to_4h_bias3[v_ax2]) + T.writes(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T_add_intermediate[v_ax0, v_ax1, v_ax2] = NT_matmul_intermediate[v_ax0, v_ax1, v_ax2] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias3[v_ax2] + for ax0, ax1, ax2 in T.grid(T.int64(1), seq_len, T.int64(2048)): + with T.block("T_multiply"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T.writes(T_multiply[v_ax0, v_ax1, v_ax2]) + T_multiply[v_ax0, v_ax1, v_ax2] = T_add_intermediate[v_ax0, v_ax1, v_ax2] * T.float32(0.70710678118654757) + for i0, i1, i2 in T.grid(T.int64(1), seq_len, T.int64(2048)): + with T.block("compute"): + v_i0, v_i1, v_i2 = T.axis.remap("SSS", [i0, i1, i2]) + T.reads(T_multiply[v_i0, v_i1, v_i2]) + T.writes(compute[v_i0, v_i1, v_i2]) + compute[v_i0, v_i1, v_i2] = T.erf(T_multiply[v_i0, v_i1, v_i2]) + for ax0, ax1, ax2 in T.grid(T.int64(1), seq_len, T.int64(2048)): + with T.block("T_multiply_1"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(compute[v_ax0, v_ax1, v_ax2]) + T.writes(T_multiply_1[v_ax0, v_ax1, v_ax2]) + T_multiply_1[v_ax0, v_ax1, v_ax2] = compute[v_ax0, v_ax1, v_ax2] * T.float32(0.5) + for ax0, ax1, ax2 in T.grid(T.int64(1), seq_len, T.int64(2048)): + with T.block("T_add_1"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_multiply_1[v_ax0, v_ax1, v_ax2]) + T.writes(T_add[v_ax0, v_ax1, v_ax2]) + T_add[v_ax0, v_ax1, v_ax2] = T.float32(0.5) + T_multiply_1[v_ax0, v_ax1, v_ax2] + for ax0, ax1, ax2 in T.grid(T.int64(1), seq_len, T.int64(2048)): + with T.block("T_multiply_2"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_add_intermediate[v_ax0, v_ax1, v_ax2], T_add[v_ax0, v_ax1, v_ax2]) + T.writes(T_multiply_intermediate[v_ax0, v_ax1, v_ax2]) + T_multiply_intermediate[v_ax0, v_ax1, v_ax2] = T_add_intermediate[v_ax0, v_ax1, v_ax2] * T_add[v_ax0, v_ax1, v_ax2] + for i0, i1, i2 in T.grid(T.int64(1), seq_len, T.int64(2048)): + with T.block("compute_1"): + v_i0, v_i1, v_i2 = T.axis.remap("SSS", [i0, i1, i2]) + T.reads(T_multiply_intermediate[v_i0, v_i1, v_i2]) + T.writes(compute_intermediate[v_i0, v_i1, v_i2]) + compute_intermediate[v_i0, v_i1, v_i2] = T.Cast("float16", T_multiply_intermediate[v_i0, v_i1, v_i2]) + + @T.prim_func(private=True) + def fused_NT_matmul8_add8_cast4(p_astype26: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_weight3: T.Buffer((T.int64(512), T.int64(2048)), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: T.Buffer((T.int64(512),), "float32"), p_output0: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + seq_len = T.int64() + astype26 = T.match_buffer(p_astype26, (T.int64(1), seq_len, T.int64(2048)), "float16") + compute_intermediate = T.match_buffer(p_output0, (T.int64(1), seq_len, T.int64(512)), "float16") + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((T.int64(1), seq_len, T.int64(512))) + T_add_intermediate = T.alloc_buffer((T.int64(1), seq_len, T.int64(512))) + for i0, i1, i2, k in T.grid(T.int64(1), seq_len, T.int64(512), T.int64(2048)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(astype26[v_i0, v_i1, v_k], gpt_neox_layers_0_mlp_dense_4h_to_h_weight3[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float32(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + T.Cast("float32", astype26[v_i0, v_i1, v_k]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_4h_to_h_weight3[v_i2, v_k]) + for ax0, ax1, ax2 in T.grid(T.int64(1), seq_len, T.int64(512)): + with T.block("T_add"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(NT_matmul_intermediate[v_ax0, v_ax1, v_ax2], gpt_neox_layers_0_mlp_dense_4h_to_h_bias3[v_ax2]) + T.writes(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T_add_intermediate[v_ax0, v_ax1, v_ax2] = NT_matmul_intermediate[v_ax0, v_ax1, v_ax2] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias3[v_ax2] + for i0, i1, i2 in T.grid(T.int64(1), seq_len, T.int64(512)): + with T.block("compute"): + v_i0, v_i1, v_i2 = T.axis.remap("SSS", [i0, i1, i2]) + T.reads(T_add_intermediate[v_i0, v_i1, v_i2]) + T.writes(compute_intermediate[v_i0, v_i1, v_i2]) + compute_intermediate[v_i0, v_i1, v_i2] = T.Cast("float16", T_add_intermediate[v_i0, v_i1, v_i2]) + + @T.prim_func(private=True) + def fused_NT_matmul9_cast5(p_take1: T.handle, p_embed_out_weight3: T.handle, p_output0: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + batch_size = T.int64() + take1 = T.match_buffer(p_take1, (T.int64(1), batch_size, T.int64(512)), "float16") + vocab_size = T.int64() + embed_out_weight3 = T.match_buffer(p_embed_out_weight3, (vocab_size, T.int64(512)), "float16") + compute_intermediate = T.match_buffer(p_output0, (T.int64(1), batch_size, vocab_size)) + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((T.int64(1), batch_size, vocab_size), "float16") + for i0, i1, i2, k in T.grid(T.int64(1), batch_size, vocab_size, T.int64(512)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(take1[v_i0, v_i1, v_k], embed_out_weight3[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float16(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + take1[v_i0, v_i1, v_k] * embed_out_weight3[v_i2, v_k] + for i0, i1, i2 in T.grid(T.int64(1), batch_size, vocab_size): + with T.block("compute"): + v_i0, v_i1, v_i2 = T.axis.remap("SSS", [i0, i1, i2]) + T.reads(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + T.writes(compute_intermediate[v_i0, v_i1, v_i2]) + compute_intermediate[v_i0, v_i1, v_i2] = T.Cast("float32", NT_matmul_intermediate[v_i0, v_i1, v_i2]) + + @T.prim_func(private=True) + def fused_NT_matmul_add(p_layer_norm39: T.handle, gpt_neox_layers_0_attention_query_key_value_weight4: T.Buffer((T.int64(1536), T.int64(512)), "float16"), gpt_neox_layers_0_attention_query_key_value_bias4: T.Buffer((T.int64(1536),), "float16"), p_output0: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + batch_size = T.int64() + layer_norm39 = T.match_buffer(p_layer_norm39, (batch_size, T.int64(1), T.int64(512)), "float16") + T_add_intermediate = T.match_buffer(p_output0, (batch_size, T.int64(1), T.int64(1536)), "float16") + # with T.block("root"): + NT_matmul_intermediate = T.alloc_buffer((batch_size, T.int64(1), T.int64(1536)), "float16") + for i0, i1, i2, k in T.grid(batch_size, T.int64(1), T.int64(1536), T.int64(512)): + with T.block("NT_matmul"): + v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) + T.reads(layer_norm39[v_i0, v_i1, v_k], gpt_neox_layers_0_attention_query_key_value_weight4[v_i2, v_k]) + T.writes(NT_matmul_intermediate[v_i0, v_i1, v_i2]) + with T.init(): + NT_matmul_intermediate[v_i0, v_i1, v_i2] = T.float16(0.0) + NT_matmul_intermediate[v_i0, v_i1, v_i2] = NT_matmul_intermediate[v_i0, v_i1, v_i2] + layer_norm39[v_i0, v_i1, v_k] * gpt_neox_layers_0_attention_query_key_value_weight4[v_i2, v_k] + for ax0, ax1, ax2 in T.grid(batch_size, T.int64(1), T.int64(1536)): + with T.block("T_add"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(NT_matmul_intermediate[v_ax0, v_ax1, v_ax2], gpt_neox_layers_0_attention_query_key_value_bias4[v_ax2]) + T.writes(T_add_intermediate[v_ax0, v_ax1, v_ax2]) + T_add_intermediate[v_ax0, v_ax1, v_ax2] = NT_matmul_intermediate[v_ax0, v_ax1, v_ax2] + gpt_neox_layers_0_attention_query_key_value_bias4[v_ax2] + + @T.prim_func(private=True) + def fused_reshape10_reshape11(lv7: T.Buffer((T.int64(1), T.int64(8), T.int64(64)), "float16"), T_reshape_intermediate_1: T.Buffer((T.int64(1), T.int64(1), T.int64(512)), "float16")): + T.func_attr({"tir.noalias": T.bool(True)}) + # with T.block("root"): + T_reshape_intermediate = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(8), T.int64(64)), "float16") + for ax0, ax1, ax2, ax3 in T.grid(T.int64(1), T.int64(1), T.int64(8), T.int64(64)): + with T.block("T_reshape"): + v_ax0, v_ax1, v_ax2, v_ax3 = T.axis.remap("SSSS", [ax0, ax1, ax2, ax3]) + T.reads(lv7[T.int64(0), (v_ax3 // T.int64(64) + v_ax2) % T.int64(8), v_ax3 % T.int64(64)]) + T.writes(T_reshape_intermediate[v_ax0, v_ax1, v_ax2, v_ax3]) + T_reshape_intermediate[v_ax0, v_ax1, v_ax2, v_ax3] = lv7[T.int64(0), (v_ax3 // T.int64(64) + v_ax2) % T.int64(8), v_ax3 % T.int64(64)] + for ax0, ax1, ax2 in T.grid(T.int64(1), T.int64(1), T.int64(512)): + with T.block("T_reshape_1"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_reshape_intermediate[T.int64(0), T.int64(0), v_ax2 % T.int64(512) // T.int64(64), v_ax2 % T.int64(64)]) + T.writes(T_reshape_intermediate_1[v_ax0, v_ax1, v_ax2]) + T_reshape_intermediate_1[v_ax0, v_ax1, v_ax2] = T_reshape_intermediate[T.int64(0), T.int64(0), v_ax2 % T.int64(512) // T.int64(64), v_ax2 % T.int64(64)] + + @T.prim_func(private=True) + def fused_reshape8_reshape9(add36: T.Buffer((T.int64(1), T.int64(1), T.int64(1536)), "float16"), T_reshape_intermediate_1: T.Buffer((T.int64(1), T.int64(24), T.int64(64)), "float16")): + T.func_attr({"tir.noalias": T.bool(True)}) + # with T.block("root"): + T_reshape_intermediate = T.alloc_buffer((T.int64(1), T.int64(1), T.int64(24), T.int64(64)), "float16") + for ax0, ax1, ax2, ax3 in T.grid(T.int64(1), T.int64(1), T.int64(24), T.int64(64)): + with T.block("T_reshape"): + v_ax0, v_ax1, v_ax2, v_ax3 = T.axis.remap("SSSS", [ax0, ax1, ax2, ax3]) + T.reads(add36[T.int64(0), T.int64(0), (v_ax2 * T.int64(64) + v_ax3) % T.int64(1536)]) + T.writes(T_reshape_intermediate[v_ax0, v_ax1, v_ax2, v_ax3]) + T_reshape_intermediate[v_ax0, v_ax1, v_ax2, v_ax3] = add36[T.int64(0), T.int64(0), (v_ax2 * T.int64(64) + v_ax3) % T.int64(1536)] + for ax0, ax1, ax2 in T.grid(T.int64(1), T.int64(24), T.int64(64)): + with T.block("T_reshape_1"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(T_reshape_intermediate[T.int64(0), T.int64(0), (v_ax2 // T.int64(64) + v_ax1) % T.int64(24), v_ax2 % T.int64(64)]) + T.writes(T_reshape_intermediate_1[v_ax0, v_ax1, v_ax2]) + T_reshape_intermediate_1[v_ax0, v_ax1, v_ax2] = T_reshape_intermediate[T.int64(0), T.int64(0), (v_ax2 // T.int64(64) + v_ax1) % T.int64(24), v_ax2 % T.int64(64)] + + @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.noalias": T.bool(True)}) + seq_len = T.int32() + qkv = T.match_buffer(var_qkv, (seq_len, 24, 64), "float16") + position_map = T.match_buffer(var_position_map, (seq_len,), "int32", offset_factor=1) + q = T.match_buffer(var_q, (seq_len, 8, 64), "float16") + k = T.match_buffer(var_k, (seq_len, 8, 64), "float16") + v = T.match_buffer(var_v, (seq_len, 8, 64), "float16") + # with T.block("root"): + for iters_0, iters_1, iters_2 in T.grid(seq_len, 24, 64): + with T.block("llama_fused_rope"): + s, h, d = T.axis.remap("SSS", [iters_0, iters_1, iters_2]) + T.reads(position_map[s], qkv[s, h, d - 8:d - 8 + 17]) + T.writes(q[s, h, d], k[s, h - 8, d], v[s, h - 16, d]) + if h < 8: + freq = T.float32() + q[s, h, d] = T.if_then_else(apply_rope > 0 and d < 16, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", qkv[s, h, d]) + T.sin(freq) * T.Cast("float32", T.if_then_else(d < 8, qkv[s, h, d + 8] * T.float16(-1.0), qkv[s, h, d - 8]))), where={freq: T.Cast("float32", position_map[s]) / T.pow(T.float32(10000.0), T.Cast("float32", d * 2 % 16) / T.float32(16.0))}), qkv[s, h, d]) + else: + if h < 16: + freq = T.float32() + k[s, h - 8, d] = T.if_then_else(apply_rope > 0 and d < 16, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", qkv[s, h, d]) + T.sin(freq) * T.Cast("float32", T.if_then_else(d < 8, qkv[s, h, d + 8] * T.float16(-1.0), qkv[s, h, d - 8]))), where={freq: T.Cast("float32", position_map[s]) / T.pow(T.float32(10000.0), T.Cast("float32", d * 2 % 16) / T.float32(16.0))}), qkv[s, h, d]) + else: + v[s, h - 16, d] = qkv[s, h, d] + + @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.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 b, j in T.grid(batch_size, n): + with T.block("gather_2d"): + vb, vj = T.axis.remap("SS", [b, j]) + T.reads(src[indices[vb], vj], indices[vb]) + T.writes(dst[vb, vj]) + dst[vb, vj] = src[indices[vb], vj] + + @T.prim_func(private=True) + def index(var_layer_norm12: T.handle, index: T.Buffer((T.int64(1), T.int64(1), T.int64(512)), "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.noalias": T.bool(True)}) + seq_len = T.int64() + layer_norm12 = T.match_buffer(var_layer_norm12, (T.int64(1), seq_len, T.int64(512)), "float16") + # with T.block("root"): + for i, _, k in T.grid(T.int64(1), T.int64(1), T.int64(512)): + with T.block("index"): + v_i, v__, v_k = T.axis.remap("SSS", [i, _, k]) + T.reads(layer_norm12[v_i, seq_len - T.int64(1), v_k]) + T.writes(index[v_i, v__, v_k]) + index[v_i, v__, v_k] = layer_norm12[v_i, seq_len - T.int64(1), v_k] + + @T.prim_func(private=True) + def layer_norm(var_input_embeds: T.handle, gpt_neox_layers_0_input_layernorm_weight4: T.Buffer((T.int64(512),), "float16"), gpt_neox_layers_0_input_layernorm_bias4: T.Buffer((T.int64(512),), "float16"), var_T_layer_norm: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + batch_size = T.int64() + input_embeds = T.match_buffer(var_input_embeds, (batch_size, T.int64(1), T.int64(512)), "float16") + T_layer_norm = T.match_buffer(var_T_layer_norm, (batch_size, T.int64(1), T.int64(512)), "float16") + # with T.block("root"): + input_embeds_red_temp_v0 = T.alloc_buffer((batch_size, T.int64(1))) + input_embeds_red_temp_v1 = T.alloc_buffer((batch_size, T.int64(1))) + for ax0, ax1, k2 in T.grid(batch_size, T.int64(1), T.int64(512)): + with T.block("input_embeds_red_temp"): + v_ax0, v_ax1, v_k2 = T.axis.remap("SSR", [ax0, ax1, k2]) + T.reads(input_embeds[v_ax0, v_ax1, v_k2]) + T.writes(input_embeds_red_temp_v0[v_ax0, v_ax1], input_embeds_red_temp_v1[v_ax0, v_ax1]) + with T.init(): + input_embeds_red_temp_v0[v_ax0, v_ax1] = T.float32(0.0) + input_embeds_red_temp_v1[v_ax0, v_ax1] = T.float32(0.0) + v_input_embeds_red_temp_v0: T.float32 = input_embeds_red_temp_v0[v_ax0, v_ax1] + T.Cast("float32", input_embeds[v_ax0, v_ax1, v_k2]) + v_input_embeds_red_temp_v1: T.float32 = input_embeds_red_temp_v1[v_ax0, v_ax1] + T.Cast("float32", input_embeds[v_ax0, v_ax1, v_k2]) * T.Cast("float32", input_embeds[v_ax0, v_ax1, v_k2]) + input_embeds_red_temp_v0[v_ax0, v_ax1] = v_input_embeds_red_temp_v0 + input_embeds_red_temp_v1[v_ax0, v_ax1] = v_input_embeds_red_temp_v1 + for ax0, ax1, ax2 in T.grid(batch_size, T.int64(1), T.int64(512)): + with T.block("T_layer_norm"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(input_embeds[v_ax0, v_ax1, v_ax2], input_embeds_red_temp_v0[v_ax0, v_ax1], input_embeds_red_temp_v1[v_ax0, v_ax1], gpt_neox_layers_0_input_layernorm_weight4[v_ax2], gpt_neox_layers_0_input_layernorm_bias4[v_ax2]) + T.writes(T_layer_norm[v_ax0, v_ax1, v_ax2]) + T_layer_norm[v_ax0, v_ax1, v_ax2] = T.Cast("float16", (T.Cast("float32", input_embeds[v_ax0, v_ax1, v_ax2]) - input_embeds_red_temp_v0[v_ax0, v_ax1] * T.float32(0.001953125)) * T.rsqrt(input_embeds_red_temp_v1[v_ax0, v_ax1] * T.float32(0.001953125) - input_embeds_red_temp_v0[v_ax0, v_ax1] * T.float32(0.001953125) * (input_embeds_red_temp_v0[v_ax0, v_ax1] * T.float32(0.001953125)) + T.float32(1.0000000000000001e-05))) * gpt_neox_layers_0_input_layernorm_weight4[v_ax2] + gpt_neox_layers_0_input_layernorm_bias4[v_ax2] + + @T.prim_func(private=True) + def layer_norm1(var_input_embeds: T.handle, gpt_neox_layers_0_input_layernorm_weight3: T.Buffer((T.int64(512),), "float16"), gpt_neox_layers_0_input_layernorm_bias3: T.Buffer((T.int64(512),), "float16"), var_T_layer_norm: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + seq_len = T.int64() + input_embeds = T.match_buffer(var_input_embeds, (T.int64(1), seq_len, T.int64(512)), "float16") + T_layer_norm = T.match_buffer(var_T_layer_norm, (T.int64(1), seq_len, T.int64(512)), "float16") + # with T.block("root"): + input_embeds_red_temp_v0 = T.alloc_buffer((T.int64(1), seq_len)) + input_embeds_red_temp_v1 = T.alloc_buffer((T.int64(1), seq_len)) + for ax0, ax1, k2 in T.grid(T.int64(1), seq_len, T.int64(512)): + with T.block("input_embeds_red_temp"): + v_ax0, v_ax1, v_k2 = T.axis.remap("SSR", [ax0, ax1, k2]) + T.reads(input_embeds[v_ax0, v_ax1, v_k2]) + T.writes(input_embeds_red_temp_v0[v_ax0, v_ax1], input_embeds_red_temp_v1[v_ax0, v_ax1]) + with T.init(): + input_embeds_red_temp_v0[v_ax0, v_ax1] = T.float32(0.0) + input_embeds_red_temp_v1[v_ax0, v_ax1] = T.float32(0.0) + v_input_embeds_red_temp_v0: T.float32 = input_embeds_red_temp_v0[v_ax0, v_ax1] + T.Cast("float32", input_embeds[v_ax0, v_ax1, v_k2]) + v_input_embeds_red_temp_v1: T.float32 = input_embeds_red_temp_v1[v_ax0, v_ax1] + T.Cast("float32", input_embeds[v_ax0, v_ax1, v_k2]) * T.Cast("float32", input_embeds[v_ax0, v_ax1, v_k2]) + input_embeds_red_temp_v0[v_ax0, v_ax1] = v_input_embeds_red_temp_v0 + input_embeds_red_temp_v1[v_ax0, v_ax1] = v_input_embeds_red_temp_v1 + for ax0, ax1, ax2 in T.grid(T.int64(1), seq_len, T.int64(512)): + with T.block("T_layer_norm"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(input_embeds[v_ax0, v_ax1, v_ax2], input_embeds_red_temp_v0[v_ax0, v_ax1], input_embeds_red_temp_v1[v_ax0, v_ax1], gpt_neox_layers_0_input_layernorm_weight3[v_ax2], gpt_neox_layers_0_input_layernorm_bias3[v_ax2]) + T.writes(T_layer_norm[v_ax0, v_ax1, v_ax2]) + T_layer_norm[v_ax0, v_ax1, v_ax2] = T.Cast("float16", (T.Cast("float32", input_embeds[v_ax0, v_ax1, v_ax2]) - input_embeds_red_temp_v0[v_ax0, v_ax1] * T.float32(0.001953125)) * T.rsqrt(input_embeds_red_temp_v1[v_ax0, v_ax1] * T.float32(0.001953125) - input_embeds_red_temp_v0[v_ax0, v_ax1] * T.float32(0.001953125) * (input_embeds_red_temp_v0[v_ax0, v_ax1] * T.float32(0.001953125)) + T.float32(1.0000000000000001e-05))) * gpt_neox_layers_0_input_layernorm_weight3[v_ax2] + gpt_neox_layers_0_input_layernorm_bias3[v_ax2] + + @T.prim_func(private=True) + def layer_norm2(input_embed: T.Buffer((T.int64(1), T.int64(1), T.int64(512)), "float16"), gpt_neox_layers_0_input_layernorm_weight2: T.Buffer((T.int64(512),), "float16"), gpt_neox_layers_0_input_layernorm_bias2: T.Buffer((T.int64(512),), "float16"), T_layer_norm: T.Buffer((T.int64(1), T.int64(1), T.int64(512)), "float16")): + T.func_attr({"tir.noalias": T.bool(True)}) + # with T.block("root"): + input_embed_red_temp_v0 = T.alloc_buffer((T.int64(1), T.int64(1))) + input_embed_red_temp_v1 = T.alloc_buffer((T.int64(1), T.int64(1))) + for ax0, ax1, k2 in T.grid(T.int64(1), T.int64(1), T.int64(512)): + with T.block("input_embed_red_temp"): + v_ax0, v_ax1, v_k2 = T.axis.remap("SSR", [ax0, ax1, k2]) + T.reads(input_embed[v_ax0, v_ax1, v_k2]) + T.writes(input_embed_red_temp_v0[v_ax0, v_ax1], input_embed_red_temp_v1[v_ax0, v_ax1]) + with T.init(): + input_embed_red_temp_v0[v_ax0, v_ax1] = T.float32(0.0) + input_embed_red_temp_v1[v_ax0, v_ax1] = T.float32(0.0) + v_input_embed_red_temp_v0: T.float32 = input_embed_red_temp_v0[v_ax0, v_ax1] + T.Cast("float32", input_embed[v_ax0, v_ax1, v_k2]) + v_input_embed_red_temp_v1: T.float32 = input_embed_red_temp_v1[v_ax0, v_ax1] + T.Cast("float32", input_embed[v_ax0, v_ax1, v_k2]) * T.Cast("float32", input_embed[v_ax0, v_ax1, v_k2]) + input_embed_red_temp_v0[v_ax0, v_ax1] = v_input_embed_red_temp_v0 + input_embed_red_temp_v1[v_ax0, v_ax1] = v_input_embed_red_temp_v1 + for ax0, ax1, ax2 in T.grid(T.int64(1), T.int64(1), T.int64(512)): + with T.block("T_layer_norm"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(input_embed[v_ax0, v_ax1, v_ax2], input_embed_red_temp_v0[v_ax0, v_ax1], input_embed_red_temp_v1[v_ax0, v_ax1], gpt_neox_layers_0_input_layernorm_weight2[v_ax2], gpt_neox_layers_0_input_layernorm_bias2[v_ax2]) + T.writes(T_layer_norm[v_ax0, v_ax1, v_ax2]) + T_layer_norm[v_ax0, v_ax1, v_ax2] = T.Cast("float16", (T.Cast("float32", input_embed[v_ax0, v_ax1, v_ax2]) - input_embed_red_temp_v0[v_ax0, v_ax1] * T.float32(0.001953125)) * T.rsqrt(input_embed_red_temp_v1[v_ax0, v_ax1] * T.float32(0.001953125) - input_embed_red_temp_v0[v_ax0, v_ax1] * T.float32(0.001953125) * (input_embed_red_temp_v0[v_ax0, v_ax1] * T.float32(0.001953125)) + T.float32(1.0000000000000001e-05))) * gpt_neox_layers_0_input_layernorm_weight2[v_ax2] + gpt_neox_layers_0_input_layernorm_bias2[v_ax2] + + @T.prim_func + def merge_state_inplace(v: T.handle, s: T.handle, v_other: T.handle, s_other: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + N, H, D = T.int32(is_size_var=True), T.int32(is_size_var=True), T.int32(is_size_var=True) + V = T.match_buffer(v, (N, H, D), "float16") + S = T.match_buffer(s, (N, H)) + V_other = T.match_buffer(v_other, (N, H, D), "float16") + S_other = T.match_buffer(s_other, (N, H)) + # with T.block("root"): + for bx in T.thread_binding(N, thread="blockIdx.x"): + for by in T.thread_binding(1, thread="blockIdx.y"): + for ty in T.thread_binding(8, thread="threadIdx.y"): + for tx in T.thread_binding(16, thread="threadIdx.x"): + with T.block("merge"): + T.reads(S[bx, ty + by * 8], S_other[bx, ty + by * 8], V[bx, ty + by * 8, tx * 4:tx * 4 + 4], V_other[bx, ty + by * 8, tx * 4:tx * 4 + 4]) + T.writes(V[bx, ty + by * 8, tx * 4:tx * 4 + 4], S[bx, ty + by * 8]) + 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 * 8] + s_other_val[0] = S_other[bx, ty + by * 8] + 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 * 8, tx * 4 + vec] + for vec in T.vectorized(4): + v_other_vec[vec] = V_other[bx, ty + by * 8, 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 * 8, tx * 4 + vec] = v_vec[vec] + S[bx, ty + by * 8] = T.log2(s_val[0] + s_other_val[0]) + s_max[0] + + @T.prim_func(private=True) + def reshape(var_add108: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + batch_size = T.int64() + add108 = T.match_buffer(var_add108, (batch_size, T.int64(1), T.int64(1536)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, T.int64(1), T.int64(24), T.int64(64)), "float16") + # with T.block("root"): + for ax0, ax1, ax2, ax3 in T.grid(batch_size, T.int64(1), T.int64(24), T.int64(64)): + with T.block("T_reshape"): + v_ax0, v_ax1, v_ax2, v_ax3 = T.axis.remap("SSSS", [ax0, ax1, ax2, ax3]) + T.reads(add108[((v_ax2 * T.int64(64) + v_ax3) // T.int64(1536) + v_ax0 + v_ax1) % batch_size, T.int64(0), (v_ax2 * T.int64(64) + v_ax3) % T.int64(1536)]) + T.writes(T_reshape[v_ax0, v_ax1, v_ax2, v_ax3]) + T_reshape[v_ax0, v_ax1, v_ax2, v_ax3] = add108[((v_ax2 * T.int64(64) + v_ax3) // T.int64(1536) + v_ax0 + v_ax1) % batch_size, T.int64(0), (v_ax2 * T.int64(64) + v_ax3) % T.int64(1536)] + + @T.prim_func(private=True) + def reshape1(var_reshape72: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + batch_size = T.int64() + reshape72 = T.match_buffer(var_reshape72, (batch_size, T.int64(1), T.int64(24), T.int64(64)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, T.int64(24), T.int64(64)), "float16") + # with T.block("root"): + for ax0, ax1, ax2 in T.grid(batch_size, T.int64(24), T.int64(64)): + with T.block("T_reshape"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(reshape72[((v_ax2 // T.int64(64) + v_ax1) // T.int64(24) + v_ax0) % batch_size, T.int64(0), (v_ax2 // T.int64(64) + v_ax1) % T.int64(24), v_ax2 % T.int64(64)]) + T.writes(T_reshape[v_ax0, v_ax1, v_ax2]) + T_reshape[v_ax0, v_ax1, v_ax2] = reshape72[((v_ax2 // T.int64(64) + v_ax1) // T.int64(24) + v_ax0) % batch_size, T.int64(0), (v_ax2 // T.int64(64) + v_ax1) % T.int64(24), v_ax2 % T.int64(64)] + + @T.prim_func(private=True) + def reshape2(var_lv19: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + batch_size = T.int64() + lv19 = T.match_buffer(var_lv19, (batch_size, T.int64(8), T.int64(64)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, T.int64(1), T.int64(8), T.int64(64)), "float16") + # with T.block("root"): + for ax0, ax1, ax2, ax3 in T.grid(batch_size, T.int64(1), T.int64(8), T.int64(64)): + with T.block("T_reshape"): + v_ax0, v_ax1, v_ax2, v_ax3 = T.axis.remap("SSSS", [ax0, ax1, ax2, ax3]) + T.reads(lv19[((v_ax3 // T.int64(64) + v_ax2) // T.int64(8) + v_ax0 + v_ax1) % batch_size, (v_ax3 // T.int64(64) + v_ax2) % T.int64(8), v_ax3 % T.int64(64)]) + T.writes(T_reshape[v_ax0, v_ax1, v_ax2, v_ax3]) + T_reshape[v_ax0, v_ax1, v_ax2, v_ax3] = lv19[((v_ax3 // T.int64(64) + v_ax2) // T.int64(8) + v_ax0 + v_ax1) % batch_size, (v_ax3 // T.int64(64) + v_ax2) % T.int64(8), v_ax3 % T.int64(64)] + + @T.prim_func(private=True) + def reshape3(var_reshape74: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + batch_size = T.int64() + reshape74 = T.match_buffer(var_reshape74, (batch_size, T.int64(1), T.int64(8), T.int64(64)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (batch_size, T.int64(1), T.int64(512)), "float16") + # with T.block("root"): + for ax0, ax1, ax2 in T.grid(batch_size, T.int64(1), T.int64(512)): + with T.block("T_reshape"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(reshape74[(v_ax2 // T.int64(512) + v_ax0 + v_ax1) % batch_size, T.int64(0), v_ax2 % T.int64(512) // T.int64(64), v_ax2 % T.int64(64)]) + T.writes(T_reshape[v_ax0, v_ax1, v_ax2]) + T_reshape[v_ax0, v_ax1, v_ax2] = reshape74[(v_ax2 // T.int64(512) + v_ax0 + v_ax1) % batch_size, T.int64(0), v_ax2 % T.int64(512) // T.int64(64), v_ax2 % T.int64(64)] + + @T.prim_func(private=True) + def reshape4(var_add72: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + seq_len = T.int64() + add72 = T.match_buffer(var_add72, (T.int64(1), seq_len, T.int64(1536)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (T.int64(1), seq_len, T.int64(24), T.int64(64)), "float16") + # with T.block("root"): + for ax0, ax1, ax2, ax3 in T.grid(T.int64(1), seq_len, T.int64(24), T.int64(64)): + with T.block("T_reshape"): + v_ax0, v_ax1, v_ax2, v_ax3 = T.axis.remap("SSSS", [ax0, ax1, ax2, ax3]) + T.reads(add72[T.int64(0), ((v_ax2 * T.int64(64) + v_ax3) // T.int64(1536) + v_ax0 * seq_len + v_ax1) % seq_len, (v_ax2 * T.int64(64) + v_ax3) % T.int64(1536)]) + T.writes(T_reshape[v_ax0, v_ax1, v_ax2, v_ax3]) + T_reshape[v_ax0, v_ax1, v_ax2, v_ax3] = add72[T.int64(0), ((v_ax2 * T.int64(64) + v_ax3) // T.int64(1536) + v_ax0 * seq_len + v_ax1) % seq_len, (v_ax2 * T.int64(64) + v_ax3) % T.int64(1536)] + + @T.prim_func(private=True) + def reshape5(var_reshape48: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + seq_len = T.int64() + reshape48 = T.match_buffer(var_reshape48, (T.int64(1), seq_len, T.int64(24), T.int64(64)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (seq_len, T.int64(24), T.int64(64)), "float16") + # with T.block("root"): + for ax0, ax1, ax2 in T.grid(seq_len, T.int64(24), T.int64(64)): + with T.block("T_reshape"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(reshape48[T.int64(0), ((v_ax2 // T.int64(64) + v_ax1) // T.int64(24) + v_ax0) % seq_len, (v_ax2 // T.int64(64) + v_ax1) % T.int64(24), v_ax2 % T.int64(64)]) + T.writes(T_reshape[v_ax0, v_ax1, v_ax2]) + T_reshape[v_ax0, v_ax1, v_ax2] = reshape48[T.int64(0), ((v_ax2 // T.int64(64) + v_ax1) // T.int64(24) + v_ax0) % seq_len, (v_ax2 // T.int64(64) + v_ax1) % T.int64(24), v_ax2 % T.int64(64)] + + @T.prim_func(private=True) + def reshape6(var_lv13: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + seq_len = T.int64() + lv13 = T.match_buffer(var_lv13, (seq_len, T.int64(8), T.int64(64)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (T.int64(1), seq_len, T.int64(8), T.int64(64)), "float16") + # with T.block("root"): + for ax0, ax1, ax2, ax3 in T.grid(T.int64(1), seq_len, T.int64(8), T.int64(64)): + with T.block("T_reshape"): + v_ax0, v_ax1, v_ax2, v_ax3 = T.axis.remap("SSSS", [ax0, ax1, ax2, ax3]) + T.reads(lv13[((v_ax3 // T.int64(64) + v_ax2) // T.int64(8) + v_ax0 * seq_len + v_ax1) % seq_len, (v_ax3 // T.int64(64) + v_ax2) % T.int64(8), v_ax3 % T.int64(64)]) + T.writes(T_reshape[v_ax0, v_ax1, v_ax2, v_ax3]) + T_reshape[v_ax0, v_ax1, v_ax2, v_ax3] = lv13[((v_ax3 // T.int64(64) + v_ax2) // T.int64(8) + v_ax0 * seq_len + v_ax1) % seq_len, (v_ax3 // T.int64(64) + v_ax2) % T.int64(8), v_ax3 % T.int64(64)] + + @T.prim_func(private=True) + def reshape7(var_reshape50: T.handle, var_T_reshape: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + seq_len = T.int64() + reshape50 = T.match_buffer(var_reshape50, (T.int64(1), seq_len, T.int64(8), T.int64(64)), "float16") + T_reshape = T.match_buffer(var_T_reshape, (T.int64(1), seq_len, T.int64(512)), "float16") + # with T.block("root"): + for ax0, ax1, ax2 in T.grid(T.int64(1), seq_len, T.int64(512)): + with T.block("T_reshape"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(reshape50[T.int64(0), (v_ax2 // T.int64(512) + v_ax0 * seq_len + v_ax1) % seq_len, v_ax2 % T.int64(512) // T.int64(64), v_ax2 % T.int64(64)]) + T.writes(T_reshape[v_ax0, v_ax1, v_ax2]) + T_reshape[v_ax0, v_ax1, v_ax2] = reshape50[T.int64(0), (v_ax2 // T.int64(512) + v_ax0 * seq_len + v_ax1) % seq_len, v_ax2 % T.int64(512) // T.int64(64), v_ax2 % T.int64(64)] + + @T.prim_func + def scatter_probs(var_src: T.handle, var_indices: T.handle, var_dst: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.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 b, j in T.grid(batch_size, n): + with T.block("scatter_2d"): + vb, vj = T.axis.remap("SS", [b, j]) + T.reads(src[vb, vj], indices[vb]) + T.writes(dst[indices[vb], vj]) + dst[indices[vb], vj] = src[vb, vj] + + @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_norm38: T.handle, var_logit_positions: T.handle, var_T_take: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + seq_len = T.int64() + layer_norm38 = T.match_buffer(var_layer_norm38, (T.int64(1), seq_len, T.int64(512)), "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(512)), "float16") + # with T.block("root"): + for ax0, ax1, ax2 in T.grid(T.int64(1), batch_size, T.int64(512)): + with T.block("T_take"): + v_ax0, v_ax1, v_ax2 = T.axis.remap("SSS", [ax0, ax1, ax2]) + T.reads(layer_norm38[v_ax0, logit_positions[v_ax1], v_ax2], logit_positions[v_ax1]) + T.writes(T_take[v_ax0, v_ax1, v_ax2]) + T_take[v_ax0, v_ax1, v_ax2] = layer_norm38[v_ax0, logit_positions[v_ax1], v_ax2] + + @T.prim_func(private=True) + def take1(var_gpt_neox_embed_in_weight: T.handle, var_input_ids: T.handle, var_T_take: T.handle): + T.func_attr({"tir.noalias": T.bool(True)}) + vocab_size = T.int64() + gpt_neox_embed_in_weight = T.match_buffer(var_gpt_neox_embed_in_weight, (vocab_size, T.int64(512)), "float16") + seq_len = T.int64() + input_ids = T.match_buffer(var_input_ids, (seq_len,), "int32") + T_take = T.match_buffer(var_T_take, (seq_len, T.int64(512)), "float16") + # with T.block("root"): + for ax0, ax1 in T.grid(seq_len, T.int64(512)): + with T.block("T_take"): + v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1]) + T.reads(gpt_neox_embed_in_weight[input_ids[v_ax0], v_ax1], input_ids[v_ax0]) + T.writes(T_take[v_ax0, v_ax1]) + T_take[v_ax0, v_ax1] = gpt_neox_embed_in_weight[input_ids[v_ax0], v_ax1] + + @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.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, 64), "float16", offset_factor=1) + seqlen = T.int64(is_size_var=True) + position_map = T.match_buffer(var_position_map, (seqlen,), "int32", offset_factor=1) + k_data = T.match_buffer(var_k_data, (6, seqlen, 8, 64), "float16") + v_data = T.match_buffer(var_v_data, (6, seqlen, 8, 64), "float16") + # with T.block("root"): + for p, h, d in T.grid(seqlen, 8, 64): + with T.block("copy0"): + vp, vh, vd = T.axis.remap("SSS", [p, h, d]) + 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.noalias": T.bool(True)}) + num_pages = T.int64() + pages = T.match_buffer(var_pages, (num_pages, 2, 8, 16, 64), "float16", offset_factor=1) + ntoken = T.int64(is_size_var=True) + k_data = T.match_buffer(var_k_data, (ntoken, 8, 64), "float16") + v_data = T.match_buffer(var_v_data, (ntoken, 8, 64), "float16") + position_map = T.match_buffer(var_position_map, (ntoken,), "int32", offset_factor=1) + # with T.block("root"): + for global_pos, h, f in T.grid(ntoken, 8, 64): + if position_map[global_pos] != -1: + with T.block("k_transpose_append"): + vgpos, vh, vf = T.axis.remap("SSS", [global_pos, h, f]) + 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, vh, vf = T.axis.remap("SSS", [global_pos, h, f]) + 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, 64), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 8, 32, 64), "float16") + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 8, 64), "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((32, 64), "float16", scope="shared") + K_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 64), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 32), scope="shared") + S_local = T.alloc_buffer((32, 32), scope="local") + O_local = T.alloc_buffer((32, 64), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 32 + length_info[b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = T.float32(0.0) + T.tvm_storage_sync("shared") + for li_lj_fused_0 in range(4): + for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for li_lj_fused_3 in T.vectorized(4): + with T.block("Q_load"): + i = T.axis.spatial(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 64) + j = T.axis.spatial(64, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = q_indptr_val + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + freq = T.float32() + Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 32, q[cur_L, cur_H_qo, j + 32] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 32]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 64) / T.float32(64.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 31) // 32): + L_kv_start: T.int32 = iterator_1 * 32 + for lz_ly_fused_0 in range(4): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 32] + page_offset: T.int32 = seq_offset % 32 + K_smem[i, j] = pages[page_no, 0, by, page_offset, j] + else: + K_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for lz_ly_fused_0 in range(4): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 64) + j = T.axis.spatial(64, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 64) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 32] + page_offset: T.int32 = seq_offset % 32 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:64], K_smem[0:32, 0:64]) + T.writes(S_local[0:32, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 4): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(8, 2, 4, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + k = T.axis.reduce(64, lk_0 * 8 + lk_1) + T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) + T.writes(S_local[i, j]) + S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.18033688011112042) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 4): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 32: + row_: T.int32 = LH_start + row + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:64]) + T.writes(O_local[0:32, 0:64]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(4, 4): + with T.block("O_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(64, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 4 + lj_1_init) + T.reads() + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, lk_1, li_1, lj_1 in T.grid(4, 8, 4, 4): + with T.block("O_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + k = T.axis.reduce(32, lk_0 * 8 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 4): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(64, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 4 + lj_1) + T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) + T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) + for li_0 in range(1): + for li_1 in T.thread_binding(4, thread="threadIdx.y"): + for li_2 in T.thread_binding(32, thread="threadIdx.x"): + with T.block("lse_store"): + i = T.axis.spatial(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @R.function + def alloc_embedding_tensor() -> R.Tensor((2048, 512), dtype="float16"): + R.func_attr({"relax.memory_plan_dynamic_func_output": True}) + gv: R.Tensor((2048, 512), dtype="float16") = R.builtin.alloc_tensor(R.shape([2048, 512]), R.dtype("float16"), R.prim_value(0), R.str("global")) + return gv + + @R.function + def batch_decode(input_embeds: R.Tensor(("batch_size", 1, 512), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor(("vocab_size", 512), 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((512,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias4: R.Tensor((512,), dtype="float16") = packed_params[2] + gpt_neox_layers_0_post_attention_layernorm_weight4: R.Tensor((512,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias4: R.Tensor((512,), dtype="float16") = packed_params[4] + gpt_neox_layers_0_attention_query_key_value_weight4: R.Tensor((1536, 512), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias4: R.Tensor((1536,), dtype="float16") = packed_params[6] + gpt_neox_layers_0_attention_dense_weight4: R.Tensor((512, 512), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias4: R.Tensor((512,), dtype="float16") = packed_params[8] + gpt_neox_layers_0_mlp_dense_h_to_4h_weight4: R.Tensor((2048, 512), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[10] + gpt_neox_layers_0_mlp_dense_4h_to_h_weight4: R.Tensor((512, 2048), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: R.Tensor((512,), dtype="float32") = packed_params[12] + gpt_neox_layers_1_input_layernorm_weight4: R.Tensor((512,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias4: R.Tensor((512,), dtype="float16") = packed_params[14] + gpt_neox_layers_1_post_attention_layernorm_weight4: R.Tensor((512,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias4: R.Tensor((512,), dtype="float16") = packed_params[16] + gpt_neox_layers_1_attention_query_key_value_weight4: R.Tensor((1536, 512), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias4: R.Tensor((1536,), dtype="float16") = packed_params[18] + gpt_neox_layers_1_attention_dense_weight4: R.Tensor((512, 512), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias4: R.Tensor((512,), dtype="float16") = packed_params[20] + gpt_neox_layers_1_mlp_dense_h_to_4h_weight4: R.Tensor((2048, 512), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[22] + gpt_neox_layers_1_mlp_dense_4h_to_h_weight4: R.Tensor((512, 2048), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias4: R.Tensor((512,), dtype="float32") = packed_params[24] + gpt_neox_layers_2_input_layernorm_weight4: R.Tensor((512,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias4: R.Tensor((512,), dtype="float16") = packed_params[26] + gpt_neox_layers_2_post_attention_layernorm_weight4: R.Tensor((512,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias4: R.Tensor((512,), dtype="float16") = packed_params[28] + gpt_neox_layers_2_attention_query_key_value_weight4: R.Tensor((1536, 512), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias4: R.Tensor((1536,), dtype="float16") = packed_params[30] + gpt_neox_layers_2_attention_dense_weight4: R.Tensor((512, 512), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias4: R.Tensor((512,), dtype="float16") = packed_params[32] + gpt_neox_layers_2_mlp_dense_h_to_4h_weight4: R.Tensor((2048, 512), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[34] + gpt_neox_layers_2_mlp_dense_4h_to_h_weight4: R.Tensor((512, 2048), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias4: R.Tensor((512,), dtype="float32") = packed_params[36] + gpt_neox_layers_3_input_layernorm_weight4: R.Tensor((512,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias4: R.Tensor((512,), dtype="float16") = packed_params[38] + gpt_neox_layers_3_post_attention_layernorm_weight4: R.Tensor((512,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias4: R.Tensor((512,), dtype="float16") = packed_params[40] + gpt_neox_layers_3_attention_query_key_value_weight4: R.Tensor((1536, 512), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias4: R.Tensor((1536,), dtype="float16") = packed_params[42] + gpt_neox_layers_3_attention_dense_weight4: R.Tensor((512, 512), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias4: R.Tensor((512,), dtype="float16") = packed_params[44] + gpt_neox_layers_3_mlp_dense_h_to_4h_weight4: R.Tensor((2048, 512), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[46] + gpt_neox_layers_3_mlp_dense_4h_to_h_weight4: R.Tensor((512, 2048), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias4: R.Tensor((512,), dtype="float32") = packed_params[48] + gpt_neox_layers_4_input_layernorm_weight4: R.Tensor((512,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias4: R.Tensor((512,), dtype="float16") = packed_params[50] + gpt_neox_layers_4_post_attention_layernorm_weight4: R.Tensor((512,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias4: R.Tensor((512,), dtype="float16") = packed_params[52] + gpt_neox_layers_4_attention_query_key_value_weight4: R.Tensor((1536, 512), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias4: R.Tensor((1536,), dtype="float16") = packed_params[54] + gpt_neox_layers_4_attention_dense_weight4: R.Tensor((512, 512), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias4: R.Tensor((512,), dtype="float16") = packed_params[56] + gpt_neox_layers_4_mlp_dense_h_to_4h_weight4: R.Tensor((2048, 512), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[58] + gpt_neox_layers_4_mlp_dense_4h_to_h_weight4: R.Tensor((512, 2048), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias4: R.Tensor((512,), dtype="float32") = packed_params[60] + gpt_neox_layers_5_input_layernorm_weight4: R.Tensor((512,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias4: R.Tensor((512,), dtype="float16") = packed_params[62] + gpt_neox_layers_5_post_attention_layernorm_weight4: R.Tensor((512,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias4: R.Tensor((512,), dtype="float16") = packed_params[64] + gpt_neox_layers_5_attention_query_key_value_weight4: R.Tensor((1536, 512), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias4: R.Tensor((1536,), dtype="float16") = packed_params[66] + gpt_neox_layers_5_attention_dense_weight4: R.Tensor((512, 512), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias4: R.Tensor((512,), dtype="float16") = packed_params[68] + gpt_neox_layers_5_mlp_dense_h_to_4h_weight4: R.Tensor((2048, 512), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[70] + gpt_neox_layers_5_mlp_dense_4h_to_h_weight4: R.Tensor((512, 2048), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias4: R.Tensor((512,), dtype="float32") = packed_params[72] + gpt_neox_final_layer_norm_weight4: R.Tensor((512,), dtype="float16") = packed_params[73] + gpt_neox_final_layer_norm_bias4: R.Tensor((512,), dtype="float16") = packed_params[74] + embed_out_weight4: R.Tensor((vocab_size, 512), dtype="float16") = packed_params[75] + layer_norm39 = 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, 512), dtype="float16")) + lv = R.call_tir(cls.fused_NT_matmul_add, (layer_norm39, gpt_neox_layers_0_attention_query_key_value_weight4, gpt_neox_layers_0_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 1536), dtype="float16")) + reshape72 = R.call_tir(cls.reshape, (lv,), out_sinfo=R.Tensor((batch_size, 1, 24, 64), dtype="float16")) + reshape73 = R.call_tir(cls.reshape1, (reshape72,), out_sinfo=R.Tensor((batch_size, 24, 64), dtype="float16")) + lv19 = 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)), reshape73), out_sinfo=R.Tensor((batch_size, 8, 64), dtype="float16")) + reshape74 = R.call_tir(cls.reshape2, (lv19,), out_sinfo=R.Tensor((batch_size, 1, 8, 64), dtype="float16")) + reshape75 = R.call_tir(cls.reshape3, (reshape74,), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + layer_norm40 = 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, 512), dtype="float16")) + lv1 = R.call_tir(cls.fused_NT_matmul2_add2_gelu_cast, (layer_norm40, gpt_neox_layers_0_mlp_dense_h_to_4h_weight4, gpt_neox_layers_0_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + lv2 = R.call_tir(cls.fused_NT_matmul3_add3_cast1, (lv1, gpt_neox_layers_0_mlp_dense_4h_to_h_weight4, gpt_neox_layers_0_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv3 = R.call_tir(cls.fused_NT_matmul1_add1_add4_add4, (reshape75, gpt_neox_layers_0_attention_dense_weight4, gpt_neox_layers_0_attention_dense_bias4, lv2, input_embeds), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + layer_norm41 = R.call_tir(cls.layer_norm, (lv3, gpt_neox_layers_1_input_layernorm_weight4, gpt_neox_layers_1_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv4 = R.call_tir(cls.fused_NT_matmul_add, (layer_norm41, gpt_neox_layers_1_attention_query_key_value_weight4, gpt_neox_layers_1_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 1536), dtype="float16")) + reshape76 = R.call_tir(cls.reshape, (lv4,), out_sinfo=R.Tensor((batch_size, 1, 24, 64), dtype="float16")) + reshape77 = R.call_tir(cls.reshape1, (reshape76,), out_sinfo=R.Tensor((batch_size, 24, 64), dtype="float16")) + lv20 = 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)), reshape77), out_sinfo=R.Tensor((batch_size, 8, 64), dtype="float16")) + reshape78 = R.call_tir(cls.reshape2, (lv20,), out_sinfo=R.Tensor((batch_size, 1, 8, 64), dtype="float16")) + reshape79 = R.call_tir(cls.reshape3, (reshape78,), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + layer_norm42 = R.call_tir(cls.layer_norm, (lv3, gpt_neox_layers_1_post_attention_layernorm_weight4, gpt_neox_layers_1_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv5 = R.call_tir(cls.fused_NT_matmul2_add2_gelu_cast, (layer_norm42, gpt_neox_layers_1_mlp_dense_h_to_4h_weight4, gpt_neox_layers_1_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + lv6 = R.call_tir(cls.fused_NT_matmul3_add3_cast1, (lv5, gpt_neox_layers_1_mlp_dense_4h_to_h_weight4, gpt_neox_layers_1_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv7 = R.call_tir(cls.fused_NT_matmul1_add1_add4_add4, (reshape79, gpt_neox_layers_1_attention_dense_weight4, gpt_neox_layers_1_attention_dense_bias4, lv6, lv3), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + layer_norm43 = R.call_tir(cls.layer_norm, (lv7, gpt_neox_layers_2_input_layernorm_weight4, gpt_neox_layers_2_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv8 = R.call_tir(cls.fused_NT_matmul_add, (layer_norm43, gpt_neox_layers_2_attention_query_key_value_weight4, gpt_neox_layers_2_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 1536), dtype="float16")) + reshape80 = R.call_tir(cls.reshape, (lv8,), out_sinfo=R.Tensor((batch_size, 1, 24, 64), dtype="float16")) + reshape81 = R.call_tir(cls.reshape1, (reshape80,), out_sinfo=R.Tensor((batch_size, 24, 64), dtype="float16")) + lv21 = 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)), reshape81), out_sinfo=R.Tensor((batch_size, 8, 64), dtype="float16")) + reshape82 = R.call_tir(cls.reshape2, (lv21,), out_sinfo=R.Tensor((batch_size, 1, 8, 64), dtype="float16")) + reshape83 = R.call_tir(cls.reshape3, (reshape82,), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + layer_norm44 = R.call_tir(cls.layer_norm, (lv7, gpt_neox_layers_2_post_attention_layernorm_weight4, gpt_neox_layers_2_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv9 = R.call_tir(cls.fused_NT_matmul2_add2_gelu_cast, (layer_norm44, gpt_neox_layers_2_mlp_dense_h_to_4h_weight4, gpt_neox_layers_2_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + lv10 = R.call_tir(cls.fused_NT_matmul3_add3_cast1, (lv9, gpt_neox_layers_2_mlp_dense_4h_to_h_weight4, gpt_neox_layers_2_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv11 = R.call_tir(cls.fused_NT_matmul1_add1_add4_add4, (reshape83, gpt_neox_layers_2_attention_dense_weight4, gpt_neox_layers_2_attention_dense_bias4, lv10, lv7), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + layer_norm45 = R.call_tir(cls.layer_norm, (lv11, gpt_neox_layers_3_input_layernorm_weight4, gpt_neox_layers_3_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv12 = R.call_tir(cls.fused_NT_matmul_add, (layer_norm45, gpt_neox_layers_3_attention_query_key_value_weight4, gpt_neox_layers_3_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 1536), dtype="float16")) + reshape84 = R.call_tir(cls.reshape, (lv12,), out_sinfo=R.Tensor((batch_size, 1, 24, 64), dtype="float16")) + reshape85 = R.call_tir(cls.reshape1, (reshape84,), out_sinfo=R.Tensor((batch_size, 24, 64), dtype="float16")) + lv22 = 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)), reshape85), out_sinfo=R.Tensor((batch_size, 8, 64), dtype="float16")) + reshape86 = R.call_tir(cls.reshape2, (lv22,), out_sinfo=R.Tensor((batch_size, 1, 8, 64), dtype="float16")) + reshape87 = R.call_tir(cls.reshape3, (reshape86,), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + layer_norm46 = R.call_tir(cls.layer_norm, (lv11, gpt_neox_layers_3_post_attention_layernorm_weight4, gpt_neox_layers_3_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv13 = R.call_tir(cls.fused_NT_matmul2_add2_gelu_cast, (layer_norm46, gpt_neox_layers_3_mlp_dense_h_to_4h_weight4, gpt_neox_layers_3_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + lv14 = R.call_tir(cls.fused_NT_matmul3_add3_cast1, (lv13, gpt_neox_layers_3_mlp_dense_4h_to_h_weight4, gpt_neox_layers_3_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv15 = R.call_tir(cls.fused_NT_matmul1_add1_add4_add4, (reshape87, gpt_neox_layers_3_attention_dense_weight4, gpt_neox_layers_3_attention_dense_bias4, lv14, lv11), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + layer_norm47 = R.call_tir(cls.layer_norm, (lv15, gpt_neox_layers_4_input_layernorm_weight4, gpt_neox_layers_4_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv16 = R.call_tir(cls.fused_NT_matmul_add, (layer_norm47, gpt_neox_layers_4_attention_query_key_value_weight4, gpt_neox_layers_4_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 1536), dtype="float16")) + reshape88 = R.call_tir(cls.reshape, (lv16,), out_sinfo=R.Tensor((batch_size, 1, 24, 64), dtype="float16")) + reshape89 = R.call_tir(cls.reshape1, (reshape88,), out_sinfo=R.Tensor((batch_size, 24, 64), dtype="float16")) + lv23 = 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)), reshape89), out_sinfo=R.Tensor((batch_size, 8, 64), dtype="float16")) + reshape90 = R.call_tir(cls.reshape2, (lv23,), out_sinfo=R.Tensor((batch_size, 1, 8, 64), dtype="float16")) + reshape91 = R.call_tir(cls.reshape3, (reshape90,), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + layer_norm48 = R.call_tir(cls.layer_norm, (lv15, gpt_neox_layers_4_post_attention_layernorm_weight4, gpt_neox_layers_4_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv17 = R.call_tir(cls.fused_NT_matmul2_add2_gelu_cast, (layer_norm48, gpt_neox_layers_4_mlp_dense_h_to_4h_weight4, gpt_neox_layers_4_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + lv18 = R.call_tir(cls.fused_NT_matmul3_add3_cast1, (lv17, gpt_neox_layers_4_mlp_dense_4h_to_h_weight4, gpt_neox_layers_4_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv19_1 = R.call_tir(cls.fused_NT_matmul1_add1_add4_add4, (reshape91, gpt_neox_layers_4_attention_dense_weight4, gpt_neox_layers_4_attention_dense_bias4, lv18, lv15), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + layer_norm49 = R.call_tir(cls.layer_norm, (lv19_1, gpt_neox_layers_5_input_layernorm_weight4, gpt_neox_layers_5_input_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv20_1 = R.call_tir(cls.fused_NT_matmul_add, (layer_norm49, gpt_neox_layers_5_attention_query_key_value_weight4, gpt_neox_layers_5_attention_query_key_value_bias4), out_sinfo=R.Tensor((batch_size, 1, 1536), dtype="float16")) + reshape92 = R.call_tir(cls.reshape, (lv20_1,), out_sinfo=R.Tensor((batch_size, 1, 24, 64), dtype="float16")) + reshape93 = R.call_tir(cls.reshape1, (reshape92,), out_sinfo=R.Tensor((batch_size, 24, 64), dtype="float16")) + lv24 = 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)), reshape93), out_sinfo=R.Tensor((batch_size, 8, 64), dtype="float16")) + reshape94 = R.call_tir(cls.reshape2, (lv24,), out_sinfo=R.Tensor((batch_size, 1, 8, 64), dtype="float16")) + reshape95 = R.call_tir(cls.reshape3, (reshape94,), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + layer_norm50 = R.call_tir(cls.layer_norm, (lv19_1, gpt_neox_layers_5_post_attention_layernorm_weight4, gpt_neox_layers_5_post_attention_layernorm_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv21_1 = R.call_tir(cls.fused_NT_matmul2_add2_gelu_cast, (layer_norm50, gpt_neox_layers_5_mlp_dense_h_to_4h_weight4, gpt_neox_layers_5_mlp_dense_h_to_4h_bias4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + lv22_1 = R.call_tir(cls.fused_NT_matmul3_add3_cast1, (lv21_1, gpt_neox_layers_5_mlp_dense_4h_to_h_weight4, gpt_neox_layers_5_mlp_dense_4h_to_h_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv23_1 = R.call_tir(cls.fused_NT_matmul1_add1_add4_add4, (reshape95, gpt_neox_layers_5_attention_dense_weight4, gpt_neox_layers_5_attention_dense_bias4, lv22_1, lv19_1), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + layer_norm51 = R.call_tir(cls.layer_norm, (lv23_1, gpt_neox_final_layer_norm_weight4, gpt_neox_final_layer_norm_bias4), out_sinfo=R.Tensor((batch_size, 1, 512), dtype="float16")) + lv24_1 = R.call_tir(cls.fused_NT_matmul4_cast2, (layer_norm51, embed_out_weight4), 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) = lv24_1, paged_kv_cache + R.output(gv4) + return gv4 + + @R.function + def batch_prefill(input_embeds: R.Tensor((1, "seq_len", 512), dtype="float16"), logit_positions: R.Tensor(("batch_size",), dtype="int32"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor(("vocab_size", 512), 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((512,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias3: R.Tensor((512,), dtype="float16") = packed_params[2] + gpt_neox_layers_0_post_attention_layernorm_weight3: R.Tensor((512,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias3: R.Tensor((512,), dtype="float16") = packed_params[4] + gpt_neox_layers_0_attention_query_key_value_weight3: R.Tensor((1536, 512), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias3: R.Tensor((1536,), dtype="float16") = packed_params[6] + gpt_neox_layers_0_attention_dense_weight3: R.Tensor((512, 512), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias3: R.Tensor((512,), dtype="float16") = packed_params[8] + gpt_neox_layers_0_mlp_dense_h_to_4h_weight3: R.Tensor((2048, 512), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[10] + gpt_neox_layers_0_mlp_dense_4h_to_h_weight3: R.Tensor((512, 2048), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: R.Tensor((512,), dtype="float32") = packed_params[12] + gpt_neox_layers_1_input_layernorm_weight3: R.Tensor((512,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias3: R.Tensor((512,), dtype="float16") = packed_params[14] + gpt_neox_layers_1_post_attention_layernorm_weight3: R.Tensor((512,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias3: R.Tensor((512,), dtype="float16") = packed_params[16] + gpt_neox_layers_1_attention_query_key_value_weight3: R.Tensor((1536, 512), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias3: R.Tensor((1536,), dtype="float16") = packed_params[18] + gpt_neox_layers_1_attention_dense_weight3: R.Tensor((512, 512), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias3: R.Tensor((512,), dtype="float16") = packed_params[20] + gpt_neox_layers_1_mlp_dense_h_to_4h_weight3: R.Tensor((2048, 512), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[22] + gpt_neox_layers_1_mlp_dense_4h_to_h_weight3: R.Tensor((512, 2048), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias3: R.Tensor((512,), dtype="float32") = packed_params[24] + gpt_neox_layers_2_input_layernorm_weight3: R.Tensor((512,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias3: R.Tensor((512,), dtype="float16") = packed_params[26] + gpt_neox_layers_2_post_attention_layernorm_weight3: R.Tensor((512,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias3: R.Tensor((512,), dtype="float16") = packed_params[28] + gpt_neox_layers_2_attention_query_key_value_weight3: R.Tensor((1536, 512), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias3: R.Tensor((1536,), dtype="float16") = packed_params[30] + gpt_neox_layers_2_attention_dense_weight3: R.Tensor((512, 512), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias3: R.Tensor((512,), dtype="float16") = packed_params[32] + gpt_neox_layers_2_mlp_dense_h_to_4h_weight3: R.Tensor((2048, 512), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[34] + gpt_neox_layers_2_mlp_dense_4h_to_h_weight3: R.Tensor((512, 2048), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias3: R.Tensor((512,), dtype="float32") = packed_params[36] + gpt_neox_layers_3_input_layernorm_weight3: R.Tensor((512,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias3: R.Tensor((512,), dtype="float16") = packed_params[38] + gpt_neox_layers_3_post_attention_layernorm_weight3: R.Tensor((512,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias3: R.Tensor((512,), dtype="float16") = packed_params[40] + gpt_neox_layers_3_attention_query_key_value_weight3: R.Tensor((1536, 512), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias3: R.Tensor((1536,), dtype="float16") = packed_params[42] + gpt_neox_layers_3_attention_dense_weight3: R.Tensor((512, 512), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias3: R.Tensor((512,), dtype="float16") = packed_params[44] + gpt_neox_layers_3_mlp_dense_h_to_4h_weight3: R.Tensor((2048, 512), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[46] + gpt_neox_layers_3_mlp_dense_4h_to_h_weight3: R.Tensor((512, 2048), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias3: R.Tensor((512,), dtype="float32") = packed_params[48] + gpt_neox_layers_4_input_layernorm_weight3: R.Tensor((512,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias3: R.Tensor((512,), dtype="float16") = packed_params[50] + gpt_neox_layers_4_post_attention_layernorm_weight3: R.Tensor((512,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias3: R.Tensor((512,), dtype="float16") = packed_params[52] + gpt_neox_layers_4_attention_query_key_value_weight3: R.Tensor((1536, 512), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias3: R.Tensor((1536,), dtype="float16") = packed_params[54] + gpt_neox_layers_4_attention_dense_weight3: R.Tensor((512, 512), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias3: R.Tensor((512,), dtype="float16") = packed_params[56] + gpt_neox_layers_4_mlp_dense_h_to_4h_weight3: R.Tensor((2048, 512), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[58] + gpt_neox_layers_4_mlp_dense_4h_to_h_weight3: R.Tensor((512, 2048), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias3: R.Tensor((512,), dtype="float32") = packed_params[60] + gpt_neox_layers_5_input_layernorm_weight3: R.Tensor((512,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias3: R.Tensor((512,), dtype="float16") = packed_params[62] + gpt_neox_layers_5_post_attention_layernorm_weight3: R.Tensor((512,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias3: R.Tensor((512,), dtype="float16") = packed_params[64] + gpt_neox_layers_5_attention_query_key_value_weight3: R.Tensor((1536, 512), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias3: R.Tensor((1536,), dtype="float16") = packed_params[66] + gpt_neox_layers_5_attention_dense_weight3: R.Tensor((512, 512), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias3: R.Tensor((512,), dtype="float16") = packed_params[68] + gpt_neox_layers_5_mlp_dense_h_to_4h_weight3: R.Tensor((2048, 512), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[70] + gpt_neox_layers_5_mlp_dense_4h_to_h_weight3: R.Tensor((512, 2048), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias3: R.Tensor((512,), dtype="float32") = packed_params[72] + gpt_neox_final_layer_norm_weight3: R.Tensor((512,), dtype="float16") = packed_params[73] + gpt_neox_final_layer_norm_bias3: R.Tensor((512,), dtype="float16") = packed_params[74] + embed_out_weight3: R.Tensor((vocab_size, 512), dtype="float16") = packed_params[75] + layer_norm26 = 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, 512), dtype="float16")) + lv25 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm26, gpt_neox_layers_0_attention_query_key_value_weight3, gpt_neox_layers_0_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape48 = R.call_tir(cls.reshape4, (lv25,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape49 = R.call_tir(cls.reshape5, (reshape48,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv13 = 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)), reshape49), out_sinfo=R.Tensor((seq_len, 8, 64), dtype="float16")) + reshape50 = R.call_tir(cls.reshape6, (lv13,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape51 = R.call_tir(cls.reshape7, (reshape50,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm27 = 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, 512), dtype="float16")) + lv26 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm27, gpt_neox_layers_0_mlp_dense_h_to_4h_weight3, gpt_neox_layers_0_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv27 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv26, gpt_neox_layers_0_mlp_dense_4h_to_h_weight3, gpt_neox_layers_0_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv28 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape51, gpt_neox_layers_0_attention_dense_weight3, gpt_neox_layers_0_attention_dense_bias3, lv27, input_embeds), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm28 = R.call_tir(cls.layer_norm1, (lv28, gpt_neox_layers_1_input_layernorm_weight3, gpt_neox_layers_1_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv29 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm28, gpt_neox_layers_1_attention_query_key_value_weight3, gpt_neox_layers_1_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape52 = R.call_tir(cls.reshape4, (lv29,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape53 = R.call_tir(cls.reshape5, (reshape52,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv14 = 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)), reshape53), out_sinfo=R.Tensor((seq_len, 8, 64), dtype="float16")) + reshape54 = R.call_tir(cls.reshape6, (lv14,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape55 = R.call_tir(cls.reshape7, (reshape54,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm29 = R.call_tir(cls.layer_norm1, (lv28, gpt_neox_layers_1_post_attention_layernorm_weight3, gpt_neox_layers_1_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv30 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm29, gpt_neox_layers_1_mlp_dense_h_to_4h_weight3, gpt_neox_layers_1_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv31 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv30, gpt_neox_layers_1_mlp_dense_4h_to_h_weight3, gpt_neox_layers_1_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv32 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape55, gpt_neox_layers_1_attention_dense_weight3, gpt_neox_layers_1_attention_dense_bias3, lv31, lv28), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm30 = R.call_tir(cls.layer_norm1, (lv32, gpt_neox_layers_2_input_layernorm_weight3, gpt_neox_layers_2_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv33 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm30, gpt_neox_layers_2_attention_query_key_value_weight3, gpt_neox_layers_2_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape56 = R.call_tir(cls.reshape4, (lv33,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape57 = R.call_tir(cls.reshape5, (reshape56,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv15 = 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)), reshape57), out_sinfo=R.Tensor((seq_len, 8, 64), dtype="float16")) + reshape58 = R.call_tir(cls.reshape6, (lv15,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape59 = R.call_tir(cls.reshape7, (reshape58,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm31 = R.call_tir(cls.layer_norm1, (lv32, gpt_neox_layers_2_post_attention_layernorm_weight3, gpt_neox_layers_2_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv34 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm31, gpt_neox_layers_2_mlp_dense_h_to_4h_weight3, gpt_neox_layers_2_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv35 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv34, gpt_neox_layers_2_mlp_dense_4h_to_h_weight3, gpt_neox_layers_2_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv36 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape59, gpt_neox_layers_2_attention_dense_weight3, gpt_neox_layers_2_attention_dense_bias3, lv35, lv32), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm32 = R.call_tir(cls.layer_norm1, (lv36, gpt_neox_layers_3_input_layernorm_weight3, gpt_neox_layers_3_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv37 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm32, gpt_neox_layers_3_attention_query_key_value_weight3, gpt_neox_layers_3_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape60 = R.call_tir(cls.reshape4, (lv37,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape61 = R.call_tir(cls.reshape5, (reshape60,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv16 = 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)), reshape61), out_sinfo=R.Tensor((seq_len, 8, 64), dtype="float16")) + reshape62 = R.call_tir(cls.reshape6, (lv16,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape63 = R.call_tir(cls.reshape7, (reshape62,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm33 = R.call_tir(cls.layer_norm1, (lv36, gpt_neox_layers_3_post_attention_layernorm_weight3, gpt_neox_layers_3_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv38 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm33, gpt_neox_layers_3_mlp_dense_h_to_4h_weight3, gpt_neox_layers_3_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv39 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv38, gpt_neox_layers_3_mlp_dense_4h_to_h_weight3, gpt_neox_layers_3_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv40 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape63, gpt_neox_layers_3_attention_dense_weight3, gpt_neox_layers_3_attention_dense_bias3, lv39, lv36), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm34 = R.call_tir(cls.layer_norm1, (lv40, gpt_neox_layers_4_input_layernorm_weight3, gpt_neox_layers_4_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv41 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm34, gpt_neox_layers_4_attention_query_key_value_weight3, gpt_neox_layers_4_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape64 = R.call_tir(cls.reshape4, (lv41,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape65 = R.call_tir(cls.reshape5, (reshape64,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv17 = 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)), reshape65), out_sinfo=R.Tensor((seq_len, 8, 64), dtype="float16")) + reshape66 = R.call_tir(cls.reshape6, (lv17,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape67 = R.call_tir(cls.reshape7, (reshape66,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm35 = R.call_tir(cls.layer_norm1, (lv40, gpt_neox_layers_4_post_attention_layernorm_weight3, gpt_neox_layers_4_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv42 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm35, gpt_neox_layers_4_mlp_dense_h_to_4h_weight3, gpt_neox_layers_4_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv43 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv42, gpt_neox_layers_4_mlp_dense_4h_to_h_weight3, gpt_neox_layers_4_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv44 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape67, gpt_neox_layers_4_attention_dense_weight3, gpt_neox_layers_4_attention_dense_bias3, lv43, lv40), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm36 = R.call_tir(cls.layer_norm1, (lv44, gpt_neox_layers_5_input_layernorm_weight3, gpt_neox_layers_5_input_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv45 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm36, gpt_neox_layers_5_attention_query_key_value_weight3, gpt_neox_layers_5_attention_query_key_value_bias3), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape68 = R.call_tir(cls.reshape4, (lv45,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape69 = R.call_tir(cls.reshape5, (reshape68,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv18 = 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)), reshape69), out_sinfo=R.Tensor((seq_len, 8, 64), dtype="float16")) + reshape70 = R.call_tir(cls.reshape6, (lv18,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape71 = R.call_tir(cls.reshape7, (reshape70,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm37 = R.call_tir(cls.layer_norm1, (lv44, gpt_neox_layers_5_post_attention_layernorm_weight3, gpt_neox_layers_5_post_attention_layernorm_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv46 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm37, gpt_neox_layers_5_mlp_dense_h_to_4h_weight3, gpt_neox_layers_5_mlp_dense_h_to_4h_bias3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv47 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv46, gpt_neox_layers_5_mlp_dense_4h_to_h_weight3, gpt_neox_layers_5_mlp_dense_4h_to_h_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv48 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape71, gpt_neox_layers_5_attention_dense_weight3, gpt_neox_layers_5_attention_dense_bias3, lv47, lv44), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm38 = R.call_tir(cls.layer_norm1, (lv48, gpt_neox_final_layer_norm_weight3, gpt_neox_final_layer_norm_bias3), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + take1 = R.call_tir(cls.take, (layer_norm38, logit_positions), out_sinfo=R.Tensor((1, batch_size, 512), dtype="float16")) + lv49 = R.call_tir(cls.fused_NT_matmul9_cast5, (take1, embed_out_weight3), 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) = lv49, paged_kv_cache + R.output(gv3) + return gv3 + + @R.function + def batch_verify(input_embeds: R.Tensor((1, "seq_len", 512), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor(("vocab_size", 512), 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((512,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias5: R.Tensor((512,), dtype="float16") = packed_params[2] + gpt_neox_layers_0_post_attention_layernorm_weight5: R.Tensor((512,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias5: R.Tensor((512,), dtype="float16") = packed_params[4] + gpt_neox_layers_0_attention_query_key_value_weight5: R.Tensor((1536, 512), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias5: R.Tensor((1536,), dtype="float16") = packed_params[6] + gpt_neox_layers_0_attention_dense_weight5: R.Tensor((512, 512), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias5: R.Tensor((512,), dtype="float16") = packed_params[8] + gpt_neox_layers_0_mlp_dense_h_to_4h_weight5: R.Tensor((2048, 512), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[10] + gpt_neox_layers_0_mlp_dense_4h_to_h_weight5: R.Tensor((512, 2048), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias5: R.Tensor((512,), dtype="float32") = packed_params[12] + gpt_neox_layers_1_input_layernorm_weight5: R.Tensor((512,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias5: R.Tensor((512,), dtype="float16") = packed_params[14] + gpt_neox_layers_1_post_attention_layernorm_weight5: R.Tensor((512,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias5: R.Tensor((512,), dtype="float16") = packed_params[16] + gpt_neox_layers_1_attention_query_key_value_weight5: R.Tensor((1536, 512), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias5: R.Tensor((1536,), dtype="float16") = packed_params[18] + gpt_neox_layers_1_attention_dense_weight5: R.Tensor((512, 512), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias5: R.Tensor((512,), dtype="float16") = packed_params[20] + gpt_neox_layers_1_mlp_dense_h_to_4h_weight5: R.Tensor((2048, 512), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[22] + gpt_neox_layers_1_mlp_dense_4h_to_h_weight5: R.Tensor((512, 2048), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias5: R.Tensor((512,), dtype="float32") = packed_params[24] + gpt_neox_layers_2_input_layernorm_weight5: R.Tensor((512,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias5: R.Tensor((512,), dtype="float16") = packed_params[26] + gpt_neox_layers_2_post_attention_layernorm_weight5: R.Tensor((512,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias5: R.Tensor((512,), dtype="float16") = packed_params[28] + gpt_neox_layers_2_attention_query_key_value_weight5: R.Tensor((1536, 512), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias5: R.Tensor((1536,), dtype="float16") = packed_params[30] + gpt_neox_layers_2_attention_dense_weight5: R.Tensor((512, 512), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias5: R.Tensor((512,), dtype="float16") = packed_params[32] + gpt_neox_layers_2_mlp_dense_h_to_4h_weight5: R.Tensor((2048, 512), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[34] + gpt_neox_layers_2_mlp_dense_4h_to_h_weight5: R.Tensor((512, 2048), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias5: R.Tensor((512,), dtype="float32") = packed_params[36] + gpt_neox_layers_3_input_layernorm_weight5: R.Tensor((512,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias5: R.Tensor((512,), dtype="float16") = packed_params[38] + gpt_neox_layers_3_post_attention_layernorm_weight5: R.Tensor((512,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias5: R.Tensor((512,), dtype="float16") = packed_params[40] + gpt_neox_layers_3_attention_query_key_value_weight5: R.Tensor((1536, 512), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias5: R.Tensor((1536,), dtype="float16") = packed_params[42] + gpt_neox_layers_3_attention_dense_weight5: R.Tensor((512, 512), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias5: R.Tensor((512,), dtype="float16") = packed_params[44] + gpt_neox_layers_3_mlp_dense_h_to_4h_weight5: R.Tensor((2048, 512), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[46] + gpt_neox_layers_3_mlp_dense_4h_to_h_weight5: R.Tensor((512, 2048), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias5: R.Tensor((512,), dtype="float32") = packed_params[48] + gpt_neox_layers_4_input_layernorm_weight5: R.Tensor((512,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias5: R.Tensor((512,), dtype="float16") = packed_params[50] + gpt_neox_layers_4_post_attention_layernorm_weight5: R.Tensor((512,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias5: R.Tensor((512,), dtype="float16") = packed_params[52] + gpt_neox_layers_4_attention_query_key_value_weight5: R.Tensor((1536, 512), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias5: R.Tensor((1536,), dtype="float16") = packed_params[54] + gpt_neox_layers_4_attention_dense_weight5: R.Tensor((512, 512), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias5: R.Tensor((512,), dtype="float16") = packed_params[56] + gpt_neox_layers_4_mlp_dense_h_to_4h_weight5: R.Tensor((2048, 512), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[58] + gpt_neox_layers_4_mlp_dense_4h_to_h_weight5: R.Tensor((512, 2048), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias5: R.Tensor((512,), dtype="float32") = packed_params[60] + gpt_neox_layers_5_input_layernorm_weight5: R.Tensor((512,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias5: R.Tensor((512,), dtype="float16") = packed_params[62] + gpt_neox_layers_5_post_attention_layernorm_weight5: R.Tensor((512,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias5: R.Tensor((512,), dtype="float16") = packed_params[64] + gpt_neox_layers_5_attention_query_key_value_weight5: R.Tensor((1536, 512), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias5: R.Tensor((1536,), dtype="float16") = packed_params[66] + gpt_neox_layers_5_attention_dense_weight5: R.Tensor((512, 512), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias5: R.Tensor((512,), dtype="float16") = packed_params[68] + gpt_neox_layers_5_mlp_dense_h_to_4h_weight5: R.Tensor((2048, 512), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[70] + gpt_neox_layers_5_mlp_dense_4h_to_h_weight5: R.Tensor((512, 2048), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias5: R.Tensor((512,), dtype="float32") = packed_params[72] + gpt_neox_final_layer_norm_weight5: R.Tensor((512,), dtype="float16") = packed_params[73] + gpt_neox_final_layer_norm_bias5: R.Tensor((512,), dtype="float16") = packed_params[74] + embed_out_weight5: R.Tensor((vocab_size, 512), dtype="float16") = packed_params[75] + layer_norm52 = 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, 512), dtype="float16")) + lv50 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm52, gpt_neox_layers_0_attention_query_key_value_weight5, gpt_neox_layers_0_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape96 = R.call_tir(cls.reshape4, (lv50,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape97 = R.call_tir(cls.reshape5, (reshape96,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv25 = 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)), reshape97), out_sinfo=R.Tensor((seq_len, 8, 64), dtype="float16")) + reshape98 = R.call_tir(cls.reshape6, (lv25,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape99 = R.call_tir(cls.reshape7, (reshape98,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm53 = 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, 512), dtype="float16")) + lv51 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm53, gpt_neox_layers_0_mlp_dense_h_to_4h_weight5, gpt_neox_layers_0_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv52 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv51, gpt_neox_layers_0_mlp_dense_4h_to_h_weight5, gpt_neox_layers_0_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv53 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape99, gpt_neox_layers_0_attention_dense_weight5, gpt_neox_layers_0_attention_dense_bias5, lv52, input_embeds), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm54 = R.call_tir(cls.layer_norm1, (lv53, gpt_neox_layers_1_input_layernorm_weight5, gpt_neox_layers_1_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv54 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm54, gpt_neox_layers_1_attention_query_key_value_weight5, gpt_neox_layers_1_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape100 = R.call_tir(cls.reshape4, (lv54,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape101 = R.call_tir(cls.reshape5, (reshape100,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv26 = 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)), reshape101), out_sinfo=R.Tensor((seq_len, 8, 64), dtype="float16")) + reshape102 = R.call_tir(cls.reshape6, (lv26,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape103 = R.call_tir(cls.reshape7, (reshape102,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm55 = R.call_tir(cls.layer_norm1, (lv53, gpt_neox_layers_1_post_attention_layernorm_weight5, gpt_neox_layers_1_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv55 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm55, gpt_neox_layers_1_mlp_dense_h_to_4h_weight5, gpt_neox_layers_1_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv56 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv55, gpt_neox_layers_1_mlp_dense_4h_to_h_weight5, gpt_neox_layers_1_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv57 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape103, gpt_neox_layers_1_attention_dense_weight5, gpt_neox_layers_1_attention_dense_bias5, lv56, lv53), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm56 = R.call_tir(cls.layer_norm1, (lv57, gpt_neox_layers_2_input_layernorm_weight5, gpt_neox_layers_2_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv58 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm56, gpt_neox_layers_2_attention_query_key_value_weight5, gpt_neox_layers_2_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape104 = R.call_tir(cls.reshape4, (lv58,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape105 = R.call_tir(cls.reshape5, (reshape104,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv27 = 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)), reshape105), out_sinfo=R.Tensor((seq_len, 8, 64), dtype="float16")) + reshape106 = R.call_tir(cls.reshape6, (lv27,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape107 = R.call_tir(cls.reshape7, (reshape106,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm57 = R.call_tir(cls.layer_norm1, (lv57, gpt_neox_layers_2_post_attention_layernorm_weight5, gpt_neox_layers_2_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv59 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm57, gpt_neox_layers_2_mlp_dense_h_to_4h_weight5, gpt_neox_layers_2_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv60 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv59, gpt_neox_layers_2_mlp_dense_4h_to_h_weight5, gpt_neox_layers_2_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv61 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape107, gpt_neox_layers_2_attention_dense_weight5, gpt_neox_layers_2_attention_dense_bias5, lv60, lv57), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm58 = R.call_tir(cls.layer_norm1, (lv61, gpt_neox_layers_3_input_layernorm_weight5, gpt_neox_layers_3_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv62 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm58, gpt_neox_layers_3_attention_query_key_value_weight5, gpt_neox_layers_3_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape108 = R.call_tir(cls.reshape4, (lv62,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape109 = R.call_tir(cls.reshape5, (reshape108,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv28 = 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)), reshape109), out_sinfo=R.Tensor((seq_len, 8, 64), dtype="float16")) + reshape110 = R.call_tir(cls.reshape6, (lv28,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape111 = R.call_tir(cls.reshape7, (reshape110,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm59 = R.call_tir(cls.layer_norm1, (lv61, gpt_neox_layers_3_post_attention_layernorm_weight5, gpt_neox_layers_3_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv63 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm59, gpt_neox_layers_3_mlp_dense_h_to_4h_weight5, gpt_neox_layers_3_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv64 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv63, gpt_neox_layers_3_mlp_dense_4h_to_h_weight5, gpt_neox_layers_3_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv65 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape111, gpt_neox_layers_3_attention_dense_weight5, gpt_neox_layers_3_attention_dense_bias5, lv64, lv61), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm60 = R.call_tir(cls.layer_norm1, (lv65, gpt_neox_layers_4_input_layernorm_weight5, gpt_neox_layers_4_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv66 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm60, gpt_neox_layers_4_attention_query_key_value_weight5, gpt_neox_layers_4_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape112 = R.call_tir(cls.reshape4, (lv66,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape113 = R.call_tir(cls.reshape5, (reshape112,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv29 = 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)), reshape113), out_sinfo=R.Tensor((seq_len, 8, 64), dtype="float16")) + reshape114 = R.call_tir(cls.reshape6, (lv29,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape115 = R.call_tir(cls.reshape7, (reshape114,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm61 = R.call_tir(cls.layer_norm1, (lv65, gpt_neox_layers_4_post_attention_layernorm_weight5, gpt_neox_layers_4_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv67 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm61, gpt_neox_layers_4_mlp_dense_h_to_4h_weight5, gpt_neox_layers_4_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv68 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv67, gpt_neox_layers_4_mlp_dense_4h_to_h_weight5, gpt_neox_layers_4_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv69 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape115, gpt_neox_layers_4_attention_dense_weight5, gpt_neox_layers_4_attention_dense_bias5, lv68, lv65), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm62 = R.call_tir(cls.layer_norm1, (lv69, gpt_neox_layers_5_input_layernorm_weight5, gpt_neox_layers_5_input_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv70 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm62, gpt_neox_layers_5_attention_query_key_value_weight5, gpt_neox_layers_5_attention_query_key_value_bias5), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape116 = R.call_tir(cls.reshape4, (lv70,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape117 = R.call_tir(cls.reshape5, (reshape116,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv30 = 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)), reshape117), out_sinfo=R.Tensor((seq_len, 8, 64), dtype="float16")) + reshape118 = R.call_tir(cls.reshape6, (lv30,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape119 = R.call_tir(cls.reshape7, (reshape118,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm63 = R.call_tir(cls.layer_norm1, (lv69, gpt_neox_layers_5_post_attention_layernorm_weight5, gpt_neox_layers_5_post_attention_layernorm_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv71 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm63, gpt_neox_layers_5_mlp_dense_h_to_4h_weight5, gpt_neox_layers_5_mlp_dense_h_to_4h_bias5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv72 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv71, gpt_neox_layers_5_mlp_dense_4h_to_h_weight5, gpt_neox_layers_5_mlp_dense_4h_to_h_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv73 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape119, gpt_neox_layers_5_attention_dense_weight5, gpt_neox_layers_5_attention_dense_bias5, lv72, lv69), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm64 = R.call_tir(cls.layer_norm1, (lv73, gpt_neox_final_layer_norm_weight5, gpt_neox_final_layer_norm_bias5), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv74 = R.call_tir(cls.fused_NT_matmul9_cast5, (layer_norm64, embed_out_weight5), 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) = lv74, 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, 6]), R.prim_value(8), R.prim_value(8), R.prim_value(64), R.prim_value(1), R.prim_value(1), R.prim_value(10000), R.const(0.0, "float16"), cls.tir_kv_cache_transpose_append, cls.batch_prefill_paged_kv, cls.batch_decode_paged_kv, cls.batch_prefill_paged_kv_sliding_window, cls.batch_decode_paged_kv_sliding_window, cls.batch_prefill_ragged_kv, cls.merge_state_inplace, cls.fused_rope, cls.copy_single_page, cls.tir_kv_cache_debug_get_kv, cls.compact_kv_copy, cls.batch_tree_attn, cls.tree_attn_paged_kv, R.prim_value(0), R.prim_value(0), sinfo_args=(R.Object,)) + return paged_kv_cache + + @R.function + def decode(input_embed: R.Tensor((1, 1, 512), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor(("vocab_size", 512), 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((512,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias2: R.Tensor((512,), dtype="float16") = packed_params[2] + gpt_neox_layers_0_post_attention_layernorm_weight2: R.Tensor((512,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias2: R.Tensor((512,), dtype="float16") = packed_params[4] + gpt_neox_layers_0_attention_query_key_value_weight2: R.Tensor((1536, 512), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias2: R.Tensor((1536,), dtype="float16") = packed_params[6] + gpt_neox_layers_0_attention_dense_weight2: R.Tensor((512, 512), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias2: R.Tensor((512,), dtype="float16") = packed_params[8] + gpt_neox_layers_0_mlp_dense_h_to_4h_weight2: R.Tensor((2048, 512), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[10] + gpt_neox_layers_0_mlp_dense_4h_to_h_weight2: R.Tensor((512, 2048), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: R.Tensor((512,), dtype="float32") = packed_params[12] + gpt_neox_layers_1_input_layernorm_weight2: R.Tensor((512,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias2: R.Tensor((512,), dtype="float16") = packed_params[14] + gpt_neox_layers_1_post_attention_layernorm_weight2: R.Tensor((512,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias2: R.Tensor((512,), dtype="float16") = packed_params[16] + gpt_neox_layers_1_attention_query_key_value_weight2: R.Tensor((1536, 512), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias2: R.Tensor((1536,), dtype="float16") = packed_params[18] + gpt_neox_layers_1_attention_dense_weight2: R.Tensor((512, 512), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias2: R.Tensor((512,), dtype="float16") = packed_params[20] + gpt_neox_layers_1_mlp_dense_h_to_4h_weight2: R.Tensor((2048, 512), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[22] + gpt_neox_layers_1_mlp_dense_4h_to_h_weight2: R.Tensor((512, 2048), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias2: R.Tensor((512,), dtype="float32") = packed_params[24] + gpt_neox_layers_2_input_layernorm_weight2: R.Tensor((512,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias2: R.Tensor((512,), dtype="float16") = packed_params[26] + gpt_neox_layers_2_post_attention_layernorm_weight2: R.Tensor((512,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias2: R.Tensor((512,), dtype="float16") = packed_params[28] + gpt_neox_layers_2_attention_query_key_value_weight2: R.Tensor((1536, 512), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias2: R.Tensor((1536,), dtype="float16") = packed_params[30] + gpt_neox_layers_2_attention_dense_weight2: R.Tensor((512, 512), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias2: R.Tensor((512,), dtype="float16") = packed_params[32] + gpt_neox_layers_2_mlp_dense_h_to_4h_weight2: R.Tensor((2048, 512), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[34] + gpt_neox_layers_2_mlp_dense_4h_to_h_weight2: R.Tensor((512, 2048), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias2: R.Tensor((512,), dtype="float32") = packed_params[36] + gpt_neox_layers_3_input_layernorm_weight2: R.Tensor((512,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias2: R.Tensor((512,), dtype="float16") = packed_params[38] + gpt_neox_layers_3_post_attention_layernorm_weight2: R.Tensor((512,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias2: R.Tensor((512,), dtype="float16") = packed_params[40] + gpt_neox_layers_3_attention_query_key_value_weight2: R.Tensor((1536, 512), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias2: R.Tensor((1536,), dtype="float16") = packed_params[42] + gpt_neox_layers_3_attention_dense_weight2: R.Tensor((512, 512), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias2: R.Tensor((512,), dtype="float16") = packed_params[44] + gpt_neox_layers_3_mlp_dense_h_to_4h_weight2: R.Tensor((2048, 512), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[46] + gpt_neox_layers_3_mlp_dense_4h_to_h_weight2: R.Tensor((512, 2048), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias2: R.Tensor((512,), dtype="float32") = packed_params[48] + gpt_neox_layers_4_input_layernorm_weight2: R.Tensor((512,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias2: R.Tensor((512,), dtype="float16") = packed_params[50] + gpt_neox_layers_4_post_attention_layernorm_weight2: R.Tensor((512,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias2: R.Tensor((512,), dtype="float16") = packed_params[52] + gpt_neox_layers_4_attention_query_key_value_weight2: R.Tensor((1536, 512), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias2: R.Tensor((1536,), dtype="float16") = packed_params[54] + gpt_neox_layers_4_attention_dense_weight2: R.Tensor((512, 512), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias2: R.Tensor((512,), dtype="float16") = packed_params[56] + gpt_neox_layers_4_mlp_dense_h_to_4h_weight2: R.Tensor((2048, 512), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[58] + gpt_neox_layers_4_mlp_dense_4h_to_h_weight2: R.Tensor((512, 2048), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias2: R.Tensor((512,), dtype="float32") = packed_params[60] + gpt_neox_layers_5_input_layernorm_weight2: R.Tensor((512,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias2: R.Tensor((512,), dtype="float16") = packed_params[62] + gpt_neox_layers_5_post_attention_layernorm_weight2: R.Tensor((512,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias2: R.Tensor((512,), dtype="float16") = packed_params[64] + gpt_neox_layers_5_attention_query_key_value_weight2: R.Tensor((1536, 512), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias2: R.Tensor((1536,), dtype="float16") = packed_params[66] + gpt_neox_layers_5_attention_dense_weight2: R.Tensor((512, 512), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias2: R.Tensor((512,), dtype="float16") = packed_params[68] + gpt_neox_layers_5_mlp_dense_h_to_4h_weight2: R.Tensor((2048, 512), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[70] + gpt_neox_layers_5_mlp_dense_4h_to_h_weight2: R.Tensor((512, 2048), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias2: R.Tensor((512,), dtype="float32") = packed_params[72] + gpt_neox_final_layer_norm_weight2: R.Tensor((512,), dtype="float16") = packed_params[73] + gpt_neox_final_layer_norm_bias2: R.Tensor((512,), dtype="float16") = packed_params[74] + embed_out_weight2: R.Tensor((vocab_size, 512), dtype="float16") = packed_params[75] + layer_norm13 = 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, 512), dtype="float16")) + lv75 = R.call_tir(cls.fused_NT_matmul10_add10, (layer_norm13, gpt_neox_layers_0_attention_query_key_value_weight2, gpt_neox_layers_0_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 1536), dtype="float16")) + lv76 = R.call_tir(cls.fused_reshape8_reshape9, (lv75,), out_sinfo=R.Tensor((1, 24, 64), dtype="float16")) + lv7 = 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)), lv76), out_sinfo=R.Tensor((1, 8, 64), dtype="float16")) + lv77 = R.call_tir(cls.fused_reshape10_reshape11, (lv7,), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + layer_norm14 = 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, 512), dtype="float16")) + lv78 = R.call_tir(cls.fused_NT_matmul12_add12_gelu2_cast6, (layer_norm14, gpt_neox_layers_0_mlp_dense_h_to_4h_weight2, gpt_neox_layers_0_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv79 = R.call_tir(cls.fused_NT_matmul13_add13_cast7, (lv78, gpt_neox_layers_0_mlp_dense_4h_to_h_weight2, gpt_neox_layers_0_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv80 = R.call_tir(cls.fused_NT_matmul11_add11_add14_add14, (lv77, gpt_neox_layers_0_attention_dense_weight2, gpt_neox_layers_0_attention_dense_bias2, lv79, input_embed), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + layer_norm15 = R.call_tir(cls.layer_norm2, (lv80, gpt_neox_layers_1_input_layernorm_weight2, gpt_neox_layers_1_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv81 = R.call_tir(cls.fused_NT_matmul10_add10, (layer_norm15, gpt_neox_layers_1_attention_query_key_value_weight2, gpt_neox_layers_1_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 1536), dtype="float16")) + lv82 = R.call_tir(cls.fused_reshape8_reshape9, (lv81,), out_sinfo=R.Tensor((1, 24, 64), dtype="float16")) + lv8 = 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)), lv82), out_sinfo=R.Tensor((1, 8, 64), dtype="float16")) + lv83 = R.call_tir(cls.fused_reshape10_reshape11, (lv8,), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + layer_norm16 = R.call_tir(cls.layer_norm2, (lv80, gpt_neox_layers_1_post_attention_layernorm_weight2, gpt_neox_layers_1_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv84 = R.call_tir(cls.fused_NT_matmul12_add12_gelu2_cast6, (layer_norm16, gpt_neox_layers_1_mlp_dense_h_to_4h_weight2, gpt_neox_layers_1_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv85 = R.call_tir(cls.fused_NT_matmul13_add13_cast7, (lv84, gpt_neox_layers_1_mlp_dense_4h_to_h_weight2, gpt_neox_layers_1_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv86 = R.call_tir(cls.fused_NT_matmul11_add11_add14_add14, (lv83, gpt_neox_layers_1_attention_dense_weight2, gpt_neox_layers_1_attention_dense_bias2, lv85, lv80), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + layer_norm17 = R.call_tir(cls.layer_norm2, (lv86, gpt_neox_layers_2_input_layernorm_weight2, gpt_neox_layers_2_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv87 = R.call_tir(cls.fused_NT_matmul10_add10, (layer_norm17, gpt_neox_layers_2_attention_query_key_value_weight2, gpt_neox_layers_2_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 1536), dtype="float16")) + lv88 = R.call_tir(cls.fused_reshape8_reshape9, (lv87,), out_sinfo=R.Tensor((1, 24, 64), dtype="float16")) + lv9 = 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)), lv88), out_sinfo=R.Tensor((1, 8, 64), dtype="float16")) + lv89 = R.call_tir(cls.fused_reshape10_reshape11, (lv9,), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + layer_norm18 = R.call_tir(cls.layer_norm2, (lv86, gpt_neox_layers_2_post_attention_layernorm_weight2, gpt_neox_layers_2_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv90 = R.call_tir(cls.fused_NT_matmul12_add12_gelu2_cast6, (layer_norm18, gpt_neox_layers_2_mlp_dense_h_to_4h_weight2, gpt_neox_layers_2_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv91 = R.call_tir(cls.fused_NT_matmul13_add13_cast7, (lv90, gpt_neox_layers_2_mlp_dense_4h_to_h_weight2, gpt_neox_layers_2_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv92 = R.call_tir(cls.fused_NT_matmul11_add11_add14_add14, (lv89, gpt_neox_layers_2_attention_dense_weight2, gpt_neox_layers_2_attention_dense_bias2, lv91, lv86), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + layer_norm19 = R.call_tir(cls.layer_norm2, (lv92, gpt_neox_layers_3_input_layernorm_weight2, gpt_neox_layers_3_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv93 = R.call_tir(cls.fused_NT_matmul10_add10, (layer_norm19, gpt_neox_layers_3_attention_query_key_value_weight2, gpt_neox_layers_3_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 1536), dtype="float16")) + lv94 = R.call_tir(cls.fused_reshape8_reshape9, (lv93,), out_sinfo=R.Tensor((1, 24, 64), dtype="float16")) + lv10 = 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)), lv94), out_sinfo=R.Tensor((1, 8, 64), dtype="float16")) + lv95 = R.call_tir(cls.fused_reshape10_reshape11, (lv10,), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + layer_norm20 = R.call_tir(cls.layer_norm2, (lv92, gpt_neox_layers_3_post_attention_layernorm_weight2, gpt_neox_layers_3_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv96 = R.call_tir(cls.fused_NT_matmul12_add12_gelu2_cast6, (layer_norm20, gpt_neox_layers_3_mlp_dense_h_to_4h_weight2, gpt_neox_layers_3_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv97 = R.call_tir(cls.fused_NT_matmul13_add13_cast7, (lv96, gpt_neox_layers_3_mlp_dense_4h_to_h_weight2, gpt_neox_layers_3_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv98 = R.call_tir(cls.fused_NT_matmul11_add11_add14_add14, (lv95, gpt_neox_layers_3_attention_dense_weight2, gpt_neox_layers_3_attention_dense_bias2, lv97, lv92), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + layer_norm21 = R.call_tir(cls.layer_norm2, (lv98, gpt_neox_layers_4_input_layernorm_weight2, gpt_neox_layers_4_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv99 = R.call_tir(cls.fused_NT_matmul10_add10, (layer_norm21, gpt_neox_layers_4_attention_query_key_value_weight2, gpt_neox_layers_4_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 1536), dtype="float16")) + lv100 = R.call_tir(cls.fused_reshape8_reshape9, (lv99,), out_sinfo=R.Tensor((1, 24, 64), dtype="float16")) + lv11 = 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)), lv100), out_sinfo=R.Tensor((1, 8, 64), dtype="float16")) + lv101 = R.call_tir(cls.fused_reshape10_reshape11, (lv11,), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + layer_norm22 = R.call_tir(cls.layer_norm2, (lv98, gpt_neox_layers_4_post_attention_layernorm_weight2, gpt_neox_layers_4_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv102 = R.call_tir(cls.fused_NT_matmul12_add12_gelu2_cast6, (layer_norm22, gpt_neox_layers_4_mlp_dense_h_to_4h_weight2, gpt_neox_layers_4_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv103 = R.call_tir(cls.fused_NT_matmul13_add13_cast7, (lv102, gpt_neox_layers_4_mlp_dense_4h_to_h_weight2, gpt_neox_layers_4_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv104 = R.call_tir(cls.fused_NT_matmul11_add11_add14_add14, (lv101, gpt_neox_layers_4_attention_dense_weight2, gpt_neox_layers_4_attention_dense_bias2, lv103, lv98), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + layer_norm23 = R.call_tir(cls.layer_norm2, (lv104, gpt_neox_layers_5_input_layernorm_weight2, gpt_neox_layers_5_input_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv105 = R.call_tir(cls.fused_NT_matmul10_add10, (layer_norm23, gpt_neox_layers_5_attention_query_key_value_weight2, gpt_neox_layers_5_attention_query_key_value_bias2), out_sinfo=R.Tensor((1, 1, 1536), dtype="float16")) + lv106 = R.call_tir(cls.fused_reshape8_reshape9, (lv105,), out_sinfo=R.Tensor((1, 24, 64), dtype="float16")) + lv12 = 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)), lv106), out_sinfo=R.Tensor((1, 8, 64), dtype="float16")) + lv107 = R.call_tir(cls.fused_reshape10_reshape11, (lv12,), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + layer_norm24 = R.call_tir(cls.layer_norm2, (lv104, gpt_neox_layers_5_post_attention_layernorm_weight2, gpt_neox_layers_5_post_attention_layernorm_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv108 = R.call_tir(cls.fused_NT_matmul12_add12_gelu2_cast6, (layer_norm24, gpt_neox_layers_5_mlp_dense_h_to_4h_weight2, gpt_neox_layers_5_mlp_dense_h_to_4h_bias2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv109 = R.call_tir(cls.fused_NT_matmul13_add13_cast7, (lv108, gpt_neox_layers_5_mlp_dense_4h_to_h_weight2, gpt_neox_layers_5_mlp_dense_4h_to_h_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv110 = R.call_tir(cls.fused_NT_matmul11_add11_add14_add14, (lv107, gpt_neox_layers_5_attention_dense_weight2, gpt_neox_layers_5_attention_dense_bias2, lv109, lv104), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + layer_norm25 = R.call_tir(cls.layer_norm2, (lv110, gpt_neox_final_layer_norm_weight2, gpt_neox_final_layer_norm_bias2), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv111 = R.call_tir(cls.fused_NT_matmul14_cast8, (layer_norm25, embed_out_weight2), out_sinfo=R.Tensor((1, 1, vocab_size), dtype="float32")) + gv2: R.Tuple(R.Tensor((1, 1, vocab_size), dtype="float32"), R.Object) = lv111, 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", 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor(("vocab_size", 512), dtype="float16"))) -> R.Tensor(("seq_len", 512), 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_weight: R.Tensor((vocab_size, 512), dtype="float16") = packed_params[0] + gv = R.call_tir(cls.take1, (gpt_neox_embed_in_weight, input_ids), out_sinfo=R.Tensor((seq_len, 512), dtype="float16")) + R.output(gv) + return gv + + @R.function + def prefill(input_embed: R.Tensor((1, "seq_len", 512), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((1536, 512), dtype="float16"), R.Tensor((1536,), dtype="float16"), R.Tensor((512, 512), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor((2048, 512), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((512, 2048), dtype="float16"), R.Tensor((512,), dtype="float32"), R.Tensor((512,), dtype="float16"), R.Tensor((512,), dtype="float16"), R.Tensor(("vocab_size", 512), 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((512,), dtype="float16") = packed_params[1] + gpt_neox_layers_0_input_layernorm_bias1: R.Tensor((512,), dtype="float16") = packed_params[2] + gpt_neox_layers_0_post_attention_layernorm_weight1: R.Tensor((512,), dtype="float16") = packed_params[3] + gpt_neox_layers_0_post_attention_layernorm_bias1: R.Tensor((512,), dtype="float16") = packed_params[4] + gpt_neox_layers_0_attention_query_key_value_weight1: R.Tensor((1536, 512), dtype="float16") = packed_params[5] + gpt_neox_layers_0_attention_query_key_value_bias1: R.Tensor((1536,), dtype="float16") = packed_params[6] + gpt_neox_layers_0_attention_dense_weight1: R.Tensor((512, 512), dtype="float16") = packed_params[7] + gpt_neox_layers_0_attention_dense_bias1: R.Tensor((512,), dtype="float16") = packed_params[8] + gpt_neox_layers_0_mlp_dense_h_to_4h_weight1: R.Tensor((2048, 512), dtype="float16") = packed_params[9] + gpt_neox_layers_0_mlp_dense_h_to_4h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[10] + gpt_neox_layers_0_mlp_dense_4h_to_h_weight1: R.Tensor((512, 2048), dtype="float16") = packed_params[11] + gpt_neox_layers_0_mlp_dense_4h_to_h_bias1: R.Tensor((512,), dtype="float32") = packed_params[12] + gpt_neox_layers_1_input_layernorm_weight1: R.Tensor((512,), dtype="float16") = packed_params[13] + gpt_neox_layers_1_input_layernorm_bias1: R.Tensor((512,), dtype="float16") = packed_params[14] + gpt_neox_layers_1_post_attention_layernorm_weight1: R.Tensor((512,), dtype="float16") = packed_params[15] + gpt_neox_layers_1_post_attention_layernorm_bias1: R.Tensor((512,), dtype="float16") = packed_params[16] + gpt_neox_layers_1_attention_query_key_value_weight1: R.Tensor((1536, 512), dtype="float16") = packed_params[17] + gpt_neox_layers_1_attention_query_key_value_bias1: R.Tensor((1536,), dtype="float16") = packed_params[18] + gpt_neox_layers_1_attention_dense_weight1: R.Tensor((512, 512), dtype="float16") = packed_params[19] + gpt_neox_layers_1_attention_dense_bias1: R.Tensor((512,), dtype="float16") = packed_params[20] + gpt_neox_layers_1_mlp_dense_h_to_4h_weight1: R.Tensor((2048, 512), dtype="float16") = packed_params[21] + gpt_neox_layers_1_mlp_dense_h_to_4h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[22] + gpt_neox_layers_1_mlp_dense_4h_to_h_weight1: R.Tensor((512, 2048), dtype="float16") = packed_params[23] + gpt_neox_layers_1_mlp_dense_4h_to_h_bias1: R.Tensor((512,), dtype="float32") = packed_params[24] + gpt_neox_layers_2_input_layernorm_weight1: R.Tensor((512,), dtype="float16") = packed_params[25] + gpt_neox_layers_2_input_layernorm_bias1: R.Tensor((512,), dtype="float16") = packed_params[26] + gpt_neox_layers_2_post_attention_layernorm_weight1: R.Tensor((512,), dtype="float16") = packed_params[27] + gpt_neox_layers_2_post_attention_layernorm_bias1: R.Tensor((512,), dtype="float16") = packed_params[28] + gpt_neox_layers_2_attention_query_key_value_weight1: R.Tensor((1536, 512), dtype="float16") = packed_params[29] + gpt_neox_layers_2_attention_query_key_value_bias1: R.Tensor((1536,), dtype="float16") = packed_params[30] + gpt_neox_layers_2_attention_dense_weight1: R.Tensor((512, 512), dtype="float16") = packed_params[31] + gpt_neox_layers_2_attention_dense_bias1: R.Tensor((512,), dtype="float16") = packed_params[32] + gpt_neox_layers_2_mlp_dense_h_to_4h_weight1: R.Tensor((2048, 512), dtype="float16") = packed_params[33] + gpt_neox_layers_2_mlp_dense_h_to_4h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[34] + gpt_neox_layers_2_mlp_dense_4h_to_h_weight1: R.Tensor((512, 2048), dtype="float16") = packed_params[35] + gpt_neox_layers_2_mlp_dense_4h_to_h_bias1: R.Tensor((512,), dtype="float32") = packed_params[36] + gpt_neox_layers_3_input_layernorm_weight1: R.Tensor((512,), dtype="float16") = packed_params[37] + gpt_neox_layers_3_input_layernorm_bias1: R.Tensor((512,), dtype="float16") = packed_params[38] + gpt_neox_layers_3_post_attention_layernorm_weight1: R.Tensor((512,), dtype="float16") = packed_params[39] + gpt_neox_layers_3_post_attention_layernorm_bias1: R.Tensor((512,), dtype="float16") = packed_params[40] + gpt_neox_layers_3_attention_query_key_value_weight1: R.Tensor((1536, 512), dtype="float16") = packed_params[41] + gpt_neox_layers_3_attention_query_key_value_bias1: R.Tensor((1536,), dtype="float16") = packed_params[42] + gpt_neox_layers_3_attention_dense_weight1: R.Tensor((512, 512), dtype="float16") = packed_params[43] + gpt_neox_layers_3_attention_dense_bias1: R.Tensor((512,), dtype="float16") = packed_params[44] + gpt_neox_layers_3_mlp_dense_h_to_4h_weight1: R.Tensor((2048, 512), dtype="float16") = packed_params[45] + gpt_neox_layers_3_mlp_dense_h_to_4h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[46] + gpt_neox_layers_3_mlp_dense_4h_to_h_weight1: R.Tensor((512, 2048), dtype="float16") = packed_params[47] + gpt_neox_layers_3_mlp_dense_4h_to_h_bias1: R.Tensor((512,), dtype="float32") = packed_params[48] + gpt_neox_layers_4_input_layernorm_weight1: R.Tensor((512,), dtype="float16") = packed_params[49] + gpt_neox_layers_4_input_layernorm_bias1: R.Tensor((512,), dtype="float16") = packed_params[50] + gpt_neox_layers_4_post_attention_layernorm_weight1: R.Tensor((512,), dtype="float16") = packed_params[51] + gpt_neox_layers_4_post_attention_layernorm_bias1: R.Tensor((512,), dtype="float16") = packed_params[52] + gpt_neox_layers_4_attention_query_key_value_weight1: R.Tensor((1536, 512), dtype="float16") = packed_params[53] + gpt_neox_layers_4_attention_query_key_value_bias1: R.Tensor((1536,), dtype="float16") = packed_params[54] + gpt_neox_layers_4_attention_dense_weight1: R.Tensor((512, 512), dtype="float16") = packed_params[55] + gpt_neox_layers_4_attention_dense_bias1: R.Tensor((512,), dtype="float16") = packed_params[56] + gpt_neox_layers_4_mlp_dense_h_to_4h_weight1: R.Tensor((2048, 512), dtype="float16") = packed_params[57] + gpt_neox_layers_4_mlp_dense_h_to_4h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[58] + gpt_neox_layers_4_mlp_dense_4h_to_h_weight1: R.Tensor((512, 2048), dtype="float16") = packed_params[59] + gpt_neox_layers_4_mlp_dense_4h_to_h_bias1: R.Tensor((512,), dtype="float32") = packed_params[60] + gpt_neox_layers_5_input_layernorm_weight1: R.Tensor((512,), dtype="float16") = packed_params[61] + gpt_neox_layers_5_input_layernorm_bias1: R.Tensor((512,), dtype="float16") = packed_params[62] + gpt_neox_layers_5_post_attention_layernorm_weight1: R.Tensor((512,), dtype="float16") = packed_params[63] + gpt_neox_layers_5_post_attention_layernorm_bias1: R.Tensor((512,), dtype="float16") = packed_params[64] + gpt_neox_layers_5_attention_query_key_value_weight1: R.Tensor((1536, 512), dtype="float16") = packed_params[65] + gpt_neox_layers_5_attention_query_key_value_bias1: R.Tensor((1536,), dtype="float16") = packed_params[66] + gpt_neox_layers_5_attention_dense_weight1: R.Tensor((512, 512), dtype="float16") = packed_params[67] + gpt_neox_layers_5_attention_dense_bias1: R.Tensor((512,), dtype="float16") = packed_params[68] + gpt_neox_layers_5_mlp_dense_h_to_4h_weight1: R.Tensor((2048, 512), dtype="float16") = packed_params[69] + gpt_neox_layers_5_mlp_dense_h_to_4h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[70] + gpt_neox_layers_5_mlp_dense_4h_to_h_weight1: R.Tensor((512, 2048), dtype="float16") = packed_params[71] + gpt_neox_layers_5_mlp_dense_4h_to_h_bias1: R.Tensor((512,), dtype="float32") = packed_params[72] + gpt_neox_final_layer_norm_weight1: R.Tensor((512,), dtype="float16") = packed_params[73] + gpt_neox_final_layer_norm_bias1: R.Tensor((512,), dtype="float16") = packed_params[74] + embed_out_weight1: R.Tensor((vocab_size, 512), dtype="float16") = packed_params[75] + 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, 512), dtype="float16")) + lv112 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm, gpt_neox_layers_0_attention_query_key_value_weight1, gpt_neox_layers_0_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape = R.call_tir(cls.reshape4, (lv112,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape1 = R.call_tir(cls.reshape5, (reshape,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv = 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, 64), dtype="float16")) + reshape2 = R.call_tir(cls.reshape6, (lv,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape3 = R.call_tir(cls.reshape7, (reshape2,), out_sinfo=R.Tensor((1, seq_len, 512), 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, 512), dtype="float16")) + lv113 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm1, gpt_neox_layers_0_mlp_dense_h_to_4h_weight1, gpt_neox_layers_0_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv114 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv113, gpt_neox_layers_0_mlp_dense_4h_to_h_weight1, gpt_neox_layers_0_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv115 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape3, gpt_neox_layers_0_attention_dense_weight1, gpt_neox_layers_0_attention_dense_bias1, lv114, input_embed), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm2 = R.call_tir(cls.layer_norm1, (lv115, gpt_neox_layers_1_input_layernorm_weight1, gpt_neox_layers_1_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv116 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm2, gpt_neox_layers_1_attention_query_key_value_weight1, gpt_neox_layers_1_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape4 = R.call_tir(cls.reshape4, (lv116,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape5 = R.call_tir(cls.reshape5, (reshape4,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv1 = 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, 64), dtype="float16")) + reshape6 = R.call_tir(cls.reshape6, (lv1,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape7 = R.call_tir(cls.reshape7, (reshape6,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm3 = R.call_tir(cls.layer_norm1, (lv115, gpt_neox_layers_1_post_attention_layernorm_weight1, gpt_neox_layers_1_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv117 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm3, gpt_neox_layers_1_mlp_dense_h_to_4h_weight1, gpt_neox_layers_1_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv118 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv117, gpt_neox_layers_1_mlp_dense_4h_to_h_weight1, gpt_neox_layers_1_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv119 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape7, gpt_neox_layers_1_attention_dense_weight1, gpt_neox_layers_1_attention_dense_bias1, lv118, lv115), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm4 = R.call_tir(cls.layer_norm1, (lv119, gpt_neox_layers_2_input_layernorm_weight1, gpt_neox_layers_2_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv120 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm4, gpt_neox_layers_2_attention_query_key_value_weight1, gpt_neox_layers_2_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape8 = R.call_tir(cls.reshape4, (lv120,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape9 = R.call_tir(cls.reshape5, (reshape8,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv2 = 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, 64), dtype="float16")) + reshape10 = R.call_tir(cls.reshape6, (lv2,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape11 = R.call_tir(cls.reshape7, (reshape10,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm5 = R.call_tir(cls.layer_norm1, (lv119, gpt_neox_layers_2_post_attention_layernorm_weight1, gpt_neox_layers_2_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv121 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm5, gpt_neox_layers_2_mlp_dense_h_to_4h_weight1, gpt_neox_layers_2_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv122 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv121, gpt_neox_layers_2_mlp_dense_4h_to_h_weight1, gpt_neox_layers_2_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv123 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape11, gpt_neox_layers_2_attention_dense_weight1, gpt_neox_layers_2_attention_dense_bias1, lv122, lv119), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm6 = R.call_tir(cls.layer_norm1, (lv123, gpt_neox_layers_3_input_layernorm_weight1, gpt_neox_layers_3_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv124 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm6, gpt_neox_layers_3_attention_query_key_value_weight1, gpt_neox_layers_3_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape12 = R.call_tir(cls.reshape4, (lv124,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape13 = R.call_tir(cls.reshape5, (reshape12,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv3 = 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, 64), dtype="float16")) + reshape14 = R.call_tir(cls.reshape6, (lv3,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape15 = R.call_tir(cls.reshape7, (reshape14,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm7 = R.call_tir(cls.layer_norm1, (lv123, gpt_neox_layers_3_post_attention_layernorm_weight1, gpt_neox_layers_3_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv125 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm7, gpt_neox_layers_3_mlp_dense_h_to_4h_weight1, gpt_neox_layers_3_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv126 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv125, gpt_neox_layers_3_mlp_dense_4h_to_h_weight1, gpt_neox_layers_3_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv127 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape15, gpt_neox_layers_3_attention_dense_weight1, gpt_neox_layers_3_attention_dense_bias1, lv126, lv123), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm8 = R.call_tir(cls.layer_norm1, (lv127, gpt_neox_layers_4_input_layernorm_weight1, gpt_neox_layers_4_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv128 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm8, gpt_neox_layers_4_attention_query_key_value_weight1, gpt_neox_layers_4_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape16 = R.call_tir(cls.reshape4, (lv128,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape17 = R.call_tir(cls.reshape5, (reshape16,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv4 = 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, 64), dtype="float16")) + reshape18 = R.call_tir(cls.reshape6, (lv4,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape19 = R.call_tir(cls.reshape7, (reshape18,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm9 = R.call_tir(cls.layer_norm1, (lv127, gpt_neox_layers_4_post_attention_layernorm_weight1, gpt_neox_layers_4_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv129 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm9, gpt_neox_layers_4_mlp_dense_h_to_4h_weight1, gpt_neox_layers_4_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv130 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv129, gpt_neox_layers_4_mlp_dense_4h_to_h_weight1, gpt_neox_layers_4_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv131 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape19, gpt_neox_layers_4_attention_dense_weight1, gpt_neox_layers_4_attention_dense_bias1, lv130, lv127), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm10 = R.call_tir(cls.layer_norm1, (lv131, gpt_neox_layers_5_input_layernorm_weight1, gpt_neox_layers_5_input_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv132 = R.call_tir(cls.fused_NT_matmul5_add5, (layer_norm10, gpt_neox_layers_5_attention_query_key_value_weight1, gpt_neox_layers_5_attention_query_key_value_bias1), out_sinfo=R.Tensor((1, seq_len, 1536), dtype="float16")) + reshape20 = R.call_tir(cls.reshape4, (lv132,), out_sinfo=R.Tensor((1, seq_len, 24, 64), dtype="float16")) + reshape21 = R.call_tir(cls.reshape5, (reshape20,), out_sinfo=R.Tensor((seq_len, 24, 64), dtype="float16")) + lv5 = 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, 64), dtype="float16")) + reshape22 = R.call_tir(cls.reshape6, (lv5,), out_sinfo=R.Tensor((1, seq_len, 8, 64), dtype="float16")) + reshape23 = R.call_tir(cls.reshape7, (reshape22,), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm11 = R.call_tir(cls.layer_norm1, (lv131, gpt_neox_layers_5_post_attention_layernorm_weight1, gpt_neox_layers_5_post_attention_layernorm_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv133 = R.call_tir(cls.fused_NT_matmul7_add7_gelu1_cast3, (layer_norm11, gpt_neox_layers_5_mlp_dense_h_to_4h_weight1, gpt_neox_layers_5_mlp_dense_h_to_4h_bias1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + lv134 = R.call_tir(cls.fused_NT_matmul8_add8_cast4, (lv133, gpt_neox_layers_5_mlp_dense_4h_to_h_weight1, gpt_neox_layers_5_mlp_dense_4h_to_h_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv135 = R.call_tir(cls.fused_NT_matmul6_add6_add9_add9, (reshape23, gpt_neox_layers_5_attention_dense_weight1, gpt_neox_layers_5_attention_dense_bias1, lv134, lv131), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + layer_norm12 = R.call_tir(cls.layer_norm1, (lv135, gpt_neox_final_layer_norm_weight1, gpt_neox_final_layer_norm_bias1), out_sinfo=R.Tensor((1, seq_len, 512), dtype="float16")) + lv6 = R.call_tir(cls.index, (layer_norm12,), out_sinfo=R.Tensor((1, 1, 512), dtype="float16")) + lv136 = R.call_tir(cls.fused_NT_matmul14_cast8, (lv6, embed_out_weight1), out_sinfo=R.Tensor((1, 1, vocab_size), dtype="float32")) + gv1: R.Tuple(R.Tensor((1, 1, vocab_size), dtype="float32"), R.Object) = lv136, 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