# 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(private=True) def NT_matmul(var_layer_norm99: T.handle, gpt_neox_layers_0_attention_query_key_value_weight4: T.Buffer((T.int64(6144), T.int64(2048)), "float16"), var_NT_matmul: T.handle): T.func_attr({"tir.noalias": T.bool(True)}) batch_size = T.int64() layer_norm99 = T.match_buffer(var_layer_norm99, (batch_size, T.int64(1), T.int64(2048)), "float16") NT_matmul = T.match_buffer(var_NT_matmul, (batch_size, T.int64(1), T.int64(6144)), "float16") # with T.block("root"): for i0, i1, i2, k in T.grid(batch_size, T.int64(1), T.int64(6144), 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(layer_norm99[v_i0, v_i1, v_k], gpt_neox_layers_0_attention_query_key_value_weight4[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float16(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + layer_norm99[v_i0, v_i1, v_k] * gpt_neox_layers_0_attention_query_key_value_weight4[v_i2, v_k] @T.prim_func(private=True) def NT_matmul1(var_reshape195: T.handle, gpt_neox_layers_0_attention_dense_weight4: T.Buffer((T.int64(2048), T.int64(2048)), "float16"), var_NT_matmul: T.handle): T.func_attr({"tir.noalias": T.bool(True)}) batch_size = T.int64() reshape195 = T.match_buffer(var_reshape195, (batch_size, T.int64(1), T.int64(2048)), "float16") NT_matmul = T.match_buffer(var_NT_matmul, (batch_size, T.int64(1), T.int64(2048)), "float16") # with T.block("root"): for i0, i1, i2, k in T.grid(batch_size, T.int64(1), T.int64(2048), 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(reshape195[v_i0, v_i1, v_k], gpt_neox_layers_0_attention_dense_weight4[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float16(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + reshape195[v_i0, v_i1, v_k] * gpt_neox_layers_0_attention_dense_weight4[v_i2, v_k] @T.prim_func(private=True) def NT_matmul10(layer_norm33: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16"), gpt_neox_layers_0_attention_query_key_value_weight2: T.Buffer((T.int64(6144), T.int64(2048)), "float16"), NT_matmul: T.Buffer((T.int64(1), T.int64(1), T.int64(6144)), "float16")): T.func_attr({"tir.noalias": T.bool(True)}) # with T.block("root"): for i0, i1, i2, k in T.grid(T.int64(1), T.int64(1), T.int64(6144), 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(layer_norm33[v_i0, v_i1, v_k], gpt_neox_layers_0_attention_query_key_value_weight2[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float16(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + layer_norm33[v_i0, v_i1, v_k] * gpt_neox_layers_0_attention_query_key_value_weight2[v_i2, v_k] @T.prim_func(private=True) def NT_matmul11(reshape67: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16"), gpt_neox_layers_0_attention_dense_weight2: T.Buffer((T.int64(2048), T.int64(2048)), "float16"), NT_matmul: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16")): T.func_attr({"tir.noalias": T.bool(True)}) # with T.block("root"): for i0, i1, i2, k in T.grid(T.int64(1), T.int64(1), T.int64(2048), 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(reshape67[v_i0, v_i1, v_k], gpt_neox_layers_0_attention_dense_weight2[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float16(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + reshape67[v_i0, v_i1, v_k] * gpt_neox_layers_0_attention_dense_weight2[v_i2, v_k] @T.prim_func(private=True) def NT_matmul12(layer_norm34: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16"), gpt_neox_layers_0_mlp_dense_h_to_4h_weight2: T.Buffer((T.int64(8192), T.int64(2048)), "float16"), NT_matmul: T.Buffer((T.int64(1), T.int64(1), T.int64(8192)), "float32")): T.func_attr({"tir.noalias": T.bool(True)}) # with T.block("root"): for i0, i1, i2, k in T.grid(T.int64(1), T.int64(1), T.int64(8192), 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(layer_norm34[v_i0, v_i1, v_k], gpt_neox_layers_0_mlp_dense_h_to_4h_weight2[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float32(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + T.Cast("float32", layer_norm34[v_i0, v_i1, v_k]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_h_to_4h_weight2[v_i2, v_k]) @T.prim_func(private=True) def NT_matmul13(astype33: T.Buffer((T.int64(1), T.int64(1), T.int64(8192)), "float16"), gpt_neox_layers_0_mlp_dense_4h_to_h_weight2: T.Buffer((T.int64(2048), T.int64(8192)), "float16"), NT_matmul: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float32")): T.func_attr({"tir.noalias": T.bool(True)}) # with T.block("root"): for i0, i1, i2, k in T.grid(T.int64(1), T.int64(1), T.int64(2048), T.int64(8192)): with T.block("NT_matmul"): v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) T.reads(astype33[v_i0, v_i1, v_k], gpt_neox_layers_0_mlp_dense_4h_to_h_weight2[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float32(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + T.Cast("float32", astype33[v_i0, v_i1, v_k]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_4h_to_h_weight2[v_i2, v_k]) @T.prim_func(private=True) def NT_matmul14(layer_norm65: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16"), var_embed_out_weight2: T.handle, var_NT_matmul: T.handle): T.func_attr({"tir.noalias": T.bool(True)}) vocab_size = T.int64() embed_out_weight2 = T.match_buffer(var_embed_out_weight2, (vocab_size, T.int64(2048)), "float16") NT_matmul = T.match_buffer(var_NT_matmul, (T.int64(1), T.int64(1), vocab_size), "float16") # with T.block("root"): for i0, i1, i2, k in T.grid(T.int64(1), T.int64(1), vocab_size, 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(layer_norm65[v_i0, v_i1, v_k], embed_out_weight2[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float16(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + layer_norm65[v_i0, v_i1, v_k] * embed_out_weight2[v_i2, v_k] @T.prim_func(private=True) def NT_matmul2(var_layer_norm100: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_weight4: T.Buffer((T.int64(8192), T.int64(2048)), "float16"), var_NT_matmul: T.handle): T.func_attr({"tir.noalias": T.bool(True)}) batch_size = T.int64() layer_norm100 = T.match_buffer(var_layer_norm100, (batch_size, T.int64(1), T.int64(2048)), "float16") NT_matmul = T.match_buffer(var_NT_matmul, (batch_size, T.int64(1), T.int64(8192))) # with T.block("root"): for i0, i1, i2, k in T.grid(batch_size, T.int64(1), T.int64(8192), 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(layer_norm100[v_i0, v_i1, v_k], gpt_neox_layers_0_mlp_dense_h_to_4h_weight4[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float32(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + T.Cast("float32", layer_norm100[v_i0, v_i1, v_k]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_h_to_4h_weight4[v_i2, v_k]) @T.prim_func(private=True) def NT_matmul3(var_astype99: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_weight4: T.Buffer((T.int64(2048), T.int64(8192)), "float16"), var_NT_matmul: T.handle): T.func_attr({"tir.noalias": T.bool(True)}) batch_size = T.int64() astype99 = T.match_buffer(var_astype99, (batch_size, T.int64(1), T.int64(8192)), "float16") NT_matmul = T.match_buffer(var_NT_matmul, (batch_size, T.int64(1), T.int64(2048))) # with T.block("root"): for i0, i1, i2, k in T.grid(batch_size, T.int64(1), T.int64(2048), T.int64(8192)): with T.block("NT_matmul"): v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) T.reads(astype99[v_i0, v_i1, v_k], gpt_neox_layers_0_mlp_dense_4h_to_h_weight4[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float32(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + T.Cast("float32", astype99[v_i0, v_i1, v_k]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_4h_to_h_weight4[v_i2, v_k]) @T.prim_func(private=True) def NT_matmul4(var_layer_norm131: T.handle, var_embed_out_weight4: T.handle, var_NT_matmul: T.handle): T.func_attr({"tir.noalias": T.bool(True)}) batch_size = T.int64() layer_norm131 = T.match_buffer(var_layer_norm131, (batch_size, T.int64(1), T.int64(2048)), "float16") vocab_size = T.int64() embed_out_weight4 = T.match_buffer(var_embed_out_weight4, (vocab_size, T.int64(2048)), "float16") NT_matmul = T.match_buffer(var_NT_matmul, (batch_size, T.int64(1), vocab_size), "float16") # with T.block("root"): for i0, i1, i2, k in T.grid(batch_size, T.int64(1), vocab_size, 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(layer_norm131[v_i0, v_i1, v_k], embed_out_weight4[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float16(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + layer_norm131[v_i0, v_i1, v_k] * embed_out_weight4[v_i2, v_k] @T.prim_func(private=True) def NT_matmul5(var_layer_norm66: T.handle, gpt_neox_layers_0_attention_query_key_value_weight3: T.Buffer((T.int64(6144), T.int64(2048)), "float16"), var_NT_matmul: T.handle): T.func_attr({"tir.noalias": T.bool(True)}) seq_len = T.int64() layer_norm66 = T.match_buffer(var_layer_norm66, (T.int64(1), seq_len, T.int64(2048)), "float16") NT_matmul = T.match_buffer(var_NT_matmul, (T.int64(1), seq_len, T.int64(6144)), "float16") # with T.block("root"): for i0, i1, i2, k in T.grid(T.int64(1), seq_len, T.int64(6144), 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(layer_norm66[v_i0, v_i1, v_k], gpt_neox_layers_0_attention_query_key_value_weight3[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float16(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + layer_norm66[v_i0, v_i1, v_k] * gpt_neox_layers_0_attention_query_key_value_weight3[v_i2, v_k] @T.prim_func(private=True) def NT_matmul6(var_reshape131: T.handle, gpt_neox_layers_0_attention_dense_weight3: T.Buffer((T.int64(2048), T.int64(2048)), "float16"), var_NT_matmul: T.handle): T.func_attr({"tir.noalias": T.bool(True)}) seq_len = T.int64() reshape131 = T.match_buffer(var_reshape131, (T.int64(1), seq_len, T.int64(2048)), "float16") NT_matmul = T.match_buffer(var_NT_matmul, (T.int64(1), seq_len, T.int64(2048)), "float16") # with T.block("root"): for i0, i1, i2, k in T.grid(T.int64(1), seq_len, T.int64(2048), 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(reshape131[v_i0, v_i1, v_k], gpt_neox_layers_0_attention_dense_weight3[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float16(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + reshape131[v_i0, v_i1, v_k] * gpt_neox_layers_0_attention_dense_weight3[v_i2, v_k] @T.prim_func(private=True) def NT_matmul7(var_layer_norm67: T.handle, gpt_neox_layers_0_mlp_dense_h_to_4h_weight3: T.Buffer((T.int64(8192), T.int64(2048)), "float16"), var_NT_matmul: T.handle): T.func_attr({"tir.noalias": T.bool(True)}) seq_len = T.int64() layer_norm67 = T.match_buffer(var_layer_norm67, (T.int64(1), seq_len, T.int64(2048)), "float16") NT_matmul = T.match_buffer(var_NT_matmul, (T.int64(1), seq_len, T.int64(8192))) # with T.block("root"): for i0, i1, i2, k in T.grid(T.int64(1), seq_len, T.int64(8192), 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(layer_norm67[v_i0, v_i1, v_k], gpt_neox_layers_0_mlp_dense_h_to_4h_weight3[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float32(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + T.Cast("float32", layer_norm67[v_i0, v_i1, v_k]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_h_to_4h_weight3[v_i2, v_k]) @T.prim_func(private=True) def NT_matmul8(var_astype66: T.handle, gpt_neox_layers_0_mlp_dense_4h_to_h_weight3: T.Buffer((T.int64(2048), T.int64(8192)), "float16"), var_NT_matmul: T.handle): T.func_attr({"tir.noalias": T.bool(True)}) seq_len = T.int64() astype66 = T.match_buffer(var_astype66, (T.int64(1), seq_len, T.int64(8192)), "float16") NT_matmul = T.match_buffer(var_NT_matmul, (T.int64(1), seq_len, T.int64(2048))) # with T.block("root"): for i0, i1, i2, k in T.grid(T.int64(1), seq_len, T.int64(2048), T.int64(8192)): with T.block("NT_matmul"): v_i0, v_i1, v_i2, v_k = T.axis.remap("SSSR", [i0, i1, i2, k]) T.reads(astype66[v_i0, v_i1, v_k], gpt_neox_layers_0_mlp_dense_4h_to_h_weight3[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float32(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + T.Cast("float32", astype66[v_i0, v_i1, v_k]) * T.Cast("float32", gpt_neox_layers_0_mlp_dense_4h_to_h_weight3[v_i2, v_k]) @T.prim_func(private=True) def NT_matmul9(var_take1: T.handle, var_embed_out_weight3: T.handle, var_NT_matmul: T.handle): T.func_attr({"tir.noalias": T.bool(True)}) batch_size = T.int64() take1 = T.match_buffer(var_take1, (T.int64(1), batch_size, T.int64(2048)), "float16") vocab_size = T.int64() embed_out_weight3 = T.match_buffer(var_embed_out_weight3, (vocab_size, T.int64(2048)), "float16") NT_matmul = T.match_buffer(var_NT_matmul, (T.int64(1), batch_size, vocab_size), "float16") # with T.block("root"): for i0, i1, i2, k in T.grid(T.int64(1), batch_size, vocab_size, 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(take1[v_i0, v_i1, v_k], embed_out_weight3[v_i2, v_k]) T.writes(NT_matmul[v_i0, v_i1, v_i2]) with T.init(): NT_matmul[v_i0, v_i1, v_i2] = T.float16(0.0) NT_matmul[v_i0, v_i1, v_i2] = NT_matmul[v_i0, v_i1, v_i2] + take1[v_i0, v_i1, v_k] * embed_out_weight3[v_i2, v_k] @T.prim_func def apply_bitmask_inplace(var_logits: T.handle, var_seq_ids: T.handle, var_bitmask: T.handle): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": T.bool(True), "tir.noalias": T.bool(True)}) batch_size, vocab_size = T.int32(is_size_var=True), T.int32(is_size_var=True) logits = T.match_buffer(var_logits, (batch_size, vocab_size)) num_seq = T.int32(is_size_var=True) seq_ids = T.match_buffer(var_seq_ids, (num_seq,), "int32") bitmask = T.match_buffer(var_bitmask, (batch_size, (vocab_size + 31) // 32), "int32") # with T.block("root"): for fused_s_v_0 in T.thread_binding((num_seq * vocab_size + 255) // 256, thread="blockIdx.x"): for fused_s_v_1 in T.thread_binding(256, thread="threadIdx.x"): with T.block("block"): vs = T.axis.spatial(num_seq, (fused_s_v_0 * 256 + fused_s_v_1) // vocab_size) vv = T.axis.spatial(vocab_size, (fused_s_v_0 * 256 + fused_s_v_1) % vocab_size) T.where(fused_s_v_0 * 256 + fused_s_v_1 < num_seq * vocab_size) T.reads(bitmask[seq_ids[vs], vv // 32], seq_ids[vs], logits[seq_ids[vs], vv]) T.writes(logits[seq_ids[vs], vv]) logits[seq_ids[vs], vv] = T.if_then_else(T.bitwise_and(T.shift_right(bitmask[seq_ids[vs], vv // 32], vv % 32), 1) == 1, logits[seq_ids[vs], vv], T.float32(-340282346638528859811704183484516925440.0)) @T.prim_func def apply_logit_bias_inplace(var_logits: T.handle, var_pos2seq_id: T.handle, var_token_ids: T.handle, var_logit_bias: T.handle): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": T.bool(True), "tir.noalias": T.bool(True)}) batch_size, vocab_size = T.int32(is_size_var=True), T.int32(is_size_var=True) logits = T.match_buffer(var_logits, (batch_size, vocab_size)) num_token = T.int32(is_size_var=True) pos2seq_id = T.match_buffer(var_pos2seq_id, (num_token,), "int32") token_ids = T.match_buffer(var_token_ids, (num_token,), "int32") logit_bias = T.match_buffer(var_logit_bias, (num_token,)) # with T.block("root"): for p0 in T.thread_binding((num_token + 255) // 256, thread="blockIdx.x"): for p1 in T.thread_binding(256, thread="threadIdx.x"): with T.block("block"): vp = T.axis.spatial(num_token, p0 * 256 + p1) T.where(p0 * 256 + p1 < num_token) T.reads(logits[pos2seq_id[vp], token_ids[vp]], pos2seq_id[vp], token_ids[vp], logit_bias[vp]) T.writes(logits[pos2seq_id[vp], token_ids[vp]]) logits[pos2seq_id[vp], token_ids[vp]] = logits[pos2seq_id[vp], token_ids[vp]] + logit_bias[vp] @T.prim_func def apply_penalty_inplace(var_logits: T.handle, var_seq_ids: T.handle, var_pos2seq_id: T.handle, var_token_ids: T.handle, var_token_cnt: T.handle, var_penalties: T.handle): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": T.bool(True), "tir.noalias": T.bool(True)}) batch_size, vocab_size = T.int32(is_size_var=True), T.int32(is_size_var=True) logits = T.match_buffer(var_logits, (batch_size, vocab_size)) num_seq = T.int32(is_size_var=True) seq_ids = T.match_buffer(var_seq_ids, (num_seq,), "int32") num_token = T.int32(is_size_var=True) pos2seq_id = T.match_buffer(var_pos2seq_id, (num_token,), "int32") token_ids = T.match_buffer(var_token_ids, (num_token,), "int32") token_cnt = T.match_buffer(var_token_cnt, (num_token,), "int32") penalties = T.match_buffer(var_penalties, (num_seq, 3)) # with T.block("root"): for p0 in T.thread_binding((num_token + 255) // 256, thread="blockIdx.x"): for p1 in T.thread_binding(256, thread="threadIdx.x"): with T.block("block"): vp = T.axis.spatial(num_token, p0 * 256 + p1) T.where(p0 * 256 + p1 < num_token) T.reads(logits[seq_ids[pos2seq_id[vp]], token_ids[vp]], seq_ids[pos2seq_id[vp]], pos2seq_id[vp], token_ids[vp], penalties[pos2seq_id[vp], 0:3], token_cnt[vp]) T.writes(logits[seq_ids[pos2seq_id[vp]], token_ids[vp]]) logits[seq_ids[pos2seq_id[vp]], token_ids[vp]] = logits[seq_ids[pos2seq_id[vp]], token_ids[vp]] - (penalties[pos2seq_id[vp], 0] + T.Cast("float32", token_cnt[vp]) * penalties[pos2seq_id[vp], 1]) logits[seq_ids[pos2seq_id[vp]], token_ids[vp]] = T.if_then_else(logits[seq_ids[pos2seq_id[vp]], token_ids[vp]] < T.float32(0.0), logits[seq_ids[pos2seq_id[vp]], token_ids[vp]] * penalties[pos2seq_id[vp], 2], logits[seq_ids[pos2seq_id[vp]], token_ids[vp]] / penalties[pos2seq_id[vp], 2]) @T.prim_func def batch_decode_paged_kv(_0: T.int32, Q_handle: T.handle, pages_handle: T.handle, page_table_indptr_handle: T.handle, page_table_values_handle: T.handle, var_length_info: T.handle, k_rope_pos_offset_handle: T.handle, q_rope_position_handle: T.handle, output_handle: T.handle, lse_handle: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) B = T.int32(is_size_var=True) Q = T.match_buffer(Q_handle, (B, 8, 256), "float16") max_num_pages = T.int32(is_size_var=True) pages = T.match_buffer(pages_handle, (max_num_pages, 2, 8, 16, 256), "float16", offset_factor=1) page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) nnz_pages = T.int32(is_size_var=True) page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) length_info = T.match_buffer(var_length_info, (B,), "int32", offset_factor=1) k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) output = T.match_buffer(output_handle, (B, 8, 256), "float16") lse = T.match_buffer(lse_handle, (B, 8)) # with T.block("root"): sm_scale: T.float32 = T.float32(0.090168440055560212) for bx in T.thread_binding(B, thread="blockIdx.x"): for fused_by_bz in T.thread_binding(8, thread="blockIdx.y"): for ty in T.thread_binding(1, thread="threadIdx.y"): for tx in T.thread_binding(64, thread="threadIdx.x"): for tz in T.thread_binding(4, thread="threadIdx.z"): with T.block("attn"): T.reads(page_table_indptr[bx:bx + 2], length_info[bx], q_rope_position[bx], Q[bx, fused_by_bz // 8 + ty + fused_by_bz % 8, tx * 4 - 128:tx * 4 - 128 + 260]) T.writes(output[bx, fused_by_bz % 8 + fused_by_bz // 8 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 8 + fused_by_bz // 8 + ty]) Q_local = T.alloc_buffer((4,), "float16", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") K_smem = T.alloc_buffer((4, 256), "float16", scope="shared") V_smem = T.alloc_buffer((4, 256), "float16", scope="shared") O_allreduce = T.alloc_buffer((4, 1, 256), scope="shared") md_allreduce = T.alloc_buffer((4, 1, 2), scope="shared") S_reduce_local = T.alloc_buffer((1,), scope="local") t0 = T.alloc_buffer((1,), scope="local") S_local = T.alloc_buffer((1,), scope="local") QK_local = T.alloc_buffer((4,), scope="local") V_local = T.alloc_buffer((4,), "float16", scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_prev = T.alloc_buffer((1,), scope="local") other_m = T.alloc_buffer((1,), scope="local") other_d = T.alloc_buffer((1,), scope="local") exp_mprev = T.alloc_buffer((1,), scope="local") exp_otherm = T.alloc_buffer((1,), scope="local") other_o = T.alloc_buffer((4,), scope="local") st_m = T.alloc_buffer((1,), scope="local") st_d = T.alloc_buffer((1,), scope="local") O_local = T.alloc_buffer((4,), scope="local") by: T.int32 = fused_by_bz % 8 bz: T.int32 = fused_by_bz // 8 batch_idx: T.int32 = bx cur_page_indptr_begin: T.int32 = page_table_indptr[batch_idx] cur_page_indptr_end: T.int32 = page_table_indptr[batch_idx + 1] kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[batch_idx], 0) st_m[0] = T.float32(-50000.0) st_d[0] = T.float32(1.0) for vec in T.vectorized(4): O_local[vec] = T.float32(0.0) for vec in T.vectorized(4): freq = T.float32() Q_local[vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", Q[bx, by + bz + ty, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 128, Q[bx, by + bz + ty, tx * 4 + vec + 128] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 128]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 256) / T.float32(256.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) for iterator in range((kv_chunk_len[0] + 3) // 4): tile_start_s: T.int32 = tz + ty tile_start_g: T.int32 = iterator * 4 + tz + ty for j in range(1): with T.block("KV_load"): T.reads() T.writes() row_g: T.int32 = tile_start_g + j if row_g < kv_chunk_len[0]: seq_offset: T.int32 = row_g page_no: T.int32 = page_table_values[cur_page_indptr_begin + seq_offset // 16] page_offset: T.int32 = seq_offset % 16 for vec in T.vectorized(4): freq = T.float32() K_smem[tile_start_s + j, tx * 4 + vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 128, pages[page_no, 0, by, page_offset, tx * 4 + vec + 128] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 128]))), where={freq: T.Cast("float32", k_rope_pos_offset[batch_idx] + row_g) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 256) / T.float32(256.0))}), pages[page_no, 0, by, page_offset, tx * 4 + vec]) V_smem[tile_start_s + j, tx * 4 + vec] = pages[page_no, 1, by, page_offset, tx * 4 + vec] else: for vec in T.vectorized(4): K_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) V_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) T.tvm_storage_sync("shared") m_prev[0] = st_m[0] for j in range(1): for vec in T.vectorized(4): QK_local[vec] = T.Cast("float32", Q_local[vec]) * T.Cast("float32", K_smem[tz + j, tx * 4 + vec]) * attn_score_scaling_factor * sm_scale S_reduce_local[0] = T.float32(0.0) for vec in T.unroll(4): S_reduce_local[0] = S_reduce_local[0] + QK_local[vec] with T.block("block_cross_thread"): T.reads(S_reduce_local[0]) T.writes(t0[0]) T.attr(T.comm_reducer(lambda x0, y0: x0 + y0, [T.float32(0.0)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))) T.tvm_thread_allreduce(T.uint32(1), S_reduce_local[0], T.bool(True), t0[0], tx) S_local[j] = T.float32(-50000.0) if iterator * 4 + tz + j < kv_chunk_len[0]: S_local[j] = t0[0] st_m[0] = T.max(st_m[0], S_local[j]) o_scale: T.float32 = T.exp2(m_prev[0] - st_m[0]) st_d[0] = st_d[0] * o_scale for j in range(1): S_local[j] = T.exp2(S_local[j] - st_m[0]) st_d[0] = st_d[0] + S_local[j] for j in T.vectorized(4): O_local[j] = O_local[j] * o_scale for j in range(1): for vec in T.vectorized(4): V_local[vec] = V_smem[tz + j, tx * 4 + vec] for vec in T.vectorized(4): O_local[vec] = O_local[vec] + T.Cast("float32", V_local[vec]) * S_local[j] for vec in T.vectorized(4): O_allreduce[tz, ty, tx * 4 + vec] = O_local[vec] md_allreduce[tz, ty, 0] = st_m[0] md_allreduce[tz, ty, 1] = st_d[0] T.tvm_storage_sync("shared") st_m[0] = T.float32(-50000.0) st_d[0] = T.float32(1.0) for vec in T.vectorized(4): O_local[vec] = T.float32(0.0) for j in range(4): m_prev[0] = st_m[0] d_prev[0] = st_d[0] other_m[0] = md_allreduce[j, ty, 0] other_d[0] = md_allreduce[j, ty, 1] for vec in T.vectorized(4): other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] st_m[0] = T.max(st_m[0], other_m[0]) st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) for vec in T.vectorized(4): O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] for vec in T.vectorized(4): O_local[vec] = O_local[vec] / st_d[0] for vec in T.vectorized(4): output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) @T.prim_func def batch_decode_paged_kv_sliding_window(_0: T.int32, Q_handle: T.handle, pages_handle: T.handle, page_table_indptr_handle: T.handle, page_table_values_handle: T.handle, var_length_info: T.handle, k_rope_pos_offset_handle: T.handle, q_rope_position_handle: T.handle, output_handle: T.handle, lse_handle: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) B = T.int32(is_size_var=True) Q = T.match_buffer(Q_handle, (B, 8, 256), "float16") max_num_pages = T.int32(is_size_var=True) pages = T.match_buffer(pages_handle, (max_num_pages, 2, 8, 16, 256), "float16", offset_factor=1) page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) nnz_pages = T.int32(is_size_var=True) page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) length_info = T.match_buffer(var_length_info, (3, B), "int32", offset_factor=1) k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) output = T.match_buffer(output_handle, (B, 8, 256), "float16") lse = T.match_buffer(lse_handle, (B, 8)) # with T.block("root"): sm_scale: T.float32 = T.float32(0.090168440055560212) for bx in T.thread_binding(B, thread="blockIdx.x"): for fused_by_bz in T.thread_binding(8, thread="blockIdx.y"): for ty in T.thread_binding(1, thread="threadIdx.y"): for tx in T.thread_binding(64, thread="threadIdx.x"): for tz in T.thread_binding(4, thread="threadIdx.z"): with T.block("attn"): T.reads(page_table_indptr[bx:bx + 2], length_info[0:3, bx], q_rope_position[bx], Q[bx, fused_by_bz // 8 + ty + fused_by_bz % 8, tx * 4 - 128:tx * 4 - 128 + 260]) T.writes(output[bx, fused_by_bz % 8 + fused_by_bz // 8 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 8 + fused_by_bz // 8 + ty]) Q_local = T.alloc_buffer((4,), "float16", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") K_smem = T.alloc_buffer((4, 256), "float16", scope="shared") V_smem = T.alloc_buffer((4, 256), "float16", scope="shared") O_allreduce = T.alloc_buffer((4, 1, 256), scope="shared") md_allreduce = T.alloc_buffer((4, 1, 2), scope="shared") S_reduce_local = T.alloc_buffer((1,), scope="local") t0 = T.alloc_buffer((1,), scope="local") S_local = T.alloc_buffer((1,), scope="local") QK_local = T.alloc_buffer((4,), scope="local") V_local = T.alloc_buffer((4,), "float16", scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_prev = T.alloc_buffer((1,), scope="local") other_m = T.alloc_buffer((1,), scope="local") other_d = T.alloc_buffer((1,), scope="local") exp_mprev = T.alloc_buffer((1,), scope="local") exp_otherm = T.alloc_buffer((1,), scope="local") other_o = T.alloc_buffer((4,), scope="local") st_m = T.alloc_buffer((1,), scope="local") st_d = T.alloc_buffer((1,), scope="local") O_local = T.alloc_buffer((4,), scope="local") by: T.int32 = fused_by_bz % 8 bz: T.int32 = fused_by_bz // 8 batch_idx: T.int32 = bx cur_page_indptr_begin: T.int32 = page_table_indptr[batch_idx] cur_page_indptr_end: T.int32 = page_table_indptr[batch_idx + 1] kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[0, batch_idx] - length_info[1, batch_idx] + length_info[2, batch_idx], 0) st_m[0] = T.float32(-50000.0) st_d[0] = T.float32(1.0) for vec in T.vectorized(4): O_local[vec] = T.float32(0.0) for vec in T.vectorized(4): freq = T.float32() Q_local[vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", Q[bx, by + bz + ty, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 128, Q[bx, by + bz + ty, tx * 4 + vec + 128] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 128]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 256) / T.float32(256.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) for iterator in range((kv_chunk_len[0] + 3) // 4): tile_start_s: T.int32 = tz + ty tile_start_g: T.int32 = iterator * 4 + tz + ty for j in range(1): with T.block("KV_load"): T.reads() T.writes() row_g: T.int32 = tile_start_g + j if row_g < kv_chunk_len[0]: seq_offset: T.int32 = T.if_then_else(row_g < length_info[2, batch_idx], row_g, row_g - length_info[2, batch_idx] + length_info[1, batch_idx]) page_no: T.int32 = page_table_values[cur_page_indptr_begin + seq_offset // 16] page_offset: T.int32 = seq_offset % 16 for vec in T.vectorized(4): freq = T.float32() K_smem[tile_start_s + j, tx * 4 + vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 128, pages[page_no, 0, by, page_offset, tx * 4 + vec + 128] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 128]))), where={freq: T.Cast("float32", k_rope_pos_offset[batch_idx] + row_g) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 256) / T.float32(256.0))}), pages[page_no, 0, by, page_offset, tx * 4 + vec]) V_smem[tile_start_s + j, tx * 4 + vec] = pages[page_no, 1, by, page_offset, tx * 4 + vec] else: for vec in T.vectorized(4): K_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) V_smem[tile_start_s + j, tx * 4 + vec] = T.float16(0.0) T.tvm_storage_sync("shared") m_prev[0] = st_m[0] for j in range(1): for vec in T.vectorized(4): QK_local[vec] = T.Cast("float32", Q_local[vec]) * T.Cast("float32", K_smem[tz + j, tx * 4 + vec]) * attn_score_scaling_factor * sm_scale S_reduce_local[0] = T.float32(0.0) for vec in T.unroll(4): S_reduce_local[0] = S_reduce_local[0] + QK_local[vec] with T.block("block_cross_thread"): T.reads(S_reduce_local[0]) T.writes(t0[0]) T.attr(T.comm_reducer(lambda x0, y0: x0 + y0, [T.float32(0.0)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))) T.tvm_thread_allreduce(T.uint32(1), S_reduce_local[0], T.bool(True), t0[0], tx) S_local[j] = T.float32(-50000.0) if iterator * 4 + tz + j < kv_chunk_len[0]: S_local[j] = t0[0] st_m[0] = T.max(st_m[0], S_local[j]) o_scale: T.float32 = T.exp2(m_prev[0] - st_m[0]) st_d[0] = st_d[0] * o_scale for j in range(1): S_local[j] = T.exp2(S_local[j] - st_m[0]) st_d[0] = st_d[0] + S_local[j] for j in T.vectorized(4): O_local[j] = O_local[j] * o_scale for j in range(1): for vec in T.vectorized(4): V_local[vec] = V_smem[tz + j, tx * 4 + vec] for vec in T.vectorized(4): O_local[vec] = O_local[vec] + T.Cast("float32", V_local[vec]) * S_local[j] for vec in T.vectorized(4): O_allreduce[tz, ty, tx * 4 + vec] = O_local[vec] md_allreduce[tz, ty, 0] = st_m[0] md_allreduce[tz, ty, 1] = st_d[0] T.tvm_storage_sync("shared") st_m[0] = T.float32(-50000.0) st_d[0] = T.float32(1.0) for vec in T.vectorized(4): O_local[vec] = T.float32(0.0) for j in range(4): m_prev[0] = st_m[0] d_prev[0] = st_d[0] other_m[0] = md_allreduce[j, ty, 0] other_d[0] = md_allreduce[j, ty, 1] for vec in T.vectorized(4): other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] st_m[0] = T.max(st_m[0], other_m[0]) st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) for vec in T.vectorized(4): O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] for vec in T.vectorized(4): O_local[vec] = O_local[vec] / st_d[0] for vec in T.vectorized(4): output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) @T.prim_func def batch_prefill_paged_kv(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) total_len = T.int32(is_size_var=True) q = T.match_buffer(var_q, (total_len, 8, 256), "float16") batch_size = T.int32(is_size_var=True) q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) max_num_pages = T.int32(is_size_var=True) pages = T.match_buffer(var_pages, (max_num_pages, 2, 8, 16, 256), "float16", offset_factor=1) page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) nnz_pages = T.int32(is_size_var=True) page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) output = T.match_buffer(var_output, (total_len, 8, 256), "float16") lse = T.match_buffer(var_lse, (total_len, 8)) # with T.block("root"): for lbx in T.thread_binding(16, thread="blockIdx.x"): for lby in T.thread_binding(8, thread="blockIdx.y"): for lty in T.thread_binding(4, thread="threadIdx.y"): for ltx in T.thread_binding(32, thread="threadIdx.x"): with T.block("attn"): bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) T.reads() T.writes() tile_id = T.alloc_buffer((1,), "int32", scope="local") batch_idx = T.alloc_buffer((1,), "int32", scope="local") batch_tiles = T.alloc_buffer((1,), "int32", scope="local") batch_rows = T.alloc_buffer((1,), "int32", scope="local") iterator = T.alloc_buffer((1,), "int32", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") Q_smem = T.alloc_buffer((16, 256), "float16", scope="shared") K_smem = T.alloc_buffer((16, 256), "float16", scope="shared") V_smem = T.alloc_buffer((16, 256), "float16", scope="shared") S_smem = T.alloc_buffer((16, 16), scope="shared") S_local = T.alloc_buffer((16, 16), scope="local") O_local = T.alloc_buffer((16, 256), scope="local") m_smem = T.alloc_buffer((16,), scope="shared") m_prev_smem = T.alloc_buffer((16,), scope="shared") d_smem = T.alloc_buffer((16,), scope="shared") m_new = T.alloc_buffer((1,), scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_new = T.alloc_buffer((1,), scope="local") tile_id[0] = bx batch_idx[0] = 0 batch_rows[0] = q_indptr[1] - q_indptr[0] batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 while T.tvm_thread_invariant(batch_idx[0] < batch_size): while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: tile_id[0] = tile_id[0] - batch_tiles[0] batch_idx[0] = batch_idx[0] + 1 if batch_idx[0] < batch_size: b_idx: T.int32 = batch_idx[0] batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 if T.tvm_thread_invariant(batch_idx[0] < batch_size): b_idx: T.int32 = batch_idx[0] LH_start: T.int32 = tile_id[0] * 16 q_indptr_val: T.int32 = q_indptr[b_idx] cur_page_indptr_begin: T.int32 = page_indptr[b_idx] cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[b_idx], 0) T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: m_smem[row] = T.float32(-50000.0) d_smem[row] = T.float32(1.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 8): with T.block("O_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) T.reads() T.writes(O_local[i, j]) O_local[i, j] = T.float32(0.0) T.tvm_storage_sync("shared") for li_lj_fused_0 in range(8): for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for li_lj_fused_3 in T.vectorized(4): with T.block("Q_load"): i = T.axis.spatial(16, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 256) j = T.axis.spatial(256, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 256) T.reads() T.writes() cur_L: T.int32 = q_indptr_val + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: freq = T.float32() Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), q[cur_L, cur_H_qo, j]) else: Q_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for iterator_1 in range((kv_chunk_len[0] + 15) // 16): L_kv_start: T.int32 = iterator_1 * 16 for lz_ly_fused_0 in range(8): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("K_load"): i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: seq_offset: T.int32 = cur_L page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] page_offset: T.int32 = seq_offset % 16 freq = T.float32() K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, pages[page_no, 0, by, page_offset, j + 128] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 128]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), pages[page_no, 0, by, page_offset, j]) else: K_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for lz_ly_fused_0 in range(8): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("V_load"): i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: seq_offset: T.int32 = cur_L page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] page_offset: T.int32 = seq_offset % 16 V_smem[i, j] = pages[page_no, 1, by, page_offset, j] else: V_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") with T.block(""): T.reads(Q_smem[0:16, 0:256], K_smem[0:16, 0:256]) T.writes(S_local[0:16, 0:16]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(1, 2): with T.block("S_gemm_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 + li_1_init) j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) T.reads() T.writes(S_local[i, j]) S_local[i, j] = T.float32(0.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, li_1, lj_1, lk_1 in T.grid(32, 1, 2, 8): with T.block("S_gemm_update"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) k = T.axis.reduce(256, lk_0 * 8 + lk_1) T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) T.writes(S_local[i, j]) S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.090168440055560212) T.tvm_storage_sync("shared") for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(1, 2): with T.block("S_store"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) T.reads(S_local[i, j]) T.writes(S_smem[i, j]) S_smem[i, j] = S_local[i, j] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: with T.block("update1"): T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) T.writes(m_prev[i], m_new[i], d_new[i]) m_prev[i] = m_smem[row] m_new[i] = m_smem[row] row_: T.int32 = LH_start + row for j in range(16): if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): m_new[i] = T.max(m_new[i], S_smem[row, j]) d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx with T.block("update"): T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) T.writes(S_smem[row, 0:16]) for j in range(16): if row < 16: row_: T.int32 = LH_start + row if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) else: S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: with T.block("update"): T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) for j in range(16): d_new[i] = d_new[i] + S_smem[row, j] m_smem[row] = m_new[i] d_smem[row] = d_new[i] m_prev_smem[row] = m_prev[i] T.tvm_storage_sync("shared") with T.block(""): T.reads(m_prev_smem[0:16], m_smem[0:16], S_smem[0:16, 0:16], V_smem[0:16, 0:256]) T.writes(O_local[0:16, 0:256]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(4, 8): with T.block("O_gemm_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + lj_1_init) T.reads() T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, lk_1, li_1, lj_1 in T.grid(2, 8, 4, 8): with T.block("O_gemm_update"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) k = T.axis.reduce(16, lk_0 * 8 + lk_1) T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 8): with T.block("O_store"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) for li_0 in range(1): for li_1 in T.thread_binding(4, thread="threadIdx.y"): for li_2 in T.thread_binding(32, thread="threadIdx.x"): with T.block("lse_store"): i = T.axis.spatial(16, li_0 * 128 + li_1 * 32 + li_2) T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) tile_id[0] = tile_id[0] + 16 @T.prim_func def batch_prefill_paged_kv_sliding_window(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) total_len = T.int32(is_size_var=True) q = T.match_buffer(var_q, (total_len, 8, 256), "float16") batch_size = T.int32(is_size_var=True) q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) max_num_pages = T.int32(is_size_var=True) pages = T.match_buffer(var_pages, (max_num_pages, 2, 8, 16, 256), "float16", offset_factor=1) page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) nnz_pages = T.int32(is_size_var=True) page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) length_info = T.match_buffer(var_length_info, (3, batch_size), "int32", offset_factor=1) k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) output = T.match_buffer(var_output, (total_len, 8, 256), "float16") lse = T.match_buffer(var_lse, (total_len, 8)) # with T.block("root"): for lbx in T.thread_binding(16, thread="blockIdx.x"): for lby in T.thread_binding(8, thread="blockIdx.y"): for lty in T.thread_binding(4, thread="threadIdx.y"): for ltx in T.thread_binding(32, thread="threadIdx.x"): with T.block("attn"): bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) T.reads() T.writes() tile_id = T.alloc_buffer((1,), "int32", scope="local") batch_idx = T.alloc_buffer((1,), "int32", scope="local") batch_tiles = T.alloc_buffer((1,), "int32", scope="local") batch_rows = T.alloc_buffer((1,), "int32", scope="local") iterator = T.alloc_buffer((1,), "int32", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") Q_smem = T.alloc_buffer((16, 256), "float16", scope="shared") K_smem = T.alloc_buffer((16, 256), "float16", scope="shared") V_smem = T.alloc_buffer((16, 256), "float16", scope="shared") S_smem = T.alloc_buffer((16, 16), scope="shared") S_local = T.alloc_buffer((16, 16), scope="local") O_local = T.alloc_buffer((16, 256), scope="local") m_smem = T.alloc_buffer((16,), scope="shared") m_prev_smem = T.alloc_buffer((16,), scope="shared") d_smem = T.alloc_buffer((16,), scope="shared") m_new = T.alloc_buffer((1,), scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_new = T.alloc_buffer((1,), scope="local") tile_id[0] = bx batch_idx[0] = 0 batch_rows[0] = q_indptr[1] - q_indptr[0] batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 while T.tvm_thread_invariant(batch_idx[0] < batch_size): while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: tile_id[0] = tile_id[0] - batch_tiles[0] batch_idx[0] = batch_idx[0] + 1 if batch_idx[0] < batch_size: b_idx: T.int32 = batch_idx[0] batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 if T.tvm_thread_invariant(batch_idx[0] < batch_size): b_idx: T.int32 = batch_idx[0] LH_start: T.int32 = tile_id[0] * 16 q_indptr_val: T.int32 = q_indptr[b_idx] cur_page_indptr_begin: T.int32 = page_indptr[b_idx] cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[0, b_idx] - length_info[1, b_idx] + length_info[2, b_idx], 0) T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: m_smem[row] = T.float32(-50000.0) d_smem[row] = T.float32(1.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 8): with T.block("O_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) T.reads() T.writes(O_local[i, j]) O_local[i, j] = T.float32(0.0) T.tvm_storage_sync("shared") for li_lj_fused_0 in range(8): for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for li_lj_fused_3 in T.vectorized(4): with T.block("Q_load"): i = T.axis.spatial(16, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 256) j = T.axis.spatial(256, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 256) T.reads() T.writes() cur_L: T.int32 = q_indptr_val + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: freq = T.float32() Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), q[cur_L, cur_H_qo, j]) else: Q_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for iterator_1 in range((kv_chunk_len[0] + 15) // 16): L_kv_start: T.int32 = iterator_1 * 16 for lz_ly_fused_0 in range(8): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("K_load"): i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: seq_offset: T.int32 = T.if_then_else(cur_L < length_info[2, b_idx], cur_L, cur_L - length_info[2, b_idx] + length_info[1, b_idx]) page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] page_offset: T.int32 = seq_offset % 16 freq = T.float32() K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", pages[page_no, 0, by, page_offset, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, pages[page_no, 0, by, page_offset, j + 128] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 128]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), pages[page_no, 0, by, page_offset, j]) else: K_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for lz_ly_fused_0 in range(8): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("V_load"): i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: seq_offset: T.int32 = T.if_then_else(cur_L < length_info[2, b_idx], cur_L, cur_L - length_info[2, b_idx] + length_info[1, b_idx]) page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] page_offset: T.int32 = seq_offset % 16 V_smem[i, j] = pages[page_no, 1, by, page_offset, j] else: V_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") with T.block(""): T.reads(Q_smem[0:16, 0:256], K_smem[0:16, 0:256]) T.writes(S_local[0:16, 0:16]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(1, 2): with T.block("S_gemm_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 + li_1_init) j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) T.reads() T.writes(S_local[i, j]) S_local[i, j] = T.float32(0.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, li_1, lj_1, lk_1 in T.grid(32, 1, 2, 8): with T.block("S_gemm_update"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) k = T.axis.reduce(256, lk_0 * 8 + lk_1) T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) T.writes(S_local[i, j]) S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.090168440055560212) T.tvm_storage_sync("shared") for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(1, 2): with T.block("S_store"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) T.reads(S_local[i, j]) T.writes(S_smem[i, j]) S_smem[i, j] = S_local[i, j] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: with T.block("update1"): T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) T.writes(m_prev[i], m_new[i], d_new[i]) m_prev[i] = m_smem[row] m_new[i] = m_smem[row] row_: T.int32 = LH_start + row for j in range(16): if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): m_new[i] = T.max(m_new[i], S_smem[row, j]) d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx with T.block("update"): T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) T.writes(S_smem[row, 0:16]) for j in range(16): if row < 16: row_: T.int32 = LH_start + row if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) else: S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: with T.block("update"): T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) for j in range(16): d_new[i] = d_new[i] + S_smem[row, j] m_smem[row] = m_new[i] d_smem[row] = d_new[i] m_prev_smem[row] = m_prev[i] T.tvm_storage_sync("shared") with T.block(""): T.reads(m_prev_smem[0:16], m_smem[0:16], S_smem[0:16, 0:16], V_smem[0:16, 0:256]) T.writes(O_local[0:16, 0:256]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(4, 8): with T.block("O_gemm_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + lj_1_init) T.reads() T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, lk_1, li_1, lj_1 in T.grid(2, 8, 4, 8): with T.block("O_gemm_update"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) k = T.axis.reduce(16, lk_0 * 8 + lk_1) T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 8): with T.block("O_store"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) for li_0 in range(1): for li_1 in T.thread_binding(4, thread="threadIdx.y"): for li_2 in T.thread_binding(32, thread="threadIdx.x"): with T.block("lse_store"): i = T.axis.spatial(16, li_0 * 128 + li_1 * 32 + li_2) T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) tile_id[0] = tile_id[0] + 16 @T.prim_func def batch_prefill_ragged_kv(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_k_rope_pos_offset: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) qo_len = T.int32(is_size_var=True) q = T.match_buffer(var_q, (qo_len, 8, 256), "float16") batch_size = T.int32(is_size_var=True) q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) kv_len = T.int32(is_size_var=True) k = T.match_buffer(var_k, (kv_len, 8, 256), "float16") v = T.match_buffer(var_v, (kv_len, 8, 256), "float16") kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) output = T.match_buffer(var_output, (qo_len, 8, 256), "float16") lse = T.match_buffer(var_lse, (qo_len, 8)) # with T.block("root"): for lbx in T.thread_binding(8, thread="blockIdx.x"): for lby in T.thread_binding(8, thread="blockIdx.y"): for lty in T.thread_binding(4, thread="threadIdx.y"): for ltx in T.thread_binding(32, thread="threadIdx.x"): with T.block("attn"): bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) T.reads() T.writes() tile_id = T.alloc_buffer((1,), "int32", scope="local") batch_idx = T.alloc_buffer((1,), "int32", scope="local") batch_tiles = T.alloc_buffer((1,), "int32", scope="local") batch_rows = T.alloc_buffer((1,), "int32", scope="local") iterator = T.alloc_buffer((1,), "int32", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") Q_smem = T.alloc_buffer((16, 256), "float16", scope="shared") K_smem = T.alloc_buffer((256, 16), "float16", scope="shared") V_smem = T.alloc_buffer((16, 256), "float16", scope="shared") S_smem = T.alloc_buffer((16, 16), scope="shared") S_local = T.alloc_buffer((16, 16), scope="local") O_local = T.alloc_buffer((16, 256), scope="local") m_smem = T.alloc_buffer((16,), scope="shared") m_prev_smem = T.alloc_buffer((16,), scope="shared") d_smem = T.alloc_buffer((16,), scope="shared") m_new = T.alloc_buffer((1,), scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_new = T.alloc_buffer((1,), scope="local") tile_id[0] = bx batch_idx[0] = 0 batch_rows[0] = q_indptr[1] - q_indptr[0] batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 while T.tvm_thread_invariant(batch_idx[0] < batch_size): while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: tile_id[0] = tile_id[0] - batch_tiles[0] batch_idx[0] = batch_idx[0] + 1 if batch_idx[0] < batch_size: b_idx: T.int32 = batch_idx[0] batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 if T.tvm_thread_invariant(batch_idx[0] < batch_size): b_idx: T.int32 = batch_idx[0] q_indptr_val: T.int32 = q_indptr[b_idx] LH_start: T.int32 = tile_id[0] * 16 kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: m_smem[row] = T.float32(-50000.0) d_smem[row] = T.float32(1.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1 in range(4): for lj_1_0 in T.unroll(1): for lj_1_1 in T.vectorized(8): with T.block("O_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1_0 * 8 + lj_1_1) T.reads() T.writes(O_local[i, j]) O_local[i, j] = T.float32(0.0) T.tvm_storage_sync("shared") for li_1_lj_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_1_lj_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_0, lj_0_0 in T.grid(2, 2): for lj_1 in T.vectorized(8): with T.block("Q_load"): i = T.axis.spatial(16, li_0 * 8 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) // 16) j = T.axis.spatial(256, lj_0_0 * 128 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) % 16 * 8 + lj_1) T.reads() T.writes() cur_L: T.int32 = q_indptr_val + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: freq = T.float32() Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), q[cur_L, cur_H_qo, j]) else: Q_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for iterator_1 in range((kv_chunk_len[0] + 15) // 16): L_kv_start: T.int32 = iterator_1 * 16 L_kv_base: T.int32 = kv_indptr[b_idx] for lz_1_ly_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for lz_1_ly_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lz_0, ly_0_0 in T.grid(2, 2): for ly_1 in T.vectorized(8): with T.block("K_load"): i = T.axis.spatial(16, lz_0 * 8 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 16) j = T.axis.spatial(256, ly_0_0 * 128 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 16 * 8 + ly_1) T.reads(kv_chunk_len[0], k_rope_pos_offset[b_idx], k[L_kv_base + L_kv_start + i, by, j - 128:j - 128 + 257]) T.writes(K_smem[j, i]) cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: freq = T.float32() K_smem[j, i] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", k[L_kv_base + cur_L, by, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, k[L_kv_base + cur_L, by, j + 128] * T.float16(-1.0), k[L_kv_base + cur_L, by, j - 128]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), k[L_kv_base + cur_L, by, j]) else: K_smem[j, i] = T.float16(0.0) T.tvm_storage_sync("shared") for lz_1_ly_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for lz_1_ly_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lz_0, ly_0_0 in T.grid(2, 2): for ly_1 in T.vectorized(8): with T.block("V_load"): i = T.axis.spatial(16, lz_0 * 8 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 16) j = T.axis.spatial(256, ly_0_0 * 128 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 16 * 8 + ly_1) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: V_smem[i, j] = v[L_kv_base + cur_L, by, j] else: V_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") with T.block(""): T.reads(Q_smem[0:16, 0:256], K_smem[0:256, 0:16]) T.writes(S_local[0:16, 0:16]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init in T.unroll(1): for lj_1_0_init in T.unroll(1): for lj_1_1_init in T.vectorized(2): with T.block("S_gemm_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 + li_1_init) j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_0_init * 2 + lj_1_1_init) T.reads() T.writes(S_local[i, j]) S_local[i, j] = T.float32(0.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0 in range(16): for li_1 in T.unroll(1): for lj_1_0 in T.unroll(1): for lj_1_1 in T.vectorized(2): for lk_1 in range(16): with T.block("S_gemm_update"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1_0 * 2 + lj_1_1) k_1 = T.axis.reduce(256, lk_0 * 16 + lk_1) T.reads(S_local[i, j], Q_smem[i, k_1], K_smem[k_1, j]) T.writes(S_local[i, j]) S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k_1]) * T.Cast("float32", K_smem[k_1, j]) * attn_score_scaling_factor * T.float32(0.090168440055560212) T.tvm_storage_sync("shared") for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1 in range(1): for lj_1_0 in T.unroll(1): for lj_1_1 in T.vectorized(2): with T.block("S_store"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 + li_1) j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1_0 * 2 + lj_1_1) T.reads(S_local[i, j]) T.writes(S_smem[i, j]) S_smem[i, j] = S_local[i, j] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: with T.block("update1"): T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) T.writes(m_prev[i], m_new[i], d_new[i]) m_prev[i] = m_smem[row] m_new[i] = m_smem[row] row_: T.int32 = LH_start + row for j in range(16): if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): m_new[i] = T.max(m_new[i], S_smem[row, j]) d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx with T.block("update"): T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) T.writes(S_smem[row, 0:16]) for j in range(16): if row < 16: row_: T.int32 = LH_start + row if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) else: S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: with T.block("update"): T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) for j in range(16): d_new[i] = d_new[i] + S_smem[row, j] m_smem[row] = m_new[i] d_smem[row] = d_new[i] m_prev_smem[row] = m_prev[i] T.tvm_storage_sync("shared") with T.block(""): T.reads(m_prev_smem[0:16], m_smem[0:16], S_smem[0:16, 0:16], V_smem[0:16, 0:256]) T.writes(O_local[0:16, 0:256]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init in T.unroll(4): for lj_1_0_init in T.unroll(1): for lj_1_1_init in T.vectorized(8): with T.block("O_gemm_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + lj_1_0_init * 8 + lj_1_1_init) T.reads() T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, lk_1 in T.grid(1, 16): for li_1 in T.unroll(4): for lj_1_0 in T.unroll(1): for lj_1_1 in T.vectorized(8): with T.block("O_gemm_update"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1_0 * 8 + lj_1_1) k_1 = T.axis.reduce(16, lk_0 * 16 + lk_1) T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k_1], V_smem[k_1, j]) T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] + S_smem[i, k_1] * T.Cast("float32", V_smem[k_1, j]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1 in range(4): for lj_1_0 in T.unroll(1): for lj_1_1 in T.vectorized(8): with T.block("O_store"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1_0 * 8 + lj_1_1) T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) for li_0 in range(1): for li_1 in T.thread_binding(4, thread="threadIdx.y"): for li_2 in T.thread_binding(32, thread="threadIdx.x"): with T.block("lse_store"): i = T.axis.spatial(16, li_0 * 128 + li_1 * 32 + li_2) T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) tile_id[0] = tile_id[0] + 8 @T.prim_func def batch_tree_attn(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_mn_indptr: T.handle, var_mask: T.handle, var_output: T.handle, var_lse: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32, batch_size: T.int32): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) qo_len = T.int32(is_size_var=True) q = T.match_buffer(var_q, (qo_len, 8, 256), "float16") q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) kv_len = T.int32(is_size_var=True) k = T.match_buffer(var_k, (kv_len, 8, 256), "float16") v = T.match_buffer(var_v, (kv_len, 8, 256), "float16") kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) mn_indptr = T.match_buffer(var_mn_indptr, (batch_size + 1,), "int32", offset_factor=1) tree_size = T.int32(is_size_var=True) mask = T.match_buffer(var_mask, (tree_size, 2), "int32", offset_factor=1) output = T.match_buffer(var_output, (qo_len, 8, 256), "float16") lse = T.match_buffer(var_lse, (qo_len, 8)) # with T.block("root"): for lbx in T.thread_binding(16, thread="blockIdx.x"): for lby in T.thread_binding(8, thread="blockIdx.y"): for lty in T.thread_binding(4, thread="threadIdx.y"): for ltx in T.thread_binding(32, thread="threadIdx.x"): with T.block("attn"): bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) T.reads() T.writes() tile_id = T.alloc_buffer((1,), "int32", scope="local") batch_idx = T.alloc_buffer((1,), "int32", scope="local") batch_tiles = T.alloc_buffer((1,), "int32", scope="local") batch_rows = T.alloc_buffer((1,), "int32", scope="local") iterator = T.alloc_buffer((1,), "int32", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") Q_smem = T.alloc_buffer((16, 256), "float16", scope="shared") K_smem = T.alloc_buffer((32, 256), "float16", scope="shared") V_smem = T.alloc_buffer((32, 256), "float16", scope="shared") S_smem = T.alloc_buffer((16, 32), scope="shared") S_local = T.alloc_buffer((16, 32), scope="local") O_local = T.alloc_buffer((16, 256), scope="local") m_smem = T.alloc_buffer((16,), scope="shared") m_prev_smem = T.alloc_buffer((16,), scope="shared") d_smem = T.alloc_buffer((16,), scope="shared") m_new = T.alloc_buffer((1,), scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_new = T.alloc_buffer((1,), scope="local") tile_id[0] = bx batch_idx[0] = 0 batch_rows[0] = q_indptr[1] - q_indptr[0] batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 while T.tvm_thread_invariant(batch_idx[0] < batch_size): while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: tile_id[0] = tile_id[0] - batch_tiles[0] batch_idx[0] = batch_idx[0] + 1 if batch_idx[0] < batch_size: b_idx: T.int32 = batch_idx[0] batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 if T.tvm_thread_invariant(batch_idx[0] < batch_size): b_idx: T.int32 = batch_idx[0] LH_start: T.int32 = tile_id[0] * 16 q_indptr_val: T.int32 = q_indptr[b_idx] kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: m_smem[row] = T.float32(-50000.0) d_smem[row] = T.float32(1.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 8): with T.block("O_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) T.reads() T.writes(O_local[i, j]) O_local[i, j] = T.float32(0.0) T.tvm_storage_sync("shared") for li_lj_fused_0 in range(8): for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for li_lj_fused_3 in T.vectorized(4): with T.block("Q_load"): i = T.axis.spatial(16, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 256) j = T.axis.spatial(256, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 256) T.reads() T.writes() cur_L: T.int32 = q_indptr_val + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: freq = T.float32() Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), q[cur_L, cur_H_qo, j]) else: Q_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for iterator_1 in range((kv_chunk_len[0] + 31) // 32): L_kv_start: T.int32 = iterator_1 * 32 L_kv_base: T.int32 = kv_indptr[b_idx] for lz_ly_fused_0 in range(16): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("KV_load"): i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) T.reads() T.writes() cur_L: T.int32 = L_kv_base + L_kv_start + i if L_kv_start + i < kv_chunk_len[0]: freq = T.float32() K_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", k[cur_L, by, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, k[cur_L, by, j + 128] * T.float16(-1.0), k[cur_L, by, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), k[cur_L, by, j]) V_smem[i, j] = v[cur_L, by, j] else: K_smem[i, j] = T.float16(0.0) V_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") with T.block(""): T.reads(Q_smem[0:16, 0:256], K_smem[0:32, 0:256]) T.writes(S_local[0:16, 0:32]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(2, 2): with T.block("S_gemm_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 2 + li_1_init) j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 2 + lj_1_init) T.reads() T.writes(S_local[i, j]) S_local[i, j] = T.float32(0.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, li_1, lj_1, lk_1 in T.grid(32, 2, 2, 8): with T.block("S_gemm_update"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 2 + li_1) j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 2 + lj_1) k_1 = T.axis.reduce(256, lk_0 * 8 + lk_1) T.reads(S_local[i, j], Q_smem[i, k_1], K_smem[j, k_1]) T.writes(S_local[i, j]) S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k_1]) * T.Cast("float32", K_smem[j, k_1]) * attn_score_scaling_factor * T.float32(0.090168440055560212) T.tvm_storage_sync("shared") for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(2, 2): with T.block("S_store"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 2 + li_1) j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 2 + lj_1) T.reads(S_local[i, j]) T.writes(S_smem[i, j]) S_smem[i, j] = S_local[i, j] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: with T.block("update1"): T.reads(m_smem[row], kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) T.writes(m_prev[i], m_new[i], d_new[i]) m_prev[i] = m_smem[row] m_new[i] = m_smem[row] row_: T.int32 = LH_start + row for j in range(32): if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): m_new[i] = T.max(m_new[i], S_smem[row, j]) d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx with T.block("update"): T.reads(kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) T.writes(S_smem[row, 0:32]) for j in range(32): if row < 16: row_: T.int32 = LH_start + row if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) else: S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: with T.block("update"): T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) for j in range(32): d_new[i] = d_new[i] + S_smem[row, j] m_smem[row] = m_new[i] d_smem[row] = d_new[i] m_prev_smem[row] = m_prev[i] T.tvm_storage_sync("shared") with T.block(""): T.reads(m_prev_smem[0:16], m_smem[0:16], S_smem[0:16, 0:32], V_smem[0:32, 0:256]) T.writes(O_local[0:16, 0:256]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(4, 8): with T.block("O_gemm_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + lj_1_init) T.reads() T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, lk_1, li_1, lj_1 in T.grid(4, 8, 4, 8): with T.block("O_gemm_update"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) k_1 = T.axis.reduce(32, lk_0 * 8 + lk_1) T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k_1], V_smem[k_1, j]) T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] + S_smem[i, k_1] * T.Cast("float32", V_smem[k_1, j]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 8): with T.block("O_store"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) for li_0 in range(1): for li_1 in T.thread_binding(4, thread="threadIdx.y"): for li_2 in T.thread_binding(32, thread="threadIdx.x"): with T.block("lse_store"): i = T.axis.spatial(16, li_0 * 128 + li_1 * 32 + li_2) T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) tile_id[0] = tile_id[0] + 16 @T.prim_func def chunk_lse(var_A: T.handle, var_temperature: T.handle, var_chunked_sum: T.handle, var_chunked_max: T.handle): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.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, 256), "float16", offset_factor=1) copy_length_indptr = T.match_buffer(var_copy_length_indptr, (batch_size + 1,), "int32", offset_factor=1) total_copy_length = T.int32() copy_src_dst_pos = T.match_buffer(var_copy_src_dst_pos, (2, total_copy_length), "int32", offset_factor=1) with T.block("root"): T.reads() T.writes() for bhd_o in T.thread_binding(batch_size * 8, thread="blockIdx.x"): for bhd_i in T.thread_binding(256, thread="threadIdx.x"): b: T.int32 = (bhd_o * 256 + bhd_i) // 2048 h: T.int32 = (bhd_o * 256 + bhd_i) // 256 % 8 d: T.int32 = (bhd_o * 256 + bhd_i) % 256 if bhd_o * 256 + bhd_i < batch_size * 8 * 256: for i in range(copy_length_indptr[b + 1] - copy_length_indptr[b]): src_pos: T.int32 = copy_src_dst_pos[0, copy_length_indptr[b] + i] dst_pos: T.int32 = copy_src_dst_pos[1, copy_length_indptr[b] + i] pages[dst_pos // 16, 0, h, dst_pos % 16, d] = pages[src_pos // 16, 0, h, src_pos % 16, d] pages[dst_pos // 16, 1, h, dst_pos % 16, d] = pages[src_pos // 16, 1, h, src_pos % 16, d] @T.prim_func def copy_single_page(var_pages: T.handle, src_page_id: T.int64, tgt_page_id: T.int64, copy_length: T.int64): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) num_pages, page_size = T.int32(), T.int64() pages = T.match_buffer(var_pages, (num_pages, 2, 8, page_size, 256), "float16", offset_factor=1) # with T.block("root"): for b in T.thread_binding(copy_length * T.int64(8), thread="blockIdx.x"): for t in T.thread_binding(256, thread="threadIdx.x"): with T.block("copy"): vh = T.axis.spatial(8, T.Cast("int32", (b * T.int64(256) + T.Cast("int64", t)) // (copy_length * T.int64(256)))) vp = T.axis.spatial(copy_length, (b * T.int64(256) + T.Cast("int64", t)) % (copy_length * T.int64(256)) // T.int64(256)) vd = T.axis.spatial(256, T.Cast("int32", (b * T.int64(256) + T.Cast("int64", t)) % T.int64(256))) T.where(b * T.int64(256) + T.Cast("int64", t) < copy_length * T.int64(8) * T.int64(256)) T.reads(pages[src_page_id, 0:2, vh, vp, vd]) T.writes(pages[tgt_page_id, 0:2, vh, vp, vd]) pages[tgt_page_id, 0, vh, vp, vd] = pages[src_page_id, 0, vh, vp, vd] pages[tgt_page_id, 1, vh, vp, vd] = pages[src_page_id, 1, vh, vp, vd] @T.prim_func 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({"op_pattern": 8, "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, 256), "float16") position_map = T.match_buffer(var_position_map, (seq_len,), "int32", offset_factor=1) q = T.match_buffer(var_q, (seq_len, 8, 256), "float16") k = T.match_buffer(var_k, (seq_len, 8, 256), "float16") v = T.match_buffer(var_v, (seq_len, 8, 256), "float16") # with T.block("root"): for iters_0, iters_1, iters_2 in T.grid(seq_len, 24, 256): 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 - 32:d - 32 + 65]) 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 < 64, 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 < 32, qkv[s, h, d + 32] * T.float16(-1.0), qkv[s, h, d - 32]))), where={freq: T.Cast("float32", position_map[s]) / T.pow(T.float32(10000.0), T.Cast("float32", d * 2 % 64) / T.float32(64.0))}), 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 < 64, 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 < 32, qkv[s, h, d + 32] * T.float16(-1.0), qkv[s, h, d - 32]))), where={freq: T.Cast("float32", position_map[s]) / T.pow(T.float32(10000.0), T.Cast("float32", d * 2 % 64) / T.float32(64.0))}), 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_norm32: T.handle, index: T.Buffer((T.int64(1), T.int64(1), T.int64(2048)), "float16")): T.func_attr({"target": T.target({"keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.noalias": T.bool(True)}) seq_len = T.int64() layer_norm32 = T.match_buffer(var_layer_norm32, (T.int64(1), seq_len, T.int64(2048)), "float16") # with T.block("root"): for i, _, k in T.grid(T.int64(1), T.int64(1), T.int64(2048)): with T.block("index"): v_i, v__, v_k = T.axis.remap("SSS", [i, _, k]) T.reads(layer_norm32[v_i, seq_len - T.int64(1), v_k]) T.writes(index[v_i, v__, v_k]) index[v_i, v__, v_k] = layer_norm32[v_i, seq_len - T.int64(1), v_k] @T.prim_func def merge_state_inplace(v: T.handle, s: T.handle, v_other: T.handle, s_other: T.handle): T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) N, H, D = T.int32(is_size_var=True), T.int32(is_size_var=True), T.int32(is_size_var=True) V = T.match_buffer(v, (N, H, D), "float16") S = T.match_buffer(s, (N, H)) V_other = T.match_buffer(v_other, (N, H, D), "float16") S_other = T.match_buffer(s_other, (N, H)) # with T.block("root"): for bx in T.thread_binding(N, thread="blockIdx.x"): for by in T.thread_binding(2, thread="blockIdx.y"): for ty in T.thread_binding(4, thread="threadIdx.y"): for tx in T.thread_binding(64, thread="threadIdx.x"): with T.block("merge"): T.reads(S[bx, ty + by * 4], S_other[bx, ty + by * 4], V[bx, ty + by * 4, tx * 4:tx * 4 + 4], V_other[bx, ty + by * 4, tx * 4:tx * 4 + 4]) T.writes(V[bx, ty + by * 4, tx * 4:tx * 4 + 4], S[bx, ty + by * 4]) s_val = T.alloc_buffer((1,), scope="local") s_other_val = T.alloc_buffer((1,), scope="local") s_max = T.alloc_buffer((1,), scope="local") scale = T.alloc_buffer((1,), scope="local") other_scale = T.alloc_buffer((1,), scope="local") v_vec = T.alloc_buffer((4,), "float16", scope="local") v_other_vec = T.alloc_buffer((4,), "float16", scope="local") s_val[0] = S[bx, ty + by * 4] s_other_val[0] = S_other[bx, ty + by * 4] s_max[0] = T.max(s_val[0], s_other_val[0]) s_val[0] = T.exp2(s_val[0] - s_max[0]) s_other_val[0] = T.exp2(s_other_val[0] - s_max[0]) scale[0] = s_val[0] / (s_val[0] + s_other_val[0]) other_scale[0] = s_other_val[0] / (s_val[0] + s_other_val[0]) for vec in T.vectorized(4): v_vec[vec] = V[bx, ty + by * 4, tx * 4 + vec] for vec in T.vectorized(4): v_other_vec[vec] = V_other[bx, ty + by * 4, tx * 4 + vec] for vec in range(4): v_vec[vec] = T.Cast("float16", T.Cast("float32", v_vec[vec]) * scale[0] + T.Cast("float32", v_other_vec[vec]) * other_scale[0]) for vec in T.vectorized(4): V[bx, ty + by * 4, tx * 4 + vec] = v_vec[vec] S[bx, ty + by * 4] = T.log2(s_val[0] + s_other_val[0]) + s_max[0] @T.prim_func 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 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, 256), "float16", offset_factor=1) seqlen = T.int64(is_size_var=True) position_map = T.match_buffer(var_position_map, (seqlen,), "int32", offset_factor=1) k_data = T.match_buffer(var_k_data, (16, seqlen, 8, 256), "float16") v_data = T.match_buffer(var_v_data, (16, seqlen, 8, 256), "float16") # with T.block("root"): for p, h, d in T.grid(seqlen, 8, 256): 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, 256), "float16", offset_factor=1) ntoken = T.int64(is_size_var=True) k_data = T.match_buffer(var_k_data, (ntoken, 8, 256), "float16") v_data = T.match_buffer(var_v_data, (ntoken, 8, 256), "float16") position_map = T.match_buffer(var_position_map, (ntoken,), "int32", offset_factor=1) # with T.block("root"): for global_pos, h, f in T.grid(ntoken, 8, 256): 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, 256), "float16") batch_size = T.int32(is_size_var=True) q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) max_num_pages = T.int32(is_size_var=True) pages = T.match_buffer(var_pages, (max_num_pages, 2, 8, 32, 256), "float16") page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) nnz_pages = T.int32(is_size_var=True) page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) output = T.match_buffer(var_output, (total_len, 8, 256), "float16") lse = T.match_buffer(var_lse, (total_len, 8)) tree_order_indptr = T.match_buffer(tree_order_indptr_handle, (batch_size + 1,), "int32", offset_factor=1) total_tree_order_len = T.int32(is_size_var=True) tree_order = T.match_buffer(tree_order_handle, (total_tree_order_len, 2), "int32", offset_factor=1) # with T.block("root"): assert rotary_mode == 0, "Inline rotary mode is not supported in tree attention." for lbx in T.thread_binding(16, thread="blockIdx.x"): for lby in T.thread_binding(8, thread="blockIdx.y"): for lty in T.thread_binding(4, thread="threadIdx.y"): for ltx in T.thread_binding(32, thread="threadIdx.x"): with T.block("attn"): bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) T.reads() T.writes() tile_id = T.alloc_buffer((1,), "int32", scope="local") batch_idx = T.alloc_buffer((1,), "int32", scope="local") batch_tiles = T.alloc_buffer((1,), "int32", scope="local") batch_rows = T.alloc_buffer((1,), "int32", scope="local") iterator = T.alloc_buffer((1,), "int32", scope="local") kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") Q_smem = T.alloc_buffer((16, 256), "float16", scope="shared") K_smem = T.alloc_buffer((32, 256), "float16", scope="shared") V_smem = T.alloc_buffer((32, 256), "float16", scope="shared") S_smem = T.alloc_buffer((16, 32), scope="shared") S_local = T.alloc_buffer((16, 32), scope="local") O_local = T.alloc_buffer((16, 256), scope="local") m_smem = T.alloc_buffer((16,), scope="shared") m_prev_smem = T.alloc_buffer((16,), scope="shared") d_smem = T.alloc_buffer((16,), scope="shared") m_new = T.alloc_buffer((1,), scope="local") m_prev = T.alloc_buffer((1,), scope="local") d_new = T.alloc_buffer((1,), scope="local") tile_id[0] = bx batch_idx[0] = 0 batch_rows[0] = q_indptr[1] - q_indptr[0] batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 while T.tvm_thread_invariant(batch_idx[0] < batch_size): while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: tile_id[0] = tile_id[0] - batch_tiles[0] batch_idx[0] = batch_idx[0] + 1 if batch_idx[0] < batch_size: b_idx: T.int32 = batch_idx[0] batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] batch_tiles[0] = (batch_rows[0] + 16 - 1) // 16 if T.tvm_thread_invariant(batch_idx[0] < batch_size): b_idx: T.int32 = batch_idx[0] LH_start: T.int32 = tile_id[0] * 16 q_indptr_val: T.int32 = q_indptr[b_idx] cur_page_indptr_begin: T.int32 = page_indptr[b_idx] cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 32 + length_info[b_idx], 0) T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: m_smem[row] = T.float32(-50000.0) d_smem[row] = T.float32(1.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 8): with T.block("O_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) T.reads() T.writes(O_local[i, j]) O_local[i, j] = T.float32(0.0) T.tvm_storage_sync("shared") for li_lj_fused_0 in range(8): for li_lj_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for li_lj_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for li_lj_fused_3 in T.vectorized(4): with T.block("Q_load"): i = T.axis.spatial(16, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 256) j = T.axis.spatial(256, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 256) T.reads() T.writes() cur_L: T.int32 = q_indptr_val + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: freq = T.float32() Q_smem[i, j] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", q[cur_L, cur_H_qo, j]) + T.sin(freq) * T.Cast("float32", T.if_then_else(j < 128, q[cur_L, cur_H_qo, j + 128] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 128]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 256) / T.float32(256.0))}), q[cur_L, cur_H_qo, j]) else: Q_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for iterator_1 in range((kv_chunk_len[0] + 31) // 32): L_kv_start: T.int32 = iterator_1 * 32 for lz_ly_fused_0 in range(16): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("K_load"): i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: seq_offset: T.int32 = cur_L page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 32] page_offset: T.int32 = seq_offset % 32 K_smem[i, j] = pages[page_no, 0, by, page_offset, j] else: K_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") for lz_ly_fused_0 in range(16): for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): for lz_ly_fused_3 in T.vectorized(4): with T.block("V_load"): i = T.axis.spatial(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 256) j = T.axis.spatial(256, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 256) T.reads() T.writes() cur_L: T.int32 = L_kv_start + i if cur_L < kv_chunk_len[0]: seq_offset: T.int32 = cur_L page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 32] page_offset: T.int32 = seq_offset % 32 V_smem[i, j] = pages[page_no, 1, by, page_offset, j] else: V_smem[i, j] = T.float16(0.0) T.tvm_storage_sync("shared") with T.block(""): T.reads(Q_smem[0:16, 0:256], K_smem[0:32, 0:256]) T.writes(S_local[0:16, 0:32]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(2, 2): with T.block("S_gemm_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 2 + li_1_init) j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 2 + lj_1_init) T.reads() T.writes(S_local[i, j]) S_local[i, j] = T.float32(0.0) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, li_1, lj_1, lk_1 in T.grid(32, 2, 2, 8): with T.block("S_gemm_update"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 2 + li_1) j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 2 + lj_1) k = T.axis.reduce(256, lk_0 * 8 + lk_1) T.reads(S_local[i, j], Q_smem[i, k], K_smem[j, k]) T.writes(S_local[i, j]) S_local[i, j] = S_local[i, j] + T.Cast("float32", Q_smem[i, k]) * T.Cast("float32", K_smem[j, k]) * attn_score_scaling_factor * T.float32(0.090168440055560212) T.tvm_storage_sync("shared") for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(2, 2): with T.block("S_store"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 2 + li_1) j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 2 + lj_1) T.reads(S_local[i, j]) T.writes(S_smem[i, j]) S_smem[i, j] = S_local[i, j] T.tvm_storage_sync("shared") for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: with T.block("update1"): T.reads(m_smem[row], kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) T.writes(m_prev[i], m_new[i], d_new[i]) m_prev[i] = m_smem[row] m_new[i] = m_smem[row] row_: T.int32 = LH_start + row for j in range(32): if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): m_new[i] = T.max(m_new[i], S_smem[row, j]) d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx with T.block("update"): T.reads(kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) T.writes(S_smem[row, 0:32]) for j in range(32): if row < 16: row_: T.int32 = LH_start + row if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) else: S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) for i in range(1): row: T.int32 = i * 32 * 4 + ty * 32 + tx if row < 16: with T.block("update"): T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) for j in range(32): d_new[i] = d_new[i] + S_smem[row, j] m_smem[row] = m_new[i] d_smem[row] = d_new[i] m_prev_smem[row] = m_prev[i] T.tvm_storage_sync("shared") with T.block(""): T.reads(m_prev_smem[0:16], m_smem[0:16], S_smem[0:16, 0:32], V_smem[0:32, 0:256]) T.writes(O_local[0:16, 0:256]) for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): for li_1_init, lj_1_init in T.grid(4, 8): with T.block("O_gemm_init"): i = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 32 * 4 + li_1_init) j = T.axis.spatial(256, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 32 * 8 + lj_1_init) T.reads() T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] * T.exp2(m_prev_smem[i] - m_smem[i]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for lk_0, lk_1, li_1, lj_1 in T.grid(4, 8, 4, 8): with T.block("O_gemm_update"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) k = T.axis.reduce(32, lk_0 * 8 + lk_1) T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k], V_smem[k, j]) T.writes(O_local[i, j]) O_local[i, j] = O_local[i, j] + S_smem[i, k] * T.Cast("float32", V_smem[k, j]) for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): for li_1, lj_1 in T.grid(4, 8): with T.block("O_store"): i = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 32 * 4 + li_1) j = T.axis.spatial(256, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 32 * 8 + lj_1) T.reads(q_indptr[b_idx:b_idx + 2], O_local[i, j], d_smem[i]) T.writes(output[q_indptr[b_idx] + (LH_start + i), by, j]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: output[cur_L, cur_H_qo, j] = T.Cast("float16", O_local[i, j] / d_smem[i]) for li_0 in range(1): for li_1 in T.thread_binding(4, thread="threadIdx.y"): for li_2 in T.thread_binding(32, thread="threadIdx.x"): with T.block("lse_store"): i = T.axis.spatial(16, li_0 * 128 + li_1 * 32 + li_2) T.where((li_0 * 4 + li_1) * 32 + li_2 < 16) T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) cur_H_qo: T.int32 = by if cur_L < q_indptr[b_idx + 1]: lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) tile_id[0] = tile_id[0] + 16 @R.function def alloc_embedding_tensor() -> R.Tensor((2048, 2048), dtype="float16"): R.func_attr({"relax.memory_plan_dynamic_func_output": True}) gv: R.Tensor((2048, 2048), dtype="float16") = R.builtin.alloc_tensor(R.shape([2048, 2048]), R.dtype("float16"), R.prim_value(0), R.str("global")) return gv @R.function def batch_decode(input_embeds: R.Tensor(("batch_size", 1, 2048), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 2048), dtype="float16"))) -> R.Tuple(R.Tensor(("batch_size", 1, "vocab_size"), dtype="float32"), R.Object): batch_size = T.int64() vocab_size = T.int64() R.func_attr({"num_input": 2, "pipeline_parallel_stages": 1, "relax.memory_plan_dynamic_func_output": True, "relax.rewrite_cuda_graph.capture_symbolic_vars": ["batch_size"], "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) cls = Module with R.dataflow(): gpt_neox_layers_0_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[1] gpt_neox_layers_0_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[2] gpt_neox_layers_0_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[3] gpt_neox_layers_0_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[4] gpt_neox_layers_0_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[5] gpt_neox_layers_0_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[6] gpt_neox_layers_0_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[8] gpt_neox_layers_0_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[9] gpt_neox_layers_0_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[10] gpt_neox_layers_0_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[11] gpt_neox_layers_0_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[12] gpt_neox_layers_1_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[13] gpt_neox_layers_1_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[14] gpt_neox_layers_1_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[15] gpt_neox_layers_1_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[16] gpt_neox_layers_1_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[17] gpt_neox_layers_1_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[18] gpt_neox_layers_1_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[19] gpt_neox_layers_1_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[20] gpt_neox_layers_1_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[21] gpt_neox_layers_1_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[22] gpt_neox_layers_1_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[23] gpt_neox_layers_1_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[24] gpt_neox_layers_2_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[25] gpt_neox_layers_2_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[26] gpt_neox_layers_2_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[27] gpt_neox_layers_2_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[28] gpt_neox_layers_2_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[29] gpt_neox_layers_2_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[30] gpt_neox_layers_2_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[31] gpt_neox_layers_2_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[32] gpt_neox_layers_2_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[33] gpt_neox_layers_2_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[34] gpt_neox_layers_2_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[35] gpt_neox_layers_2_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[36] gpt_neox_layers_3_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[37] gpt_neox_layers_3_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[38] gpt_neox_layers_3_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[39] gpt_neox_layers_3_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[40] gpt_neox_layers_3_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[41] gpt_neox_layers_3_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[42] gpt_neox_layers_3_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[43] gpt_neox_layers_3_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[44] gpt_neox_layers_3_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[45] gpt_neox_layers_3_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[46] gpt_neox_layers_3_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[47] gpt_neox_layers_3_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[48] gpt_neox_layers_4_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[49] gpt_neox_layers_4_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[50] gpt_neox_layers_4_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[51] gpt_neox_layers_4_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[52] gpt_neox_layers_4_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[53] gpt_neox_layers_4_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[54] gpt_neox_layers_4_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[55] gpt_neox_layers_4_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[56] gpt_neox_layers_4_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[57] gpt_neox_layers_4_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[58] gpt_neox_layers_4_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[59] gpt_neox_layers_4_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[60] gpt_neox_layers_5_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[61] gpt_neox_layers_5_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[62] gpt_neox_layers_5_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[63] gpt_neox_layers_5_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[64] gpt_neox_layers_5_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[65] gpt_neox_layers_5_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[66] gpt_neox_layers_5_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[67] gpt_neox_layers_5_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[68] gpt_neox_layers_5_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[69] gpt_neox_layers_5_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[70] gpt_neox_layers_5_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[71] gpt_neox_layers_5_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[72] gpt_neox_layers_6_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[73] gpt_neox_layers_6_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[74] gpt_neox_layers_6_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[75] gpt_neox_layers_6_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[76] gpt_neox_layers_6_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[77] gpt_neox_layers_6_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[78] gpt_neox_layers_6_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[79] gpt_neox_layers_6_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[80] gpt_neox_layers_6_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[81] gpt_neox_layers_6_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[82] gpt_neox_layers_6_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[83] gpt_neox_layers_6_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[84] gpt_neox_layers_7_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[85] gpt_neox_layers_7_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[86] gpt_neox_layers_7_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[87] gpt_neox_layers_7_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[88] gpt_neox_layers_7_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[89] gpt_neox_layers_7_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[90] gpt_neox_layers_7_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[91] gpt_neox_layers_7_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[92] gpt_neox_layers_7_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[93] gpt_neox_layers_7_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[94] gpt_neox_layers_7_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[95] gpt_neox_layers_7_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[96] gpt_neox_layers_8_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[97] gpt_neox_layers_8_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[98] gpt_neox_layers_8_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[99] gpt_neox_layers_8_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[100] gpt_neox_layers_8_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[101] gpt_neox_layers_8_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[102] gpt_neox_layers_8_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[103] gpt_neox_layers_8_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[104] gpt_neox_layers_8_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[105] gpt_neox_layers_8_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[106] gpt_neox_layers_8_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[107] gpt_neox_layers_8_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[108] gpt_neox_layers_9_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[109] gpt_neox_layers_9_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[110] gpt_neox_layers_9_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[111] gpt_neox_layers_9_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[112] gpt_neox_layers_9_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[113] gpt_neox_layers_9_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[114] gpt_neox_layers_9_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[115] gpt_neox_layers_9_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[116] gpt_neox_layers_9_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[117] gpt_neox_layers_9_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[118] gpt_neox_layers_9_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[119] gpt_neox_layers_9_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[120] gpt_neox_layers_10_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[121] gpt_neox_layers_10_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[122] gpt_neox_layers_10_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[123] gpt_neox_layers_10_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[124] gpt_neox_layers_10_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[125] gpt_neox_layers_10_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[126] gpt_neox_layers_10_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[127] gpt_neox_layers_10_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[128] gpt_neox_layers_10_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[129] gpt_neox_layers_10_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[130] gpt_neox_layers_10_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[131] gpt_neox_layers_10_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[132] gpt_neox_layers_11_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[133] gpt_neox_layers_11_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[134] gpt_neox_layers_11_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[135] gpt_neox_layers_11_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[136] gpt_neox_layers_11_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[137] gpt_neox_layers_11_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[138] gpt_neox_layers_11_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[139] gpt_neox_layers_11_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[140] gpt_neox_layers_11_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[141] gpt_neox_layers_11_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[142] gpt_neox_layers_11_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[143] gpt_neox_layers_11_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[144] gpt_neox_layers_12_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[145] gpt_neox_layers_12_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[146] gpt_neox_layers_12_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[147] gpt_neox_layers_12_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[148] gpt_neox_layers_12_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[149] gpt_neox_layers_12_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[150] gpt_neox_layers_12_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[151] gpt_neox_layers_12_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[152] gpt_neox_layers_12_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[153] gpt_neox_layers_12_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[154] gpt_neox_layers_12_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[155] gpt_neox_layers_12_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[156] gpt_neox_layers_13_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[157] gpt_neox_layers_13_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[158] gpt_neox_layers_13_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[159] gpt_neox_layers_13_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[160] gpt_neox_layers_13_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[161] gpt_neox_layers_13_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[162] gpt_neox_layers_13_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[163] gpt_neox_layers_13_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[164] gpt_neox_layers_13_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[165] gpt_neox_layers_13_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[166] gpt_neox_layers_13_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[167] gpt_neox_layers_13_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[168] gpt_neox_layers_14_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[169] gpt_neox_layers_14_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[170] gpt_neox_layers_14_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[171] gpt_neox_layers_14_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[172] gpt_neox_layers_14_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[173] gpt_neox_layers_14_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[174] gpt_neox_layers_14_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[175] gpt_neox_layers_14_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[176] gpt_neox_layers_14_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[177] gpt_neox_layers_14_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[178] gpt_neox_layers_14_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[179] gpt_neox_layers_14_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[180] gpt_neox_layers_15_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[181] gpt_neox_layers_15_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[182] gpt_neox_layers_15_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[183] gpt_neox_layers_15_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[184] gpt_neox_layers_15_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[185] gpt_neox_layers_15_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[186] gpt_neox_layers_15_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[187] gpt_neox_layers_15_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[188] gpt_neox_layers_15_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[189] gpt_neox_layers_15_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[190] gpt_neox_layers_15_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[191] gpt_neox_layers_15_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[192] gpt_neox_final_layer_norm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[193] gpt_neox_final_layer_norm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[194] embed_out_weight4: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[195] layer_norm99: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(input_embeds, gpt_neox_layers_0_input_layernorm_weight4, gpt_neox_layers_0_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv = R.call_tir(cls.NT_matmul, (layer_norm99, gpt_neox_layers_0_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add288: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv, gpt_neox_layers_0_attention_query_key_value_bias4) reshape192: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add288, R.shape([batch_size, 1, 24, 256])) reshape193: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape192, R.shape([batch_size, 24, 256])) lv49 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), reshape193), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape194: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv49, R.shape([batch_size, 1, 8, 256])) reshape195: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape194, R.shape([batch_size, 1, 2048])) lv1 = R.call_tir(cls.NT_matmul1, (reshape195, gpt_neox_layers_0_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add289: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv1, gpt_neox_layers_0_attention_dense_bias4) layer_norm100: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(input_embeds, gpt_neox_layers_0_post_attention_layernorm_weight4, gpt_neox_layers_0_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv2 = R.call_tir(cls.NT_matmul2, (layer_norm100, gpt_neox_layers_0_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add290: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv2, gpt_neox_layers_0_mlp_dense_h_to_4h_bias4) gelu48: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add290) astype99: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu48, dtype="float16") lv3 = R.call_tir(cls.NT_matmul3, (astype99, gpt_neox_layers_0_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add291: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv3, gpt_neox_layers_0_mlp_dense_4h_to_h_bias4) astype100: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add291, dtype="float16") add292: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype100, add289) add293: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add292, input_embeds) layer_norm101: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add293, gpt_neox_layers_1_input_layernorm_weight4, gpt_neox_layers_1_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv4 = R.call_tir(cls.NT_matmul, (layer_norm101, gpt_neox_layers_1_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add294: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv4, gpt_neox_layers_1_attention_query_key_value_bias4) reshape196: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add294, R.shape([batch_size, 1, 24, 256])) reshape197: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape196, R.shape([batch_size, 24, 256])) lv50 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), reshape197), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape198: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv50, R.shape([batch_size, 1, 8, 256])) reshape199: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape198, R.shape([batch_size, 1, 2048])) lv5 = R.call_tir(cls.NT_matmul1, (reshape199, gpt_neox_layers_1_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add295: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv5, gpt_neox_layers_1_attention_dense_bias4) layer_norm102: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add293, gpt_neox_layers_1_post_attention_layernorm_weight4, gpt_neox_layers_1_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv6 = R.call_tir(cls.NT_matmul2, (layer_norm102, gpt_neox_layers_1_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add296: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv6, gpt_neox_layers_1_mlp_dense_h_to_4h_bias4) gelu49: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add296) astype101: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu49, dtype="float16") lv7 = R.call_tir(cls.NT_matmul3, (astype101, gpt_neox_layers_1_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add297: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv7, gpt_neox_layers_1_mlp_dense_4h_to_h_bias4) astype102: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add297, dtype="float16") add298: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype102, add295) add299: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add298, add293) layer_norm103: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add299, gpt_neox_layers_2_input_layernorm_weight4, gpt_neox_layers_2_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv8 = R.call_tir(cls.NT_matmul, (layer_norm103, gpt_neox_layers_2_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add300: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv8, gpt_neox_layers_2_attention_query_key_value_bias4) reshape200: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add300, R.shape([batch_size, 1, 24, 256])) reshape201: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape200, R.shape([batch_size, 24, 256])) lv51 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), reshape201), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape202: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv51, R.shape([batch_size, 1, 8, 256])) reshape203: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape202, R.shape([batch_size, 1, 2048])) lv9 = R.call_tir(cls.NT_matmul1, (reshape203, gpt_neox_layers_2_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add301: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv9, gpt_neox_layers_2_attention_dense_bias4) layer_norm104: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add299, gpt_neox_layers_2_post_attention_layernorm_weight4, gpt_neox_layers_2_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv10 = R.call_tir(cls.NT_matmul2, (layer_norm104, gpt_neox_layers_2_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add302: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv10, gpt_neox_layers_2_mlp_dense_h_to_4h_bias4) gelu50: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add302) astype103: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu50, dtype="float16") lv11 = R.call_tir(cls.NT_matmul3, (astype103, gpt_neox_layers_2_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add303: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv11, gpt_neox_layers_2_mlp_dense_4h_to_h_bias4) astype104: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add303, dtype="float16") add304: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype104, add301) add305: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add304, add299) layer_norm105: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add305, gpt_neox_layers_3_input_layernorm_weight4, gpt_neox_layers_3_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv12 = R.call_tir(cls.NT_matmul, (layer_norm105, gpt_neox_layers_3_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add306: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv12, gpt_neox_layers_3_attention_query_key_value_bias4) reshape204: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add306, R.shape([batch_size, 1, 24, 256])) reshape205: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape204, R.shape([batch_size, 24, 256])) lv52 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), reshape205), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape206: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv52, R.shape([batch_size, 1, 8, 256])) reshape207: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape206, R.shape([batch_size, 1, 2048])) lv13 = R.call_tir(cls.NT_matmul1, (reshape207, gpt_neox_layers_3_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add307: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv13, gpt_neox_layers_3_attention_dense_bias4) layer_norm106: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add305, gpt_neox_layers_3_post_attention_layernorm_weight4, gpt_neox_layers_3_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv14 = R.call_tir(cls.NT_matmul2, (layer_norm106, gpt_neox_layers_3_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add308: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv14, gpt_neox_layers_3_mlp_dense_h_to_4h_bias4) gelu51: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add308) astype105: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu51, dtype="float16") lv15 = R.call_tir(cls.NT_matmul3, (astype105, gpt_neox_layers_3_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add309: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv15, gpt_neox_layers_3_mlp_dense_4h_to_h_bias4) astype106: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add309, dtype="float16") add310: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype106, add307) add311: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add310, add305) layer_norm107: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add311, gpt_neox_layers_4_input_layernorm_weight4, gpt_neox_layers_4_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv16 = R.call_tir(cls.NT_matmul, (layer_norm107, gpt_neox_layers_4_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add312: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv16, gpt_neox_layers_4_attention_query_key_value_bias4) reshape208: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add312, R.shape([batch_size, 1, 24, 256])) reshape209: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape208, R.shape([batch_size, 24, 256])) lv53 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), reshape209), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape210: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv53, R.shape([batch_size, 1, 8, 256])) reshape211: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape210, R.shape([batch_size, 1, 2048])) lv17 = R.call_tir(cls.NT_matmul1, (reshape211, gpt_neox_layers_4_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add313: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv17, gpt_neox_layers_4_attention_dense_bias4) layer_norm108: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add311, gpt_neox_layers_4_post_attention_layernorm_weight4, gpt_neox_layers_4_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv18 = R.call_tir(cls.NT_matmul2, (layer_norm108, gpt_neox_layers_4_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add314: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv18, gpt_neox_layers_4_mlp_dense_h_to_4h_bias4) gelu52: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add314) astype107: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu52, dtype="float16") lv19 = R.call_tir(cls.NT_matmul3, (astype107, gpt_neox_layers_4_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add315: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv19, gpt_neox_layers_4_mlp_dense_4h_to_h_bias4) astype108: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add315, dtype="float16") add316: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype108, add313) add317: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add316, add311) layer_norm109: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add317, gpt_neox_layers_5_input_layernorm_weight4, gpt_neox_layers_5_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv20 = R.call_tir(cls.NT_matmul, (layer_norm109, gpt_neox_layers_5_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add318: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv20, gpt_neox_layers_5_attention_query_key_value_bias4) reshape212: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add318, R.shape([batch_size, 1, 24, 256])) reshape213: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape212, R.shape([batch_size, 24, 256])) lv54 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), reshape213), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape214: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv54, R.shape([batch_size, 1, 8, 256])) reshape215: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape214, R.shape([batch_size, 1, 2048])) lv21 = R.call_tir(cls.NT_matmul1, (reshape215, gpt_neox_layers_5_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add319: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv21, gpt_neox_layers_5_attention_dense_bias4) layer_norm110: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add317, gpt_neox_layers_5_post_attention_layernorm_weight4, gpt_neox_layers_5_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv22 = R.call_tir(cls.NT_matmul2, (layer_norm110, gpt_neox_layers_5_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add320: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv22, gpt_neox_layers_5_mlp_dense_h_to_4h_bias4) gelu53: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add320) astype109: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu53, dtype="float16") lv23 = R.call_tir(cls.NT_matmul3, (astype109, gpt_neox_layers_5_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add321: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv23, gpt_neox_layers_5_mlp_dense_4h_to_h_bias4) astype110: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add321, dtype="float16") add322: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype110, add319) add323: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add322, add317) layer_norm111: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add323, gpt_neox_layers_6_input_layernorm_weight4, gpt_neox_layers_6_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv24 = R.call_tir(cls.NT_matmul, (layer_norm111, gpt_neox_layers_6_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add324: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv24, gpt_neox_layers_6_attention_query_key_value_bias4) reshape216: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add324, R.shape([batch_size, 1, 24, 256])) reshape217: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape216, R.shape([batch_size, 24, 256])) lv55 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), reshape217), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape218: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv55, R.shape([batch_size, 1, 8, 256])) reshape219: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape218, R.shape([batch_size, 1, 2048])) lv25 = R.call_tir(cls.NT_matmul1, (reshape219, gpt_neox_layers_6_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add325: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv25, gpt_neox_layers_6_attention_dense_bias4) layer_norm112: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add323, gpt_neox_layers_6_post_attention_layernorm_weight4, gpt_neox_layers_6_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv26 = R.call_tir(cls.NT_matmul2, (layer_norm112, gpt_neox_layers_6_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add326: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv26, gpt_neox_layers_6_mlp_dense_h_to_4h_bias4) gelu54: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add326) astype111: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu54, dtype="float16") lv27 = R.call_tir(cls.NT_matmul3, (astype111, gpt_neox_layers_6_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add327: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv27, gpt_neox_layers_6_mlp_dense_4h_to_h_bias4) astype112: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add327, dtype="float16") add328: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype112, add325) add329: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add328, add323) layer_norm113: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add329, gpt_neox_layers_7_input_layernorm_weight4, gpt_neox_layers_7_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv28 = R.call_tir(cls.NT_matmul, (layer_norm113, gpt_neox_layers_7_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add330: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv28, gpt_neox_layers_7_attention_query_key_value_bias4) reshape220: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add330, R.shape([batch_size, 1, 24, 256])) reshape221: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape220, R.shape([batch_size, 24, 256])) lv56 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), reshape221), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape222: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv56, R.shape([batch_size, 1, 8, 256])) reshape223: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape222, R.shape([batch_size, 1, 2048])) lv29 = R.call_tir(cls.NT_matmul1, (reshape223, gpt_neox_layers_7_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add331: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv29, gpt_neox_layers_7_attention_dense_bias4) layer_norm114: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add329, gpt_neox_layers_7_post_attention_layernorm_weight4, gpt_neox_layers_7_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv30 = R.call_tir(cls.NT_matmul2, (layer_norm114, gpt_neox_layers_7_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add332: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv30, gpt_neox_layers_7_mlp_dense_h_to_4h_bias4) gelu55: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add332) astype113: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu55, dtype="float16") lv31 = R.call_tir(cls.NT_matmul3, (astype113, gpt_neox_layers_7_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add333: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv31, gpt_neox_layers_7_mlp_dense_4h_to_h_bias4) astype114: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add333, dtype="float16") add334: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype114, add331) add335: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add334, add329) layer_norm115: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add335, gpt_neox_layers_8_input_layernorm_weight4, gpt_neox_layers_8_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv32 = R.call_tir(cls.NT_matmul, (layer_norm115, gpt_neox_layers_8_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add336: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv32, gpt_neox_layers_8_attention_query_key_value_bias4) reshape224: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add336, R.shape([batch_size, 1, 24, 256])) reshape225: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape224, R.shape([batch_size, 24, 256])) lv57 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), reshape225), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape226: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv57, R.shape([batch_size, 1, 8, 256])) reshape227: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape226, R.shape([batch_size, 1, 2048])) lv33 = R.call_tir(cls.NT_matmul1, (reshape227, gpt_neox_layers_8_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add337: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv33, gpt_neox_layers_8_attention_dense_bias4) layer_norm116: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add335, gpt_neox_layers_8_post_attention_layernorm_weight4, gpt_neox_layers_8_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv34 = R.call_tir(cls.NT_matmul2, (layer_norm116, gpt_neox_layers_8_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add338: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv34, gpt_neox_layers_8_mlp_dense_h_to_4h_bias4) gelu56: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add338) astype115: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu56, dtype="float16") lv35 = R.call_tir(cls.NT_matmul3, (astype115, gpt_neox_layers_8_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add339: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv35, gpt_neox_layers_8_mlp_dense_4h_to_h_bias4) astype116: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add339, dtype="float16") add340: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype116, add337) add341: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add340, add335) layer_norm117: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add341, gpt_neox_layers_9_input_layernorm_weight4, gpt_neox_layers_9_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv36 = R.call_tir(cls.NT_matmul, (layer_norm117, gpt_neox_layers_9_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add342: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv36, gpt_neox_layers_9_attention_query_key_value_bias4) reshape228: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add342, R.shape([batch_size, 1, 24, 256])) reshape229: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape228, R.shape([batch_size, 24, 256])) lv58 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), reshape229), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape230: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv58, R.shape([batch_size, 1, 8, 256])) reshape231: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape230, R.shape([batch_size, 1, 2048])) lv37 = R.call_tir(cls.NT_matmul1, (reshape231, gpt_neox_layers_9_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add343: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv37, gpt_neox_layers_9_attention_dense_bias4) layer_norm118: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add341, gpt_neox_layers_9_post_attention_layernorm_weight4, gpt_neox_layers_9_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv38 = R.call_tir(cls.NT_matmul2, (layer_norm118, gpt_neox_layers_9_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add344: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv38, gpt_neox_layers_9_mlp_dense_h_to_4h_bias4) gelu57: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add344) astype117: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu57, dtype="float16") lv39 = R.call_tir(cls.NT_matmul3, (astype117, gpt_neox_layers_9_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add345: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv39, gpt_neox_layers_9_mlp_dense_4h_to_h_bias4) astype118: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add345, dtype="float16") add346: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype118, add343) add347: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add346, add341) layer_norm119: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add347, gpt_neox_layers_10_input_layernorm_weight4, gpt_neox_layers_10_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv40 = R.call_tir(cls.NT_matmul, (layer_norm119, gpt_neox_layers_10_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add348: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv40, gpt_neox_layers_10_attention_query_key_value_bias4) reshape232: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add348, R.shape([batch_size, 1, 24, 256])) reshape233: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape232, R.shape([batch_size, 24, 256])) lv59 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), reshape233), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape234: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv59, R.shape([batch_size, 1, 8, 256])) reshape235: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape234, R.shape([batch_size, 1, 2048])) lv41 = R.call_tir(cls.NT_matmul1, (reshape235, gpt_neox_layers_10_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add349: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv41, gpt_neox_layers_10_attention_dense_bias4) layer_norm120: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add347, gpt_neox_layers_10_post_attention_layernorm_weight4, gpt_neox_layers_10_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv42 = R.call_tir(cls.NT_matmul2, (layer_norm120, gpt_neox_layers_10_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add350: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv42, gpt_neox_layers_10_mlp_dense_h_to_4h_bias4) gelu58: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add350) astype119: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu58, dtype="float16") lv43 = R.call_tir(cls.NT_matmul3, (astype119, gpt_neox_layers_10_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add351: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv43, gpt_neox_layers_10_mlp_dense_4h_to_h_bias4) astype120: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add351, dtype="float16") add352: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype120, add349) add353: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add352, add347) layer_norm121: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add353, gpt_neox_layers_11_input_layernorm_weight4, gpt_neox_layers_11_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv44 = R.call_tir(cls.NT_matmul, (layer_norm121, gpt_neox_layers_11_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add354: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv44, gpt_neox_layers_11_attention_query_key_value_bias4) reshape236: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add354, R.shape([batch_size, 1, 24, 256])) reshape237: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape236, R.shape([batch_size, 24, 256])) lv60 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), reshape237), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape238: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv60, R.shape([batch_size, 1, 8, 256])) reshape239: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape238, R.shape([batch_size, 1, 2048])) lv45 = R.call_tir(cls.NT_matmul1, (reshape239, gpt_neox_layers_11_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add355: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv45, gpt_neox_layers_11_attention_dense_bias4) layer_norm122: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add353, gpt_neox_layers_11_post_attention_layernorm_weight4, gpt_neox_layers_11_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv46 = R.call_tir(cls.NT_matmul2, (layer_norm122, gpt_neox_layers_11_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add356: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv46, gpt_neox_layers_11_mlp_dense_h_to_4h_bias4) gelu59: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add356) astype121: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu59, dtype="float16") lv47 = R.call_tir(cls.NT_matmul3, (astype121, gpt_neox_layers_11_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add357: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv47, gpt_neox_layers_11_mlp_dense_4h_to_h_bias4) astype122: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add357, dtype="float16") add358: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype122, add355) add359: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add358, add353) layer_norm123: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add359, gpt_neox_layers_12_input_layernorm_weight4, gpt_neox_layers_12_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv48 = R.call_tir(cls.NT_matmul, (layer_norm123, gpt_neox_layers_12_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add360: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv48, gpt_neox_layers_12_attention_query_key_value_bias4) reshape240: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add360, R.shape([batch_size, 1, 24, 256])) reshape241: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape240, R.shape([batch_size, 24, 256])) lv61 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape241), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape242: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv61, R.shape([batch_size, 1, 8, 256])) reshape243: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape242, R.shape([batch_size, 1, 2048])) lv49_1 = R.call_tir(cls.NT_matmul1, (reshape243, gpt_neox_layers_12_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add361: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv49_1, gpt_neox_layers_12_attention_dense_bias4) layer_norm124: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add359, gpt_neox_layers_12_post_attention_layernorm_weight4, gpt_neox_layers_12_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv50_1 = R.call_tir(cls.NT_matmul2, (layer_norm124, gpt_neox_layers_12_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add362: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv50_1, gpt_neox_layers_12_mlp_dense_h_to_4h_bias4) gelu60: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add362) astype123: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu60, dtype="float16") lv51_1 = R.call_tir(cls.NT_matmul3, (astype123, gpt_neox_layers_12_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add363: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv51_1, gpt_neox_layers_12_mlp_dense_4h_to_h_bias4) astype124: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add363, dtype="float16") add364: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype124, add361) add365: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add364, add359) layer_norm125: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add365, gpt_neox_layers_13_input_layernorm_weight4, gpt_neox_layers_13_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv52_1 = R.call_tir(cls.NT_matmul, (layer_norm125, gpt_neox_layers_13_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add366: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv52_1, gpt_neox_layers_13_attention_query_key_value_bias4) reshape244: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add366, R.shape([batch_size, 1, 24, 256])) reshape245: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape244, R.shape([batch_size, 24, 256])) lv62 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape245), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape246: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv62, R.shape([batch_size, 1, 8, 256])) reshape247: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape246, R.shape([batch_size, 1, 2048])) lv53_1 = R.call_tir(cls.NT_matmul1, (reshape247, gpt_neox_layers_13_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add367: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv53_1, gpt_neox_layers_13_attention_dense_bias4) layer_norm126: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add365, gpt_neox_layers_13_post_attention_layernorm_weight4, gpt_neox_layers_13_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv54_1 = R.call_tir(cls.NT_matmul2, (layer_norm126, gpt_neox_layers_13_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add368: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv54_1, gpt_neox_layers_13_mlp_dense_h_to_4h_bias4) gelu61: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add368) astype125: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu61, dtype="float16") lv55_1 = R.call_tir(cls.NT_matmul3, (astype125, gpt_neox_layers_13_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add369: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv55_1, gpt_neox_layers_13_mlp_dense_4h_to_h_bias4) astype126: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add369, dtype="float16") add370: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype126, add367) add371: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add370, add365) layer_norm127: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add371, gpt_neox_layers_14_input_layernorm_weight4, gpt_neox_layers_14_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv56_1 = R.call_tir(cls.NT_matmul, (layer_norm127, gpt_neox_layers_14_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add372: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv56_1, gpt_neox_layers_14_attention_query_key_value_bias4) reshape248: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add372, R.shape([batch_size, 1, 24, 256])) reshape249: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape248, R.shape([batch_size, 24, 256])) lv63 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape249), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape250: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv63, R.shape([batch_size, 1, 8, 256])) reshape251: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape250, R.shape([batch_size, 1, 2048])) lv57_1 = R.call_tir(cls.NT_matmul1, (reshape251, gpt_neox_layers_14_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add373: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv57_1, gpt_neox_layers_14_attention_dense_bias4) layer_norm128: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add371, gpt_neox_layers_14_post_attention_layernorm_weight4, gpt_neox_layers_14_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv58_1 = R.call_tir(cls.NT_matmul2, (layer_norm128, gpt_neox_layers_14_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add374: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv58_1, gpt_neox_layers_14_mlp_dense_h_to_4h_bias4) gelu62: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add374) astype127: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu62, dtype="float16") lv59_1 = R.call_tir(cls.NT_matmul3, (astype127, gpt_neox_layers_14_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add375: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv59_1, gpt_neox_layers_14_mlp_dense_4h_to_h_bias4) astype128: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add375, dtype="float16") add376: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype128, add373) add377: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add376, add371) layer_norm129: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add377, gpt_neox_layers_15_input_layernorm_weight4, gpt_neox_layers_15_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv60_1 = R.call_tir(cls.NT_matmul, (layer_norm129, gpt_neox_layers_15_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) add378: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv60_1, gpt_neox_layers_15_attention_query_key_value_bias4) reshape252: R.Tensor((batch_size, 1, 24, 256), dtype="float16") = R.reshape(add378, R.shape([batch_size, 1, 24, 256])) reshape253: R.Tensor((batch_size, 24, 256), dtype="float16") = R.reshape(reshape252, R.shape([batch_size, 24, 256])) lv64 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape253), out_sinfo=R.Tensor((batch_size, 8, 256), dtype="float16")) reshape254: R.Tensor((batch_size, 1, 8, 256), dtype="float16") = R.reshape(lv64, R.shape([batch_size, 1, 8, 256])) reshape255: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape254, R.shape([batch_size, 1, 2048])) lv61_1 = R.call_tir(cls.NT_matmul1, (reshape255, gpt_neox_layers_15_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) add379: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv61_1, gpt_neox_layers_15_attention_dense_bias4) layer_norm130: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add377, gpt_neox_layers_15_post_attention_layernorm_weight4, gpt_neox_layers_15_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv62_1 = R.call_tir(cls.NT_matmul2, (layer_norm130, gpt_neox_layers_15_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) add380: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv62_1, gpt_neox_layers_15_mlp_dense_h_to_4h_bias4) gelu63: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add380) astype129: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu63, dtype="float16") lv63_1 = R.call_tir(cls.NT_matmul3, (astype129, gpt_neox_layers_15_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) add381: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv63_1, gpt_neox_layers_15_mlp_dense_4h_to_h_bias4) astype130: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add381, dtype="float16") add382: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype130, add379) add383: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add382, add377) layer_norm131: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add383, gpt_neox_final_layer_norm_weight4, gpt_neox_final_layer_norm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv64_1 = R.call_tir(cls.NT_matmul4, (layer_norm131, embed_out_weight4), out_sinfo=R.Tensor((batch_size, 1, vocab_size), dtype="float16")) astype131: R.Tensor((batch_size, 1, vocab_size), dtype="float32") = R.astype(lv64_1, dtype="float32") gv4: R.Tuple(R.Tensor((batch_size, 1, vocab_size), dtype="float32"), R.Object) = astype131, paged_kv_cache R.output(gv4) return gv4 @R.function def batch_prefill(input_embeds: R.Tensor((1, "seq_len", 2048), dtype="float16"), logit_positions: R.Tensor(("batch_size",), dtype="int32"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 2048), dtype="float16"))) -> R.Tuple(R.Tensor((1, "batch_size", "vocab_size"), dtype="float32"), R.Object): batch_size = T.int64() vocab_size = T.int64() seq_len = T.int64() R.func_attr({"num_input": 3, "pipeline_parallel_stages": 1, "relax.memory_plan_dynamic_func_output": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) cls = Module with R.dataflow(): gpt_neox_layers_0_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[1] gpt_neox_layers_0_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[2] gpt_neox_layers_0_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[3] gpt_neox_layers_0_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[4] gpt_neox_layers_0_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[5] gpt_neox_layers_0_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[6] gpt_neox_layers_0_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[8] gpt_neox_layers_0_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[9] gpt_neox_layers_0_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[10] gpt_neox_layers_0_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[11] gpt_neox_layers_0_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[12] gpt_neox_layers_1_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[13] gpt_neox_layers_1_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[14] gpt_neox_layers_1_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[15] gpt_neox_layers_1_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[16] gpt_neox_layers_1_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[17] gpt_neox_layers_1_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[18] gpt_neox_layers_1_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[19] gpt_neox_layers_1_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[20] gpt_neox_layers_1_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[21] gpt_neox_layers_1_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[22] gpt_neox_layers_1_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[23] gpt_neox_layers_1_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[24] gpt_neox_layers_2_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[25] gpt_neox_layers_2_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[26] gpt_neox_layers_2_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[27] gpt_neox_layers_2_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[28] gpt_neox_layers_2_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[29] gpt_neox_layers_2_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[30] gpt_neox_layers_2_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[31] gpt_neox_layers_2_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[32] gpt_neox_layers_2_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[33] gpt_neox_layers_2_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[34] gpt_neox_layers_2_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[35] gpt_neox_layers_2_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[36] gpt_neox_layers_3_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[37] gpt_neox_layers_3_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[38] gpt_neox_layers_3_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[39] gpt_neox_layers_3_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[40] gpt_neox_layers_3_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[41] gpt_neox_layers_3_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[42] gpt_neox_layers_3_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[43] gpt_neox_layers_3_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[44] gpt_neox_layers_3_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[45] gpt_neox_layers_3_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[46] gpt_neox_layers_3_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[47] gpt_neox_layers_3_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[48] gpt_neox_layers_4_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[49] gpt_neox_layers_4_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[50] gpt_neox_layers_4_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[51] gpt_neox_layers_4_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[52] gpt_neox_layers_4_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[53] gpt_neox_layers_4_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[54] gpt_neox_layers_4_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[55] gpt_neox_layers_4_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[56] gpt_neox_layers_4_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[57] gpt_neox_layers_4_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[58] gpt_neox_layers_4_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[59] gpt_neox_layers_4_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[60] gpt_neox_layers_5_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[61] gpt_neox_layers_5_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[62] gpt_neox_layers_5_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[63] gpt_neox_layers_5_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[64] gpt_neox_layers_5_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[65] gpt_neox_layers_5_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[66] gpt_neox_layers_5_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[67] gpt_neox_layers_5_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[68] gpt_neox_layers_5_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[69] gpt_neox_layers_5_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[70] gpt_neox_layers_5_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[71] gpt_neox_layers_5_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[72] gpt_neox_layers_6_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[73] gpt_neox_layers_6_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[74] gpt_neox_layers_6_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[75] gpt_neox_layers_6_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[76] gpt_neox_layers_6_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[77] gpt_neox_layers_6_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[78] gpt_neox_layers_6_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[79] gpt_neox_layers_6_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[80] gpt_neox_layers_6_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[81] gpt_neox_layers_6_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[82] gpt_neox_layers_6_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[83] gpt_neox_layers_6_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[84] gpt_neox_layers_7_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[85] gpt_neox_layers_7_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[86] gpt_neox_layers_7_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[87] gpt_neox_layers_7_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[88] gpt_neox_layers_7_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[89] gpt_neox_layers_7_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[90] gpt_neox_layers_7_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[91] gpt_neox_layers_7_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[92] gpt_neox_layers_7_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[93] gpt_neox_layers_7_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[94] gpt_neox_layers_7_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[95] gpt_neox_layers_7_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[96] gpt_neox_layers_8_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[97] gpt_neox_layers_8_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[98] gpt_neox_layers_8_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[99] gpt_neox_layers_8_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[100] gpt_neox_layers_8_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[101] gpt_neox_layers_8_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[102] gpt_neox_layers_8_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[103] gpt_neox_layers_8_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[104] gpt_neox_layers_8_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[105] gpt_neox_layers_8_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[106] gpt_neox_layers_8_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[107] gpt_neox_layers_8_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[108] gpt_neox_layers_9_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[109] gpt_neox_layers_9_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[110] gpt_neox_layers_9_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[111] gpt_neox_layers_9_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[112] gpt_neox_layers_9_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[113] gpt_neox_layers_9_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[114] gpt_neox_layers_9_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[115] gpt_neox_layers_9_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[116] gpt_neox_layers_9_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[117] gpt_neox_layers_9_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[118] gpt_neox_layers_9_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[119] gpt_neox_layers_9_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[120] gpt_neox_layers_10_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[121] gpt_neox_layers_10_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[122] gpt_neox_layers_10_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[123] gpt_neox_layers_10_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[124] gpt_neox_layers_10_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[125] gpt_neox_layers_10_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[126] gpt_neox_layers_10_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[127] gpt_neox_layers_10_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[128] gpt_neox_layers_10_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[129] gpt_neox_layers_10_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[130] gpt_neox_layers_10_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[131] gpt_neox_layers_10_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[132] gpt_neox_layers_11_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[133] gpt_neox_layers_11_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[134] gpt_neox_layers_11_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[135] gpt_neox_layers_11_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[136] gpt_neox_layers_11_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[137] gpt_neox_layers_11_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[138] gpt_neox_layers_11_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[139] gpt_neox_layers_11_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[140] gpt_neox_layers_11_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[141] gpt_neox_layers_11_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[142] gpt_neox_layers_11_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[143] gpt_neox_layers_11_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[144] gpt_neox_layers_12_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[145] gpt_neox_layers_12_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[146] gpt_neox_layers_12_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[147] gpt_neox_layers_12_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[148] gpt_neox_layers_12_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[149] gpt_neox_layers_12_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[150] gpt_neox_layers_12_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[151] gpt_neox_layers_12_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[152] gpt_neox_layers_12_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[153] gpt_neox_layers_12_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[154] gpt_neox_layers_12_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[155] gpt_neox_layers_12_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[156] gpt_neox_layers_13_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[157] gpt_neox_layers_13_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[158] gpt_neox_layers_13_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[159] gpt_neox_layers_13_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[160] gpt_neox_layers_13_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[161] gpt_neox_layers_13_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[162] gpt_neox_layers_13_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[163] gpt_neox_layers_13_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[164] gpt_neox_layers_13_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[165] gpt_neox_layers_13_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[166] gpt_neox_layers_13_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[167] gpt_neox_layers_13_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[168] gpt_neox_layers_14_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[169] gpt_neox_layers_14_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[170] gpt_neox_layers_14_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[171] gpt_neox_layers_14_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[172] gpt_neox_layers_14_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[173] gpt_neox_layers_14_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[174] gpt_neox_layers_14_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[175] gpt_neox_layers_14_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[176] gpt_neox_layers_14_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[177] gpt_neox_layers_14_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[178] gpt_neox_layers_14_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[179] gpt_neox_layers_14_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[180] gpt_neox_layers_15_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[181] gpt_neox_layers_15_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[182] gpt_neox_layers_15_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[183] gpt_neox_layers_15_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[184] gpt_neox_layers_15_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[185] gpt_neox_layers_15_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[186] gpt_neox_layers_15_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[187] gpt_neox_layers_15_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[188] gpt_neox_layers_15_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[189] gpt_neox_layers_15_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[190] gpt_neox_layers_15_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[191] gpt_neox_layers_15_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[192] gpt_neox_final_layer_norm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[193] gpt_neox_final_layer_norm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[194] embed_out_weight3: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[195] layer_norm66: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(input_embeds, gpt_neox_layers_0_input_layernorm_weight3, gpt_neox_layers_0_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv65 = R.call_tir(cls.NT_matmul5, (layer_norm66, gpt_neox_layers_0_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add192: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv65, gpt_neox_layers_0_attention_query_key_value_bias3) reshape128: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add192, R.shape([1, seq_len, 24, 256])) reshape129: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape128, R.shape([seq_len, 24, 256])) lv33 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), reshape129), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape130: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv33, R.shape([1, seq_len, 8, 256])) reshape131: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape130, R.shape([1, seq_len, 2048])) lv66 = R.call_tir(cls.NT_matmul6, (reshape131, gpt_neox_layers_0_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add193: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv66, gpt_neox_layers_0_attention_dense_bias3) layer_norm67: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(input_embeds, gpt_neox_layers_0_post_attention_layernorm_weight3, gpt_neox_layers_0_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv67 = R.call_tir(cls.NT_matmul7, (layer_norm67, gpt_neox_layers_0_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add194: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv67, gpt_neox_layers_0_mlp_dense_h_to_4h_bias3) gelu32: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add194) astype66: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu32, dtype="float16") lv68 = R.call_tir(cls.NT_matmul8, (astype66, gpt_neox_layers_0_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add195: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv68, gpt_neox_layers_0_mlp_dense_4h_to_h_bias3) astype67: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add195, dtype="float16") add196: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype67, add193) add197: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add196, input_embeds) layer_norm68: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add197, gpt_neox_layers_1_input_layernorm_weight3, gpt_neox_layers_1_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv69 = R.call_tir(cls.NT_matmul5, (layer_norm68, gpt_neox_layers_1_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add198: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv69, gpt_neox_layers_1_attention_query_key_value_bias3) reshape132: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add198, R.shape([1, seq_len, 24, 256])) reshape133: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape132, R.shape([seq_len, 24, 256])) lv34 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), reshape133), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape134: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv34, R.shape([1, seq_len, 8, 256])) reshape135: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape134, R.shape([1, seq_len, 2048])) lv70 = R.call_tir(cls.NT_matmul6, (reshape135, gpt_neox_layers_1_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add199: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv70, gpt_neox_layers_1_attention_dense_bias3) layer_norm69: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add197, gpt_neox_layers_1_post_attention_layernorm_weight3, gpt_neox_layers_1_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv71 = R.call_tir(cls.NT_matmul7, (layer_norm69, gpt_neox_layers_1_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add200: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv71, gpt_neox_layers_1_mlp_dense_h_to_4h_bias3) gelu33: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add200) astype68: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu33, dtype="float16") lv72 = R.call_tir(cls.NT_matmul8, (astype68, gpt_neox_layers_1_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add201: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv72, gpt_neox_layers_1_mlp_dense_4h_to_h_bias3) astype69: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add201, dtype="float16") add202: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype69, add199) add203: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add202, add197) layer_norm70: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add203, gpt_neox_layers_2_input_layernorm_weight3, gpt_neox_layers_2_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv73 = R.call_tir(cls.NT_matmul5, (layer_norm70, gpt_neox_layers_2_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add204: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv73, gpt_neox_layers_2_attention_query_key_value_bias3) reshape136: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add204, R.shape([1, seq_len, 24, 256])) reshape137: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape136, R.shape([seq_len, 24, 256])) lv35 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), reshape137), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape138: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv35, R.shape([1, seq_len, 8, 256])) reshape139: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape138, R.shape([1, seq_len, 2048])) lv74 = R.call_tir(cls.NT_matmul6, (reshape139, gpt_neox_layers_2_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add205: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv74, gpt_neox_layers_2_attention_dense_bias3) layer_norm71: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add203, gpt_neox_layers_2_post_attention_layernorm_weight3, gpt_neox_layers_2_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv75 = R.call_tir(cls.NT_matmul7, (layer_norm71, gpt_neox_layers_2_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add206: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv75, gpt_neox_layers_2_mlp_dense_h_to_4h_bias3) gelu34: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add206) astype70: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu34, dtype="float16") lv76 = R.call_tir(cls.NT_matmul8, (astype70, gpt_neox_layers_2_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add207: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv76, gpt_neox_layers_2_mlp_dense_4h_to_h_bias3) astype71: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add207, dtype="float16") add208: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype71, add205) add209: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add208, add203) layer_norm72: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add209, gpt_neox_layers_3_input_layernorm_weight3, gpt_neox_layers_3_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv77 = R.call_tir(cls.NT_matmul5, (layer_norm72, gpt_neox_layers_3_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add210: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv77, gpt_neox_layers_3_attention_query_key_value_bias3) reshape140: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add210, R.shape([1, seq_len, 24, 256])) reshape141: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape140, R.shape([seq_len, 24, 256])) lv36 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), reshape141), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape142: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv36, R.shape([1, seq_len, 8, 256])) reshape143: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape142, R.shape([1, seq_len, 2048])) lv78 = R.call_tir(cls.NT_matmul6, (reshape143, gpt_neox_layers_3_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add211: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv78, gpt_neox_layers_3_attention_dense_bias3) layer_norm73: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add209, gpt_neox_layers_3_post_attention_layernorm_weight3, gpt_neox_layers_3_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv79 = R.call_tir(cls.NT_matmul7, (layer_norm73, gpt_neox_layers_3_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add212: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv79, gpt_neox_layers_3_mlp_dense_h_to_4h_bias3) gelu35: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add212) astype72: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu35, dtype="float16") lv80 = R.call_tir(cls.NT_matmul8, (astype72, gpt_neox_layers_3_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add213: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv80, gpt_neox_layers_3_mlp_dense_4h_to_h_bias3) astype73: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add213, dtype="float16") add214: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype73, add211) add215: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add214, add209) layer_norm74: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add215, gpt_neox_layers_4_input_layernorm_weight3, gpt_neox_layers_4_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv81 = R.call_tir(cls.NT_matmul5, (layer_norm74, gpt_neox_layers_4_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add216: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv81, gpt_neox_layers_4_attention_query_key_value_bias3) reshape144: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add216, R.shape([1, seq_len, 24, 256])) reshape145: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape144, R.shape([seq_len, 24, 256])) lv37 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), reshape145), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape146: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv37, R.shape([1, seq_len, 8, 256])) reshape147: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape146, R.shape([1, seq_len, 2048])) lv82 = R.call_tir(cls.NT_matmul6, (reshape147, gpt_neox_layers_4_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add217: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv82, gpt_neox_layers_4_attention_dense_bias3) layer_norm75: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add215, gpt_neox_layers_4_post_attention_layernorm_weight3, gpt_neox_layers_4_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv83 = R.call_tir(cls.NT_matmul7, (layer_norm75, gpt_neox_layers_4_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add218: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv83, gpt_neox_layers_4_mlp_dense_h_to_4h_bias3) gelu36: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add218) astype74: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu36, dtype="float16") lv84 = R.call_tir(cls.NT_matmul8, (astype74, gpt_neox_layers_4_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add219: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv84, gpt_neox_layers_4_mlp_dense_4h_to_h_bias3) astype75: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add219, dtype="float16") add220: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype75, add217) add221: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add220, add215) layer_norm76: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add221, gpt_neox_layers_5_input_layernorm_weight3, gpt_neox_layers_5_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv85 = R.call_tir(cls.NT_matmul5, (layer_norm76, gpt_neox_layers_5_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add222: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv85, gpt_neox_layers_5_attention_query_key_value_bias3) reshape148: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add222, R.shape([1, seq_len, 24, 256])) reshape149: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape148, R.shape([seq_len, 24, 256])) lv38 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), reshape149), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape150: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv38, R.shape([1, seq_len, 8, 256])) reshape151: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape150, R.shape([1, seq_len, 2048])) lv86 = R.call_tir(cls.NT_matmul6, (reshape151, gpt_neox_layers_5_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add223: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv86, gpt_neox_layers_5_attention_dense_bias3) layer_norm77: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add221, gpt_neox_layers_5_post_attention_layernorm_weight3, gpt_neox_layers_5_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv87 = R.call_tir(cls.NT_matmul7, (layer_norm77, gpt_neox_layers_5_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add224: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv87, gpt_neox_layers_5_mlp_dense_h_to_4h_bias3) gelu37: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add224) astype76: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu37, dtype="float16") lv88 = R.call_tir(cls.NT_matmul8, (astype76, gpt_neox_layers_5_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add225: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv88, gpt_neox_layers_5_mlp_dense_4h_to_h_bias3) astype77: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add225, dtype="float16") add226: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype77, add223) add227: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add226, add221) layer_norm78: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add227, gpt_neox_layers_6_input_layernorm_weight3, gpt_neox_layers_6_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv89 = R.call_tir(cls.NT_matmul5, (layer_norm78, gpt_neox_layers_6_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add228: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv89, gpt_neox_layers_6_attention_query_key_value_bias3) reshape152: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add228, R.shape([1, seq_len, 24, 256])) reshape153: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape152, R.shape([seq_len, 24, 256])) lv39 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), reshape153), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape154: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv39, R.shape([1, seq_len, 8, 256])) reshape155: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape154, R.shape([1, seq_len, 2048])) lv90 = R.call_tir(cls.NT_matmul6, (reshape155, gpt_neox_layers_6_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add229: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv90, gpt_neox_layers_6_attention_dense_bias3) layer_norm79: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add227, gpt_neox_layers_6_post_attention_layernorm_weight3, gpt_neox_layers_6_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv91 = R.call_tir(cls.NT_matmul7, (layer_norm79, gpt_neox_layers_6_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add230: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv91, gpt_neox_layers_6_mlp_dense_h_to_4h_bias3) gelu38: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add230) astype78: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu38, dtype="float16") lv92 = R.call_tir(cls.NT_matmul8, (astype78, gpt_neox_layers_6_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add231: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv92, gpt_neox_layers_6_mlp_dense_4h_to_h_bias3) astype79: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add231, dtype="float16") add232: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype79, add229) add233: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add232, add227) layer_norm80: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add233, gpt_neox_layers_7_input_layernorm_weight3, gpt_neox_layers_7_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv93 = R.call_tir(cls.NT_matmul5, (layer_norm80, gpt_neox_layers_7_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add234: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv93, gpt_neox_layers_7_attention_query_key_value_bias3) reshape156: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add234, R.shape([1, seq_len, 24, 256])) reshape157: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape156, R.shape([seq_len, 24, 256])) lv40 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), reshape157), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape158: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv40, R.shape([1, seq_len, 8, 256])) reshape159: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape158, R.shape([1, seq_len, 2048])) lv94 = R.call_tir(cls.NT_matmul6, (reshape159, gpt_neox_layers_7_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add235: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv94, gpt_neox_layers_7_attention_dense_bias3) layer_norm81: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add233, gpt_neox_layers_7_post_attention_layernorm_weight3, gpt_neox_layers_7_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv95 = R.call_tir(cls.NT_matmul7, (layer_norm81, gpt_neox_layers_7_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add236: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv95, gpt_neox_layers_7_mlp_dense_h_to_4h_bias3) gelu39: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add236) astype80: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu39, dtype="float16") lv96 = R.call_tir(cls.NT_matmul8, (astype80, gpt_neox_layers_7_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add237: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv96, gpt_neox_layers_7_mlp_dense_4h_to_h_bias3) astype81: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add237, dtype="float16") add238: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype81, add235) add239: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add238, add233) layer_norm82: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add239, gpt_neox_layers_8_input_layernorm_weight3, gpt_neox_layers_8_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv97 = R.call_tir(cls.NT_matmul5, (layer_norm82, gpt_neox_layers_8_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add240: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv97, gpt_neox_layers_8_attention_query_key_value_bias3) reshape160: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add240, R.shape([1, seq_len, 24, 256])) reshape161: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape160, R.shape([seq_len, 24, 256])) lv41 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), reshape161), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape162: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv41, R.shape([1, seq_len, 8, 256])) reshape163: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape162, R.shape([1, seq_len, 2048])) lv98 = R.call_tir(cls.NT_matmul6, (reshape163, gpt_neox_layers_8_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add241: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv98, gpt_neox_layers_8_attention_dense_bias3) layer_norm83: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add239, gpt_neox_layers_8_post_attention_layernorm_weight3, gpt_neox_layers_8_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv99 = R.call_tir(cls.NT_matmul7, (layer_norm83, gpt_neox_layers_8_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add242: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv99, gpt_neox_layers_8_mlp_dense_h_to_4h_bias3) gelu40: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add242) astype82: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu40, dtype="float16") lv100 = R.call_tir(cls.NT_matmul8, (astype82, gpt_neox_layers_8_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add243: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv100, gpt_neox_layers_8_mlp_dense_4h_to_h_bias3) astype83: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add243, dtype="float16") add244: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype83, add241) add245: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add244, add239) layer_norm84: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add245, gpt_neox_layers_9_input_layernorm_weight3, gpt_neox_layers_9_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv101 = R.call_tir(cls.NT_matmul5, (layer_norm84, gpt_neox_layers_9_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add246: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv101, gpt_neox_layers_9_attention_query_key_value_bias3) reshape164: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add246, R.shape([1, seq_len, 24, 256])) reshape165: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape164, R.shape([seq_len, 24, 256])) lv42 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), reshape165), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape166: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv42, R.shape([1, seq_len, 8, 256])) reshape167: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape166, R.shape([1, seq_len, 2048])) lv102 = R.call_tir(cls.NT_matmul6, (reshape167, gpt_neox_layers_9_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add247: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv102, gpt_neox_layers_9_attention_dense_bias3) layer_norm85: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add245, gpt_neox_layers_9_post_attention_layernorm_weight3, gpt_neox_layers_9_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv103 = R.call_tir(cls.NT_matmul7, (layer_norm85, gpt_neox_layers_9_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add248: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv103, gpt_neox_layers_9_mlp_dense_h_to_4h_bias3) gelu41: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add248) astype84: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu41, dtype="float16") lv104 = R.call_tir(cls.NT_matmul8, (astype84, gpt_neox_layers_9_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add249: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv104, gpt_neox_layers_9_mlp_dense_4h_to_h_bias3) astype85: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add249, dtype="float16") add250: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype85, add247) add251: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add250, add245) layer_norm86: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add251, gpt_neox_layers_10_input_layernorm_weight3, gpt_neox_layers_10_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv105 = R.call_tir(cls.NT_matmul5, (layer_norm86, gpt_neox_layers_10_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add252: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv105, gpt_neox_layers_10_attention_query_key_value_bias3) reshape168: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add252, R.shape([1, seq_len, 24, 256])) reshape169: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape168, R.shape([seq_len, 24, 256])) lv43 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), reshape169), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape170: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv43, R.shape([1, seq_len, 8, 256])) reshape171: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape170, R.shape([1, seq_len, 2048])) lv106 = R.call_tir(cls.NT_matmul6, (reshape171, gpt_neox_layers_10_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add253: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv106, gpt_neox_layers_10_attention_dense_bias3) layer_norm87: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add251, gpt_neox_layers_10_post_attention_layernorm_weight3, gpt_neox_layers_10_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv107 = R.call_tir(cls.NT_matmul7, (layer_norm87, gpt_neox_layers_10_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add254: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv107, gpt_neox_layers_10_mlp_dense_h_to_4h_bias3) gelu42: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add254) astype86: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu42, dtype="float16") lv108 = R.call_tir(cls.NT_matmul8, (astype86, gpt_neox_layers_10_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add255: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv108, gpt_neox_layers_10_mlp_dense_4h_to_h_bias3) astype87: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add255, dtype="float16") add256: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype87, add253) add257: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add256, add251) layer_norm88: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add257, gpt_neox_layers_11_input_layernorm_weight3, gpt_neox_layers_11_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv109 = R.call_tir(cls.NT_matmul5, (layer_norm88, gpt_neox_layers_11_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add258: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv109, gpt_neox_layers_11_attention_query_key_value_bias3) reshape172: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add258, R.shape([1, seq_len, 24, 256])) reshape173: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape172, R.shape([seq_len, 24, 256])) lv44 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), reshape173), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape174: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv44, R.shape([1, seq_len, 8, 256])) reshape175: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape174, R.shape([1, seq_len, 2048])) lv110 = R.call_tir(cls.NT_matmul6, (reshape175, gpt_neox_layers_11_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add259: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv110, gpt_neox_layers_11_attention_dense_bias3) layer_norm89: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add257, gpt_neox_layers_11_post_attention_layernorm_weight3, gpt_neox_layers_11_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv111 = R.call_tir(cls.NT_matmul7, (layer_norm89, gpt_neox_layers_11_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add260: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv111, gpt_neox_layers_11_mlp_dense_h_to_4h_bias3) gelu43: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add260) astype88: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu43, dtype="float16") lv112 = R.call_tir(cls.NT_matmul8, (astype88, gpt_neox_layers_11_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add261: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv112, gpt_neox_layers_11_mlp_dense_4h_to_h_bias3) astype89: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add261, dtype="float16") add262: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype89, add259) add263: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add262, add257) layer_norm90: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add263, gpt_neox_layers_12_input_layernorm_weight3, gpt_neox_layers_12_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv113 = R.call_tir(cls.NT_matmul5, (layer_norm90, gpt_neox_layers_12_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add264: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv113, gpt_neox_layers_12_attention_query_key_value_bias3) reshape176: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add264, R.shape([1, seq_len, 24, 256])) reshape177: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape176, R.shape([seq_len, 24, 256])) lv45 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape177), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape178: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv45, R.shape([1, seq_len, 8, 256])) reshape179: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape178, R.shape([1, seq_len, 2048])) lv114 = R.call_tir(cls.NT_matmul6, (reshape179, gpt_neox_layers_12_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add265: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv114, gpt_neox_layers_12_attention_dense_bias3) layer_norm91: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add263, gpt_neox_layers_12_post_attention_layernorm_weight3, gpt_neox_layers_12_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv115 = R.call_tir(cls.NT_matmul7, (layer_norm91, gpt_neox_layers_12_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add266: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv115, gpt_neox_layers_12_mlp_dense_h_to_4h_bias3) gelu44: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add266) astype90: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu44, dtype="float16") lv116 = R.call_tir(cls.NT_matmul8, (astype90, gpt_neox_layers_12_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add267: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv116, gpt_neox_layers_12_mlp_dense_4h_to_h_bias3) astype91: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add267, dtype="float16") add268: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype91, add265) add269: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add268, add263) layer_norm92: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add269, gpt_neox_layers_13_input_layernorm_weight3, gpt_neox_layers_13_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv117 = R.call_tir(cls.NT_matmul5, (layer_norm92, gpt_neox_layers_13_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add270: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv117, gpt_neox_layers_13_attention_query_key_value_bias3) reshape180: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add270, R.shape([1, seq_len, 24, 256])) reshape181: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape180, R.shape([seq_len, 24, 256])) lv46 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape181), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape182: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv46, R.shape([1, seq_len, 8, 256])) reshape183: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape182, R.shape([1, seq_len, 2048])) lv118 = R.call_tir(cls.NT_matmul6, (reshape183, gpt_neox_layers_13_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add271: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv118, gpt_neox_layers_13_attention_dense_bias3) layer_norm93: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add269, gpt_neox_layers_13_post_attention_layernorm_weight3, gpt_neox_layers_13_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv119 = R.call_tir(cls.NT_matmul7, (layer_norm93, gpt_neox_layers_13_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add272: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv119, gpt_neox_layers_13_mlp_dense_h_to_4h_bias3) gelu45: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add272) astype92: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu45, dtype="float16") lv120 = R.call_tir(cls.NT_matmul8, (astype92, gpt_neox_layers_13_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add273: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv120, gpt_neox_layers_13_mlp_dense_4h_to_h_bias3) astype93: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add273, dtype="float16") add274: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype93, add271) add275: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add274, add269) layer_norm94: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add275, gpt_neox_layers_14_input_layernorm_weight3, gpt_neox_layers_14_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv121 = R.call_tir(cls.NT_matmul5, (layer_norm94, gpt_neox_layers_14_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add276: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv121, gpt_neox_layers_14_attention_query_key_value_bias3) reshape184: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add276, R.shape([1, seq_len, 24, 256])) reshape185: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape184, R.shape([seq_len, 24, 256])) lv47 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape185), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape186: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv47, R.shape([1, seq_len, 8, 256])) reshape187: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape186, R.shape([1, seq_len, 2048])) lv122 = R.call_tir(cls.NT_matmul6, (reshape187, gpt_neox_layers_14_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add277: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv122, gpt_neox_layers_14_attention_dense_bias3) layer_norm95: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add275, gpt_neox_layers_14_post_attention_layernorm_weight3, gpt_neox_layers_14_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv123 = R.call_tir(cls.NT_matmul7, (layer_norm95, gpt_neox_layers_14_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add278: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv123, gpt_neox_layers_14_mlp_dense_h_to_4h_bias3) gelu46: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add278) astype94: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu46, dtype="float16") lv124 = R.call_tir(cls.NT_matmul8, (astype94, gpt_neox_layers_14_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add279: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv124, gpt_neox_layers_14_mlp_dense_4h_to_h_bias3) astype95: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add279, dtype="float16") add280: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype95, add277) add281: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add280, add275) layer_norm96: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add281, gpt_neox_layers_15_input_layernorm_weight3, gpt_neox_layers_15_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv125 = R.call_tir(cls.NT_matmul5, (layer_norm96, gpt_neox_layers_15_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add282: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv125, gpt_neox_layers_15_attention_query_key_value_bias3) reshape188: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add282, R.shape([1, seq_len, 24, 256])) reshape189: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape188, R.shape([seq_len, 24, 256])) lv48 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape189), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape190: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv48, R.shape([1, seq_len, 8, 256])) reshape191: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape190, R.shape([1, seq_len, 2048])) lv126 = R.call_tir(cls.NT_matmul6, (reshape191, gpt_neox_layers_15_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add283: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv126, gpt_neox_layers_15_attention_dense_bias3) layer_norm97: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add281, gpt_neox_layers_15_post_attention_layernorm_weight3, gpt_neox_layers_15_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv127 = R.call_tir(cls.NT_matmul7, (layer_norm97, gpt_neox_layers_15_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add284: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv127, gpt_neox_layers_15_mlp_dense_h_to_4h_bias3) gelu47: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add284) astype96: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu47, dtype="float16") lv128 = R.call_tir(cls.NT_matmul8, (astype96, gpt_neox_layers_15_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add285: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv128, gpt_neox_layers_15_mlp_dense_4h_to_h_bias3) astype97: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add285, dtype="float16") add286: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype97, add283) add287: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add286, add281) layer_norm98: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add287, gpt_neox_final_layer_norm_weight3, gpt_neox_final_layer_norm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) take1: R.Tensor((1, batch_size, 2048), dtype="float16") = R.take(layer_norm98, logit_positions, axis=1) lv129 = R.call_tir(cls.NT_matmul9, (take1, embed_out_weight3), out_sinfo=R.Tensor((1, batch_size, vocab_size), dtype="float16")) astype98: R.Tensor((1, batch_size, vocab_size), dtype="float32") = R.astype(lv129, dtype="float32") gv3: R.Tuple(R.Tensor((1, batch_size, vocab_size), dtype="float32"), R.Object) = astype98, paged_kv_cache R.output(gv3) return gv3 @R.function def batch_verify(input_embeds: R.Tensor((1, "seq_len", 2048), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 2048), dtype="float16"))) -> R.Tuple(R.Tensor((1, "seq_len", "vocab_size"), dtype="float32"), R.Object): seq_len = T.int64() vocab_size = T.int64() R.func_attr({"num_input": 2, "pipeline_parallel_stages": 1, "relax.memory_plan_dynamic_func_output": True, "relax.rewrite_cuda_graph.capture_symbolic_vars": ["batch_size", "seq_len"], "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) cls = Module with R.dataflow(): gpt_neox_layers_0_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[1] gpt_neox_layers_0_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[2] gpt_neox_layers_0_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[3] gpt_neox_layers_0_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[4] gpt_neox_layers_0_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[5] gpt_neox_layers_0_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[6] gpt_neox_layers_0_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[8] gpt_neox_layers_0_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[9] gpt_neox_layers_0_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[10] gpt_neox_layers_0_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[11] gpt_neox_layers_0_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[12] gpt_neox_layers_1_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[13] gpt_neox_layers_1_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[14] gpt_neox_layers_1_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[15] gpt_neox_layers_1_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[16] gpt_neox_layers_1_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[17] gpt_neox_layers_1_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[18] gpt_neox_layers_1_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[19] gpt_neox_layers_1_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[20] gpt_neox_layers_1_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[21] gpt_neox_layers_1_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[22] gpt_neox_layers_1_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[23] gpt_neox_layers_1_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[24] gpt_neox_layers_2_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[25] gpt_neox_layers_2_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[26] gpt_neox_layers_2_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[27] gpt_neox_layers_2_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[28] gpt_neox_layers_2_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[29] gpt_neox_layers_2_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[30] gpt_neox_layers_2_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[31] gpt_neox_layers_2_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[32] gpt_neox_layers_2_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[33] gpt_neox_layers_2_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[34] gpt_neox_layers_2_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[35] gpt_neox_layers_2_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[36] gpt_neox_layers_3_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[37] gpt_neox_layers_3_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[38] gpt_neox_layers_3_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[39] gpt_neox_layers_3_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[40] gpt_neox_layers_3_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[41] gpt_neox_layers_3_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[42] gpt_neox_layers_3_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[43] gpt_neox_layers_3_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[44] gpt_neox_layers_3_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[45] gpt_neox_layers_3_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[46] gpt_neox_layers_3_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[47] gpt_neox_layers_3_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[48] gpt_neox_layers_4_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[49] gpt_neox_layers_4_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[50] gpt_neox_layers_4_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[51] gpt_neox_layers_4_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[52] gpt_neox_layers_4_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[53] gpt_neox_layers_4_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[54] gpt_neox_layers_4_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[55] gpt_neox_layers_4_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[56] gpt_neox_layers_4_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[57] gpt_neox_layers_4_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[58] gpt_neox_layers_4_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[59] gpt_neox_layers_4_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[60] gpt_neox_layers_5_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[61] gpt_neox_layers_5_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[62] gpt_neox_layers_5_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[63] gpt_neox_layers_5_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[64] gpt_neox_layers_5_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[65] gpt_neox_layers_5_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[66] gpt_neox_layers_5_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[67] gpt_neox_layers_5_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[68] gpt_neox_layers_5_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[69] gpt_neox_layers_5_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[70] gpt_neox_layers_5_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[71] gpt_neox_layers_5_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[72] gpt_neox_layers_6_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[73] gpt_neox_layers_6_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[74] gpt_neox_layers_6_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[75] gpt_neox_layers_6_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[76] gpt_neox_layers_6_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[77] gpt_neox_layers_6_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[78] gpt_neox_layers_6_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[79] gpt_neox_layers_6_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[80] gpt_neox_layers_6_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[81] gpt_neox_layers_6_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[82] gpt_neox_layers_6_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[83] gpt_neox_layers_6_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[84] gpt_neox_layers_7_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[85] gpt_neox_layers_7_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[86] gpt_neox_layers_7_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[87] gpt_neox_layers_7_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[88] gpt_neox_layers_7_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[89] gpt_neox_layers_7_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[90] gpt_neox_layers_7_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[91] gpt_neox_layers_7_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[92] gpt_neox_layers_7_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[93] gpt_neox_layers_7_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[94] gpt_neox_layers_7_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[95] gpt_neox_layers_7_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[96] gpt_neox_layers_8_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[97] gpt_neox_layers_8_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[98] gpt_neox_layers_8_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[99] gpt_neox_layers_8_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[100] gpt_neox_layers_8_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[101] gpt_neox_layers_8_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[102] gpt_neox_layers_8_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[103] gpt_neox_layers_8_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[104] gpt_neox_layers_8_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[105] gpt_neox_layers_8_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[106] gpt_neox_layers_8_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[107] gpt_neox_layers_8_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[108] gpt_neox_layers_9_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[109] gpt_neox_layers_9_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[110] gpt_neox_layers_9_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[111] gpt_neox_layers_9_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[112] gpt_neox_layers_9_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[113] gpt_neox_layers_9_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[114] gpt_neox_layers_9_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[115] gpt_neox_layers_9_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[116] gpt_neox_layers_9_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[117] gpt_neox_layers_9_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[118] gpt_neox_layers_9_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[119] gpt_neox_layers_9_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[120] gpt_neox_layers_10_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[121] gpt_neox_layers_10_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[122] gpt_neox_layers_10_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[123] gpt_neox_layers_10_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[124] gpt_neox_layers_10_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[125] gpt_neox_layers_10_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[126] gpt_neox_layers_10_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[127] gpt_neox_layers_10_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[128] gpt_neox_layers_10_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[129] gpt_neox_layers_10_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[130] gpt_neox_layers_10_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[131] gpt_neox_layers_10_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[132] gpt_neox_layers_11_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[133] gpt_neox_layers_11_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[134] gpt_neox_layers_11_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[135] gpt_neox_layers_11_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[136] gpt_neox_layers_11_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[137] gpt_neox_layers_11_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[138] gpt_neox_layers_11_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[139] gpt_neox_layers_11_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[140] gpt_neox_layers_11_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[141] gpt_neox_layers_11_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[142] gpt_neox_layers_11_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[143] gpt_neox_layers_11_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[144] gpt_neox_layers_12_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[145] gpt_neox_layers_12_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[146] gpt_neox_layers_12_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[147] gpt_neox_layers_12_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[148] gpt_neox_layers_12_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[149] gpt_neox_layers_12_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[150] gpt_neox_layers_12_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[151] gpt_neox_layers_12_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[152] gpt_neox_layers_12_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[153] gpt_neox_layers_12_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[154] gpt_neox_layers_12_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[155] gpt_neox_layers_12_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[156] gpt_neox_layers_13_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[157] gpt_neox_layers_13_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[158] gpt_neox_layers_13_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[159] gpt_neox_layers_13_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[160] gpt_neox_layers_13_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[161] gpt_neox_layers_13_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[162] gpt_neox_layers_13_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[163] gpt_neox_layers_13_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[164] gpt_neox_layers_13_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[165] gpt_neox_layers_13_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[166] gpt_neox_layers_13_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[167] gpt_neox_layers_13_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[168] gpt_neox_layers_14_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[169] gpt_neox_layers_14_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[170] gpt_neox_layers_14_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[171] gpt_neox_layers_14_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[172] gpt_neox_layers_14_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[173] gpt_neox_layers_14_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[174] gpt_neox_layers_14_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[175] gpt_neox_layers_14_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[176] gpt_neox_layers_14_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[177] gpt_neox_layers_14_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[178] gpt_neox_layers_14_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[179] gpt_neox_layers_14_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[180] gpt_neox_layers_15_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[181] gpt_neox_layers_15_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[182] gpt_neox_layers_15_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[183] gpt_neox_layers_15_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[184] gpt_neox_layers_15_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[185] gpt_neox_layers_15_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[186] gpt_neox_layers_15_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[187] gpt_neox_layers_15_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[188] gpt_neox_layers_15_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[189] gpt_neox_layers_15_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[190] gpt_neox_layers_15_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[191] gpt_neox_layers_15_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[192] gpt_neox_final_layer_norm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[193] gpt_neox_final_layer_norm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[194] embed_out_weight5: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[195] layer_norm132: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(input_embeds, gpt_neox_layers_0_input_layernorm_weight5, gpt_neox_layers_0_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv130 = R.call_tir(cls.NT_matmul5, (layer_norm132, gpt_neox_layers_0_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add384: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv130, gpt_neox_layers_0_attention_query_key_value_bias5) reshape256: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add384, R.shape([1, seq_len, 24, 256])) reshape257: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape256, R.shape([seq_len, 24, 256])) lv65 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), reshape257), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape258: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv65, R.shape([1, seq_len, 8, 256])) reshape259: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape258, R.shape([1, seq_len, 2048])) lv131 = R.call_tir(cls.NT_matmul6, (reshape259, gpt_neox_layers_0_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add385: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv131, gpt_neox_layers_0_attention_dense_bias5) layer_norm133: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(input_embeds, gpt_neox_layers_0_post_attention_layernorm_weight5, gpt_neox_layers_0_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv132 = R.call_tir(cls.NT_matmul7, (layer_norm133, gpt_neox_layers_0_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add386: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv132, gpt_neox_layers_0_mlp_dense_h_to_4h_bias5) gelu64: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add386) astype132: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu64, dtype="float16") lv133 = R.call_tir(cls.NT_matmul8, (astype132, gpt_neox_layers_0_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add387: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv133, gpt_neox_layers_0_mlp_dense_4h_to_h_bias5) astype133: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add387, dtype="float16") add388: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype133, add385) add389: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add388, input_embeds) layer_norm134: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add389, gpt_neox_layers_1_input_layernorm_weight5, gpt_neox_layers_1_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv134 = R.call_tir(cls.NT_matmul5, (layer_norm134, gpt_neox_layers_1_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add390: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv134, gpt_neox_layers_1_attention_query_key_value_bias5) reshape260: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add390, R.shape([1, seq_len, 24, 256])) reshape261: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape260, R.shape([seq_len, 24, 256])) lv66 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), reshape261), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape262: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv66, R.shape([1, seq_len, 8, 256])) reshape263: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape262, R.shape([1, seq_len, 2048])) lv135 = R.call_tir(cls.NT_matmul6, (reshape263, gpt_neox_layers_1_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add391: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv135, gpt_neox_layers_1_attention_dense_bias5) layer_norm135: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add389, gpt_neox_layers_1_post_attention_layernorm_weight5, gpt_neox_layers_1_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv136 = R.call_tir(cls.NT_matmul7, (layer_norm135, gpt_neox_layers_1_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add392: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv136, gpt_neox_layers_1_mlp_dense_h_to_4h_bias5) gelu65: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add392) astype134: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu65, dtype="float16") lv137 = R.call_tir(cls.NT_matmul8, (astype134, gpt_neox_layers_1_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add393: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv137, gpt_neox_layers_1_mlp_dense_4h_to_h_bias5) astype135: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add393, dtype="float16") add394: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype135, add391) add395: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add394, add389) layer_norm136: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add395, gpt_neox_layers_2_input_layernorm_weight5, gpt_neox_layers_2_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv138 = R.call_tir(cls.NT_matmul5, (layer_norm136, gpt_neox_layers_2_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add396: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv138, gpt_neox_layers_2_attention_query_key_value_bias5) reshape264: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add396, R.shape([1, seq_len, 24, 256])) reshape265: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape264, R.shape([seq_len, 24, 256])) lv67 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), reshape265), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape266: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv67, R.shape([1, seq_len, 8, 256])) reshape267: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape266, R.shape([1, seq_len, 2048])) lv139 = R.call_tir(cls.NT_matmul6, (reshape267, gpt_neox_layers_2_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add397: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv139, gpt_neox_layers_2_attention_dense_bias5) layer_norm137: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add395, gpt_neox_layers_2_post_attention_layernorm_weight5, gpt_neox_layers_2_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv140 = R.call_tir(cls.NT_matmul7, (layer_norm137, gpt_neox_layers_2_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add398: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv140, gpt_neox_layers_2_mlp_dense_h_to_4h_bias5) gelu66: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add398) astype136: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu66, dtype="float16") lv141 = R.call_tir(cls.NT_matmul8, (astype136, gpt_neox_layers_2_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add399: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv141, gpt_neox_layers_2_mlp_dense_4h_to_h_bias5) astype137: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add399, dtype="float16") add400: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype137, add397) add401: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add400, add395) layer_norm138: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add401, gpt_neox_layers_3_input_layernorm_weight5, gpt_neox_layers_3_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv142 = R.call_tir(cls.NT_matmul5, (layer_norm138, gpt_neox_layers_3_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add402: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv142, gpt_neox_layers_3_attention_query_key_value_bias5) reshape268: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add402, R.shape([1, seq_len, 24, 256])) reshape269: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape268, R.shape([seq_len, 24, 256])) lv68 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), reshape269), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape270: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv68, R.shape([1, seq_len, 8, 256])) reshape271: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape270, R.shape([1, seq_len, 2048])) lv143 = R.call_tir(cls.NT_matmul6, (reshape271, gpt_neox_layers_3_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add403: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv143, gpt_neox_layers_3_attention_dense_bias5) layer_norm139: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add401, gpt_neox_layers_3_post_attention_layernorm_weight5, gpt_neox_layers_3_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv144 = R.call_tir(cls.NT_matmul7, (layer_norm139, gpt_neox_layers_3_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add404: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv144, gpt_neox_layers_3_mlp_dense_h_to_4h_bias5) gelu67: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add404) astype138: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu67, dtype="float16") lv145 = R.call_tir(cls.NT_matmul8, (astype138, gpt_neox_layers_3_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add405: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv145, gpt_neox_layers_3_mlp_dense_4h_to_h_bias5) astype139: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add405, dtype="float16") add406: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype139, add403) add407: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add406, add401) layer_norm140: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add407, gpt_neox_layers_4_input_layernorm_weight5, gpt_neox_layers_4_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv146 = R.call_tir(cls.NT_matmul5, (layer_norm140, gpt_neox_layers_4_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add408: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv146, gpt_neox_layers_4_attention_query_key_value_bias5) reshape272: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add408, R.shape([1, seq_len, 24, 256])) reshape273: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape272, R.shape([seq_len, 24, 256])) lv69 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), reshape273), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape274: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv69, R.shape([1, seq_len, 8, 256])) reshape275: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape274, R.shape([1, seq_len, 2048])) lv147 = R.call_tir(cls.NT_matmul6, (reshape275, gpt_neox_layers_4_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add409: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv147, gpt_neox_layers_4_attention_dense_bias5) layer_norm141: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add407, gpt_neox_layers_4_post_attention_layernorm_weight5, gpt_neox_layers_4_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv148 = R.call_tir(cls.NT_matmul7, (layer_norm141, gpt_neox_layers_4_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add410: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv148, gpt_neox_layers_4_mlp_dense_h_to_4h_bias5) gelu68: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add410) astype140: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu68, dtype="float16") lv149 = R.call_tir(cls.NT_matmul8, (astype140, gpt_neox_layers_4_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add411: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv149, gpt_neox_layers_4_mlp_dense_4h_to_h_bias5) astype141: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add411, dtype="float16") add412: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype141, add409) add413: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add412, add407) layer_norm142: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add413, gpt_neox_layers_5_input_layernorm_weight5, gpt_neox_layers_5_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv150 = R.call_tir(cls.NT_matmul5, (layer_norm142, gpt_neox_layers_5_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add414: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv150, gpt_neox_layers_5_attention_query_key_value_bias5) reshape276: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add414, R.shape([1, seq_len, 24, 256])) reshape277: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape276, R.shape([seq_len, 24, 256])) lv70 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), reshape277), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape278: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv70, R.shape([1, seq_len, 8, 256])) reshape279: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape278, R.shape([1, seq_len, 2048])) lv151 = R.call_tir(cls.NT_matmul6, (reshape279, gpt_neox_layers_5_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add415: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv151, gpt_neox_layers_5_attention_dense_bias5) layer_norm143: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add413, gpt_neox_layers_5_post_attention_layernorm_weight5, gpt_neox_layers_5_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv152 = R.call_tir(cls.NT_matmul7, (layer_norm143, gpt_neox_layers_5_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add416: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv152, gpt_neox_layers_5_mlp_dense_h_to_4h_bias5) gelu69: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add416) astype142: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu69, dtype="float16") lv153 = R.call_tir(cls.NT_matmul8, (astype142, gpt_neox_layers_5_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add417: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv153, gpt_neox_layers_5_mlp_dense_4h_to_h_bias5) astype143: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add417, dtype="float16") add418: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype143, add415) add419: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add418, add413) layer_norm144: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add419, gpt_neox_layers_6_input_layernorm_weight5, gpt_neox_layers_6_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv154 = R.call_tir(cls.NT_matmul5, (layer_norm144, gpt_neox_layers_6_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add420: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv154, gpt_neox_layers_6_attention_query_key_value_bias5) reshape280: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add420, R.shape([1, seq_len, 24, 256])) reshape281: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape280, R.shape([seq_len, 24, 256])) lv71 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), reshape281), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape282: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv71, R.shape([1, seq_len, 8, 256])) reshape283: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape282, R.shape([1, seq_len, 2048])) lv155 = R.call_tir(cls.NT_matmul6, (reshape283, gpt_neox_layers_6_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add421: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv155, gpt_neox_layers_6_attention_dense_bias5) layer_norm145: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add419, gpt_neox_layers_6_post_attention_layernorm_weight5, gpt_neox_layers_6_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv156 = R.call_tir(cls.NT_matmul7, (layer_norm145, gpt_neox_layers_6_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add422: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv156, gpt_neox_layers_6_mlp_dense_h_to_4h_bias5) gelu70: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add422) astype144: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu70, dtype="float16") lv157 = R.call_tir(cls.NT_matmul8, (astype144, gpt_neox_layers_6_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add423: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv157, gpt_neox_layers_6_mlp_dense_4h_to_h_bias5) astype145: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add423, dtype="float16") add424: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype145, add421) add425: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add424, add419) layer_norm146: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add425, gpt_neox_layers_7_input_layernorm_weight5, gpt_neox_layers_7_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv158 = R.call_tir(cls.NT_matmul5, (layer_norm146, gpt_neox_layers_7_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add426: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv158, gpt_neox_layers_7_attention_query_key_value_bias5) reshape284: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add426, R.shape([1, seq_len, 24, 256])) reshape285: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape284, R.shape([seq_len, 24, 256])) lv72 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), reshape285), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape286: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv72, R.shape([1, seq_len, 8, 256])) reshape287: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape286, R.shape([1, seq_len, 2048])) lv159 = R.call_tir(cls.NT_matmul6, (reshape287, gpt_neox_layers_7_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add427: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv159, gpt_neox_layers_7_attention_dense_bias5) layer_norm147: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add425, gpt_neox_layers_7_post_attention_layernorm_weight5, gpt_neox_layers_7_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv160 = R.call_tir(cls.NT_matmul7, (layer_norm147, gpt_neox_layers_7_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add428: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv160, gpt_neox_layers_7_mlp_dense_h_to_4h_bias5) gelu71: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add428) astype146: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu71, dtype="float16") lv161 = R.call_tir(cls.NT_matmul8, (astype146, gpt_neox_layers_7_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add429: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv161, gpt_neox_layers_7_mlp_dense_4h_to_h_bias5) astype147: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add429, dtype="float16") add430: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype147, add427) add431: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add430, add425) layer_norm148: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add431, gpt_neox_layers_8_input_layernorm_weight5, gpt_neox_layers_8_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv162 = R.call_tir(cls.NT_matmul5, (layer_norm148, gpt_neox_layers_8_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add432: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv162, gpt_neox_layers_8_attention_query_key_value_bias5) reshape288: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add432, R.shape([1, seq_len, 24, 256])) reshape289: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape288, R.shape([seq_len, 24, 256])) lv73 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), reshape289), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape290: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv73, R.shape([1, seq_len, 8, 256])) reshape291: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape290, R.shape([1, seq_len, 2048])) lv163 = R.call_tir(cls.NT_matmul6, (reshape291, gpt_neox_layers_8_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add433: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv163, gpt_neox_layers_8_attention_dense_bias5) layer_norm149: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add431, gpt_neox_layers_8_post_attention_layernorm_weight5, gpt_neox_layers_8_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv164 = R.call_tir(cls.NT_matmul7, (layer_norm149, gpt_neox_layers_8_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add434: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv164, gpt_neox_layers_8_mlp_dense_h_to_4h_bias5) gelu72: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add434) astype148: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu72, dtype="float16") lv165 = R.call_tir(cls.NT_matmul8, (astype148, gpt_neox_layers_8_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add435: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv165, gpt_neox_layers_8_mlp_dense_4h_to_h_bias5) astype149: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add435, dtype="float16") add436: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype149, add433) add437: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add436, add431) layer_norm150: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add437, gpt_neox_layers_9_input_layernorm_weight5, gpt_neox_layers_9_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv166 = R.call_tir(cls.NT_matmul5, (layer_norm150, gpt_neox_layers_9_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add438: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv166, gpt_neox_layers_9_attention_query_key_value_bias5) reshape292: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add438, R.shape([1, seq_len, 24, 256])) reshape293: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape292, R.shape([seq_len, 24, 256])) lv74 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), reshape293), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape294: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv74, R.shape([1, seq_len, 8, 256])) reshape295: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape294, R.shape([1, seq_len, 2048])) lv167 = R.call_tir(cls.NT_matmul6, (reshape295, gpt_neox_layers_9_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add439: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv167, gpt_neox_layers_9_attention_dense_bias5) layer_norm151: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add437, gpt_neox_layers_9_post_attention_layernorm_weight5, gpt_neox_layers_9_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv168 = R.call_tir(cls.NT_matmul7, (layer_norm151, gpt_neox_layers_9_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add440: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv168, gpt_neox_layers_9_mlp_dense_h_to_4h_bias5) gelu73: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add440) astype150: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu73, dtype="float16") lv169 = R.call_tir(cls.NT_matmul8, (astype150, gpt_neox_layers_9_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add441: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv169, gpt_neox_layers_9_mlp_dense_4h_to_h_bias5) astype151: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add441, dtype="float16") add442: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype151, add439) add443: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add442, add437) layer_norm152: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add443, gpt_neox_layers_10_input_layernorm_weight5, gpt_neox_layers_10_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv170 = R.call_tir(cls.NT_matmul5, (layer_norm152, gpt_neox_layers_10_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add444: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv170, gpt_neox_layers_10_attention_query_key_value_bias5) reshape296: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add444, R.shape([1, seq_len, 24, 256])) reshape297: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape296, R.shape([seq_len, 24, 256])) lv75 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), reshape297), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape298: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv75, R.shape([1, seq_len, 8, 256])) reshape299: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape298, R.shape([1, seq_len, 2048])) lv171 = R.call_tir(cls.NT_matmul6, (reshape299, gpt_neox_layers_10_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add445: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv171, gpt_neox_layers_10_attention_dense_bias5) layer_norm153: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add443, gpt_neox_layers_10_post_attention_layernorm_weight5, gpt_neox_layers_10_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv172 = R.call_tir(cls.NT_matmul7, (layer_norm153, gpt_neox_layers_10_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add446: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv172, gpt_neox_layers_10_mlp_dense_h_to_4h_bias5) gelu74: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add446) astype152: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu74, dtype="float16") lv173 = R.call_tir(cls.NT_matmul8, (astype152, gpt_neox_layers_10_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add447: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv173, gpt_neox_layers_10_mlp_dense_4h_to_h_bias5) astype153: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add447, dtype="float16") add448: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype153, add445) add449: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add448, add443) layer_norm154: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add449, gpt_neox_layers_11_input_layernorm_weight5, gpt_neox_layers_11_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv174 = R.call_tir(cls.NT_matmul5, (layer_norm154, gpt_neox_layers_11_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add450: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv174, gpt_neox_layers_11_attention_query_key_value_bias5) reshape300: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add450, R.shape([1, seq_len, 24, 256])) reshape301: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape300, R.shape([seq_len, 24, 256])) lv76 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), reshape301), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape302: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv76, R.shape([1, seq_len, 8, 256])) reshape303: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape302, R.shape([1, seq_len, 2048])) lv175 = R.call_tir(cls.NT_matmul6, (reshape303, gpt_neox_layers_11_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add451: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv175, gpt_neox_layers_11_attention_dense_bias5) layer_norm155: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add449, gpt_neox_layers_11_post_attention_layernorm_weight5, gpt_neox_layers_11_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv176 = R.call_tir(cls.NT_matmul7, (layer_norm155, gpt_neox_layers_11_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add452: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv176, gpt_neox_layers_11_mlp_dense_h_to_4h_bias5) gelu75: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add452) astype154: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu75, dtype="float16") lv177 = R.call_tir(cls.NT_matmul8, (astype154, gpt_neox_layers_11_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add453: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv177, gpt_neox_layers_11_mlp_dense_4h_to_h_bias5) astype155: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add453, dtype="float16") add454: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype155, add451) add455: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add454, add449) layer_norm156: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add455, gpt_neox_layers_12_input_layernorm_weight5, gpt_neox_layers_12_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv178 = R.call_tir(cls.NT_matmul5, (layer_norm156, gpt_neox_layers_12_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add456: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv178, gpt_neox_layers_12_attention_query_key_value_bias5) reshape304: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add456, R.shape([1, seq_len, 24, 256])) reshape305: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape304, R.shape([seq_len, 24, 256])) lv77 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape305), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape306: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv77, R.shape([1, seq_len, 8, 256])) reshape307: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape306, R.shape([1, seq_len, 2048])) lv179 = R.call_tir(cls.NT_matmul6, (reshape307, gpt_neox_layers_12_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add457: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv179, gpt_neox_layers_12_attention_dense_bias5) layer_norm157: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add455, gpt_neox_layers_12_post_attention_layernorm_weight5, gpt_neox_layers_12_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv180 = R.call_tir(cls.NT_matmul7, (layer_norm157, gpt_neox_layers_12_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add458: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv180, gpt_neox_layers_12_mlp_dense_h_to_4h_bias5) gelu76: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add458) astype156: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu76, dtype="float16") lv181 = R.call_tir(cls.NT_matmul8, (astype156, gpt_neox_layers_12_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add459: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv181, gpt_neox_layers_12_mlp_dense_4h_to_h_bias5) astype157: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add459, dtype="float16") add460: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype157, add457) add461: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add460, add455) layer_norm158: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add461, gpt_neox_layers_13_input_layernorm_weight5, gpt_neox_layers_13_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv182 = R.call_tir(cls.NT_matmul5, (layer_norm158, gpt_neox_layers_13_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add462: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv182, gpt_neox_layers_13_attention_query_key_value_bias5) reshape308: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add462, R.shape([1, seq_len, 24, 256])) reshape309: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape308, R.shape([seq_len, 24, 256])) lv78 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape309), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape310: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv78, R.shape([1, seq_len, 8, 256])) reshape311: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape310, R.shape([1, seq_len, 2048])) lv183 = R.call_tir(cls.NT_matmul6, (reshape311, gpt_neox_layers_13_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add463: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv183, gpt_neox_layers_13_attention_dense_bias5) layer_norm159: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add461, gpt_neox_layers_13_post_attention_layernorm_weight5, gpt_neox_layers_13_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv184 = R.call_tir(cls.NT_matmul7, (layer_norm159, gpt_neox_layers_13_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add464: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv184, gpt_neox_layers_13_mlp_dense_h_to_4h_bias5) gelu77: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add464) astype158: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu77, dtype="float16") lv185 = R.call_tir(cls.NT_matmul8, (astype158, gpt_neox_layers_13_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add465: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv185, gpt_neox_layers_13_mlp_dense_4h_to_h_bias5) astype159: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add465, dtype="float16") add466: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype159, add463) add467: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add466, add461) layer_norm160: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add467, gpt_neox_layers_14_input_layernorm_weight5, gpt_neox_layers_14_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv186 = R.call_tir(cls.NT_matmul5, (layer_norm160, gpt_neox_layers_14_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add468: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv186, gpt_neox_layers_14_attention_query_key_value_bias5) reshape312: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add468, R.shape([1, seq_len, 24, 256])) reshape313: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape312, R.shape([seq_len, 24, 256])) lv79 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape313), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape314: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv79, R.shape([1, seq_len, 8, 256])) reshape315: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape314, R.shape([1, seq_len, 2048])) lv187 = R.call_tir(cls.NT_matmul6, (reshape315, gpt_neox_layers_14_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add469: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv187, gpt_neox_layers_14_attention_dense_bias5) layer_norm161: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add467, gpt_neox_layers_14_post_attention_layernorm_weight5, gpt_neox_layers_14_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv188 = R.call_tir(cls.NT_matmul7, (layer_norm161, gpt_neox_layers_14_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add470: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv188, gpt_neox_layers_14_mlp_dense_h_to_4h_bias5) gelu78: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add470) astype160: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu78, dtype="float16") lv189 = R.call_tir(cls.NT_matmul8, (astype160, gpt_neox_layers_14_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add471: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv189, gpt_neox_layers_14_mlp_dense_4h_to_h_bias5) astype161: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add471, dtype="float16") add472: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype161, add469) add473: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add472, add467) layer_norm162: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add473, gpt_neox_layers_15_input_layernorm_weight5, gpt_neox_layers_15_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv190 = R.call_tir(cls.NT_matmul5, (layer_norm162, gpt_neox_layers_15_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add474: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv190, gpt_neox_layers_15_attention_query_key_value_bias5) reshape316: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add474, R.shape([1, seq_len, 24, 256])) reshape317: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape316, R.shape([seq_len, 24, 256])) lv80 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape317), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape318: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv80, R.shape([1, seq_len, 8, 256])) reshape319: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape318, R.shape([1, seq_len, 2048])) lv191 = R.call_tir(cls.NT_matmul6, (reshape319, gpt_neox_layers_15_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add475: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv191, gpt_neox_layers_15_attention_dense_bias5) layer_norm163: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add473, gpt_neox_layers_15_post_attention_layernorm_weight5, gpt_neox_layers_15_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv192 = R.call_tir(cls.NT_matmul7, (layer_norm163, gpt_neox_layers_15_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add476: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv192, gpt_neox_layers_15_mlp_dense_h_to_4h_bias5) gelu79: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add476) astype162: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu79, dtype="float16") lv193 = R.call_tir(cls.NT_matmul8, (astype162, gpt_neox_layers_15_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add477: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv193, gpt_neox_layers_15_mlp_dense_4h_to_h_bias5) astype163: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add477, dtype="float16") add478: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype163, add475) add479: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add478, add473) layer_norm164: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add479, gpt_neox_final_layer_norm_weight5, gpt_neox_final_layer_norm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv194 = R.call_tir(cls.NT_matmul9, (layer_norm164, embed_out_weight5), out_sinfo=R.Tensor((1, seq_len, vocab_size), dtype="float16")) astype164: R.Tensor((1, seq_len, vocab_size), dtype="float32") = R.astype(lv194, dtype="float32") gv5: R.Tuple(R.Tensor((1, seq_len, vocab_size), dtype="float32"), R.Object) = astype164, 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 gv: R.Tensor((), dtype="float16") = R.zeros(R.shape([]), dtype="float16") paged_kv_cache: R.Object = R.call_pure_packed("vm.builtin.paged_attention_kv_cache_create_reduced", R.shape([max_batch_size, max_total_seq_len, prefill_chunk_size, page_size, support_sliding_window]), R.shape([0, 16]), R.prim_value(8), R.prim_value(8), R.prim_value(256), R.prim_value(1), R.prim_value(1), R.prim_value(10000), gv, cls.tir_kv_cache_transpose_append, cls.batch_prefill_paged_kv, cls.batch_decode_paged_kv, cls.batch_prefill_paged_kv_sliding_window, cls.batch_decode_paged_kv_sliding_window, cls.batch_prefill_ragged_kv, cls.merge_state_inplace, cls.fused_rope, cls.copy_single_page, cls.tir_kv_cache_debug_get_kv, cls.compact_kv_copy, cls.batch_tree_attn, cls.tree_attn_paged_kv, R.prim_value(0), R.prim_value(0), sinfo_args=(R.Object,)) return paged_kv_cache @R.function def decode(input_embed: R.Tensor((1, 1, 2048), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 2048), dtype="float16"))) -> R.Tuple(R.Tensor((1, 1, "vocab_size"), dtype="float32"), R.Object): vocab_size = T.int64() R.func_attr({"num_input": 2, "pipeline_parallel_stages": 1, "relax.memory_plan_dynamic_func_output": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) cls = Module with R.dataflow(): gpt_neox_layers_0_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[1] gpt_neox_layers_0_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[2] gpt_neox_layers_0_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[3] gpt_neox_layers_0_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[4] gpt_neox_layers_0_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[5] gpt_neox_layers_0_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[6] gpt_neox_layers_0_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[8] gpt_neox_layers_0_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[9] gpt_neox_layers_0_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[10] gpt_neox_layers_0_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[11] gpt_neox_layers_0_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[12] gpt_neox_layers_1_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[13] gpt_neox_layers_1_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[14] gpt_neox_layers_1_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[15] gpt_neox_layers_1_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[16] gpt_neox_layers_1_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[17] gpt_neox_layers_1_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[18] gpt_neox_layers_1_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[19] gpt_neox_layers_1_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[20] gpt_neox_layers_1_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[21] gpt_neox_layers_1_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[22] gpt_neox_layers_1_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[23] gpt_neox_layers_1_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[24] gpt_neox_layers_2_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[25] gpt_neox_layers_2_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[26] gpt_neox_layers_2_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[27] gpt_neox_layers_2_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[28] gpt_neox_layers_2_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[29] gpt_neox_layers_2_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[30] gpt_neox_layers_2_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[31] gpt_neox_layers_2_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[32] gpt_neox_layers_2_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[33] gpt_neox_layers_2_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[34] gpt_neox_layers_2_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[35] gpt_neox_layers_2_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[36] gpt_neox_layers_3_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[37] gpt_neox_layers_3_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[38] gpt_neox_layers_3_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[39] gpt_neox_layers_3_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[40] gpt_neox_layers_3_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[41] gpt_neox_layers_3_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[42] gpt_neox_layers_3_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[43] gpt_neox_layers_3_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[44] gpt_neox_layers_3_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[45] gpt_neox_layers_3_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[46] gpt_neox_layers_3_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[47] gpt_neox_layers_3_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[48] gpt_neox_layers_4_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[49] gpt_neox_layers_4_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[50] gpt_neox_layers_4_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[51] gpt_neox_layers_4_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[52] gpt_neox_layers_4_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[53] gpt_neox_layers_4_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[54] gpt_neox_layers_4_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[55] gpt_neox_layers_4_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[56] gpt_neox_layers_4_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[57] gpt_neox_layers_4_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[58] gpt_neox_layers_4_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[59] gpt_neox_layers_4_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[60] gpt_neox_layers_5_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[61] gpt_neox_layers_5_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[62] gpt_neox_layers_5_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[63] gpt_neox_layers_5_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[64] gpt_neox_layers_5_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[65] gpt_neox_layers_5_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[66] gpt_neox_layers_5_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[67] gpt_neox_layers_5_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[68] gpt_neox_layers_5_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[69] gpt_neox_layers_5_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[70] gpt_neox_layers_5_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[71] gpt_neox_layers_5_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[72] gpt_neox_layers_6_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[73] gpt_neox_layers_6_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[74] gpt_neox_layers_6_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[75] gpt_neox_layers_6_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[76] gpt_neox_layers_6_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[77] gpt_neox_layers_6_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[78] gpt_neox_layers_6_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[79] gpt_neox_layers_6_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[80] gpt_neox_layers_6_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[81] gpt_neox_layers_6_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[82] gpt_neox_layers_6_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[83] gpt_neox_layers_6_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[84] gpt_neox_layers_7_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[85] gpt_neox_layers_7_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[86] gpt_neox_layers_7_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[87] gpt_neox_layers_7_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[88] gpt_neox_layers_7_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[89] gpt_neox_layers_7_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[90] gpt_neox_layers_7_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[91] gpt_neox_layers_7_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[92] gpt_neox_layers_7_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[93] gpt_neox_layers_7_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[94] gpt_neox_layers_7_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[95] gpt_neox_layers_7_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[96] gpt_neox_layers_8_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[97] gpt_neox_layers_8_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[98] gpt_neox_layers_8_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[99] gpt_neox_layers_8_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[100] gpt_neox_layers_8_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[101] gpt_neox_layers_8_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[102] gpt_neox_layers_8_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[103] gpt_neox_layers_8_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[104] gpt_neox_layers_8_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[105] gpt_neox_layers_8_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[106] gpt_neox_layers_8_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[107] gpt_neox_layers_8_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[108] gpt_neox_layers_9_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[109] gpt_neox_layers_9_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[110] gpt_neox_layers_9_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[111] gpt_neox_layers_9_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[112] gpt_neox_layers_9_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[113] gpt_neox_layers_9_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[114] gpt_neox_layers_9_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[115] gpt_neox_layers_9_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[116] gpt_neox_layers_9_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[117] gpt_neox_layers_9_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[118] gpt_neox_layers_9_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[119] gpt_neox_layers_9_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[120] gpt_neox_layers_10_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[121] gpt_neox_layers_10_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[122] gpt_neox_layers_10_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[123] gpt_neox_layers_10_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[124] gpt_neox_layers_10_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[125] gpt_neox_layers_10_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[126] gpt_neox_layers_10_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[127] gpt_neox_layers_10_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[128] gpt_neox_layers_10_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[129] gpt_neox_layers_10_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[130] gpt_neox_layers_10_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[131] gpt_neox_layers_10_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[132] gpt_neox_layers_11_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[133] gpt_neox_layers_11_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[134] gpt_neox_layers_11_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[135] gpt_neox_layers_11_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[136] gpt_neox_layers_11_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[137] gpt_neox_layers_11_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[138] gpt_neox_layers_11_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[139] gpt_neox_layers_11_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[140] gpt_neox_layers_11_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[141] gpt_neox_layers_11_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[142] gpt_neox_layers_11_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[143] gpt_neox_layers_11_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[144] gpt_neox_layers_12_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[145] gpt_neox_layers_12_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[146] gpt_neox_layers_12_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[147] gpt_neox_layers_12_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[148] gpt_neox_layers_12_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[149] gpt_neox_layers_12_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[150] gpt_neox_layers_12_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[151] gpt_neox_layers_12_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[152] gpt_neox_layers_12_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[153] gpt_neox_layers_12_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[154] gpt_neox_layers_12_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[155] gpt_neox_layers_12_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[156] gpt_neox_layers_13_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[157] gpt_neox_layers_13_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[158] gpt_neox_layers_13_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[159] gpt_neox_layers_13_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[160] gpt_neox_layers_13_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[161] gpt_neox_layers_13_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[162] gpt_neox_layers_13_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[163] gpt_neox_layers_13_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[164] gpt_neox_layers_13_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[165] gpt_neox_layers_13_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[166] gpt_neox_layers_13_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[167] gpt_neox_layers_13_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[168] gpt_neox_layers_14_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[169] gpt_neox_layers_14_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[170] gpt_neox_layers_14_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[171] gpt_neox_layers_14_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[172] gpt_neox_layers_14_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[173] gpt_neox_layers_14_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[174] gpt_neox_layers_14_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[175] gpt_neox_layers_14_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[176] gpt_neox_layers_14_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[177] gpt_neox_layers_14_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[178] gpt_neox_layers_14_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[179] gpt_neox_layers_14_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[180] gpt_neox_layers_15_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[181] gpt_neox_layers_15_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[182] gpt_neox_layers_15_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[183] gpt_neox_layers_15_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[184] gpt_neox_layers_15_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[185] gpt_neox_layers_15_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[186] gpt_neox_layers_15_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[187] gpt_neox_layers_15_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[188] gpt_neox_layers_15_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[189] gpt_neox_layers_15_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[190] gpt_neox_layers_15_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[191] gpt_neox_layers_15_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[192] gpt_neox_final_layer_norm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[193] gpt_neox_final_layer_norm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[194] embed_out_weight2: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[195] layer_norm33: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(input_embed, gpt_neox_layers_0_input_layernorm_weight2, gpt_neox_layers_0_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv195 = R.call_tir(cls.NT_matmul10, (layer_norm33, gpt_neox_layers_0_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add96: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv195, gpt_neox_layers_0_attention_query_key_value_bias2) reshape64: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add96, R.shape([1, 1, 24, 256])) reshape65: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape64, R.shape([1, 24, 256])) lv17 = 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)), reshape65), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape66: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv17, R.shape([1, 1, 8, 256])) reshape67: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape66, R.shape([1, 1, 2048])) lv196 = R.call_tir(cls.NT_matmul11, (reshape67, gpt_neox_layers_0_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add97: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv196, gpt_neox_layers_0_attention_dense_bias2) layer_norm34: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(input_embed, gpt_neox_layers_0_post_attention_layernorm_weight2, gpt_neox_layers_0_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv197 = R.call_tir(cls.NT_matmul12, (layer_norm34, gpt_neox_layers_0_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add98: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv197, gpt_neox_layers_0_mlp_dense_h_to_4h_bias2) gelu16: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add98) astype33: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu16, dtype="float16") lv198 = R.call_tir(cls.NT_matmul13, (astype33, gpt_neox_layers_0_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add99: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv198, gpt_neox_layers_0_mlp_dense_4h_to_h_bias2) astype34: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add99, dtype="float16") add100: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype34, add97) add101: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add100, input_embed) layer_norm35: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add101, gpt_neox_layers_1_input_layernorm_weight2, gpt_neox_layers_1_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv199 = R.call_tir(cls.NT_matmul10, (layer_norm35, gpt_neox_layers_1_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add102: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv199, gpt_neox_layers_1_attention_query_key_value_bias2) reshape68: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add102, R.shape([1, 1, 24, 256])) reshape69: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape68, R.shape([1, 24, 256])) lv18 = 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)), reshape69), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape70: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv18, R.shape([1, 1, 8, 256])) reshape71: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape70, R.shape([1, 1, 2048])) lv200 = R.call_tir(cls.NT_matmul11, (reshape71, gpt_neox_layers_1_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add103: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv200, gpt_neox_layers_1_attention_dense_bias2) layer_norm36: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add101, gpt_neox_layers_1_post_attention_layernorm_weight2, gpt_neox_layers_1_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv201 = R.call_tir(cls.NT_matmul12, (layer_norm36, gpt_neox_layers_1_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add104: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv201, gpt_neox_layers_1_mlp_dense_h_to_4h_bias2) gelu17: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add104) astype35: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu17, dtype="float16") lv202 = R.call_tir(cls.NT_matmul13, (astype35, gpt_neox_layers_1_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add105: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv202, gpt_neox_layers_1_mlp_dense_4h_to_h_bias2) astype36: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add105, dtype="float16") add106: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype36, add103) add107: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add106, add101) layer_norm37: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add107, gpt_neox_layers_2_input_layernorm_weight2, gpt_neox_layers_2_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv203 = R.call_tir(cls.NT_matmul10, (layer_norm37, gpt_neox_layers_2_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add108: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv203, gpt_neox_layers_2_attention_query_key_value_bias2) reshape72: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add108, R.shape([1, 1, 24, 256])) reshape73: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape72, R.shape([1, 24, 256])) lv19 = 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)), reshape73), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape74: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv19, R.shape([1, 1, 8, 256])) reshape75: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape74, R.shape([1, 1, 2048])) lv204 = R.call_tir(cls.NT_matmul11, (reshape75, gpt_neox_layers_2_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add109: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv204, gpt_neox_layers_2_attention_dense_bias2) layer_norm38: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add107, gpt_neox_layers_2_post_attention_layernorm_weight2, gpt_neox_layers_2_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv205 = R.call_tir(cls.NT_matmul12, (layer_norm38, gpt_neox_layers_2_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add110: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv205, gpt_neox_layers_2_mlp_dense_h_to_4h_bias2) gelu18: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add110) astype37: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu18, dtype="float16") lv206 = R.call_tir(cls.NT_matmul13, (astype37, gpt_neox_layers_2_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add111: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv206, gpt_neox_layers_2_mlp_dense_4h_to_h_bias2) astype38: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add111, dtype="float16") add112: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype38, add109) add113: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add112, add107) layer_norm39: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add113, gpt_neox_layers_3_input_layernorm_weight2, gpt_neox_layers_3_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv207 = R.call_tir(cls.NT_matmul10, (layer_norm39, gpt_neox_layers_3_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add114: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv207, gpt_neox_layers_3_attention_query_key_value_bias2) reshape76: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add114, R.shape([1, 1, 24, 256])) reshape77: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape76, R.shape([1, 24, 256])) lv20 = 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)), reshape77), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape78: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv20, R.shape([1, 1, 8, 256])) reshape79: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape78, R.shape([1, 1, 2048])) lv208 = R.call_tir(cls.NT_matmul11, (reshape79, gpt_neox_layers_3_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add115: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv208, gpt_neox_layers_3_attention_dense_bias2) layer_norm40: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add113, gpt_neox_layers_3_post_attention_layernorm_weight2, gpt_neox_layers_3_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv209 = R.call_tir(cls.NT_matmul12, (layer_norm40, gpt_neox_layers_3_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add116: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv209, gpt_neox_layers_3_mlp_dense_h_to_4h_bias2) gelu19: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add116) astype39: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu19, dtype="float16") lv210 = R.call_tir(cls.NT_matmul13, (astype39, gpt_neox_layers_3_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add117: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv210, gpt_neox_layers_3_mlp_dense_4h_to_h_bias2) astype40: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add117, dtype="float16") add118: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype40, add115) add119: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add118, add113) layer_norm41: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add119, gpt_neox_layers_4_input_layernorm_weight2, gpt_neox_layers_4_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv211 = R.call_tir(cls.NT_matmul10, (layer_norm41, gpt_neox_layers_4_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add120: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv211, gpt_neox_layers_4_attention_query_key_value_bias2) reshape80: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add120, R.shape([1, 1, 24, 256])) reshape81: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape80, R.shape([1, 24, 256])) lv21 = 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)), reshape81), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape82: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv21, R.shape([1, 1, 8, 256])) reshape83: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape82, R.shape([1, 1, 2048])) lv212 = R.call_tir(cls.NT_matmul11, (reshape83, gpt_neox_layers_4_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add121: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv212, gpt_neox_layers_4_attention_dense_bias2) layer_norm42: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add119, gpt_neox_layers_4_post_attention_layernorm_weight2, gpt_neox_layers_4_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv213 = R.call_tir(cls.NT_matmul12, (layer_norm42, gpt_neox_layers_4_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add122: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv213, gpt_neox_layers_4_mlp_dense_h_to_4h_bias2) gelu20: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add122) astype41: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu20, dtype="float16") lv214 = R.call_tir(cls.NT_matmul13, (astype41, gpt_neox_layers_4_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add123: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv214, gpt_neox_layers_4_mlp_dense_4h_to_h_bias2) astype42: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add123, dtype="float16") add124: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype42, add121) add125: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add124, add119) layer_norm43: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add125, gpt_neox_layers_5_input_layernorm_weight2, gpt_neox_layers_5_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv215 = R.call_tir(cls.NT_matmul10, (layer_norm43, gpt_neox_layers_5_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add126: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv215, gpt_neox_layers_5_attention_query_key_value_bias2) reshape84: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add126, R.shape([1, 1, 24, 256])) reshape85: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape84, R.shape([1, 24, 256])) lv22 = 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)), reshape85), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape86: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv22, R.shape([1, 1, 8, 256])) reshape87: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape86, R.shape([1, 1, 2048])) lv216 = R.call_tir(cls.NT_matmul11, (reshape87, gpt_neox_layers_5_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add127: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv216, gpt_neox_layers_5_attention_dense_bias2) layer_norm44: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add125, gpt_neox_layers_5_post_attention_layernorm_weight2, gpt_neox_layers_5_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv217 = R.call_tir(cls.NT_matmul12, (layer_norm44, gpt_neox_layers_5_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add128: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv217, gpt_neox_layers_5_mlp_dense_h_to_4h_bias2) gelu21: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add128) astype43: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu21, dtype="float16") lv218 = R.call_tir(cls.NT_matmul13, (astype43, gpt_neox_layers_5_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add129: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv218, gpt_neox_layers_5_mlp_dense_4h_to_h_bias2) astype44: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add129, dtype="float16") add130: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype44, add127) add131: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add130, add125) layer_norm45: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add131, gpt_neox_layers_6_input_layernorm_weight2, gpt_neox_layers_6_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv219 = R.call_tir(cls.NT_matmul10, (layer_norm45, gpt_neox_layers_6_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add132: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv219, gpt_neox_layers_6_attention_query_key_value_bias2) reshape88: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add132, R.shape([1, 1, 24, 256])) reshape89: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape88, R.shape([1, 24, 256])) lv23 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), reshape89), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape90: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv23, R.shape([1, 1, 8, 256])) reshape91: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape90, R.shape([1, 1, 2048])) lv220 = R.call_tir(cls.NT_matmul11, (reshape91, gpt_neox_layers_6_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add133: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv220, gpt_neox_layers_6_attention_dense_bias2) layer_norm46: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add131, gpt_neox_layers_6_post_attention_layernorm_weight2, gpt_neox_layers_6_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv221 = R.call_tir(cls.NT_matmul12, (layer_norm46, gpt_neox_layers_6_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add134: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv221, gpt_neox_layers_6_mlp_dense_h_to_4h_bias2) gelu22: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add134) astype45: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu22, dtype="float16") lv222 = R.call_tir(cls.NT_matmul13, (astype45, gpt_neox_layers_6_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add135: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv222, gpt_neox_layers_6_mlp_dense_4h_to_h_bias2) astype46: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add135, dtype="float16") add136: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype46, add133) add137: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add136, add131) layer_norm47: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add137, gpt_neox_layers_7_input_layernorm_weight2, gpt_neox_layers_7_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv223 = R.call_tir(cls.NT_matmul10, (layer_norm47, gpt_neox_layers_7_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add138: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv223, gpt_neox_layers_7_attention_query_key_value_bias2) reshape92: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add138, R.shape([1, 1, 24, 256])) reshape93: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape92, R.shape([1, 24, 256])) lv24 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), reshape93), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape94: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv24, R.shape([1, 1, 8, 256])) reshape95: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape94, R.shape([1, 1, 2048])) lv224 = R.call_tir(cls.NT_matmul11, (reshape95, gpt_neox_layers_7_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add139: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv224, gpt_neox_layers_7_attention_dense_bias2) layer_norm48: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add137, gpt_neox_layers_7_post_attention_layernorm_weight2, gpt_neox_layers_7_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv225 = R.call_tir(cls.NT_matmul12, (layer_norm48, gpt_neox_layers_7_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add140: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv225, gpt_neox_layers_7_mlp_dense_h_to_4h_bias2) gelu23: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add140) astype47: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu23, dtype="float16") lv226 = R.call_tir(cls.NT_matmul13, (astype47, gpt_neox_layers_7_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add141: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv226, gpt_neox_layers_7_mlp_dense_4h_to_h_bias2) astype48: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add141, dtype="float16") add142: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype48, add139) add143: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add142, add137) layer_norm49: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add143, gpt_neox_layers_8_input_layernorm_weight2, gpt_neox_layers_8_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv227 = R.call_tir(cls.NT_matmul10, (layer_norm49, gpt_neox_layers_8_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add144: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv227, gpt_neox_layers_8_attention_query_key_value_bias2) reshape96: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add144, R.shape([1, 1, 24, 256])) reshape97: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape96, R.shape([1, 24, 256])) lv25 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), reshape97), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape98: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv25, R.shape([1, 1, 8, 256])) reshape99: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape98, R.shape([1, 1, 2048])) lv228 = R.call_tir(cls.NT_matmul11, (reshape99, gpt_neox_layers_8_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add145: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv228, gpt_neox_layers_8_attention_dense_bias2) layer_norm50: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add143, gpt_neox_layers_8_post_attention_layernorm_weight2, gpt_neox_layers_8_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv229 = R.call_tir(cls.NT_matmul12, (layer_norm50, gpt_neox_layers_8_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add146: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv229, gpt_neox_layers_8_mlp_dense_h_to_4h_bias2) gelu24: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add146) astype49: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu24, dtype="float16") lv230 = R.call_tir(cls.NT_matmul13, (astype49, gpt_neox_layers_8_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add147: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv230, gpt_neox_layers_8_mlp_dense_4h_to_h_bias2) astype50: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add147, dtype="float16") add148: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype50, add145) add149: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add148, add143) layer_norm51: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add149, gpt_neox_layers_9_input_layernorm_weight2, gpt_neox_layers_9_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv231 = R.call_tir(cls.NT_matmul10, (layer_norm51, gpt_neox_layers_9_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add150: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv231, gpt_neox_layers_9_attention_query_key_value_bias2) reshape100: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add150, R.shape([1, 1, 24, 256])) reshape101: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape100, R.shape([1, 24, 256])) lv26 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), reshape101), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape102: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv26, R.shape([1, 1, 8, 256])) reshape103: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape102, R.shape([1, 1, 2048])) lv232 = R.call_tir(cls.NT_matmul11, (reshape103, gpt_neox_layers_9_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add151: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv232, gpt_neox_layers_9_attention_dense_bias2) layer_norm52: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add149, gpt_neox_layers_9_post_attention_layernorm_weight2, gpt_neox_layers_9_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv233 = R.call_tir(cls.NT_matmul12, (layer_norm52, gpt_neox_layers_9_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add152: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv233, gpt_neox_layers_9_mlp_dense_h_to_4h_bias2) gelu25: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add152) astype51: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu25, dtype="float16") lv234 = R.call_tir(cls.NT_matmul13, (astype51, gpt_neox_layers_9_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add153: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv234, gpt_neox_layers_9_mlp_dense_4h_to_h_bias2) astype52: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add153, dtype="float16") add154: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype52, add151) add155: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add154, add149) layer_norm53: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add155, gpt_neox_layers_10_input_layernorm_weight2, gpt_neox_layers_10_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv235 = R.call_tir(cls.NT_matmul10, (layer_norm53, gpt_neox_layers_10_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add156: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv235, gpt_neox_layers_10_attention_query_key_value_bias2) reshape104: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add156, R.shape([1, 1, 24, 256])) reshape105: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape104, R.shape([1, 24, 256])) lv27 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), reshape105), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape106: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv27, R.shape([1, 1, 8, 256])) reshape107: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape106, R.shape([1, 1, 2048])) lv236 = R.call_tir(cls.NT_matmul11, (reshape107, gpt_neox_layers_10_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add157: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv236, gpt_neox_layers_10_attention_dense_bias2) layer_norm54: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add155, gpt_neox_layers_10_post_attention_layernorm_weight2, gpt_neox_layers_10_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv237 = R.call_tir(cls.NT_matmul12, (layer_norm54, gpt_neox_layers_10_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add158: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv237, gpt_neox_layers_10_mlp_dense_h_to_4h_bias2) gelu26: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add158) astype53: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu26, dtype="float16") lv238 = R.call_tir(cls.NT_matmul13, (astype53, gpt_neox_layers_10_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add159: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv238, gpt_neox_layers_10_mlp_dense_4h_to_h_bias2) astype54: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add159, dtype="float16") add160: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype54, add157) add161: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add160, add155) layer_norm55: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add161, gpt_neox_layers_11_input_layernorm_weight2, gpt_neox_layers_11_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv239 = R.call_tir(cls.NT_matmul10, (layer_norm55, gpt_neox_layers_11_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add162: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv239, gpt_neox_layers_11_attention_query_key_value_bias2) reshape108: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add162, R.shape([1, 1, 24, 256])) reshape109: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape108, R.shape([1, 24, 256])) lv28 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), reshape109), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape110: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv28, R.shape([1, 1, 8, 256])) reshape111: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape110, R.shape([1, 1, 2048])) lv240 = R.call_tir(cls.NT_matmul11, (reshape111, gpt_neox_layers_11_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add163: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv240, gpt_neox_layers_11_attention_dense_bias2) layer_norm56: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add161, gpt_neox_layers_11_post_attention_layernorm_weight2, gpt_neox_layers_11_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv241 = R.call_tir(cls.NT_matmul12, (layer_norm56, gpt_neox_layers_11_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add164: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv241, gpt_neox_layers_11_mlp_dense_h_to_4h_bias2) gelu27: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add164) astype55: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu27, dtype="float16") lv242 = R.call_tir(cls.NT_matmul13, (astype55, gpt_neox_layers_11_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add165: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv242, gpt_neox_layers_11_mlp_dense_4h_to_h_bias2) astype56: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add165, dtype="float16") add166: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype56, add163) add167: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add166, add161) layer_norm57: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add167, gpt_neox_layers_12_input_layernorm_weight2, gpt_neox_layers_12_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv243 = R.call_tir(cls.NT_matmul10, (layer_norm57, gpt_neox_layers_12_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add168: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv243, gpt_neox_layers_12_attention_query_key_value_bias2) reshape112: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add168, R.shape([1, 1, 24, 256])) reshape113: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape112, R.shape([1, 24, 256])) lv29 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape113), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape114: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv29, R.shape([1, 1, 8, 256])) reshape115: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape114, R.shape([1, 1, 2048])) lv244 = R.call_tir(cls.NT_matmul11, (reshape115, gpt_neox_layers_12_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add169: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv244, gpt_neox_layers_12_attention_dense_bias2) layer_norm58: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add167, gpt_neox_layers_12_post_attention_layernorm_weight2, gpt_neox_layers_12_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv245 = R.call_tir(cls.NT_matmul12, (layer_norm58, gpt_neox_layers_12_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add170: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv245, gpt_neox_layers_12_mlp_dense_h_to_4h_bias2) gelu28: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add170) astype57: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu28, dtype="float16") lv246 = R.call_tir(cls.NT_matmul13, (astype57, gpt_neox_layers_12_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add171: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv246, gpt_neox_layers_12_mlp_dense_4h_to_h_bias2) astype58: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add171, dtype="float16") add172: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype58, add169) add173: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add172, add167) layer_norm59: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add173, gpt_neox_layers_13_input_layernorm_weight2, gpt_neox_layers_13_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv247 = R.call_tir(cls.NT_matmul10, (layer_norm59, gpt_neox_layers_13_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add174: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv247, gpt_neox_layers_13_attention_query_key_value_bias2) reshape116: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add174, R.shape([1, 1, 24, 256])) reshape117: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape116, R.shape([1, 24, 256])) lv30 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape117), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape118: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv30, R.shape([1, 1, 8, 256])) reshape119: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape118, R.shape([1, 1, 2048])) lv248 = R.call_tir(cls.NT_matmul11, (reshape119, gpt_neox_layers_13_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add175: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv248, gpt_neox_layers_13_attention_dense_bias2) layer_norm60: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add173, gpt_neox_layers_13_post_attention_layernorm_weight2, gpt_neox_layers_13_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv249 = R.call_tir(cls.NT_matmul12, (layer_norm60, gpt_neox_layers_13_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add176: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv249, gpt_neox_layers_13_mlp_dense_h_to_4h_bias2) gelu29: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add176) astype59: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu29, dtype="float16") lv250 = R.call_tir(cls.NT_matmul13, (astype59, gpt_neox_layers_13_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add177: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv250, gpt_neox_layers_13_mlp_dense_4h_to_h_bias2) astype60: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add177, dtype="float16") add178: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype60, add175) add179: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add178, add173) layer_norm61: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add179, gpt_neox_layers_14_input_layernorm_weight2, gpt_neox_layers_14_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv251 = R.call_tir(cls.NT_matmul10, (layer_norm61, gpt_neox_layers_14_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add180: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv251, gpt_neox_layers_14_attention_query_key_value_bias2) reshape120: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add180, R.shape([1, 1, 24, 256])) reshape121: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape120, R.shape([1, 24, 256])) lv31 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape121), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape122: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv31, R.shape([1, 1, 8, 256])) reshape123: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape122, R.shape([1, 1, 2048])) lv252 = R.call_tir(cls.NT_matmul11, (reshape123, gpt_neox_layers_14_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add181: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv252, gpt_neox_layers_14_attention_dense_bias2) layer_norm62: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add179, gpt_neox_layers_14_post_attention_layernorm_weight2, gpt_neox_layers_14_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv253 = R.call_tir(cls.NT_matmul12, (layer_norm62, gpt_neox_layers_14_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add182: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv253, gpt_neox_layers_14_mlp_dense_h_to_4h_bias2) gelu30: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add182) astype61: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu30, dtype="float16") lv254 = R.call_tir(cls.NT_matmul13, (astype61, gpt_neox_layers_14_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add183: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv254, gpt_neox_layers_14_mlp_dense_4h_to_h_bias2) astype62: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add183, dtype="float16") add184: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype62, add181) add185: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add184, add179) layer_norm63: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add185, gpt_neox_layers_15_input_layernorm_weight2, gpt_neox_layers_15_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv255 = R.call_tir(cls.NT_matmul10, (layer_norm63, gpt_neox_layers_15_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) add186: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv255, gpt_neox_layers_15_attention_query_key_value_bias2) reshape124: R.Tensor((1, 1, 24, 256), dtype="float16") = R.reshape(add186, R.shape([1, 1, 24, 256])) reshape125: R.Tensor((1, 24, 256), dtype="float16") = R.reshape(reshape124, R.shape([1, 24, 256])) lv32 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape125), out_sinfo=R.Tensor((1, 8, 256), dtype="float16")) reshape126: R.Tensor((1, 1, 8, 256), dtype="float16") = R.reshape(lv32, R.shape([1, 1, 8, 256])) reshape127: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape126, R.shape([1, 1, 2048])) lv256 = R.call_tir(cls.NT_matmul11, (reshape127, gpt_neox_layers_15_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) add187: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv256, gpt_neox_layers_15_attention_dense_bias2) layer_norm64: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add185, gpt_neox_layers_15_post_attention_layernorm_weight2, gpt_neox_layers_15_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv257 = R.call_tir(cls.NT_matmul12, (layer_norm64, gpt_neox_layers_15_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) add188: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv257, gpt_neox_layers_15_mlp_dense_h_to_4h_bias2) gelu31: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add188) astype63: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu31, dtype="float16") lv258 = R.call_tir(cls.NT_matmul13, (astype63, gpt_neox_layers_15_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) add189: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv258, gpt_neox_layers_15_mlp_dense_4h_to_h_bias2) astype64: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add189, dtype="float16") add190: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype64, add187) add191: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add190, add185) layer_norm65: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add191, gpt_neox_final_layer_norm_weight2, gpt_neox_final_layer_norm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv259 = R.call_tir(cls.NT_matmul14, (layer_norm65, embed_out_weight2), out_sinfo=R.Tensor((1, 1, vocab_size), dtype="float16")) astype65: R.Tensor((1, 1, vocab_size), dtype="float32") = R.astype(lv259, dtype="float32") gv2: R.Tuple(R.Tensor((1, 1, vocab_size), dtype="float32"), R.Object) = astype65, 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", 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 2048), dtype="float16"))) -> R.Tensor(("seq_len", 2048), dtype="float16"): seq_len = T.int64() vocab_size = T.int64() R.func_attr({"num_input": 1, "relax.memory_plan_dynamic_func_output": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) with R.dataflow(): gpt_neox_embed_in_weight: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[0] take: R.Tensor((seq_len, 2048), dtype="float16") = R.take(gpt_neox_embed_in_weight, input_ids, axis=0) gv: R.Tensor((seq_len, 2048), dtype="float16") = take R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul(gpt_neox_layers_0_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16"), layer_norm99: R.Tensor(("batch_size", 1, 2048), dtype="float16")) -> R.Tensor(("batch_size", 1, 6144), dtype="float16"): batch_size = T.int64() R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims195: R.Tensor((2048, 6144), dtype="float16") = R.permute_dims(gpt_neox_layers_0_attention_query_key_value_weight4, axes=None) gv: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.matmul(layer_norm99, permute_dims195, out_dtype="void") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul1(gpt_neox_layers_0_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16"), reshape195: R.Tensor(("batch_size", 1, 2048), dtype="float16")) -> R.Tensor(("batch_size", 1, 2048), dtype="float16"): batch_size = T.int64() R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims196: R.Tensor((2048, 2048), dtype="float16") = R.permute_dims(gpt_neox_layers_0_attention_dense_weight4, axes=None) gv: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.matmul(reshape195, permute_dims196, out_dtype="void") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul10(gpt_neox_layers_0_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16"), layer_norm33: R.Tensor((1, 1, 2048), dtype="float16")) -> R.Tensor((1, 1, 6144), dtype="float16"): R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims65: R.Tensor((2048, 6144), dtype="float16") = R.permute_dims(gpt_neox_layers_0_attention_query_key_value_weight2, axes=None) gv: R.Tensor((1, 1, 6144), dtype="float16") = R.matmul(layer_norm33, permute_dims65, out_dtype="void") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul11(gpt_neox_layers_0_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16"), reshape67: R.Tensor((1, 1, 2048), dtype="float16")) -> R.Tensor((1, 1, 2048), dtype="float16"): R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims66: R.Tensor((2048, 2048), dtype="float16") = R.permute_dims(gpt_neox_layers_0_attention_dense_weight2, axes=None) gv: R.Tensor((1, 1, 2048), dtype="float16") = R.matmul(reshape67, permute_dims66, out_dtype="void") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul12(gpt_neox_layers_0_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16"), layer_norm34: R.Tensor((1, 1, 2048), dtype="float16")) -> R.Tensor((1, 1, 8192), dtype="float32"): R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims67: R.Tensor((2048, 8192), dtype="float16") = R.permute_dims(gpt_neox_layers_0_mlp_dense_h_to_4h_weight2, axes=None) gv: R.Tensor((1, 1, 8192), dtype="float32") = R.matmul(layer_norm34, permute_dims67, out_dtype="float32") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul13(gpt_neox_layers_0_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16"), astype33: R.Tensor((1, 1, 8192), dtype="float16")) -> R.Tensor((1, 1, 2048), dtype="float32"): R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims68: R.Tensor((8192, 2048), dtype="float16") = R.permute_dims(gpt_neox_layers_0_mlp_dense_4h_to_h_weight2, axes=None) gv: R.Tensor((1, 1, 2048), dtype="float32") = R.matmul(astype33, permute_dims68, out_dtype="float32") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul14(embed_out_weight2: R.Tensor(("vocab_size", 2048), dtype="float16"), layer_norm65: R.Tensor((1, 1, 2048), dtype="float16")) -> R.Tensor((1, 1, "vocab_size"), dtype="float16"): vocab_size = T.int64() R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims129: R.Tensor((2048, vocab_size), dtype="float16") = R.permute_dims(embed_out_weight2, axes=None) gv: R.Tensor((1, 1, vocab_size), dtype="float16") = R.matmul(layer_norm65, permute_dims129, out_dtype="void") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul2(gpt_neox_layers_0_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16"), layer_norm100: R.Tensor(("batch_size", 1, 2048), dtype="float16")) -> R.Tensor(("batch_size", 1, 8192), dtype="float32"): batch_size = T.int64() R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims197: R.Tensor((2048, 8192), dtype="float16") = R.permute_dims(gpt_neox_layers_0_mlp_dense_h_to_4h_weight4, axes=None) gv: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.matmul(layer_norm100, permute_dims197, out_dtype="float32") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul3(gpt_neox_layers_0_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16"), astype99: R.Tensor(("batch_size", 1, 8192), dtype="float16")) -> R.Tensor(("batch_size", 1, 2048), dtype="float32"): batch_size = T.int64() R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims198: R.Tensor((8192, 2048), dtype="float16") = R.permute_dims(gpt_neox_layers_0_mlp_dense_4h_to_h_weight4, axes=None) gv: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.matmul(astype99, permute_dims198, out_dtype="float32") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul4(embed_out_weight4: R.Tensor(("vocab_size", 2048), dtype="float16"), layer_norm131: R.Tensor(("batch_size", 1, 2048), dtype="float16")) -> R.Tensor(("batch_size", 1, "vocab_size"), dtype="float16"): batch_size = T.int64() vocab_size = T.int64() R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims259: R.Tensor((2048, vocab_size), dtype="float16") = R.permute_dims(embed_out_weight4, axes=None) gv: R.Tensor((batch_size, 1, vocab_size), dtype="float16") = R.matmul(layer_norm131, permute_dims259, out_dtype="void") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul5(gpt_neox_layers_0_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16"), layer_norm66: R.Tensor((1, "seq_len", 2048), dtype="float16")) -> R.Tensor((1, "seq_len", 6144), dtype="float16"): seq_len = T.int64() R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims130: R.Tensor((2048, 6144), dtype="float16") = R.permute_dims(gpt_neox_layers_0_attention_query_key_value_weight3, axes=None) gv: R.Tensor((1, seq_len, 6144), dtype="float16") = R.matmul(layer_norm66, permute_dims130, out_dtype="void") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul6(gpt_neox_layers_0_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16"), reshape131: R.Tensor((1, "seq_len", 2048), dtype="float16")) -> R.Tensor((1, "seq_len", 2048), dtype="float16"): seq_len = T.int64() R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims131: R.Tensor((2048, 2048), dtype="float16") = R.permute_dims(gpt_neox_layers_0_attention_dense_weight3, axes=None) gv: R.Tensor((1, seq_len, 2048), dtype="float16") = R.matmul(reshape131, permute_dims131, out_dtype="void") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul7(gpt_neox_layers_0_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16"), layer_norm67: R.Tensor((1, "seq_len", 2048), dtype="float16")) -> R.Tensor((1, "seq_len", 8192), dtype="float32"): seq_len = T.int64() R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims132: R.Tensor((2048, 8192), dtype="float16") = R.permute_dims(gpt_neox_layers_0_mlp_dense_h_to_4h_weight3, axes=None) gv: R.Tensor((1, seq_len, 8192), dtype="float32") = R.matmul(layer_norm67, permute_dims132, out_dtype="float32") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul8(gpt_neox_layers_0_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16"), astype66: R.Tensor((1, "seq_len", 8192), dtype="float16")) -> R.Tensor((1, "seq_len", 2048), dtype="float32"): seq_len = T.int64() R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims133: R.Tensor((8192, 2048), dtype="float16") = R.permute_dims(gpt_neox_layers_0_mlp_dense_4h_to_h_weight3, axes=None) gv: R.Tensor((1, seq_len, 2048), dtype="float32") = R.matmul(astype66, permute_dims133, out_dtype="float32") R.output(gv) return gv @R.function(private=True) def fused_relax_permute_dims_relax_matmul9(embed_out_weight3: R.Tensor(("vocab_size", 2048), dtype="float16"), take1: R.Tensor((1, "batch_size", 2048), dtype="float16")) -> R.Tensor((1, "batch_size", "vocab_size"), dtype="float16"): batch_size = T.int64() vocab_size = T.int64() R.func_attr({"Composite": "transpose_matmul_fuse", "Primitive": 1}) with R.dataflow(): permute_dims194: R.Tensor((2048, vocab_size), dtype="float16") = R.permute_dims(embed_out_weight3, axes=None) gv: R.Tensor((1, batch_size, vocab_size), dtype="float16") = R.matmul(take1, permute_dims194, out_dtype="void") R.output(gv) return gv @R.function def prefill(input_embed: R.Tensor((1, "seq_len", 2048), dtype="float16"), paged_kv_cache: R.Object, packed_params: R.Tuple(R.Tensor(("vocab_size", 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((6144, 2048), dtype="float16"), R.Tensor((6144,), dtype="float16"), R.Tensor((2048, 2048), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor((8192, 2048), dtype="float16"), R.Tensor((8192,), dtype="float32"), R.Tensor((2048, 8192), dtype="float16"), R.Tensor((2048,), dtype="float32"), R.Tensor((2048,), dtype="float16"), R.Tensor((2048,), dtype="float16"), R.Tensor(("vocab_size", 2048), dtype="float16"))) -> R.Tuple(R.Tensor((1, 1, "vocab_size"), dtype="float32"), R.Object): vocab_size = T.int64() seq_len = T.int64() R.func_attr({"num_input": 2, "pipeline_parallel_stages": 1, "relax.memory_plan_dynamic_func_output": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) cls = Module with R.dataflow(): gpt_neox_layers_0_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[1] gpt_neox_layers_0_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[2] gpt_neox_layers_0_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[3] gpt_neox_layers_0_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[4] gpt_neox_layers_0_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[5] gpt_neox_layers_0_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[6] gpt_neox_layers_0_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[7] gpt_neox_layers_0_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[8] gpt_neox_layers_0_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[9] gpt_neox_layers_0_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[10] gpt_neox_layers_0_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[11] gpt_neox_layers_0_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[12] gpt_neox_layers_1_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[13] gpt_neox_layers_1_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[14] gpt_neox_layers_1_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[15] gpt_neox_layers_1_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[16] gpt_neox_layers_1_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[17] gpt_neox_layers_1_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[18] gpt_neox_layers_1_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[19] gpt_neox_layers_1_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[20] gpt_neox_layers_1_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[21] gpt_neox_layers_1_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[22] gpt_neox_layers_1_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[23] gpt_neox_layers_1_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[24] gpt_neox_layers_2_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[25] gpt_neox_layers_2_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[26] gpt_neox_layers_2_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[27] gpt_neox_layers_2_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[28] gpt_neox_layers_2_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[29] gpt_neox_layers_2_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[30] gpt_neox_layers_2_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[31] gpt_neox_layers_2_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[32] gpt_neox_layers_2_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[33] gpt_neox_layers_2_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[34] gpt_neox_layers_2_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[35] gpt_neox_layers_2_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[36] gpt_neox_layers_3_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[37] gpt_neox_layers_3_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[38] gpt_neox_layers_3_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[39] gpt_neox_layers_3_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[40] gpt_neox_layers_3_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[41] gpt_neox_layers_3_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[42] gpt_neox_layers_3_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[43] gpt_neox_layers_3_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[44] gpt_neox_layers_3_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[45] gpt_neox_layers_3_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[46] gpt_neox_layers_3_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[47] gpt_neox_layers_3_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[48] gpt_neox_layers_4_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[49] gpt_neox_layers_4_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[50] gpt_neox_layers_4_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[51] gpt_neox_layers_4_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[52] gpt_neox_layers_4_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[53] gpt_neox_layers_4_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[54] gpt_neox_layers_4_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[55] gpt_neox_layers_4_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[56] gpt_neox_layers_4_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[57] gpt_neox_layers_4_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[58] gpt_neox_layers_4_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[59] gpt_neox_layers_4_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[60] gpt_neox_layers_5_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[61] gpt_neox_layers_5_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[62] gpt_neox_layers_5_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[63] gpt_neox_layers_5_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[64] gpt_neox_layers_5_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[65] gpt_neox_layers_5_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[66] gpt_neox_layers_5_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[67] gpt_neox_layers_5_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[68] gpt_neox_layers_5_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[69] gpt_neox_layers_5_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[70] gpt_neox_layers_5_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[71] gpt_neox_layers_5_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[72] gpt_neox_layers_6_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[73] gpt_neox_layers_6_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[74] gpt_neox_layers_6_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[75] gpt_neox_layers_6_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[76] gpt_neox_layers_6_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[77] gpt_neox_layers_6_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[78] gpt_neox_layers_6_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[79] gpt_neox_layers_6_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[80] gpt_neox_layers_6_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[81] gpt_neox_layers_6_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[82] gpt_neox_layers_6_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[83] gpt_neox_layers_6_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[84] gpt_neox_layers_7_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[85] gpt_neox_layers_7_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[86] gpt_neox_layers_7_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[87] gpt_neox_layers_7_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[88] gpt_neox_layers_7_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[89] gpt_neox_layers_7_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[90] gpt_neox_layers_7_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[91] gpt_neox_layers_7_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[92] gpt_neox_layers_7_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[93] gpt_neox_layers_7_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[94] gpt_neox_layers_7_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[95] gpt_neox_layers_7_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[96] gpt_neox_layers_8_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[97] gpt_neox_layers_8_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[98] gpt_neox_layers_8_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[99] gpt_neox_layers_8_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[100] gpt_neox_layers_8_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[101] gpt_neox_layers_8_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[102] gpt_neox_layers_8_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[103] gpt_neox_layers_8_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[104] gpt_neox_layers_8_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[105] gpt_neox_layers_8_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[106] gpt_neox_layers_8_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[107] gpt_neox_layers_8_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[108] gpt_neox_layers_9_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[109] gpt_neox_layers_9_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[110] gpt_neox_layers_9_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[111] gpt_neox_layers_9_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[112] gpt_neox_layers_9_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[113] gpt_neox_layers_9_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[114] gpt_neox_layers_9_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[115] gpt_neox_layers_9_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[116] gpt_neox_layers_9_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[117] gpt_neox_layers_9_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[118] gpt_neox_layers_9_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[119] gpt_neox_layers_9_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[120] gpt_neox_layers_10_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[121] gpt_neox_layers_10_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[122] gpt_neox_layers_10_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[123] gpt_neox_layers_10_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[124] gpt_neox_layers_10_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[125] gpt_neox_layers_10_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[126] gpt_neox_layers_10_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[127] gpt_neox_layers_10_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[128] gpt_neox_layers_10_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[129] gpt_neox_layers_10_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[130] gpt_neox_layers_10_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[131] gpt_neox_layers_10_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[132] gpt_neox_layers_11_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[133] gpt_neox_layers_11_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[134] gpt_neox_layers_11_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[135] gpt_neox_layers_11_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[136] gpt_neox_layers_11_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[137] gpt_neox_layers_11_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[138] gpt_neox_layers_11_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[139] gpt_neox_layers_11_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[140] gpt_neox_layers_11_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[141] gpt_neox_layers_11_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[142] gpt_neox_layers_11_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[143] gpt_neox_layers_11_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[144] gpt_neox_layers_12_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[145] gpt_neox_layers_12_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[146] gpt_neox_layers_12_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[147] gpt_neox_layers_12_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[148] gpt_neox_layers_12_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[149] gpt_neox_layers_12_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[150] gpt_neox_layers_12_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[151] gpt_neox_layers_12_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[152] gpt_neox_layers_12_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[153] gpt_neox_layers_12_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[154] gpt_neox_layers_12_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[155] gpt_neox_layers_12_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[156] gpt_neox_layers_13_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[157] gpt_neox_layers_13_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[158] gpt_neox_layers_13_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[159] gpt_neox_layers_13_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[160] gpt_neox_layers_13_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[161] gpt_neox_layers_13_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[162] gpt_neox_layers_13_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[163] gpt_neox_layers_13_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[164] gpt_neox_layers_13_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[165] gpt_neox_layers_13_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[166] gpt_neox_layers_13_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[167] gpt_neox_layers_13_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[168] gpt_neox_layers_14_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[169] gpt_neox_layers_14_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[170] gpt_neox_layers_14_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[171] gpt_neox_layers_14_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[172] gpt_neox_layers_14_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[173] gpt_neox_layers_14_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[174] gpt_neox_layers_14_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[175] gpt_neox_layers_14_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[176] gpt_neox_layers_14_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[177] gpt_neox_layers_14_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[178] gpt_neox_layers_14_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[179] gpt_neox_layers_14_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[180] gpt_neox_layers_15_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[181] gpt_neox_layers_15_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[182] gpt_neox_layers_15_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[183] gpt_neox_layers_15_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[184] gpt_neox_layers_15_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[185] gpt_neox_layers_15_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[186] gpt_neox_layers_15_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[187] gpt_neox_layers_15_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[188] gpt_neox_layers_15_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[189] gpt_neox_layers_15_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[190] gpt_neox_layers_15_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[191] gpt_neox_layers_15_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[192] gpt_neox_final_layer_norm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[193] gpt_neox_final_layer_norm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[194] embed_out_weight1: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[195] layer_norm: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(input_embed, gpt_neox_layers_0_input_layernorm_weight1, gpt_neox_layers_0_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv260 = R.call_tir(cls.NT_matmul5, (layer_norm, gpt_neox_layers_0_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv260, gpt_neox_layers_0_attention_query_key_value_bias1) reshape: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add, R.shape([1, seq_len, 24, 256])) reshape1: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape, R.shape([seq_len, 24, 256])) 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, 256), dtype="float16")) reshape2: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv, R.shape([1, seq_len, 8, 256])) reshape3: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape2, R.shape([1, seq_len, 2048])) lv261 = R.call_tir(cls.NT_matmul6, (reshape3, gpt_neox_layers_0_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add1: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv261, gpt_neox_layers_0_attention_dense_bias1) layer_norm1: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(input_embed, gpt_neox_layers_0_post_attention_layernorm_weight1, gpt_neox_layers_0_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv262 = R.call_tir(cls.NT_matmul7, (layer_norm1, gpt_neox_layers_0_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add2: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv262, gpt_neox_layers_0_mlp_dense_h_to_4h_bias1) gelu: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add2) astype: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu, dtype="float16") lv263 = R.call_tir(cls.NT_matmul8, (astype, gpt_neox_layers_0_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add3: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv263, gpt_neox_layers_0_mlp_dense_4h_to_h_bias1) astype1: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add3, dtype="float16") add4: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype1, add1) add5: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add4, input_embed) layer_norm2: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add5, gpt_neox_layers_1_input_layernorm_weight1, gpt_neox_layers_1_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv264 = R.call_tir(cls.NT_matmul5, (layer_norm2, gpt_neox_layers_1_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add6: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv264, gpt_neox_layers_1_attention_query_key_value_bias1) reshape4: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add6, R.shape([1, seq_len, 24, 256])) reshape5: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape4, R.shape([seq_len, 24, 256])) 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, 256), dtype="float16")) reshape6: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv1, R.shape([1, seq_len, 8, 256])) reshape7: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape6, R.shape([1, seq_len, 2048])) lv265 = R.call_tir(cls.NT_matmul6, (reshape7, gpt_neox_layers_1_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add7: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv265, gpt_neox_layers_1_attention_dense_bias1) layer_norm3: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add5, gpt_neox_layers_1_post_attention_layernorm_weight1, gpt_neox_layers_1_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv266 = R.call_tir(cls.NT_matmul7, (layer_norm3, gpt_neox_layers_1_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add8: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv266, gpt_neox_layers_1_mlp_dense_h_to_4h_bias1) gelu1: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add8) astype2: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu1, dtype="float16") lv267 = R.call_tir(cls.NT_matmul8, (astype2, gpt_neox_layers_1_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add9: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv267, gpt_neox_layers_1_mlp_dense_4h_to_h_bias1) astype3: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add9, dtype="float16") add10: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype3, add7) add11: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add10, add5) layer_norm4: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add11, gpt_neox_layers_2_input_layernorm_weight1, gpt_neox_layers_2_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv268 = R.call_tir(cls.NT_matmul5, (layer_norm4, gpt_neox_layers_2_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add12: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv268, gpt_neox_layers_2_attention_query_key_value_bias1) reshape8: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add12, R.shape([1, seq_len, 24, 256])) reshape9: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape8, R.shape([seq_len, 24, 256])) 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, 256), dtype="float16")) reshape10: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv2, R.shape([1, seq_len, 8, 256])) reshape11: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape10, R.shape([1, seq_len, 2048])) lv269 = R.call_tir(cls.NT_matmul6, (reshape11, gpt_neox_layers_2_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add13: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv269, gpt_neox_layers_2_attention_dense_bias1) layer_norm5: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add11, gpt_neox_layers_2_post_attention_layernorm_weight1, gpt_neox_layers_2_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv270 = R.call_tir(cls.NT_matmul7, (layer_norm5, gpt_neox_layers_2_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add14: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv270, gpt_neox_layers_2_mlp_dense_h_to_4h_bias1) gelu2: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add14) astype4: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu2, dtype="float16") lv271 = R.call_tir(cls.NT_matmul8, (astype4, gpt_neox_layers_2_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add15: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv271, gpt_neox_layers_2_mlp_dense_4h_to_h_bias1) astype5: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add15, dtype="float16") add16: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype5, add13) add17: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add16, add11) layer_norm6: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add17, gpt_neox_layers_3_input_layernorm_weight1, gpt_neox_layers_3_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv272 = R.call_tir(cls.NT_matmul5, (layer_norm6, gpt_neox_layers_3_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add18: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv272, gpt_neox_layers_3_attention_query_key_value_bias1) reshape12: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add18, R.shape([1, seq_len, 24, 256])) reshape13: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape12, R.shape([seq_len, 24, 256])) 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, 256), dtype="float16")) reshape14: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv3, R.shape([1, seq_len, 8, 256])) reshape15: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape14, R.shape([1, seq_len, 2048])) lv273 = R.call_tir(cls.NT_matmul6, (reshape15, gpt_neox_layers_3_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add19: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv273, gpt_neox_layers_3_attention_dense_bias1) layer_norm7: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add17, gpt_neox_layers_3_post_attention_layernorm_weight1, gpt_neox_layers_3_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv274 = R.call_tir(cls.NT_matmul7, (layer_norm7, gpt_neox_layers_3_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add20: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv274, gpt_neox_layers_3_mlp_dense_h_to_4h_bias1) gelu3: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add20) astype6: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu3, dtype="float16") lv275 = R.call_tir(cls.NT_matmul8, (astype6, gpt_neox_layers_3_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add21: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv275, gpt_neox_layers_3_mlp_dense_4h_to_h_bias1) astype7: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add21, dtype="float16") add22: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype7, add19) add23: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add22, add17) layer_norm8: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add23, gpt_neox_layers_4_input_layernorm_weight1, gpt_neox_layers_4_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv276 = R.call_tir(cls.NT_matmul5, (layer_norm8, gpt_neox_layers_4_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add24: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv276, gpt_neox_layers_4_attention_query_key_value_bias1) reshape16: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add24, R.shape([1, seq_len, 24, 256])) reshape17: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape16, R.shape([seq_len, 24, 256])) 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, 256), dtype="float16")) reshape18: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv4, R.shape([1, seq_len, 8, 256])) reshape19: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape18, R.shape([1, seq_len, 2048])) lv277 = R.call_tir(cls.NT_matmul6, (reshape19, gpt_neox_layers_4_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add25: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv277, gpt_neox_layers_4_attention_dense_bias1) layer_norm9: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add23, gpt_neox_layers_4_post_attention_layernorm_weight1, gpt_neox_layers_4_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv278 = R.call_tir(cls.NT_matmul7, (layer_norm9, gpt_neox_layers_4_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add26: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv278, gpt_neox_layers_4_mlp_dense_h_to_4h_bias1) gelu4: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add26) astype8: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu4, dtype="float16") lv279 = R.call_tir(cls.NT_matmul8, (astype8, gpt_neox_layers_4_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add27: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv279, gpt_neox_layers_4_mlp_dense_4h_to_h_bias1) astype9: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add27, dtype="float16") add28: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype9, add25) add29: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add28, add23) layer_norm10: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add29, gpt_neox_layers_5_input_layernorm_weight1, gpt_neox_layers_5_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv280 = R.call_tir(cls.NT_matmul5, (layer_norm10, gpt_neox_layers_5_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add30: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv280, gpt_neox_layers_5_attention_query_key_value_bias1) reshape20: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add30, R.shape([1, seq_len, 24, 256])) reshape21: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape20, R.shape([seq_len, 24, 256])) 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, 256), dtype="float16")) reshape22: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv5, R.shape([1, seq_len, 8, 256])) reshape23: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape22, R.shape([1, seq_len, 2048])) lv281 = R.call_tir(cls.NT_matmul6, (reshape23, gpt_neox_layers_5_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add31: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv281, gpt_neox_layers_5_attention_dense_bias1) layer_norm11: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add29, gpt_neox_layers_5_post_attention_layernorm_weight1, gpt_neox_layers_5_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv282 = R.call_tir(cls.NT_matmul7, (layer_norm11, gpt_neox_layers_5_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add32: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv282, gpt_neox_layers_5_mlp_dense_h_to_4h_bias1) gelu5: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add32) astype10: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu5, dtype="float16") lv283 = R.call_tir(cls.NT_matmul8, (astype10, gpt_neox_layers_5_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add33: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv283, gpt_neox_layers_5_mlp_dense_4h_to_h_bias1) astype11: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add33, dtype="float16") add34: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype11, add31) add35: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add34, add29) layer_norm12: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add35, gpt_neox_layers_6_input_layernorm_weight1, gpt_neox_layers_6_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv284 = R.call_tir(cls.NT_matmul5, (layer_norm12, gpt_neox_layers_6_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add36: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv284, gpt_neox_layers_6_attention_query_key_value_bias1) reshape24: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add36, R.shape([1, seq_len, 24, 256])) reshape25: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape24, R.shape([seq_len, 24, 256])) lv6 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), reshape25), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape26: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv6, R.shape([1, seq_len, 8, 256])) reshape27: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape26, R.shape([1, seq_len, 2048])) lv285 = R.call_tir(cls.NT_matmul6, (reshape27, gpt_neox_layers_6_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add37: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv285, gpt_neox_layers_6_attention_dense_bias1) layer_norm13: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add35, gpt_neox_layers_6_post_attention_layernorm_weight1, gpt_neox_layers_6_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv286 = R.call_tir(cls.NT_matmul7, (layer_norm13, gpt_neox_layers_6_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add38: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv286, gpt_neox_layers_6_mlp_dense_h_to_4h_bias1) gelu6: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add38) astype12: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu6, dtype="float16") lv287 = R.call_tir(cls.NT_matmul8, (astype12, gpt_neox_layers_6_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add39: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv287, gpt_neox_layers_6_mlp_dense_4h_to_h_bias1) astype13: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add39, dtype="float16") add40: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype13, add37) add41: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add40, add35) layer_norm14: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add41, gpt_neox_layers_7_input_layernorm_weight1, gpt_neox_layers_7_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv288 = R.call_tir(cls.NT_matmul5, (layer_norm14, gpt_neox_layers_7_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add42: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv288, gpt_neox_layers_7_attention_query_key_value_bias1) reshape28: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add42, R.shape([1, seq_len, 24, 256])) reshape29: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape28, R.shape([seq_len, 24, 256])) lv7 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), reshape29), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape30: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv7, R.shape([1, seq_len, 8, 256])) reshape31: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape30, R.shape([1, seq_len, 2048])) lv289 = R.call_tir(cls.NT_matmul6, (reshape31, gpt_neox_layers_7_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add43: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv289, gpt_neox_layers_7_attention_dense_bias1) layer_norm15: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add41, gpt_neox_layers_7_post_attention_layernorm_weight1, gpt_neox_layers_7_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv290 = R.call_tir(cls.NT_matmul7, (layer_norm15, gpt_neox_layers_7_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add44: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv290, gpt_neox_layers_7_mlp_dense_h_to_4h_bias1) gelu7: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add44) astype14: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu7, dtype="float16") lv291 = R.call_tir(cls.NT_matmul8, (astype14, gpt_neox_layers_7_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add45: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv291, gpt_neox_layers_7_mlp_dense_4h_to_h_bias1) astype15: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add45, dtype="float16") add46: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype15, add43) add47: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add46, add41) layer_norm16: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add47, gpt_neox_layers_8_input_layernorm_weight1, gpt_neox_layers_8_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv292 = R.call_tir(cls.NT_matmul5, (layer_norm16, gpt_neox_layers_8_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add48: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv292, gpt_neox_layers_8_attention_query_key_value_bias1) reshape32: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add48, R.shape([1, seq_len, 24, 256])) reshape33: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape32, R.shape([seq_len, 24, 256])) lv8 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), reshape33), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape34: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv8, R.shape([1, seq_len, 8, 256])) reshape35: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape34, R.shape([1, seq_len, 2048])) lv293 = R.call_tir(cls.NT_matmul6, (reshape35, gpt_neox_layers_8_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add49: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv293, gpt_neox_layers_8_attention_dense_bias1) layer_norm17: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add47, gpt_neox_layers_8_post_attention_layernorm_weight1, gpt_neox_layers_8_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv294 = R.call_tir(cls.NT_matmul7, (layer_norm17, gpt_neox_layers_8_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add50: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv294, gpt_neox_layers_8_mlp_dense_h_to_4h_bias1) gelu8: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add50) astype16: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu8, dtype="float16") lv295 = R.call_tir(cls.NT_matmul8, (astype16, gpt_neox_layers_8_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add51: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv295, gpt_neox_layers_8_mlp_dense_4h_to_h_bias1) astype17: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add51, dtype="float16") add52: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype17, add49) add53: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add52, add47) layer_norm18: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add53, gpt_neox_layers_9_input_layernorm_weight1, gpt_neox_layers_9_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv296 = R.call_tir(cls.NT_matmul5, (layer_norm18, gpt_neox_layers_9_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add54: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv296, gpt_neox_layers_9_attention_query_key_value_bias1) reshape36: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add54, R.shape([1, seq_len, 24, 256])) reshape37: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape36, R.shape([seq_len, 24, 256])) lv9 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), reshape37), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape38: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv9, R.shape([1, seq_len, 8, 256])) reshape39: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape38, R.shape([1, seq_len, 2048])) lv297 = R.call_tir(cls.NT_matmul6, (reshape39, gpt_neox_layers_9_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add55: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv297, gpt_neox_layers_9_attention_dense_bias1) layer_norm19: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add53, gpt_neox_layers_9_post_attention_layernorm_weight1, gpt_neox_layers_9_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv298 = R.call_tir(cls.NT_matmul7, (layer_norm19, gpt_neox_layers_9_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add56: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv298, gpt_neox_layers_9_mlp_dense_h_to_4h_bias1) gelu9: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add56) astype18: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu9, dtype="float16") lv299 = R.call_tir(cls.NT_matmul8, (astype18, gpt_neox_layers_9_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add57: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv299, gpt_neox_layers_9_mlp_dense_4h_to_h_bias1) astype19: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add57, dtype="float16") add58: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype19, add55) add59: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add58, add53) layer_norm20: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add59, gpt_neox_layers_10_input_layernorm_weight1, gpt_neox_layers_10_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv300 = R.call_tir(cls.NT_matmul5, (layer_norm20, gpt_neox_layers_10_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add60: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv300, gpt_neox_layers_10_attention_query_key_value_bias1) reshape40: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add60, R.shape([1, seq_len, 24, 256])) reshape41: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape40, R.shape([seq_len, 24, 256])) lv10 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), reshape41), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape42: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv10, R.shape([1, seq_len, 8, 256])) reshape43: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape42, R.shape([1, seq_len, 2048])) lv301 = R.call_tir(cls.NT_matmul6, (reshape43, gpt_neox_layers_10_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add61: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv301, gpt_neox_layers_10_attention_dense_bias1) layer_norm21: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add59, gpt_neox_layers_10_post_attention_layernorm_weight1, gpt_neox_layers_10_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv302 = R.call_tir(cls.NT_matmul7, (layer_norm21, gpt_neox_layers_10_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add62: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv302, gpt_neox_layers_10_mlp_dense_h_to_4h_bias1) gelu10: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add62) astype20: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu10, dtype="float16") lv303 = R.call_tir(cls.NT_matmul8, (astype20, gpt_neox_layers_10_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add63: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv303, gpt_neox_layers_10_mlp_dense_4h_to_h_bias1) astype21: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add63, dtype="float16") add64: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype21, add61) add65: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add64, add59) layer_norm22: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add65, gpt_neox_layers_11_input_layernorm_weight1, gpt_neox_layers_11_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv304 = R.call_tir(cls.NT_matmul5, (layer_norm22, gpt_neox_layers_11_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add66: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv304, gpt_neox_layers_11_attention_query_key_value_bias1) reshape44: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add66, R.shape([1, seq_len, 24, 256])) reshape45: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape44, R.shape([seq_len, 24, 256])) lv11 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), reshape45), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape46: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv11, R.shape([1, seq_len, 8, 256])) reshape47: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape46, R.shape([1, seq_len, 2048])) lv305 = R.call_tir(cls.NT_matmul6, (reshape47, gpt_neox_layers_11_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add67: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv305, gpt_neox_layers_11_attention_dense_bias1) layer_norm23: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add65, gpt_neox_layers_11_post_attention_layernorm_weight1, gpt_neox_layers_11_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv306 = R.call_tir(cls.NT_matmul7, (layer_norm23, gpt_neox_layers_11_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add68: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv306, gpt_neox_layers_11_mlp_dense_h_to_4h_bias1) gelu11: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add68) astype22: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu11, dtype="float16") lv307 = R.call_tir(cls.NT_matmul8, (astype22, gpt_neox_layers_11_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add69: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv307, gpt_neox_layers_11_mlp_dense_4h_to_h_bias1) astype23: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add69, dtype="float16") add70: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype23, add67) add71: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add70, add65) layer_norm24: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add71, gpt_neox_layers_12_input_layernorm_weight1, gpt_neox_layers_12_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv308 = R.call_tir(cls.NT_matmul5, (layer_norm24, gpt_neox_layers_12_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add72: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv308, gpt_neox_layers_12_attention_query_key_value_bias1) reshape48: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add72, R.shape([1, seq_len, 24, 256])) reshape49: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape48, R.shape([seq_len, 24, 256])) lv12 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape49), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape50: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv12, R.shape([1, seq_len, 8, 256])) reshape51: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape50, R.shape([1, seq_len, 2048])) lv309 = R.call_tir(cls.NT_matmul6, (reshape51, gpt_neox_layers_12_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add73: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv309, gpt_neox_layers_12_attention_dense_bias1) layer_norm25: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add71, gpt_neox_layers_12_post_attention_layernorm_weight1, gpt_neox_layers_12_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv310 = R.call_tir(cls.NT_matmul7, (layer_norm25, gpt_neox_layers_12_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add74: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv310, gpt_neox_layers_12_mlp_dense_h_to_4h_bias1) gelu12: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add74) astype24: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu12, dtype="float16") lv311 = R.call_tir(cls.NT_matmul8, (astype24, gpt_neox_layers_12_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add75: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv311, gpt_neox_layers_12_mlp_dense_4h_to_h_bias1) astype25: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add75, dtype="float16") add76: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype25, add73) add77: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add76, add71) layer_norm26: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add77, gpt_neox_layers_13_input_layernorm_weight1, gpt_neox_layers_13_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv312 = R.call_tir(cls.NT_matmul5, (layer_norm26, gpt_neox_layers_13_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add78: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv312, gpt_neox_layers_13_attention_query_key_value_bias1) reshape52: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add78, R.shape([1, seq_len, 24, 256])) reshape53: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape52, R.shape([seq_len, 24, 256])) lv13 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape53), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape54: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv13, R.shape([1, seq_len, 8, 256])) reshape55: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape54, R.shape([1, seq_len, 2048])) lv313 = R.call_tir(cls.NT_matmul6, (reshape55, gpt_neox_layers_13_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add79: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv313, gpt_neox_layers_13_attention_dense_bias1) layer_norm27: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add77, gpt_neox_layers_13_post_attention_layernorm_weight1, gpt_neox_layers_13_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv314 = R.call_tir(cls.NT_matmul7, (layer_norm27, gpt_neox_layers_13_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add80: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv314, gpt_neox_layers_13_mlp_dense_h_to_4h_bias1) gelu13: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add80) astype26: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu13, dtype="float16") lv315 = R.call_tir(cls.NT_matmul8, (astype26, gpt_neox_layers_13_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add81: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv315, gpt_neox_layers_13_mlp_dense_4h_to_h_bias1) astype27: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add81, dtype="float16") add82: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype27, add79) add83: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add82, add77) layer_norm28: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add83, gpt_neox_layers_14_input_layernorm_weight1, gpt_neox_layers_14_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv316 = R.call_tir(cls.NT_matmul5, (layer_norm28, gpt_neox_layers_14_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add84: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv316, gpt_neox_layers_14_attention_query_key_value_bias1) reshape56: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add84, R.shape([1, seq_len, 24, 256])) reshape57: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape56, R.shape([seq_len, 24, 256])) lv14 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape57), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape58: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv14, R.shape([1, seq_len, 8, 256])) reshape59: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape58, R.shape([1, seq_len, 2048])) lv317 = R.call_tir(cls.NT_matmul6, (reshape59, gpt_neox_layers_14_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add85: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv317, gpt_neox_layers_14_attention_dense_bias1) layer_norm29: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add83, gpt_neox_layers_14_post_attention_layernorm_weight1, gpt_neox_layers_14_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv318 = R.call_tir(cls.NT_matmul7, (layer_norm29, gpt_neox_layers_14_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add86: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv318, gpt_neox_layers_14_mlp_dense_h_to_4h_bias1) gelu14: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add86) astype28: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu14, dtype="float16") lv319 = R.call_tir(cls.NT_matmul8, (astype28, gpt_neox_layers_14_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add87: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv319, gpt_neox_layers_14_mlp_dense_4h_to_h_bias1) astype29: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add87, dtype="float16") add88: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype29, add85) add89: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add88, add83) layer_norm30: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add89, gpt_neox_layers_15_input_layernorm_weight1, gpt_neox_layers_15_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv320 = R.call_tir(cls.NT_matmul5, (layer_norm30, gpt_neox_layers_15_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) add90: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv320, gpt_neox_layers_15_attention_query_key_value_bias1) reshape60: R.Tensor((1, seq_len, 24, 256), dtype="float16") = R.reshape(add90, R.shape([1, seq_len, 24, 256])) reshape61: R.Tensor((seq_len, 24, 256), dtype="float16") = R.reshape(reshape60, R.shape([seq_len, 24, 256])) lv15 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape61), out_sinfo=R.Tensor((seq_len, 8, 256), dtype="float16")) reshape62: R.Tensor((1, seq_len, 8, 256), dtype="float16") = R.reshape(lv15, R.shape([1, seq_len, 8, 256])) reshape63: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape62, R.shape([1, seq_len, 2048])) lv321 = R.call_tir(cls.NT_matmul6, (reshape63, gpt_neox_layers_15_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) add91: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv321, gpt_neox_layers_15_attention_dense_bias1) layer_norm31: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add89, gpt_neox_layers_15_post_attention_layernorm_weight1, gpt_neox_layers_15_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv322 = R.call_tir(cls.NT_matmul7, (layer_norm31, gpt_neox_layers_15_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) add92: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv322, gpt_neox_layers_15_mlp_dense_h_to_4h_bias1) gelu15: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add92) astype30: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu15, dtype="float16") lv323 = R.call_tir(cls.NT_matmul8, (astype30, gpt_neox_layers_15_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) add93: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv323, gpt_neox_layers_15_mlp_dense_4h_to_h_bias1) astype31: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add93, dtype="float16") add94: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype31, add91) add95: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add94, add89) layer_norm32: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add95, gpt_neox_final_layer_norm_weight1, gpt_neox_final_layer_norm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) lv16 = R.call_tir(cls.index, (layer_norm32,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) lv324 = R.call_tir(cls.NT_matmul14, (lv16, embed_out_weight1), out_sinfo=R.Tensor((1, 1, vocab_size), dtype="float16")) astype32: R.Tensor((1, 1, vocab_size), dtype="float32") = R.astype(lv324, dtype="float32") gv1: R.Tuple(R.Tensor((1, 1, vocab_size), dtype="float32"), R.Object) = astype32, 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