diff --git "a/ir_dump/debug-phase1.py" "b/ir_dump/debug-phase1.py" new file mode 100644--- /dev/null +++ "b/ir_dump/debug-phase1.py" @@ -0,0 +1,6234 @@ +# 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_norm147: 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_norm147 = T.match_buffer(var_layer_norm147, (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_norm147[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_norm147[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_reshape291: 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() + reshape291 = T.match_buffer(var_reshape291, (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(reshape291[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] + reshape291[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_norm49: 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_norm49[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_norm49[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(reshape99: 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(reshape99[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] + reshape99[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_norm50: 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_norm50[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_norm50[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(astype49: 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(astype49[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", astype49[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_norm97: 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_norm97[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_norm97[v_i0, v_i1, v_k] * embed_out_weight2[v_i2, v_k] + + @T.prim_func(private=True) + def NT_matmul2(var_layer_norm148: 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_norm148 = T.match_buffer(var_layer_norm148, (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_norm148[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_norm148[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_astype147: 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() + astype147 = T.match_buffer(var_astype147, (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(astype147[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", astype147[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_norm195: 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_norm195 = T.match_buffer(var_layer_norm195, (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_norm195[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_norm195[v_i0, v_i1, v_k] * embed_out_weight4[v_i2, v_k] + + @T.prim_func(private=True) + def NT_matmul5(var_layer_norm98: 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_norm98 = T.match_buffer(var_layer_norm98, (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_norm98[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_norm98[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_reshape195: 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() + reshape195 = T.match_buffer(var_reshape195, (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(reshape195[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] + reshape195[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_norm99: 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_norm99 = T.match_buffer(var_layer_norm99, (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_norm99[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_norm99[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_astype98: 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() + astype98 = T.match_buffer(var_astype98, (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(astype98[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", astype98[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, 16, 128), "float16") + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(pages_handle, (max_num_pages, 2, 16, 16, 128), "float16", offset_factor=1) + page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (B,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) + output = T.match_buffer(output_handle, (B, 16, 128), "float16") + lse = T.match_buffer(lse_handle, (B, 16)) + # with T.block("root"): + sm_scale: T.float32 = T.float32(0.12751743082459868) + for bx in T.thread_binding(B, thread="blockIdx.x"): + for fused_by_bz in T.thread_binding(16, thread="blockIdx.y"): + for ty in T.thread_binding(1, thread="threadIdx.y"): + for tx in T.thread_binding(32, thread="threadIdx.x"): + for tz in T.thread_binding(8, thread="threadIdx.z"): + with T.block("attn"): + T.reads(page_table_indptr[bx:bx + 2], length_info[bx], q_rope_position[bx], Q[bx, fused_by_bz // 16 + ty + fused_by_bz % 16, tx * 4 - 64:tx * 4 - 64 + 132]) + T.writes(output[bx, fused_by_bz % 16 + fused_by_bz // 16 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 16 + fused_by_bz // 16 + ty]) + Q_local = T.alloc_buffer((4,), "float16", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + K_smem = T.alloc_buffer((8, 128), "float16", scope="shared") + V_smem = T.alloc_buffer((8, 128), "float16", scope="shared") + O_allreduce = T.alloc_buffer((8, 1, 128), scope="shared") + md_allreduce = T.alloc_buffer((8, 1, 2), scope="shared") + S_reduce_local = T.alloc_buffer((1,), scope="local") + t0 = T.alloc_buffer((1,), scope="local") + S_local = T.alloc_buffer((1,), scope="local") + QK_local = T.alloc_buffer((4,), scope="local") + V_local = T.alloc_buffer((4,), "float16", scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_prev = T.alloc_buffer((1,), scope="local") + other_m = T.alloc_buffer((1,), scope="local") + other_d = T.alloc_buffer((1,), scope="local") + exp_mprev = T.alloc_buffer((1,), scope="local") + exp_otherm = T.alloc_buffer((1,), scope="local") + other_o = T.alloc_buffer((4,), scope="local") + st_m = T.alloc_buffer((1,), scope="local") + st_d = T.alloc_buffer((1,), scope="local") + O_local = T.alloc_buffer((4,), scope="local") + by: T.int32 = fused_by_bz % 16 + bz: T.int32 = fused_by_bz // 16 + batch_idx: T.int32 = bx + cur_page_indptr_begin: T.int32 = page_table_indptr[batch_idx] + cur_page_indptr_end: T.int32 = page_table_indptr[batch_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[batch_idx], 0) + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for vec in T.vectorized(4): + freq = T.float32() + Q_local[vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", Q[bx, by + bz + ty, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 64, Q[bx, by + bz + ty, tx * 4 + vec + 64] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 64]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 128) / T.float32(128.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) + for iterator in range((kv_chunk_len[0] + 7) // 8): + tile_start_s: T.int32 = tz + ty + tile_start_g: T.int32 = iterator * 8 + 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 < 64, pages[page_no, 0, by, page_offset, tx * 4 + vec + 64] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 64]))), 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 % 128) / T.float32(128.0))}), 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 * 8 + 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(8): + m_prev[0] = st_m[0] + d_prev[0] = st_d[0] + other_m[0] = md_allreduce[j, ty, 0] + other_d[0] = md_allreduce[j, ty, 1] + for vec in T.vectorized(4): + other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] + st_m[0] = T.max(st_m[0], other_m[0]) + st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) + exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) + exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] / st_d[0] + for vec in T.vectorized(4): + output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) + lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) + + @T.prim_func + def batch_decode_paged_kv_sliding_window(_0: T.int32, Q_handle: T.handle, pages_handle: T.handle, page_table_indptr_handle: T.handle, page_table_values_handle: T.handle, var_length_info: T.handle, k_rope_pos_offset_handle: T.handle, q_rope_position_handle: T.handle, output_handle: T.handle, lse_handle: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + B = T.int32(is_size_var=True) + Q = T.match_buffer(Q_handle, (B, 16, 128), "float16") + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(pages_handle, (max_num_pages, 2, 16, 16, 128), "float16", offset_factor=1) + page_table_indptr = T.match_buffer(page_table_indptr_handle, (B + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_table_values = T.match_buffer(page_table_values_handle, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (3, B), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(k_rope_pos_offset_handle, (B,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(q_rope_position_handle, (B,), "int32", offset_factor=1) + output = T.match_buffer(output_handle, (B, 16, 128), "float16") + lse = T.match_buffer(lse_handle, (B, 16)) + # with T.block("root"): + sm_scale: T.float32 = T.float32(0.12751743082459868) + for bx in T.thread_binding(B, thread="blockIdx.x"): + for fused_by_bz in T.thread_binding(16, thread="blockIdx.y"): + for ty in T.thread_binding(1, thread="threadIdx.y"): + for tx in T.thread_binding(32, thread="threadIdx.x"): + for tz in T.thread_binding(8, thread="threadIdx.z"): + with T.block("attn"): + T.reads(page_table_indptr[bx:bx + 2], length_info[0:3, bx], q_rope_position[bx], Q[bx, fused_by_bz // 16 + ty + fused_by_bz % 16, tx * 4 - 64:tx * 4 - 64 + 132]) + T.writes(output[bx, fused_by_bz % 16 + fused_by_bz // 16 + ty, tx * 4:tx * 4 + 4], lse[bx, fused_by_bz % 16 + fused_by_bz // 16 + ty]) + Q_local = T.alloc_buffer((4,), "float16", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + K_smem = T.alloc_buffer((8, 128), "float16", scope="shared") + V_smem = T.alloc_buffer((8, 128), "float16", scope="shared") + O_allreduce = T.alloc_buffer((8, 1, 128), scope="shared") + md_allreduce = T.alloc_buffer((8, 1, 2), scope="shared") + S_reduce_local = T.alloc_buffer((1,), scope="local") + t0 = T.alloc_buffer((1,), scope="local") + S_local = T.alloc_buffer((1,), scope="local") + QK_local = T.alloc_buffer((4,), scope="local") + V_local = T.alloc_buffer((4,), "float16", scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_prev = T.alloc_buffer((1,), scope="local") + other_m = T.alloc_buffer((1,), scope="local") + other_d = T.alloc_buffer((1,), scope="local") + exp_mprev = T.alloc_buffer((1,), scope="local") + exp_otherm = T.alloc_buffer((1,), scope="local") + other_o = T.alloc_buffer((4,), scope="local") + st_m = T.alloc_buffer((1,), scope="local") + st_d = T.alloc_buffer((1,), scope="local") + O_local = T.alloc_buffer((4,), scope="local") + by: T.int32 = fused_by_bz % 16 + bz: T.int32 = fused_by_bz // 16 + batch_idx: T.int32 = bx + cur_page_indptr_begin: T.int32 = page_table_indptr[batch_idx] + cur_page_indptr_end: T.int32 = page_table_indptr[batch_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[0, batch_idx] - length_info[1, batch_idx] + length_info[2, batch_idx], 0) + st_m[0] = T.float32(-50000.0) + st_d[0] = T.float32(1.0) + for vec in T.vectorized(4): + O_local[vec] = T.float32(0.0) + for vec in T.vectorized(4): + freq = T.float32() + Q_local[vec] = T.if_then_else(rotary_mode == 1, T.Let(T.Cast("float16", T.cos(freq) * T.Cast("float32", Q[bx, by + bz + ty, tx * 4 + vec]) + T.sin(freq) * T.Cast("float32", T.if_then_else(tx * 4 + vec < 64, Q[bx, by + bz + ty, tx * 4 + vec + 64] * T.float16(-1.0), Q[bx, by + bz + ty, tx * 4 + vec - 64]))), where={freq: T.Cast("float32", q_rope_position[batch_idx]) * rope_scale / T.pow(rope_theta, T.Cast("float32", (tx * 4 + vec) * 2 % 128) / T.float32(128.0))}), Q[bx, by + bz + ty, tx * 4 + vec]) + for iterator in range((kv_chunk_len[0] + 7) // 8): + tile_start_s: T.int32 = tz + ty + tile_start_g: T.int32 = iterator * 8 + 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 < 64, pages[page_no, 0, by, page_offset, tx * 4 + vec + 64] * T.float16(-1.0), pages[page_no, 0, by, page_offset, tx * 4 + vec - 64]))), 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 % 128) / T.float32(128.0))}), 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 * 8 + 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(8): + m_prev[0] = st_m[0] + d_prev[0] = st_d[0] + other_m[0] = md_allreduce[j, ty, 0] + other_d[0] = md_allreduce[j, ty, 1] + for vec in T.vectorized(4): + other_o[vec] = O_allreduce[j, ty, tx * 4 + vec] + st_m[0] = T.max(st_m[0], other_m[0]) + st_d[0] = d_prev[0] * T.exp2(m_prev[0] - st_m[0]) + other_d[0] * T.exp2(other_m[0] - st_m[0]) + exp_mprev[0] = T.exp2(m_prev[0] - st_m[0]) + exp_otherm[0] = T.exp2(other_m[0] - st_m[0]) + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] * exp_mprev[0] + other_o[vec] * exp_otherm[0] + for vec in T.vectorized(4): + O_local[vec] = O_local[vec] / st_d[0] + for vec in T.vectorized(4): + output[batch_idx, by + bz + ty, tx * 4 + vec] = T.Cast("float16", O_local[vec]) + lse[batch_idx, by + bz + ty] = st_m[0] + T.log2(st_d[0]) + + @T.prim_func + def batch_prefill_paged_kv(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + total_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (total_len, 16, 128), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 16, 16, 128), "float16", offset_factor=1) + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 16, 128), "float16") + lse = T.match_buffer(var_lse, (total_len, 16)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(16, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 128), "float16", scope="shared") + K_smem = T.alloc_buffer((16, 128), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 128), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 16), scope="shared") + S_local = T.alloc_buffer((32, 16), scope="local") + O_local = T.alloc_buffer((32, 128), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 8): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 128) + j = T.axis.spatial(128, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 128) + 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 < 64, q[cur_L, cur_H_qo, j + 64] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 64]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 128) / T.float32(128.0))}), 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(4): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 128) + j = T.axis.spatial(128, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 128) + 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 < 64, pages[page_no, 0, by, page_offset, j + 64] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 64]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 128) / T.float32(128.0))}), 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(4): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 128) + j = T.axis.spatial(128, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 128) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:128], K_smem[0:16, 0:128]) + T.writes(S_local[0:32, 0:16]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(16, 2, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + k = T.axis.reduce(128, 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(0.12751743082459868) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 2): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(16): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) + T.writes(S_smem[row, 0:16]) + for j in range(16): + if row < 32: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(16): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:16], V_smem[0:16, 0:128]) + T.writes(O_local[0:32, 0:128]) + 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(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(128, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 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(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def batch_prefill_paged_kv_sliding_window(_0: T.int32, var_q: T.handle, var_q_indptr: T.handle, var_pages: T.handle, var_page_indptr: T.handle, var_page_values: T.handle, var_length_info: T.handle, var_k_rope_pos_offset: T.handle, var_q_rope_position: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + total_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (total_len, 16, 128), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 16, 16, 128), "float16", offset_factor=1) + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (3, batch_size), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 16, 128), "float16") + lse = T.match_buffer(var_lse, (total_len, 16)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(16, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 128), "float16", scope="shared") + K_smem = T.alloc_buffer((16, 128), "float16", scope="shared") + V_smem = T.alloc_buffer((16, 128), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 16), scope="shared") + S_local = T.alloc_buffer((32, 16), scope="local") + O_local = T.alloc_buffer((32, 128), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 16 + length_info[0, b_idx] - length_info[1, b_idx] + length_info[2, b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 8): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 128) + j = T.axis.spatial(128, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 128) + 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 < 64, q[cur_L, cur_H_qo, j + 64] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 64]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 128) / T.float32(128.0))}), 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(4): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("K_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 128) + j = T.axis.spatial(128, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 128) + 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 < 64, pages[page_no, 0, by, page_offset, j + 64] * T.float16(-1.0), pages[page_no, 0, by, page_offset, j - 64]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 128) / T.float32(128.0))}), 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(4): + for lz_ly_fused_1 in T.thread_binding(4, thread="threadIdx.y"): + for lz_ly_fused_2 in T.thread_binding(32, thread="threadIdx.x"): + for lz_ly_fused_3 in T.vectorized(4): + with T.block("V_load"): + i = T.axis.spatial(16, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 128) + j = T.axis.spatial(128, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 128) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = T.if_then_else(cur_L < length_info[2, b_idx], cur_L, cur_L - length_info[2, b_idx] + length_info[1, b_idx]) + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 16] + page_offset: T.int32 = seq_offset % 16 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:128], K_smem[0:16, 0:128]) + T.writes(S_local[0:32, 0:16]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 2): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(16, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 2 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(16, 2, 2, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + k = T.axis.reduce(128, 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(0.12751743082459868) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 2): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(16, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 2 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:16], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(16): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:16], m_new[i]) + T.writes(S_smem[row, 0:16]) + for j in range(16): + if row < 32: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:16], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(16): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:16], V_smem[0:16, 0:128]) + T.writes(O_local[0:32, 0:128]) + 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(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(128, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 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(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def batch_prefill_ragged_kv(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_k_rope_pos_offset: T.handle, var_output: T.handle, var_lse: T.handle, causal: T.int32, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + qo_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (qo_len, 16, 128), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + kv_len = T.int32(is_size_var=True) + k = T.match_buffer(var_k, (kv_len, 16, 128), "float16") + v = T.match_buffer(var_v, (kv_len, 16, 128), "float16") + kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (qo_len, 16, 128), "float16") + lse = T.match_buffer(var_lse, (qo_len, 16)) + # with T.block("root"): + for lbx in T.thread_binding(8, thread="blockIdx.x"): + for lby in T.thread_binding(16, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 128), "float16", scope="shared") + K_smem = T.alloc_buffer((128, 32), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 128), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 32), scope="shared") + S_local = T.alloc_buffer((32, 32), scope="local") + O_local = T.alloc_buffer((32, 128), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + q_indptr_val: T.int32 = q_indptr[b_idx] + LH_start: T.int32 = tile_id[0] * 32 + kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(8): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, li_0 * 16 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) // 8) + j = T.axis.spatial(128, lj_0_0 * 64 + (li_1_lj_0_1_fused_0 * 32 + li_1_lj_0_1_fused_1) % 8 * 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 < 64, q[cur_L, cur_H_qo, j + 64] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 64]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 128) / T.float32(128.0))}), q[cur_L, cur_H_qo, j]) + else: + Q_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + for iterator_1 in range((kv_chunk_len[0] + 31) // 32): + L_kv_start: T.int32 = iterator_1 * 32 + L_kv_base: T.int32 = kv_indptr[b_idx] + for lz_1_ly_0_1_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for lz_1_ly_0_1_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lz_0, ly_0_0 in T.grid(2, 2): + for ly_1 in T.vectorized(8): + with T.block("K_load"): + i = T.axis.spatial(32, lz_0 * 16 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 8) + j = T.axis.spatial(128, ly_0_0 * 64 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 8 * 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 - 64:j - 64 + 129]) + 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 < 64, k[L_kv_base + cur_L, by, j + 64] * T.float16(-1.0), k[L_kv_base + cur_L, by, j - 64]))), where={freq: T.Cast("float32", k_rope_pos_offset[b_idx] + cur_L) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 128) / T.float32(128.0))}), 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(32, lz_0 * 16 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) // 8) + j = T.axis.spatial(128, ly_0_0 * 64 + (lz_1_ly_0_1_fused_0 * 32 + lz_1_ly_0_1_fused_1) % 8 * 8 + ly_1) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + V_smem[i, j] = v[L_kv_base + cur_L, by, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:128], K_smem[0:128, 0:32]) + T.writes(S_local[0:32, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init in T.unroll(2): + for lj_1_0_init in T.unroll(1): + for lj_1_1_init in T.vectorized(4): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_0_init * 4 + lj_1_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0 in range(8): + for li_1 in T.unroll(2): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + for lk_1 in range(16): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1_0 * 4 + lj_1_1) + k_1 = T.axis.reduce(128, 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(0.12751743082459868) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(2): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(4): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1_0 * 4 + lj_1_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 32: + row_: T.int32 = LH_start + row + if T.if_then_else(causal > 0, L_kv_start + j < kv_chunk_len[0] - (q_indptr[b_idx + 1] - q_indptr[b_idx]) + row_ + 1, L_kv_start + j < kv_chunk_len[0]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:128]) + T.writes(O_local[0:32, 0:128]) + 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(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(128, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 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(2, 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(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 8 + lj_1_0 * 8 + lj_1_1) + k_1 = T.axis.reduce(32, lk_0 * 16 + lk_1) + T.reads(O_local[i, j], m_prev_smem[i], m_smem[i], S_smem[i, k_1], V_smem[k_1, j]) + T.writes(O_local[i, j]) + O_local[i, j] = O_local[i, j] + S_smem[i, k_1] * T.Cast("float32", V_smem[k_1, j]) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1 in range(4): + for lj_1_0 in T.unroll(1): + for lj_1_1 in T.vectorized(8): + with T.block("O_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 8 + + @T.prim_func + def batch_tree_attn(var_q: T.handle, var_q_indptr: T.handle, var_k: T.handle, var_v: T.handle, var_kv_indptr: T.handle, var_q_rope_position: T.handle, var_mn_indptr: T.handle, var_mask: T.handle, var_output: T.handle, var_lse: T.handle, rotary_mode: T.int32, rope_scale: T.float32, rope_theta: T.float32, attn_score_scaling_factor: T.float32, batch_size: T.int32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + qo_len = T.int32(is_size_var=True) + q = T.match_buffer(var_q, (qo_len, 16, 128), "float16") + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + kv_len = T.int32(is_size_var=True) + k = T.match_buffer(var_k, (kv_len, 16, 128), "float16") + v = T.match_buffer(var_v, (kv_len, 16, 128), "float16") + kv_indptr = T.match_buffer(var_kv_indptr, (batch_size + 1,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (qo_len,), "int32", offset_factor=1) + mn_indptr = T.match_buffer(var_mn_indptr, (batch_size + 1,), "int32", offset_factor=1) + tree_size = T.int32(is_size_var=True) + mask = T.match_buffer(var_mask, (tree_size, 2), "int32", offset_factor=1) + output = T.match_buffer(var_output, (qo_len, 16, 128), "float16") + lse = T.match_buffer(var_lse, (qo_len, 16)) + # with T.block("root"): + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(16, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 128), "float16", scope="shared") + K_smem = T.alloc_buffer((32, 128), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 128), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 32), scope="shared") + S_local = T.alloc_buffer((32, 32), scope="local") + O_local = T.alloc_buffer((32, 128), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + kv_chunk_len[0] = kv_indptr[b_idx + 1] - kv_indptr[b_idx] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 8): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 128) + j = T.axis.spatial(128, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 128) + 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 < 64, q[cur_L, cur_H_qo, j + 64] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 64]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 128) / T.float32(128.0))}), 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(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("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) // 128) + j = T.axis.spatial(128, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 128) + 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 < 64, k[cur_L, by, j + 64] * T.float16(-1.0), k[cur_L, by, j - 64]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 128) / T.float32(128.0))}), k[cur_L, by, j]) + V_smem[i, j] = v[cur_L, by, j] + else: + K_smem[i, j] = T.float16(0.0) + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:128], K_smem[0:32, 0:128]) + T.writes(S_local[0:32, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 4): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(16, 2, 4, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + k_1 = T.axis.reduce(128, 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(0.12751743082459868) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 4): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], mn_indptr[b_idx:b_idx + 2], mask[T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + mn_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + mn_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 32: + row_: T.int32 = LH_start + row + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) or mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 0] and mask[mn_indptr[b_idx] + (row_ + (mn_indptr[b_idx + 1] - mn_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < mask[mn_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (mn_indptr[b_idx + 1] - mn_indptr[b_idx]))), 1]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:128]) + T.writes(O_local[0:32, 0:128]) + 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(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(128, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 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(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @T.prim_func + def chunk_lse(var_A: T.handle, var_temperature: T.handle, var_chunked_sum: T.handle, var_chunked_max: T.handle): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.noalias": T.bool(True)}) + batch_size, vocab_size = T.int64(is_size_var=True), T.int64(is_size_var=True) + A = T.match_buffer(var_A, (batch_size, vocab_size)) + temperature = T.match_buffer(var_temperature, (batch_size,)) + num_chunks = T.int64(is_size_var=True) + chunked_sum = T.match_buffer(var_chunked_sum, (batch_size, num_chunks)) + chunked_max = T.match_buffer(var_chunked_max, (batch_size, num_chunks)) + # with T.block("root"): + A_pad = T.alloc_buffer((batch_size, num_chunks, T.int64(4096))) + temp_max = T.alloc_buffer((batch_size, num_chunks)) + temp_sum = T.alloc_buffer((batch_size, num_chunks)) + for l0, l1, l2 in T.grid(batch_size, num_chunks, T.int64(4096)): + with T.block("pad"): + v0, v1, v2 = T.axis.remap("SSS", [l0, l1, l2]) + T.reads(temperature[v0], A[v0, v1 * T.int64(4096) + v2]) + T.writes(A_pad[v0, v1, v2]) + A_pad[v0, v1, v2] = T.if_then_else(v1 * T.int64(4096) + v2 < vocab_size, T.if_then_else(temperature[v0] > T.float32(1.0000000000000001e-05), A[v0, v1 * T.int64(4096) + v2] / temperature[v0], A[v0, v1 * T.int64(4096) + v2]), T.float32(-340282346638528859811704183484516925440.0)) + for l0, l1, l2 in T.grid(batch_size, num_chunks, T.int64(4096)): + with T.block("max"): + v0, v1, v2 = T.axis.remap("SSR", [l0, l1, l2]) + T.reads(A_pad[v0, v1, v2]) + T.writes(temp_max[v0, v1]) + with T.init(): + temp_max[v0, v1] = T.float32(-340282346638528859811704183484516925440.0) + temp_max[v0, v1] = T.max(temp_max[v0, v1], A_pad[v0, v1, v2]) + for l0, l1, l2 in T.grid(batch_size, num_chunks, T.int64(4096)): + with T.block("sum_exp"): + v0, v1, v2 = T.axis.remap("SSR", [l0, l1, l2]) + T.reads(temperature[v0], A_pad[v0, v1, v2], temp_max[v0, v1]) + T.writes(temp_sum[v0, v1]) + with T.init(): + temp_sum[v0, v1] = T.float32(0.0) + temp_sum[v0, v1] = temp_sum[v0, v1] + T.if_then_else(v1 * T.int64(4096) + v2 < vocab_size, T.Select(temperature[v0] > T.float32(1.0000000000000001e-05), T.exp(A_pad[v0, v1, v2] - temp_max[v0, v1]), T.Cast("float32", A_pad[v0, v1, v2] == temp_max[v0, v1])), T.float32(0.0)) + for l0, l1, l2 in T.grid(batch_size, num_chunks, T.int64(1)): + with T.block("log"): + v0, v1, v2 = T.axis.remap("SSS", [l0, l1, l2]) + T.reads(temperature[v0], temp_sum[v0, v1], temp_max[v0, v1]) + T.writes(chunked_sum[v0, v1], chunked_max[v0, v1]) + chunked_sum[v0, v1] = T.Select(temperature[v0] > T.float32(1.0000000000000001e-05), T.log(temp_sum[v0, v1]), temp_sum[v0, v1]) + chunked_max[v0, v1] = temp_max[v0, v1] + + @T.prim_func + def compact_kv_copy(var_pages: T.handle, var_copy_length_indptr: T.handle, var_copy_src_dst_pos: T.handle, batch_size: T.int32): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + num_pages = T.int32() + pages = T.match_buffer(var_pages, (num_pages, 2, 16, 16, 128), "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) // 128 % 16 + d: T.int32 = (bhd_o * 256 + bhd_i) % 128 + if bhd_o * 256 + bhd_i < batch_size * 16 * 128: + for i in range(copy_length_indptr[b + 1] - copy_length_indptr[b]): + src_pos: T.int32 = copy_src_dst_pos[0, copy_length_indptr[b] + i] + dst_pos: T.int32 = copy_src_dst_pos[1, copy_length_indptr[b] + i] + pages[dst_pos // 16, 0, h, dst_pos % 16, d] = pages[src_pos // 16, 0, h, src_pos % 16, d] + pages[dst_pos // 16, 1, h, dst_pos % 16, d] = pages[src_pos // 16, 1, h, src_pos % 16, d] + + @T.prim_func + def copy_single_page(var_pages: T.handle, src_page_id: T.int64, tgt_page_id: T.int64, copy_length: T.int64): + T.func_attr({"target": T.target({"host": {"keys": ["arm_cpu", "cpu"], "kind": "llvm", "mtriple": "aarch64-linux-android", "tag": ""}, "keys": ["opencl", "gpu"], "kind": "opencl", "max_function_args": 128, "max_num_threads": 256, "max_shared_memory_per_block": 16384, "max_threads_per_block": 256, "tag": "", "texture_spatial_limit": 16384, "thread_warp_size": 1}), "tir.is_scheduled": 1}) + num_pages, page_size = T.int32(), T.int64() + pages = T.match_buffer(var_pages, (num_pages, 2, 16, page_size, 128), "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(16, T.Cast("int32", (b * T.int64(256) + T.Cast("int64", t)) // (copy_length * T.int64(128)))) + vp = T.axis.spatial(copy_length, (b * T.int64(256) + T.Cast("int64", t)) % (copy_length * T.int64(128)) // T.int64(128)) + vd = T.axis.spatial(128, T.Cast("int32", (b * T.int64(256) + T.Cast("int64", t)) % T.int64(128))) + T.where(b * T.int64(256) + T.Cast("int64", t) < copy_length * T.int64(16) * T.int64(128)) + 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, 48, 128), "float16") + position_map = T.match_buffer(var_position_map, (seq_len,), "int32", offset_factor=1) + q = T.match_buffer(var_q, (seq_len, 16, 128), "float16") + k = T.match_buffer(var_k, (seq_len, 16, 128), "float16") + v = T.match_buffer(var_v, (seq_len, 16, 128), "float16") + # with T.block("root"): + for iters_0, iters_1, iters_2 in T.grid(seq_len, 48, 128): + 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 - 16:d - 16 + 33]) + T.writes(q[s, h, d], k[s, h - 16, d], v[s, h - 32, d]) + if h < 16: + freq = T.float32() + q[s, h, d] = T.if_then_else(apply_rope > 0 and d < 32, 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 < 16, qkv[s, h, d + 16] * T.float16(-1.0), qkv[s, h, d - 16]))), where={freq: T.Cast("float32", position_map[s]) / T.pow(T.float32(10000.0), T.Cast("float32", d * 2 % 32) / T.float32(32.0))}), qkv[s, h, d]) + else: + if h < 32: + freq = T.float32() + k[s, h - 16, d] = T.if_then_else(apply_rope > 0 and d < 32, 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 < 16, qkv[s, h, d + 16] * T.float16(-1.0), qkv[s, h, d - 16]))), where={freq: T.Cast("float32", position_map[s]) / T.pow(T.float32(10000.0), T.Cast("float32", d * 2 % 32) / T.float32(32.0))}), qkv[s, h, d]) + else: + v[s, h - 32, 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_norm48: 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_norm48 = T.match_buffer(var_layer_norm48, (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_norm48[v_i, seq_len - T.int64(1), v_k]) + T.writes(index[v_i, v__, v_k]) + index[v_i, v__, v_k] = layer_norm48[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(8, thread="threadIdx.y"): + for tx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("merge"): + T.reads(S[bx, ty + by * 8], S_other[bx, ty + by * 8], V[bx, ty + by * 8, tx * 4:tx * 4 + 4], V_other[bx, ty + by * 8, tx * 4:tx * 4 + 4]) + T.writes(V[bx, ty + by * 8, tx * 4:tx * 4 + 4], S[bx, ty + by * 8]) + s_val = T.alloc_buffer((1,), scope="local") + s_other_val = T.alloc_buffer((1,), scope="local") + s_max = T.alloc_buffer((1,), scope="local") + scale = T.alloc_buffer((1,), scope="local") + other_scale = T.alloc_buffer((1,), scope="local") + v_vec = T.alloc_buffer((4,), "float16", scope="local") + v_other_vec = T.alloc_buffer((4,), "float16", scope="local") + s_val[0] = S[bx, ty + by * 8] + s_other_val[0] = S_other[bx, ty + by * 8] + s_max[0] = T.max(s_val[0], s_other_val[0]) + s_val[0] = T.exp2(s_val[0] - s_max[0]) + s_other_val[0] = T.exp2(s_other_val[0] - s_max[0]) + scale[0] = s_val[0] / (s_val[0] + s_other_val[0]) + other_scale[0] = s_other_val[0] / (s_val[0] + s_other_val[0]) + for vec in T.vectorized(4): + v_vec[vec] = V[bx, ty + by * 8, tx * 4 + vec] + for vec in T.vectorized(4): + v_other_vec[vec] = V_other[bx, ty + by * 8, tx * 4 + vec] + for vec in range(4): + v_vec[vec] = T.Cast("float16", T.Cast("float32", v_vec[vec]) * scale[0] + T.Cast("float32", v_other_vec[vec]) * other_scale[0]) + for vec in T.vectorized(4): + V[bx, ty + by * 8, tx * 4 + vec] = v_vec[vec] + S[bx, ty + by * 8] = T.log2(s_val[0] + s_other_val[0]) + s_max[0] + + @T.prim_func + 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, 16, page_size, 128), "float16", offset_factor=1) + seqlen = T.int64(is_size_var=True) + position_map = T.match_buffer(var_position_map, (seqlen,), "int32", offset_factor=1) + k_data = T.match_buffer(var_k_data, (24, seqlen, 16, 128), "float16") + v_data = T.match_buffer(var_v_data, (24, seqlen, 16, 128), "float16") + # with T.block("root"): + for p, h, d in T.grid(seqlen, 16, 128): + 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, 16, 16, 128), "float16", offset_factor=1) + ntoken = T.int64(is_size_var=True) + k_data = T.match_buffer(var_k_data, (ntoken, 16, 128), "float16") + v_data = T.match_buffer(var_v_data, (ntoken, 16, 128), "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, 16, 128): + 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, 16, 128), "float16") + batch_size = T.int32(is_size_var=True) + q_indptr = T.match_buffer(var_q_indptr, (batch_size + 1,), "int32", offset_factor=1) + max_num_pages = T.int32(is_size_var=True) + pages = T.match_buffer(var_pages, (max_num_pages, 2, 16, 32, 128), "float16") + page_indptr = T.match_buffer(var_page_indptr, (batch_size + 1,), "int32", offset_factor=1) + nnz_pages = T.int32(is_size_var=True) + page_values = T.match_buffer(var_page_values, (nnz_pages,), "int32", offset_factor=1) + length_info = T.match_buffer(var_length_info, (batch_size,), "int32", offset_factor=1) + k_rope_pos_offset = T.match_buffer(var_k_rope_pos_offset, (batch_size,), "int32", offset_factor=1) + q_rope_position = T.match_buffer(var_q_rope_position, (total_len,), "int32", offset_factor=1) + output = T.match_buffer(var_output, (total_len, 16, 128), "float16") + lse = T.match_buffer(var_lse, (total_len, 16)) + tree_order_indptr = T.match_buffer(tree_order_indptr_handle, (batch_size + 1,), "int32", offset_factor=1) + total_tree_order_len = T.int32(is_size_var=True) + tree_order = T.match_buffer(tree_order_handle, (total_tree_order_len, 2), "int32", offset_factor=1) + # with T.block("root"): + assert rotary_mode == 0, "Inline rotary mode is not supported in tree attention." + for lbx in T.thread_binding(16, thread="blockIdx.x"): + for lby in T.thread_binding(16, thread="blockIdx.y"): + for lty in T.thread_binding(4, thread="threadIdx.y"): + for ltx in T.thread_binding(32, thread="threadIdx.x"): + with T.block("attn"): + bx, by, ty, tx = T.axis.remap("SSSS", [lbx, lby, lty, ltx]) + T.reads() + T.writes() + tile_id = T.alloc_buffer((1,), "int32", scope="local") + batch_idx = T.alloc_buffer((1,), "int32", scope="local") + batch_tiles = T.alloc_buffer((1,), "int32", scope="local") + batch_rows = T.alloc_buffer((1,), "int32", scope="local") + iterator = T.alloc_buffer((1,), "int32", scope="local") + kv_chunk_len = T.alloc_buffer((1,), "int32", scope="local") + Q_smem = T.alloc_buffer((32, 128), "float16", scope="shared") + K_smem = T.alloc_buffer((32, 128), "float16", scope="shared") + V_smem = T.alloc_buffer((32, 128), "float16", scope="shared") + S_smem = T.alloc_buffer((32, 32), scope="shared") + S_local = T.alloc_buffer((32, 32), scope="local") + O_local = T.alloc_buffer((32, 128), scope="local") + m_smem = T.alloc_buffer((32,), scope="shared") + m_prev_smem = T.alloc_buffer((32,), scope="shared") + d_smem = T.alloc_buffer((32,), scope="shared") + m_new = T.alloc_buffer((1,), scope="local") + m_prev = T.alloc_buffer((1,), scope="local") + d_new = T.alloc_buffer((1,), scope="local") + tile_id[0] = bx + batch_idx[0] = 0 + batch_rows[0] = q_indptr[1] - q_indptr[0] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + while T.tvm_thread_invariant(batch_idx[0] < batch_size): + while tile_id[0] >= batch_tiles[0] and batch_idx[0] < batch_size: + tile_id[0] = tile_id[0] - batch_tiles[0] + batch_idx[0] = batch_idx[0] + 1 + if batch_idx[0] < batch_size: + b_idx: T.int32 = batch_idx[0] + batch_rows[0] = q_indptr[b_idx + 1] - q_indptr[b_idx] + batch_tiles[0] = (batch_rows[0] + 32 - 1) // 32 + if T.tvm_thread_invariant(batch_idx[0] < batch_size): + b_idx: T.int32 = batch_idx[0] + LH_start: T.int32 = tile_id[0] * 32 + q_indptr_val: T.int32 = q_indptr[b_idx] + cur_page_indptr_begin: T.int32 = page_indptr[b_idx] + cur_page_indptr_end: T.int32 = page_indptr[b_idx + 1] + kv_chunk_len[0] = T.if_then_else(cur_page_indptr_begin != cur_page_indptr_end, (cur_page_indptr_end - cur_page_indptr_begin - 1) * 32 + length_info[b_idx], 0) + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + m_smem[row] = T.float32(-50000.0) + d_smem[row] = T.float32(1.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(4, 8): + with T.block("O_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) // 128) + j = T.axis.spatial(128, (li_lj_fused_0 * 512 + li_lj_fused_1 * 128 + li_lj_fused_2 * 4 + li_lj_fused_3) % 128) + 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 < 64, q[cur_L, cur_H_qo, j + 64] * T.float16(-1.0), q[cur_L, cur_H_qo, j - 64]))), where={freq: T.Cast("float32", q_rope_position[cur_L]) * rope_scale / T.pow(rope_theta, T.Cast("float32", j * 2 % 128) / T.float32(128.0))}), 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(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(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 128) + j = T.axis.spatial(128, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 128) + 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(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(32, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) // 128) + j = T.axis.spatial(128, (lz_ly_fused_0 * 512 + lz_ly_fused_1 * 128 + lz_ly_fused_2 * 4 + lz_ly_fused_3) % 128) + T.reads() + T.writes() + cur_L: T.int32 = L_kv_start + i + if cur_L < kv_chunk_len[0]: + seq_offset: T.int32 = cur_L + page_no: T.int32 = page_values[cur_page_indptr_begin + seq_offset // 32] + page_offset: T.int32 = seq_offset % 32 + V_smem[i, j] = pages[page_no, 1, by, page_offset, j] + else: + V_smem[i, j] = T.float16(0.0) + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(Q_smem[0:32, 0:128], K_smem[0:32, 0:128]) + T.writes(S_local[0:32, 0:32]) + for li_0_lj_0_fused_0_init in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1_init in T.thread_binding(32, thread="threadIdx.x"): + for li_1_init, lj_1_init in T.grid(2, 4): + with T.block("S_gemm_init"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 8 * 2 + li_1_init) + j = T.axis.spatial(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 8 * 4 + lj_1_init) + T.reads() + T.writes(S_local[i, j]) + S_local[i, j] = T.float32(0.0) + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for lk_0, li_1, lj_1, lk_1 in T.grid(16, 2, 4, 8): + with T.block("S_gemm_update"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + k = T.axis.reduce(128, 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(0.12751743082459868) + T.tvm_storage_sync("shared") + for li_0_lj_0_fused_0 in T.thread_binding(4, thread="threadIdx.y"): + for li_0_lj_0_fused_1 in T.thread_binding(32, thread="threadIdx.x"): + for li_1, lj_1 in T.grid(2, 4): + with T.block("S_store"): + i = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 8 * 2 + li_1) + j = T.axis.spatial(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 8 * 4 + lj_1) + T.reads(S_local[i, j]) + T.writes(S_smem[i, j]) + S_smem[i, j] = S_local[i, j] + T.tvm_storage_sync("shared") + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update1"): + T.reads(m_smem[row], kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], m_new[i], S_smem[row, 0:32], d_smem[row], m_prev[i]) + T.writes(m_prev[i], m_new[i], d_new[i]) + m_prev[i] = m_smem[row] + m_new[i] = m_smem[row] + row_: T.int32 = LH_start + row + for j in range(32): + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): + m_new[i] = T.max(m_new[i], S_smem[row, j]) + d_new[i] = d_smem[row] * T.exp2(m_prev[i] - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + with T.block("update"): + T.reads(kv_chunk_len[0], tree_order_indptr[b_idx:b_idx + 2], tree_order[T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]):T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0]) + (T.max(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] + 31 - kv_chunk_len[0]) + 1 - T.min(LH_start + row + tree_order_indptr[b_idx + 1] + q_indptr[b_idx] - q_indptr[b_idx + 1], L_kv_start + tree_order_indptr[b_idx + 1] - kv_chunk_len[0])), 0:2], q_indptr[b_idx:b_idx + 2], S_smem[row, 0:32], m_new[i]) + T.writes(S_smem[row, 0:32]) + for j in range(32): + if row < 32: + row_: T.int32 = LH_start + row + if L_kv_start + j < kv_chunk_len[0] and (L_kv_start + j < kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) or tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] >= tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 0] and tree_order[tree_order_indptr[b_idx] + (row_ + (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]) - (q_indptr[b_idx + 1] - q_indptr[b_idx])), 0] < tree_order[tree_order_indptr[b_idx] + (L_kv_start + j - (kv_chunk_len[0] - (tree_order_indptr[b_idx + 1] - tree_order_indptr[b_idx]))), 1]): + S_smem[row, j] = T.exp2(S_smem[row, j] - m_new[i]) + else: + S_smem[row, j] = T.exp2(T.float32(-50000.0) - m_new[i]) + for i in range(1): + row: T.int32 = i * 32 * 4 + ty * 32 + tx + if row < 32: + with T.block("update"): + T.reads(d_new[i], S_smem[row, 0:32], m_new[i], m_prev[i]) + T.writes(d_new[i], m_smem[row], d_smem[row], m_prev_smem[row]) + for j in range(32): + d_new[i] = d_new[i] + S_smem[row, j] + m_smem[row] = m_new[i] + d_smem[row] = d_new[i] + m_prev_smem[row] = m_prev[i] + T.tvm_storage_sync("shared") + with T.block(""): + T.reads(m_prev_smem[0:32], m_smem[0:32], S_smem[0:32, 0:32], V_smem[0:32, 0:128]) + T.writes(O_local[0:32, 0:128]) + 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(32, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) // 16 * 4 + li_1_init) + j = T.axis.spatial(128, (li_0_lj_0_fused_0_init * 32 + li_0_lj_0_fused_1_init) % 16 * 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(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) // 16 * 4 + li_1) + j = T.axis.spatial(128, (li_0_lj_0_fused_0 * 32 + li_0_lj_0_fused_1) % 16 * 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(32, li_0 * 128 + li_1 * 32 + li_2) + T.where((li_0 * 4 + li_1) * 32 + li_2 < 32) + T.reads(q_indptr[b_idx:b_idx + 2], m_smem[i], d_smem[i]) + T.writes(lse[q_indptr[b_idx] + (LH_start + i), by]) + cur_L: T.int32 = q_indptr[b_idx] + (LH_start + i) + cur_H_qo: T.int32 = by + if cur_L < q_indptr[b_idx + 1]: + lse[cur_L, cur_H_qo] = m_smem[i] + T.log2(d_smem[i]) + tile_id[0] = tile_id[0] + 16 + + @R.function + def alloc_embedding_tensor() -> R.Tensor((2048, 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((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_layers_16_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[193] + gpt_neox_layers_16_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[194] + gpt_neox_layers_16_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[195] + gpt_neox_layers_16_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[196] + gpt_neox_layers_16_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[197] + gpt_neox_layers_16_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[198] + gpt_neox_layers_16_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[199] + gpt_neox_layers_16_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[200] + gpt_neox_layers_16_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[201] + gpt_neox_layers_16_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[202] + gpt_neox_layers_16_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[203] + gpt_neox_layers_16_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[204] + gpt_neox_layers_17_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[205] + gpt_neox_layers_17_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[206] + gpt_neox_layers_17_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[207] + gpt_neox_layers_17_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[208] + gpt_neox_layers_17_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[209] + gpt_neox_layers_17_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[210] + gpt_neox_layers_17_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[211] + gpt_neox_layers_17_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[212] + gpt_neox_layers_17_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[213] + gpt_neox_layers_17_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[214] + gpt_neox_layers_17_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[215] + gpt_neox_layers_17_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[216] + gpt_neox_layers_18_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[217] + gpt_neox_layers_18_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[218] + gpt_neox_layers_18_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[219] + gpt_neox_layers_18_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[220] + gpt_neox_layers_18_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[221] + gpt_neox_layers_18_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[222] + gpt_neox_layers_18_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[223] + gpt_neox_layers_18_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[224] + gpt_neox_layers_18_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[225] + gpt_neox_layers_18_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[226] + gpt_neox_layers_18_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[227] + gpt_neox_layers_18_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[228] + gpt_neox_layers_19_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[229] + gpt_neox_layers_19_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[230] + gpt_neox_layers_19_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[231] + gpt_neox_layers_19_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[232] + gpt_neox_layers_19_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[233] + gpt_neox_layers_19_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[234] + gpt_neox_layers_19_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[235] + gpt_neox_layers_19_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[236] + gpt_neox_layers_19_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[237] + gpt_neox_layers_19_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[238] + gpt_neox_layers_19_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[239] + gpt_neox_layers_19_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[240] + gpt_neox_layers_20_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[241] + gpt_neox_layers_20_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[242] + gpt_neox_layers_20_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[243] + gpt_neox_layers_20_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[244] + gpt_neox_layers_20_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[245] + gpt_neox_layers_20_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[246] + gpt_neox_layers_20_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[247] + gpt_neox_layers_20_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[248] + gpt_neox_layers_20_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[249] + gpt_neox_layers_20_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[250] + gpt_neox_layers_20_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[251] + gpt_neox_layers_20_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[252] + gpt_neox_layers_21_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[253] + gpt_neox_layers_21_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[254] + gpt_neox_layers_21_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[255] + gpt_neox_layers_21_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[256] + gpt_neox_layers_21_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[257] + gpt_neox_layers_21_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[258] + gpt_neox_layers_21_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[259] + gpt_neox_layers_21_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[260] + gpt_neox_layers_21_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[261] + gpt_neox_layers_21_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[262] + gpt_neox_layers_21_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[263] + gpt_neox_layers_21_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[264] + gpt_neox_layers_22_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[265] + gpt_neox_layers_22_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[266] + gpt_neox_layers_22_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[267] + gpt_neox_layers_22_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[268] + gpt_neox_layers_22_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[269] + gpt_neox_layers_22_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[270] + gpt_neox_layers_22_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[271] + gpt_neox_layers_22_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[272] + gpt_neox_layers_22_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[273] + gpt_neox_layers_22_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[274] + gpt_neox_layers_22_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[275] + gpt_neox_layers_22_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[276] + gpt_neox_layers_23_input_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[277] + gpt_neox_layers_23_input_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[278] + gpt_neox_layers_23_post_attention_layernorm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[279] + gpt_neox_layers_23_post_attention_layernorm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[280] + gpt_neox_layers_23_attention_query_key_value_weight4: R.Tensor((6144, 2048), dtype="float16") = packed_params[281] + gpt_neox_layers_23_attention_query_key_value_bias4: R.Tensor((6144,), dtype="float16") = packed_params[282] + gpt_neox_layers_23_attention_dense_weight4: R.Tensor((2048, 2048), dtype="float16") = packed_params[283] + gpt_neox_layers_23_attention_dense_bias4: R.Tensor((2048,), dtype="float16") = packed_params[284] + gpt_neox_layers_23_mlp_dense_h_to_4h_weight4: R.Tensor((8192, 2048), dtype="float16") = packed_params[285] + gpt_neox_layers_23_mlp_dense_h_to_4h_bias4: R.Tensor((8192,), dtype="float32") = packed_params[286] + gpt_neox_layers_23_mlp_dense_4h_to_h_weight4: R.Tensor((2048, 8192), dtype="float16") = packed_params[287] + gpt_neox_layers_23_mlp_dense_4h_to_h_bias4: R.Tensor((2048,), dtype="float32") = packed_params[288] + gpt_neox_final_layer_norm_weight4: R.Tensor((2048,), dtype="float16") = packed_params[289] + gpt_neox_final_layer_norm_bias4: R.Tensor((2048,), dtype="float16") = packed_params[290] + embed_out_weight4: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[291] + layer_norm147: 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_norm147, gpt_neox_layers_0_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add432: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv, gpt_neox_layers_0_attention_query_key_value_bias4) + reshape288: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add432, R.shape([batch_size, 1, 48, 128])) + reshape289: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape288, R.shape([batch_size, 48, 128])) + lv73 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), reshape289), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape290: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv73, R.shape([batch_size, 1, 16, 128])) + reshape291: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape290, R.shape([batch_size, 1, 2048])) + lv1 = R.call_tir(cls.NT_matmul1, (reshape291, gpt_neox_layers_0_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add433: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv1, gpt_neox_layers_0_attention_dense_bias4) + layer_norm148: 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_norm148, gpt_neox_layers_0_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add434: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv2, gpt_neox_layers_0_mlp_dense_h_to_4h_bias4) + gelu72: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add434) + astype147: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu72, dtype="float16") + lv3 = R.call_tir(cls.NT_matmul3, (astype147, gpt_neox_layers_0_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add435: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv3, gpt_neox_layers_0_mlp_dense_4h_to_h_bias4) + astype148: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add435, dtype="float16") + add436: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype148, add433) + add437: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add436, input_embeds) + layer_norm149: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add437, 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_norm149, gpt_neox_layers_1_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add438: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv4, gpt_neox_layers_1_attention_query_key_value_bias4) + reshape292: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add438, R.shape([batch_size, 1, 48, 128])) + reshape293: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape292, R.shape([batch_size, 48, 128])) + lv74 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), reshape293), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape294: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv74, R.shape([batch_size, 1, 16, 128])) + reshape295: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape294, R.shape([batch_size, 1, 2048])) + lv5 = R.call_tir(cls.NT_matmul1, (reshape295, gpt_neox_layers_1_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add439: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv5, gpt_neox_layers_1_attention_dense_bias4) + layer_norm150: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add437, 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_norm150, gpt_neox_layers_1_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add440: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv6, gpt_neox_layers_1_mlp_dense_h_to_4h_bias4) + gelu73: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add440) + astype149: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu73, dtype="float16") + lv7 = R.call_tir(cls.NT_matmul3, (astype149, gpt_neox_layers_1_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add441: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv7, gpt_neox_layers_1_mlp_dense_4h_to_h_bias4) + astype150: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add441, dtype="float16") + add442: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype150, add439) + add443: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add442, add437) + layer_norm151: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add443, 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_norm151, gpt_neox_layers_2_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add444: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv8, gpt_neox_layers_2_attention_query_key_value_bias4) + reshape296: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add444, R.shape([batch_size, 1, 48, 128])) + reshape297: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape296, R.shape([batch_size, 48, 128])) + lv75 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), reshape297), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape298: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv75, R.shape([batch_size, 1, 16, 128])) + reshape299: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape298, R.shape([batch_size, 1, 2048])) + lv9 = R.call_tir(cls.NT_matmul1, (reshape299, gpt_neox_layers_2_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add445: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv9, gpt_neox_layers_2_attention_dense_bias4) + layer_norm152: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add443, 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_norm152, gpt_neox_layers_2_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add446: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv10, gpt_neox_layers_2_mlp_dense_h_to_4h_bias4) + gelu74: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add446) + astype151: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu74, dtype="float16") + lv11 = R.call_tir(cls.NT_matmul3, (astype151, gpt_neox_layers_2_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add447: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv11, gpt_neox_layers_2_mlp_dense_4h_to_h_bias4) + astype152: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add447, dtype="float16") + add448: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype152, add445) + add449: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add448, add443) + layer_norm153: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add449, 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_norm153, gpt_neox_layers_3_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add450: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv12, gpt_neox_layers_3_attention_query_key_value_bias4) + reshape300: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add450, R.shape([batch_size, 1, 48, 128])) + reshape301: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape300, R.shape([batch_size, 48, 128])) + lv76 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), reshape301), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape302: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv76, R.shape([batch_size, 1, 16, 128])) + reshape303: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape302, R.shape([batch_size, 1, 2048])) + lv13 = R.call_tir(cls.NT_matmul1, (reshape303, gpt_neox_layers_3_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add451: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv13, gpt_neox_layers_3_attention_dense_bias4) + layer_norm154: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add449, 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_norm154, gpt_neox_layers_3_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add452: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv14, gpt_neox_layers_3_mlp_dense_h_to_4h_bias4) + gelu75: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add452) + astype153: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu75, dtype="float16") + lv15 = R.call_tir(cls.NT_matmul3, (astype153, gpt_neox_layers_3_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add453: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv15, gpt_neox_layers_3_mlp_dense_4h_to_h_bias4) + astype154: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add453, dtype="float16") + add454: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype154, add451) + add455: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add454, add449) + layer_norm155: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add455, 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_norm155, gpt_neox_layers_4_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add456: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv16, gpt_neox_layers_4_attention_query_key_value_bias4) + reshape304: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add456, R.shape([batch_size, 1, 48, 128])) + reshape305: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape304, R.shape([batch_size, 48, 128])) + lv77 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), reshape305), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape306: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv77, R.shape([batch_size, 1, 16, 128])) + reshape307: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape306, R.shape([batch_size, 1, 2048])) + lv17 = R.call_tir(cls.NT_matmul1, (reshape307, gpt_neox_layers_4_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add457: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv17, gpt_neox_layers_4_attention_dense_bias4) + layer_norm156: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add455, 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_norm156, gpt_neox_layers_4_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add458: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv18, gpt_neox_layers_4_mlp_dense_h_to_4h_bias4) + gelu76: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add458) + astype155: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu76, dtype="float16") + lv19 = R.call_tir(cls.NT_matmul3, (astype155, gpt_neox_layers_4_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add459: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv19, gpt_neox_layers_4_mlp_dense_4h_to_h_bias4) + astype156: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add459, dtype="float16") + add460: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype156, add457) + add461: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add460, add455) + layer_norm157: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add461, 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_norm157, gpt_neox_layers_5_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add462: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv20, gpt_neox_layers_5_attention_query_key_value_bias4) + reshape308: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add462, R.shape([batch_size, 1, 48, 128])) + reshape309: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape308, R.shape([batch_size, 48, 128])) + lv78 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), reshape309), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape310: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv78, R.shape([batch_size, 1, 16, 128])) + reshape311: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape310, R.shape([batch_size, 1, 2048])) + lv21 = R.call_tir(cls.NT_matmul1, (reshape311, gpt_neox_layers_5_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add463: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv21, gpt_neox_layers_5_attention_dense_bias4) + layer_norm158: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add461, 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_norm158, gpt_neox_layers_5_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add464: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv22, gpt_neox_layers_5_mlp_dense_h_to_4h_bias4) + gelu77: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add464) + astype157: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu77, dtype="float16") + lv23 = R.call_tir(cls.NT_matmul3, (astype157, gpt_neox_layers_5_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add465: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv23, gpt_neox_layers_5_mlp_dense_4h_to_h_bias4) + astype158: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add465, dtype="float16") + add466: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype158, add463) + add467: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add466, add461) + layer_norm159: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add467, 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_norm159, gpt_neox_layers_6_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add468: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv24, gpt_neox_layers_6_attention_query_key_value_bias4) + reshape312: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add468, R.shape([batch_size, 1, 48, 128])) + reshape313: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape312, R.shape([batch_size, 48, 128])) + lv79 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), reshape313), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape314: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv79, R.shape([batch_size, 1, 16, 128])) + reshape315: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape314, R.shape([batch_size, 1, 2048])) + lv25 = R.call_tir(cls.NT_matmul1, (reshape315, gpt_neox_layers_6_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add469: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv25, gpt_neox_layers_6_attention_dense_bias4) + layer_norm160: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add467, 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_norm160, gpt_neox_layers_6_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add470: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv26, gpt_neox_layers_6_mlp_dense_h_to_4h_bias4) + gelu78: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add470) + astype159: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu78, dtype="float16") + lv27 = R.call_tir(cls.NT_matmul3, (astype159, gpt_neox_layers_6_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add471: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv27, gpt_neox_layers_6_mlp_dense_4h_to_h_bias4) + astype160: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add471, dtype="float16") + add472: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype160, add469) + add473: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add472, add467) + layer_norm161: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add473, 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_norm161, gpt_neox_layers_7_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add474: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv28, gpt_neox_layers_7_attention_query_key_value_bias4) + reshape316: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add474, R.shape([batch_size, 1, 48, 128])) + reshape317: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape316, R.shape([batch_size, 48, 128])) + lv80 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), reshape317), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape318: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv80, R.shape([batch_size, 1, 16, 128])) + reshape319: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape318, R.shape([batch_size, 1, 2048])) + lv29 = R.call_tir(cls.NT_matmul1, (reshape319, gpt_neox_layers_7_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add475: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv29, gpt_neox_layers_7_attention_dense_bias4) + layer_norm162: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add473, 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_norm162, gpt_neox_layers_7_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add476: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv30, gpt_neox_layers_7_mlp_dense_h_to_4h_bias4) + gelu79: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add476) + astype161: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu79, dtype="float16") + lv31 = R.call_tir(cls.NT_matmul3, (astype161, gpt_neox_layers_7_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add477: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv31, gpt_neox_layers_7_mlp_dense_4h_to_h_bias4) + astype162: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add477, dtype="float16") + add478: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype162, add475) + add479: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add478, add473) + layer_norm163: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add479, 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_norm163, gpt_neox_layers_8_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add480: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv32, gpt_neox_layers_8_attention_query_key_value_bias4) + reshape320: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add480, R.shape([batch_size, 1, 48, 128])) + reshape321: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape320, R.shape([batch_size, 48, 128])) + lv81 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), reshape321), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape322: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv81, R.shape([batch_size, 1, 16, 128])) + reshape323: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape322, R.shape([batch_size, 1, 2048])) + lv33 = R.call_tir(cls.NT_matmul1, (reshape323, gpt_neox_layers_8_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add481: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv33, gpt_neox_layers_8_attention_dense_bias4) + layer_norm164: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add479, 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_norm164, gpt_neox_layers_8_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add482: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv34, gpt_neox_layers_8_mlp_dense_h_to_4h_bias4) + gelu80: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add482) + astype163: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu80, dtype="float16") + lv35 = R.call_tir(cls.NT_matmul3, (astype163, gpt_neox_layers_8_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add483: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv35, gpt_neox_layers_8_mlp_dense_4h_to_h_bias4) + astype164: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add483, dtype="float16") + add484: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype164, add481) + add485: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add484, add479) + layer_norm165: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add485, 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_norm165, gpt_neox_layers_9_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add486: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv36, gpt_neox_layers_9_attention_query_key_value_bias4) + reshape324: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add486, R.shape([batch_size, 1, 48, 128])) + reshape325: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape324, R.shape([batch_size, 48, 128])) + lv82 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), reshape325), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape326: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv82, R.shape([batch_size, 1, 16, 128])) + reshape327: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape326, R.shape([batch_size, 1, 2048])) + lv37 = R.call_tir(cls.NT_matmul1, (reshape327, gpt_neox_layers_9_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add487: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv37, gpt_neox_layers_9_attention_dense_bias4) + layer_norm166: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add485, 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_norm166, gpt_neox_layers_9_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add488: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv38, gpt_neox_layers_9_mlp_dense_h_to_4h_bias4) + gelu81: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add488) + astype165: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu81, dtype="float16") + lv39 = R.call_tir(cls.NT_matmul3, (astype165, gpt_neox_layers_9_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add489: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv39, gpt_neox_layers_9_mlp_dense_4h_to_h_bias4) + astype166: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add489, dtype="float16") + add490: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype166, add487) + add491: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add490, add485) + layer_norm167: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add491, 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_norm167, gpt_neox_layers_10_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add492: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv40, gpt_neox_layers_10_attention_query_key_value_bias4) + reshape328: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add492, R.shape([batch_size, 1, 48, 128])) + reshape329: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape328, R.shape([batch_size, 48, 128])) + lv83 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), reshape329), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape330: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv83, R.shape([batch_size, 1, 16, 128])) + reshape331: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape330, R.shape([batch_size, 1, 2048])) + lv41 = R.call_tir(cls.NT_matmul1, (reshape331, gpt_neox_layers_10_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add493: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv41, gpt_neox_layers_10_attention_dense_bias4) + layer_norm168: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add491, 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_norm168, gpt_neox_layers_10_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add494: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv42, gpt_neox_layers_10_mlp_dense_h_to_4h_bias4) + gelu82: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add494) + astype167: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu82, dtype="float16") + lv43 = R.call_tir(cls.NT_matmul3, (astype167, gpt_neox_layers_10_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add495: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv43, gpt_neox_layers_10_mlp_dense_4h_to_h_bias4) + astype168: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add495, dtype="float16") + add496: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype168, add493) + add497: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add496, add491) + layer_norm169: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add497, 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_norm169, gpt_neox_layers_11_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add498: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv44, gpt_neox_layers_11_attention_query_key_value_bias4) + reshape332: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add498, R.shape([batch_size, 1, 48, 128])) + reshape333: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape332, R.shape([batch_size, 48, 128])) + lv84 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), reshape333), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape334: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv84, R.shape([batch_size, 1, 16, 128])) + reshape335: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape334, R.shape([batch_size, 1, 2048])) + lv45 = R.call_tir(cls.NT_matmul1, (reshape335, gpt_neox_layers_11_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add499: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv45, gpt_neox_layers_11_attention_dense_bias4) + layer_norm170: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add497, 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_norm170, gpt_neox_layers_11_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add500: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv46, gpt_neox_layers_11_mlp_dense_h_to_4h_bias4) + gelu83: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add500) + astype169: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu83, dtype="float16") + lv47 = R.call_tir(cls.NT_matmul3, (astype169, gpt_neox_layers_11_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add501: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv47, gpt_neox_layers_11_mlp_dense_4h_to_h_bias4) + astype170: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add501, dtype="float16") + add502: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype170, add499) + add503: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add502, add497) + layer_norm171: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add503, 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_norm171, gpt_neox_layers_12_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add504: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv48, gpt_neox_layers_12_attention_query_key_value_bias4) + reshape336: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add504, R.shape([batch_size, 1, 48, 128])) + reshape337: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape336, R.shape([batch_size, 48, 128])) + lv85 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape337), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape338: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv85, R.shape([batch_size, 1, 16, 128])) + reshape339: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape338, R.shape([batch_size, 1, 2048])) + lv49 = R.call_tir(cls.NT_matmul1, (reshape339, gpt_neox_layers_12_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add505: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv49, gpt_neox_layers_12_attention_dense_bias4) + layer_norm172: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add503, 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 = R.call_tir(cls.NT_matmul2, (layer_norm172, gpt_neox_layers_12_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add506: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv50, gpt_neox_layers_12_mlp_dense_h_to_4h_bias4) + gelu84: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add506) + astype171: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu84, dtype="float16") + lv51 = R.call_tir(cls.NT_matmul3, (astype171, gpt_neox_layers_12_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add507: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv51, gpt_neox_layers_12_mlp_dense_4h_to_h_bias4) + astype172: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add507, dtype="float16") + add508: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype172, add505) + add509: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add508, add503) + layer_norm173: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add509, 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 = R.call_tir(cls.NT_matmul, (layer_norm173, gpt_neox_layers_13_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add510: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv52, gpt_neox_layers_13_attention_query_key_value_bias4) + reshape340: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add510, R.shape([batch_size, 1, 48, 128])) + reshape341: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape340, R.shape([batch_size, 48, 128])) + lv86 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape341), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape342: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv86, R.shape([batch_size, 1, 16, 128])) + reshape343: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape342, R.shape([batch_size, 1, 2048])) + lv53 = R.call_tir(cls.NT_matmul1, (reshape343, gpt_neox_layers_13_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add511: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv53, gpt_neox_layers_13_attention_dense_bias4) + layer_norm174: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add509, 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 = R.call_tir(cls.NT_matmul2, (layer_norm174, gpt_neox_layers_13_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add512: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv54, gpt_neox_layers_13_mlp_dense_h_to_4h_bias4) + gelu85: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add512) + astype173: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu85, dtype="float16") + lv55 = R.call_tir(cls.NT_matmul3, (astype173, gpt_neox_layers_13_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add513: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv55, gpt_neox_layers_13_mlp_dense_4h_to_h_bias4) + astype174: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add513, dtype="float16") + add514: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype174, add511) + add515: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add514, add509) + layer_norm175: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add515, 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 = R.call_tir(cls.NT_matmul, (layer_norm175, gpt_neox_layers_14_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add516: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv56, gpt_neox_layers_14_attention_query_key_value_bias4) + reshape344: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add516, R.shape([batch_size, 1, 48, 128])) + reshape345: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape344, R.shape([batch_size, 48, 128])) + lv87 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape345), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape346: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv87, R.shape([batch_size, 1, 16, 128])) + reshape347: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape346, R.shape([batch_size, 1, 2048])) + lv57 = R.call_tir(cls.NT_matmul1, (reshape347, gpt_neox_layers_14_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add517: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv57, gpt_neox_layers_14_attention_dense_bias4) + layer_norm176: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add515, 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 = R.call_tir(cls.NT_matmul2, (layer_norm176, gpt_neox_layers_14_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add518: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv58, gpt_neox_layers_14_mlp_dense_h_to_4h_bias4) + gelu86: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add518) + astype175: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu86, dtype="float16") + lv59 = R.call_tir(cls.NT_matmul3, (astype175, gpt_neox_layers_14_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add519: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv59, gpt_neox_layers_14_mlp_dense_4h_to_h_bias4) + astype176: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add519, dtype="float16") + add520: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype176, add517) + add521: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add520, add515) + layer_norm177: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add521, 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 = R.call_tir(cls.NT_matmul, (layer_norm177, gpt_neox_layers_15_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add522: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv60, gpt_neox_layers_15_attention_query_key_value_bias4) + reshape348: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add522, R.shape([batch_size, 1, 48, 128])) + reshape349: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape348, R.shape([batch_size, 48, 128])) + lv88 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape349), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape350: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv88, R.shape([batch_size, 1, 16, 128])) + reshape351: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape350, R.shape([batch_size, 1, 2048])) + lv61 = R.call_tir(cls.NT_matmul1, (reshape351, gpt_neox_layers_15_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add523: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv61, gpt_neox_layers_15_attention_dense_bias4) + layer_norm178: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add521, 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 = R.call_tir(cls.NT_matmul2, (layer_norm178, gpt_neox_layers_15_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add524: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv62, gpt_neox_layers_15_mlp_dense_h_to_4h_bias4) + gelu87: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add524) + astype177: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu87, dtype="float16") + lv63 = R.call_tir(cls.NT_matmul3, (astype177, gpt_neox_layers_15_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add525: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv63, gpt_neox_layers_15_mlp_dense_4h_to_h_bias4) + astype178: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add525, dtype="float16") + add526: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype178, add523) + add527: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add526, add521) + layer_norm179: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add527, gpt_neox_layers_16_input_layernorm_weight4, gpt_neox_layers_16_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv64 = R.call_tir(cls.NT_matmul, (layer_norm179, gpt_neox_layers_16_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add528: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv64, gpt_neox_layers_16_attention_query_key_value_bias4) + reshape352: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add528, R.shape([batch_size, 1, 48, 128])) + reshape353: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape352, R.shape([batch_size, 48, 128])) + lv89 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), reshape353), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape354: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv89, R.shape([batch_size, 1, 16, 128])) + reshape355: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape354, R.shape([batch_size, 1, 2048])) + lv65 = R.call_tir(cls.NT_matmul1, (reshape355, gpt_neox_layers_16_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add529: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv65, gpt_neox_layers_16_attention_dense_bias4) + layer_norm180: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add527, gpt_neox_layers_16_post_attention_layernorm_weight4, gpt_neox_layers_16_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv66 = R.call_tir(cls.NT_matmul2, (layer_norm180, gpt_neox_layers_16_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add530: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv66, gpt_neox_layers_16_mlp_dense_h_to_4h_bias4) + gelu88: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add530) + astype179: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu88, dtype="float16") + lv67 = R.call_tir(cls.NT_matmul3, (astype179, gpt_neox_layers_16_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add531: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv67, gpt_neox_layers_16_mlp_dense_4h_to_h_bias4) + astype180: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add531, dtype="float16") + add532: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype180, add529) + add533: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add532, add527) + layer_norm181: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add533, gpt_neox_layers_17_input_layernorm_weight4, gpt_neox_layers_17_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv68 = R.call_tir(cls.NT_matmul, (layer_norm181, gpt_neox_layers_17_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add534: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv68, gpt_neox_layers_17_attention_query_key_value_bias4) + reshape356: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add534, R.shape([batch_size, 1, 48, 128])) + reshape357: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape356, R.shape([batch_size, 48, 128])) + lv90 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), reshape357), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape358: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv90, R.shape([batch_size, 1, 16, 128])) + reshape359: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape358, R.shape([batch_size, 1, 2048])) + lv69 = R.call_tir(cls.NT_matmul1, (reshape359, gpt_neox_layers_17_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add535: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv69, gpt_neox_layers_17_attention_dense_bias4) + layer_norm182: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add533, gpt_neox_layers_17_post_attention_layernorm_weight4, gpt_neox_layers_17_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv70 = R.call_tir(cls.NT_matmul2, (layer_norm182, gpt_neox_layers_17_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add536: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv70, gpt_neox_layers_17_mlp_dense_h_to_4h_bias4) + gelu89: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add536) + astype181: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu89, dtype="float16") + lv71 = R.call_tir(cls.NT_matmul3, (astype181, gpt_neox_layers_17_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add537: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv71, gpt_neox_layers_17_mlp_dense_4h_to_h_bias4) + astype182: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add537, dtype="float16") + add538: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype182, add535) + add539: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add538, add533) + layer_norm183: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add539, gpt_neox_layers_18_input_layernorm_weight4, gpt_neox_layers_18_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv72 = R.call_tir(cls.NT_matmul, (layer_norm183, gpt_neox_layers_18_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add540: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv72, gpt_neox_layers_18_attention_query_key_value_bias4) + reshape360: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add540, R.shape([batch_size, 1, 48, 128])) + reshape361: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape360, R.shape([batch_size, 48, 128])) + lv91 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), reshape361), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape362: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv91, R.shape([batch_size, 1, 16, 128])) + reshape363: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape362, R.shape([batch_size, 1, 2048])) + lv73_1 = R.call_tir(cls.NT_matmul1, (reshape363, gpt_neox_layers_18_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add541: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv73_1, gpt_neox_layers_18_attention_dense_bias4) + layer_norm184: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add539, gpt_neox_layers_18_post_attention_layernorm_weight4, gpt_neox_layers_18_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv74_1 = R.call_tir(cls.NT_matmul2, (layer_norm184, gpt_neox_layers_18_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add542: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv74_1, gpt_neox_layers_18_mlp_dense_h_to_4h_bias4) + gelu90: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add542) + astype183: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu90, dtype="float16") + lv75_1 = R.call_tir(cls.NT_matmul3, (astype183, gpt_neox_layers_18_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add543: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv75_1, gpt_neox_layers_18_mlp_dense_4h_to_h_bias4) + astype184: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add543, dtype="float16") + add544: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype184, add541) + add545: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add544, add539) + layer_norm185: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add545, gpt_neox_layers_19_input_layernorm_weight4, gpt_neox_layers_19_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv76_1 = R.call_tir(cls.NT_matmul, (layer_norm185, gpt_neox_layers_19_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add546: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv76_1, gpt_neox_layers_19_attention_query_key_value_bias4) + reshape364: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add546, R.shape([batch_size, 1, 48, 128])) + reshape365: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape364, R.shape([batch_size, 48, 128])) + lv92 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), reshape365), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape366: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv92, R.shape([batch_size, 1, 16, 128])) + reshape367: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape366, R.shape([batch_size, 1, 2048])) + lv77_1 = R.call_tir(cls.NT_matmul1, (reshape367, gpt_neox_layers_19_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add547: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv77_1, gpt_neox_layers_19_attention_dense_bias4) + layer_norm186: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add545, gpt_neox_layers_19_post_attention_layernorm_weight4, gpt_neox_layers_19_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv78_1 = R.call_tir(cls.NT_matmul2, (layer_norm186, gpt_neox_layers_19_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add548: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv78_1, gpt_neox_layers_19_mlp_dense_h_to_4h_bias4) + gelu91: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add548) + astype185: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu91, dtype="float16") + lv79_1 = R.call_tir(cls.NT_matmul3, (astype185, gpt_neox_layers_19_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add549: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv79_1, gpt_neox_layers_19_mlp_dense_4h_to_h_bias4) + astype186: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add549, dtype="float16") + add550: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype186, add547) + add551: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add550, add545) + layer_norm187: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add551, gpt_neox_layers_20_input_layernorm_weight4, gpt_neox_layers_20_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv80_1 = R.call_tir(cls.NT_matmul, (layer_norm187, gpt_neox_layers_20_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add552: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv80_1, gpt_neox_layers_20_attention_query_key_value_bias4) + reshape368: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add552, R.shape([batch_size, 1, 48, 128])) + reshape369: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape368, R.shape([batch_size, 48, 128])) + lv93 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), reshape369), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape370: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv93, R.shape([batch_size, 1, 16, 128])) + reshape371: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape370, R.shape([batch_size, 1, 2048])) + lv81_1 = R.call_tir(cls.NT_matmul1, (reshape371, gpt_neox_layers_20_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add553: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv81_1, gpt_neox_layers_20_attention_dense_bias4) + layer_norm188: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add551, gpt_neox_layers_20_post_attention_layernorm_weight4, gpt_neox_layers_20_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv82_1 = R.call_tir(cls.NT_matmul2, (layer_norm188, gpt_neox_layers_20_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add554: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv82_1, gpt_neox_layers_20_mlp_dense_h_to_4h_bias4) + gelu92: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add554) + astype187: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu92, dtype="float16") + lv83_1 = R.call_tir(cls.NT_matmul3, (astype187, gpt_neox_layers_20_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add555: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv83_1, gpt_neox_layers_20_mlp_dense_4h_to_h_bias4) + astype188: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add555, dtype="float16") + add556: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype188, add553) + add557: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add556, add551) + layer_norm189: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add557, gpt_neox_layers_21_input_layernorm_weight4, gpt_neox_layers_21_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv84_1 = R.call_tir(cls.NT_matmul, (layer_norm189, gpt_neox_layers_21_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add558: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv84_1, gpt_neox_layers_21_attention_query_key_value_bias4) + reshape372: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add558, R.shape([batch_size, 1, 48, 128])) + reshape373: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape372, R.shape([batch_size, 48, 128])) + lv94 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), reshape373), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape374: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv94, R.shape([batch_size, 1, 16, 128])) + reshape375: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape374, R.shape([batch_size, 1, 2048])) + lv85_1 = R.call_tir(cls.NT_matmul1, (reshape375, gpt_neox_layers_21_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add559: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv85_1, gpt_neox_layers_21_attention_dense_bias4) + layer_norm190: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add557, gpt_neox_layers_21_post_attention_layernorm_weight4, gpt_neox_layers_21_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv86_1 = R.call_tir(cls.NT_matmul2, (layer_norm190, gpt_neox_layers_21_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add560: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv86_1, gpt_neox_layers_21_mlp_dense_h_to_4h_bias4) + gelu93: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add560) + astype189: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu93, dtype="float16") + lv87_1 = R.call_tir(cls.NT_matmul3, (astype189, gpt_neox_layers_21_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add561: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv87_1, gpt_neox_layers_21_mlp_dense_4h_to_h_bias4) + astype190: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add561, dtype="float16") + add562: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype190, add559) + add563: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add562, add557) + layer_norm191: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add563, gpt_neox_layers_22_input_layernorm_weight4, gpt_neox_layers_22_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv88_1 = R.call_tir(cls.NT_matmul, (layer_norm191, gpt_neox_layers_22_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add564: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv88_1, gpt_neox_layers_22_attention_query_key_value_bias4) + reshape376: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add564, R.shape([batch_size, 1, 48, 128])) + reshape377: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape376, R.shape([batch_size, 48, 128])) + lv95 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), reshape377), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape378: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv95, R.shape([batch_size, 1, 16, 128])) + reshape379: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape378, R.shape([batch_size, 1, 2048])) + lv89_1 = R.call_tir(cls.NT_matmul1, (reshape379, gpt_neox_layers_22_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add565: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv89_1, gpt_neox_layers_22_attention_dense_bias4) + layer_norm192: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add563, gpt_neox_layers_22_post_attention_layernorm_weight4, gpt_neox_layers_22_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv90_1 = R.call_tir(cls.NT_matmul2, (layer_norm192, gpt_neox_layers_22_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add566: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv90_1, gpt_neox_layers_22_mlp_dense_h_to_4h_bias4) + gelu94: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add566) + astype191: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu94, dtype="float16") + lv91_1 = R.call_tir(cls.NT_matmul3, (astype191, gpt_neox_layers_22_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add567: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv91_1, gpt_neox_layers_22_mlp_dense_4h_to_h_bias4) + astype192: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add567, dtype="float16") + add568: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype192, add565) + add569: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add568, add563) + layer_norm193: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add569, gpt_neox_layers_23_input_layernorm_weight4, gpt_neox_layers_23_input_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv92_1 = R.call_tir(cls.NT_matmul, (layer_norm193, gpt_neox_layers_23_attention_query_key_value_weight4), out_sinfo=R.Tensor((batch_size, 1, 6144), dtype="float16")) + add570: R.Tensor((batch_size, 1, 6144), dtype="float16") = R.add(lv92_1, gpt_neox_layers_23_attention_query_key_value_bias4) + reshape380: R.Tensor((batch_size, 1, 48, 128), dtype="float16") = R.reshape(add570, R.shape([batch_size, 1, 48, 128])) + reshape381: R.Tensor((batch_size, 48, 128), dtype="float16") = R.reshape(reshape380, R.shape([batch_size, 48, 128])) + lv96 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), reshape381), out_sinfo=R.Tensor((batch_size, 16, 128), dtype="float16")) + reshape382: R.Tensor((batch_size, 1, 16, 128), dtype="float16") = R.reshape(lv96, R.shape([batch_size, 1, 16, 128])) + reshape383: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.reshape(reshape382, R.shape([batch_size, 1, 2048])) + lv93_1 = R.call_tir(cls.NT_matmul1, (reshape383, gpt_neox_layers_23_attention_dense_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float16")) + add571: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(lv93_1, gpt_neox_layers_23_attention_dense_bias4) + layer_norm194: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add569, gpt_neox_layers_23_post_attention_layernorm_weight4, gpt_neox_layers_23_post_attention_layernorm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv94_1 = R.call_tir(cls.NT_matmul2, (layer_norm194, gpt_neox_layers_23_mlp_dense_h_to_4h_weight4), out_sinfo=R.Tensor((batch_size, 1, 8192), dtype="float32")) + add572: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.add(lv94_1, gpt_neox_layers_23_mlp_dense_h_to_4h_bias4) + gelu95: R.Tensor((batch_size, 1, 8192), dtype="float32") = R.nn.gelu(add572) + astype193: R.Tensor((batch_size, 1, 8192), dtype="float16") = R.astype(gelu95, dtype="float16") + lv95_1 = R.call_tir(cls.NT_matmul3, (astype193, gpt_neox_layers_23_mlp_dense_4h_to_h_weight4), out_sinfo=R.Tensor((batch_size, 1, 2048), dtype="float32")) + add573: R.Tensor((batch_size, 1, 2048), dtype="float32") = R.add(lv95_1, gpt_neox_layers_23_mlp_dense_4h_to_h_bias4) + astype194: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.astype(add573, dtype="float16") + add574: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(astype194, add571) + add575: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.add(add574, add569) + layer_norm195: R.Tensor((batch_size, 1, 2048), dtype="float16") = R.nn.layer_norm(add575, gpt_neox_final_layer_norm_weight4, gpt_neox_final_layer_norm_bias4, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv96_1 = R.call_tir(cls.NT_matmul4, (layer_norm195, embed_out_weight4), out_sinfo=R.Tensor((batch_size, 1, vocab_size), dtype="float16")) + astype195: R.Tensor((batch_size, 1, vocab_size), dtype="float32") = R.astype(lv96_1, dtype="float32") + gv4: R.Tuple(R.Tensor((batch_size, 1, vocab_size), dtype="float32"), R.Object) = astype195, 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((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_layers_16_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[193] + gpt_neox_layers_16_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[194] + gpt_neox_layers_16_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[195] + gpt_neox_layers_16_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[196] + gpt_neox_layers_16_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[197] + gpt_neox_layers_16_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[198] + gpt_neox_layers_16_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[199] + gpt_neox_layers_16_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[200] + gpt_neox_layers_16_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[201] + gpt_neox_layers_16_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[202] + gpt_neox_layers_16_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[203] + gpt_neox_layers_16_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[204] + gpt_neox_layers_17_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[205] + gpt_neox_layers_17_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[206] + gpt_neox_layers_17_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[207] + gpt_neox_layers_17_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[208] + gpt_neox_layers_17_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[209] + gpt_neox_layers_17_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[210] + gpt_neox_layers_17_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[211] + gpt_neox_layers_17_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[212] + gpt_neox_layers_17_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[213] + gpt_neox_layers_17_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[214] + gpt_neox_layers_17_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[215] + gpt_neox_layers_17_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[216] + gpt_neox_layers_18_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[217] + gpt_neox_layers_18_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[218] + gpt_neox_layers_18_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[219] + gpt_neox_layers_18_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[220] + gpt_neox_layers_18_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[221] + gpt_neox_layers_18_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[222] + gpt_neox_layers_18_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[223] + gpt_neox_layers_18_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[224] + gpt_neox_layers_18_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[225] + gpt_neox_layers_18_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[226] + gpt_neox_layers_18_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[227] + gpt_neox_layers_18_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[228] + gpt_neox_layers_19_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[229] + gpt_neox_layers_19_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[230] + gpt_neox_layers_19_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[231] + gpt_neox_layers_19_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[232] + gpt_neox_layers_19_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[233] + gpt_neox_layers_19_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[234] + gpt_neox_layers_19_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[235] + gpt_neox_layers_19_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[236] + gpt_neox_layers_19_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[237] + gpt_neox_layers_19_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[238] + gpt_neox_layers_19_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[239] + gpt_neox_layers_19_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[240] + gpt_neox_layers_20_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[241] + gpt_neox_layers_20_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[242] + gpt_neox_layers_20_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[243] + gpt_neox_layers_20_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[244] + gpt_neox_layers_20_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[245] + gpt_neox_layers_20_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[246] + gpt_neox_layers_20_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[247] + gpt_neox_layers_20_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[248] + gpt_neox_layers_20_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[249] + gpt_neox_layers_20_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[250] + gpt_neox_layers_20_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[251] + gpt_neox_layers_20_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[252] + gpt_neox_layers_21_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[253] + gpt_neox_layers_21_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[254] + gpt_neox_layers_21_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[255] + gpt_neox_layers_21_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[256] + gpt_neox_layers_21_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[257] + gpt_neox_layers_21_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[258] + gpt_neox_layers_21_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[259] + gpt_neox_layers_21_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[260] + gpt_neox_layers_21_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[261] + gpt_neox_layers_21_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[262] + gpt_neox_layers_21_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[263] + gpt_neox_layers_21_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[264] + gpt_neox_layers_22_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[265] + gpt_neox_layers_22_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[266] + gpt_neox_layers_22_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[267] + gpt_neox_layers_22_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[268] + gpt_neox_layers_22_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[269] + gpt_neox_layers_22_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[270] + gpt_neox_layers_22_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[271] + gpt_neox_layers_22_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[272] + gpt_neox_layers_22_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[273] + gpt_neox_layers_22_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[274] + gpt_neox_layers_22_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[275] + gpt_neox_layers_22_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[276] + gpt_neox_layers_23_input_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[277] + gpt_neox_layers_23_input_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[278] + gpt_neox_layers_23_post_attention_layernorm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[279] + gpt_neox_layers_23_post_attention_layernorm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[280] + gpt_neox_layers_23_attention_query_key_value_weight3: R.Tensor((6144, 2048), dtype="float16") = packed_params[281] + gpt_neox_layers_23_attention_query_key_value_bias3: R.Tensor((6144,), dtype="float16") = packed_params[282] + gpt_neox_layers_23_attention_dense_weight3: R.Tensor((2048, 2048), dtype="float16") = packed_params[283] + gpt_neox_layers_23_attention_dense_bias3: R.Tensor((2048,), dtype="float16") = packed_params[284] + gpt_neox_layers_23_mlp_dense_h_to_4h_weight3: R.Tensor((8192, 2048), dtype="float16") = packed_params[285] + gpt_neox_layers_23_mlp_dense_h_to_4h_bias3: R.Tensor((8192,), dtype="float32") = packed_params[286] + gpt_neox_layers_23_mlp_dense_4h_to_h_weight3: R.Tensor((2048, 8192), dtype="float16") = packed_params[287] + gpt_neox_layers_23_mlp_dense_4h_to_h_bias3: R.Tensor((2048,), dtype="float32") = packed_params[288] + gpt_neox_final_layer_norm_weight3: R.Tensor((2048,), dtype="float16") = packed_params[289] + gpt_neox_final_layer_norm_bias3: R.Tensor((2048,), dtype="float16") = packed_params[290] + embed_out_weight3: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[291] + layer_norm98: 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) + lv97 = R.call_tir(cls.NT_matmul5, (layer_norm98, gpt_neox_layers_0_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add288: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv97, gpt_neox_layers_0_attention_query_key_value_bias3) + reshape192: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add288, R.shape([1, seq_len, 48, 128])) + reshape193: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape192, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape194: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv49, R.shape([1, seq_len, 16, 128])) + reshape195: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape194, R.shape([1, seq_len, 2048])) + lv98 = R.call_tir(cls.NT_matmul6, (reshape195, gpt_neox_layers_0_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add289: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv98, gpt_neox_layers_0_attention_dense_bias3) + layer_norm99: 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) + lv99 = R.call_tir(cls.NT_matmul7, (layer_norm99, gpt_neox_layers_0_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add290: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv99, gpt_neox_layers_0_mlp_dense_h_to_4h_bias3) + gelu48: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add290) + astype98: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu48, dtype="float16") + lv100 = R.call_tir(cls.NT_matmul8, (astype98, gpt_neox_layers_0_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add291: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv100, gpt_neox_layers_0_mlp_dense_4h_to_h_bias3) + astype99: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add291, dtype="float16") + add292: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype99, add289) + add293: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add292, input_embeds) + layer_norm100: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add293, gpt_neox_layers_1_input_layernorm_weight3, gpt_neox_layers_1_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv101 = R.call_tir(cls.NT_matmul5, (layer_norm100, gpt_neox_layers_1_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add294: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv101, gpt_neox_layers_1_attention_query_key_value_bias3) + reshape196: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add294, R.shape([1, seq_len, 48, 128])) + reshape197: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape196, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape198: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv50, R.shape([1, seq_len, 16, 128])) + reshape199: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape198, R.shape([1, seq_len, 2048])) + lv102 = R.call_tir(cls.NT_matmul6, (reshape199, gpt_neox_layers_1_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add295: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv102, gpt_neox_layers_1_attention_dense_bias3) + layer_norm101: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add293, 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) + lv103 = R.call_tir(cls.NT_matmul7, (layer_norm101, gpt_neox_layers_1_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add296: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv103, gpt_neox_layers_1_mlp_dense_h_to_4h_bias3) + gelu49: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add296) + astype100: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu49, dtype="float16") + lv104 = R.call_tir(cls.NT_matmul8, (astype100, gpt_neox_layers_1_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add297: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv104, gpt_neox_layers_1_mlp_dense_4h_to_h_bias3) + astype101: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add297, dtype="float16") + add298: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype101, add295) + add299: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add298, add293) + layer_norm102: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add299, gpt_neox_layers_2_input_layernorm_weight3, gpt_neox_layers_2_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv105 = R.call_tir(cls.NT_matmul5, (layer_norm102, gpt_neox_layers_2_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add300: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv105, gpt_neox_layers_2_attention_query_key_value_bias3) + reshape200: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add300, R.shape([1, seq_len, 48, 128])) + reshape201: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape200, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape202: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv51, R.shape([1, seq_len, 16, 128])) + reshape203: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape202, R.shape([1, seq_len, 2048])) + lv106 = R.call_tir(cls.NT_matmul6, (reshape203, gpt_neox_layers_2_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add301: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv106, gpt_neox_layers_2_attention_dense_bias3) + layer_norm103: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add299, 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) + lv107 = R.call_tir(cls.NT_matmul7, (layer_norm103, gpt_neox_layers_2_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add302: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv107, gpt_neox_layers_2_mlp_dense_h_to_4h_bias3) + gelu50: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add302) + astype102: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu50, dtype="float16") + lv108 = R.call_tir(cls.NT_matmul8, (astype102, gpt_neox_layers_2_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add303: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv108, gpt_neox_layers_2_mlp_dense_4h_to_h_bias3) + astype103: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add303, dtype="float16") + add304: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype103, add301) + add305: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add304, add299) + layer_norm104: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add305, gpt_neox_layers_3_input_layernorm_weight3, gpt_neox_layers_3_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv109 = R.call_tir(cls.NT_matmul5, (layer_norm104, gpt_neox_layers_3_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add306: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv109, gpt_neox_layers_3_attention_query_key_value_bias3) + reshape204: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add306, R.shape([1, seq_len, 48, 128])) + reshape205: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape204, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape206: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv52, R.shape([1, seq_len, 16, 128])) + reshape207: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape206, R.shape([1, seq_len, 2048])) + lv110 = R.call_tir(cls.NT_matmul6, (reshape207, gpt_neox_layers_3_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add307: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv110, gpt_neox_layers_3_attention_dense_bias3) + layer_norm105: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add305, 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) + lv111 = R.call_tir(cls.NT_matmul7, (layer_norm105, gpt_neox_layers_3_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add308: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv111, gpt_neox_layers_3_mlp_dense_h_to_4h_bias3) + gelu51: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add308) + astype104: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu51, dtype="float16") + lv112 = R.call_tir(cls.NT_matmul8, (astype104, gpt_neox_layers_3_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add309: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv112, gpt_neox_layers_3_mlp_dense_4h_to_h_bias3) + astype105: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add309, dtype="float16") + add310: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype105, add307) + add311: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add310, add305) + layer_norm106: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add311, gpt_neox_layers_4_input_layernorm_weight3, gpt_neox_layers_4_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv113 = R.call_tir(cls.NT_matmul5, (layer_norm106, gpt_neox_layers_4_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add312: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv113, gpt_neox_layers_4_attention_query_key_value_bias3) + reshape208: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add312, R.shape([1, seq_len, 48, 128])) + reshape209: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape208, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape210: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv53, R.shape([1, seq_len, 16, 128])) + reshape211: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape210, R.shape([1, seq_len, 2048])) + lv114 = R.call_tir(cls.NT_matmul6, (reshape211, gpt_neox_layers_4_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add313: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv114, gpt_neox_layers_4_attention_dense_bias3) + layer_norm107: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add311, 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) + lv115 = R.call_tir(cls.NT_matmul7, (layer_norm107, gpt_neox_layers_4_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add314: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv115, gpt_neox_layers_4_mlp_dense_h_to_4h_bias3) + gelu52: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add314) + astype106: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu52, dtype="float16") + lv116 = R.call_tir(cls.NT_matmul8, (astype106, gpt_neox_layers_4_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add315: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv116, gpt_neox_layers_4_mlp_dense_4h_to_h_bias3) + astype107: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add315, dtype="float16") + add316: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype107, add313) + add317: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add316, add311) + layer_norm108: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add317, gpt_neox_layers_5_input_layernorm_weight3, gpt_neox_layers_5_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv117 = R.call_tir(cls.NT_matmul5, (layer_norm108, gpt_neox_layers_5_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add318: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv117, gpt_neox_layers_5_attention_query_key_value_bias3) + reshape212: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add318, R.shape([1, seq_len, 48, 128])) + reshape213: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape212, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape214: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv54, R.shape([1, seq_len, 16, 128])) + reshape215: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape214, R.shape([1, seq_len, 2048])) + lv118 = R.call_tir(cls.NT_matmul6, (reshape215, gpt_neox_layers_5_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add319: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv118, gpt_neox_layers_5_attention_dense_bias3) + layer_norm109: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add317, 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) + lv119 = R.call_tir(cls.NT_matmul7, (layer_norm109, gpt_neox_layers_5_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add320: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv119, gpt_neox_layers_5_mlp_dense_h_to_4h_bias3) + gelu53: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add320) + astype108: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu53, dtype="float16") + lv120 = R.call_tir(cls.NT_matmul8, (astype108, gpt_neox_layers_5_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add321: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv120, gpt_neox_layers_5_mlp_dense_4h_to_h_bias3) + astype109: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add321, dtype="float16") + add322: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype109, add319) + add323: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add322, add317) + layer_norm110: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add323, gpt_neox_layers_6_input_layernorm_weight3, gpt_neox_layers_6_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv121 = R.call_tir(cls.NT_matmul5, (layer_norm110, gpt_neox_layers_6_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add324: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv121, gpt_neox_layers_6_attention_query_key_value_bias3) + reshape216: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add324, R.shape([1, seq_len, 48, 128])) + reshape217: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape216, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape218: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv55, R.shape([1, seq_len, 16, 128])) + reshape219: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape218, R.shape([1, seq_len, 2048])) + lv122 = R.call_tir(cls.NT_matmul6, (reshape219, gpt_neox_layers_6_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add325: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv122, gpt_neox_layers_6_attention_dense_bias3) + layer_norm111: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add323, 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) + lv123 = R.call_tir(cls.NT_matmul7, (layer_norm111, gpt_neox_layers_6_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add326: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv123, gpt_neox_layers_6_mlp_dense_h_to_4h_bias3) + gelu54: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add326) + astype110: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu54, dtype="float16") + lv124 = R.call_tir(cls.NT_matmul8, (astype110, gpt_neox_layers_6_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add327: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv124, gpt_neox_layers_6_mlp_dense_4h_to_h_bias3) + astype111: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add327, dtype="float16") + add328: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype111, add325) + add329: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add328, add323) + layer_norm112: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add329, gpt_neox_layers_7_input_layernorm_weight3, gpt_neox_layers_7_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv125 = R.call_tir(cls.NT_matmul5, (layer_norm112, gpt_neox_layers_7_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add330: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv125, gpt_neox_layers_7_attention_query_key_value_bias3) + reshape220: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add330, R.shape([1, seq_len, 48, 128])) + reshape221: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape220, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape222: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv56, R.shape([1, seq_len, 16, 128])) + reshape223: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape222, R.shape([1, seq_len, 2048])) + lv126 = R.call_tir(cls.NT_matmul6, (reshape223, gpt_neox_layers_7_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add331: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv126, gpt_neox_layers_7_attention_dense_bias3) + layer_norm113: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add329, 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) + lv127 = R.call_tir(cls.NT_matmul7, (layer_norm113, gpt_neox_layers_7_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add332: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv127, gpt_neox_layers_7_mlp_dense_h_to_4h_bias3) + gelu55: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add332) + astype112: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu55, dtype="float16") + lv128 = R.call_tir(cls.NT_matmul8, (astype112, gpt_neox_layers_7_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add333: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv128, gpt_neox_layers_7_mlp_dense_4h_to_h_bias3) + astype113: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add333, dtype="float16") + add334: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype113, add331) + add335: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add334, add329) + layer_norm114: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add335, gpt_neox_layers_8_input_layernorm_weight3, gpt_neox_layers_8_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv129 = R.call_tir(cls.NT_matmul5, (layer_norm114, gpt_neox_layers_8_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add336: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv129, gpt_neox_layers_8_attention_query_key_value_bias3) + reshape224: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add336, R.shape([1, seq_len, 48, 128])) + reshape225: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape224, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape226: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv57, R.shape([1, seq_len, 16, 128])) + reshape227: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape226, R.shape([1, seq_len, 2048])) + lv130 = R.call_tir(cls.NT_matmul6, (reshape227, gpt_neox_layers_8_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add337: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv130, gpt_neox_layers_8_attention_dense_bias3) + layer_norm115: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add335, 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) + lv131 = R.call_tir(cls.NT_matmul7, (layer_norm115, gpt_neox_layers_8_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add338: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv131, gpt_neox_layers_8_mlp_dense_h_to_4h_bias3) + gelu56: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add338) + astype114: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu56, dtype="float16") + lv132 = R.call_tir(cls.NT_matmul8, (astype114, gpt_neox_layers_8_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add339: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv132, gpt_neox_layers_8_mlp_dense_4h_to_h_bias3) + astype115: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add339, dtype="float16") + add340: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype115, add337) + add341: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add340, add335) + layer_norm116: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add341, gpt_neox_layers_9_input_layernorm_weight3, gpt_neox_layers_9_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv133 = R.call_tir(cls.NT_matmul5, (layer_norm116, gpt_neox_layers_9_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add342: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv133, gpt_neox_layers_9_attention_query_key_value_bias3) + reshape228: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add342, R.shape([1, seq_len, 48, 128])) + reshape229: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape228, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape230: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv58, R.shape([1, seq_len, 16, 128])) + reshape231: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape230, R.shape([1, seq_len, 2048])) + lv134 = R.call_tir(cls.NT_matmul6, (reshape231, gpt_neox_layers_9_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add343: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv134, gpt_neox_layers_9_attention_dense_bias3) + layer_norm117: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add341, 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) + lv135 = R.call_tir(cls.NT_matmul7, (layer_norm117, gpt_neox_layers_9_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add344: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv135, gpt_neox_layers_9_mlp_dense_h_to_4h_bias3) + gelu57: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add344) + astype116: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu57, dtype="float16") + lv136 = R.call_tir(cls.NT_matmul8, (astype116, gpt_neox_layers_9_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add345: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv136, gpt_neox_layers_9_mlp_dense_4h_to_h_bias3) + astype117: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add345, dtype="float16") + add346: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype117, add343) + add347: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add346, add341) + layer_norm118: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add347, gpt_neox_layers_10_input_layernorm_weight3, gpt_neox_layers_10_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv137 = R.call_tir(cls.NT_matmul5, (layer_norm118, gpt_neox_layers_10_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add348: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv137, gpt_neox_layers_10_attention_query_key_value_bias3) + reshape232: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add348, R.shape([1, seq_len, 48, 128])) + reshape233: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape232, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape234: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv59, R.shape([1, seq_len, 16, 128])) + reshape235: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape234, R.shape([1, seq_len, 2048])) + lv138 = R.call_tir(cls.NT_matmul6, (reshape235, gpt_neox_layers_10_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add349: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv138, gpt_neox_layers_10_attention_dense_bias3) + layer_norm119: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add347, 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) + lv139 = R.call_tir(cls.NT_matmul7, (layer_norm119, gpt_neox_layers_10_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add350: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv139, gpt_neox_layers_10_mlp_dense_h_to_4h_bias3) + gelu58: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add350) + astype118: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu58, dtype="float16") + lv140 = R.call_tir(cls.NT_matmul8, (astype118, gpt_neox_layers_10_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add351: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv140, gpt_neox_layers_10_mlp_dense_4h_to_h_bias3) + astype119: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add351, dtype="float16") + add352: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype119, add349) + add353: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add352, add347) + layer_norm120: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add353, gpt_neox_layers_11_input_layernorm_weight3, gpt_neox_layers_11_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv141 = R.call_tir(cls.NT_matmul5, (layer_norm120, gpt_neox_layers_11_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add354: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv141, gpt_neox_layers_11_attention_query_key_value_bias3) + reshape236: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add354, R.shape([1, seq_len, 48, 128])) + reshape237: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape236, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape238: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv60, R.shape([1, seq_len, 16, 128])) + reshape239: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape238, R.shape([1, seq_len, 2048])) + lv142 = R.call_tir(cls.NT_matmul6, (reshape239, gpt_neox_layers_11_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add355: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv142, gpt_neox_layers_11_attention_dense_bias3) + layer_norm121: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add353, 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) + lv143 = R.call_tir(cls.NT_matmul7, (layer_norm121, gpt_neox_layers_11_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add356: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv143, gpt_neox_layers_11_mlp_dense_h_to_4h_bias3) + gelu59: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add356) + astype120: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu59, dtype="float16") + lv144 = R.call_tir(cls.NT_matmul8, (astype120, gpt_neox_layers_11_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add357: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv144, gpt_neox_layers_11_mlp_dense_4h_to_h_bias3) + astype121: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add357, dtype="float16") + add358: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype121, add355) + add359: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add358, add353) + layer_norm122: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add359, gpt_neox_layers_12_input_layernorm_weight3, gpt_neox_layers_12_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv145 = R.call_tir(cls.NT_matmul5, (layer_norm122, gpt_neox_layers_12_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add360: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv145, gpt_neox_layers_12_attention_query_key_value_bias3) + reshape240: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add360, R.shape([1, seq_len, 48, 128])) + reshape241: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape240, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape242: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv61, R.shape([1, seq_len, 16, 128])) + reshape243: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape242, R.shape([1, seq_len, 2048])) + lv146 = R.call_tir(cls.NT_matmul6, (reshape243, gpt_neox_layers_12_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add361: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv146, gpt_neox_layers_12_attention_dense_bias3) + layer_norm123: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add359, 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) + lv147 = R.call_tir(cls.NT_matmul7, (layer_norm123, gpt_neox_layers_12_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add362: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv147, gpt_neox_layers_12_mlp_dense_h_to_4h_bias3) + gelu60: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add362) + astype122: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu60, dtype="float16") + lv148 = R.call_tir(cls.NT_matmul8, (astype122, gpt_neox_layers_12_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add363: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv148, gpt_neox_layers_12_mlp_dense_4h_to_h_bias3) + astype123: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add363, dtype="float16") + add364: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype123, add361) + add365: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add364, add359) + layer_norm124: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add365, gpt_neox_layers_13_input_layernorm_weight3, gpt_neox_layers_13_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv149 = R.call_tir(cls.NT_matmul5, (layer_norm124, gpt_neox_layers_13_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add366: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv149, gpt_neox_layers_13_attention_query_key_value_bias3) + reshape244: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add366, R.shape([1, seq_len, 48, 128])) + reshape245: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape244, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape246: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv62, R.shape([1, seq_len, 16, 128])) + reshape247: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape246, R.shape([1, seq_len, 2048])) + lv150 = R.call_tir(cls.NT_matmul6, (reshape247, gpt_neox_layers_13_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add367: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv150, gpt_neox_layers_13_attention_dense_bias3) + layer_norm125: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add365, 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) + lv151 = R.call_tir(cls.NT_matmul7, (layer_norm125, gpt_neox_layers_13_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add368: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv151, gpt_neox_layers_13_mlp_dense_h_to_4h_bias3) + gelu61: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add368) + astype124: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu61, dtype="float16") + lv152 = R.call_tir(cls.NT_matmul8, (astype124, gpt_neox_layers_13_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add369: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv152, gpt_neox_layers_13_mlp_dense_4h_to_h_bias3) + astype125: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add369, dtype="float16") + add370: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype125, add367) + add371: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add370, add365) + layer_norm126: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add371, gpt_neox_layers_14_input_layernorm_weight3, gpt_neox_layers_14_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv153 = R.call_tir(cls.NT_matmul5, (layer_norm126, gpt_neox_layers_14_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add372: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv153, gpt_neox_layers_14_attention_query_key_value_bias3) + reshape248: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add372, R.shape([1, seq_len, 48, 128])) + reshape249: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape248, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape250: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv63, R.shape([1, seq_len, 16, 128])) + reshape251: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape250, R.shape([1, seq_len, 2048])) + lv154 = R.call_tir(cls.NT_matmul6, (reshape251, gpt_neox_layers_14_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add373: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv154, gpt_neox_layers_14_attention_dense_bias3) + layer_norm127: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add371, 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) + lv155 = R.call_tir(cls.NT_matmul7, (layer_norm127, gpt_neox_layers_14_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add374: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv155, gpt_neox_layers_14_mlp_dense_h_to_4h_bias3) + gelu62: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add374) + astype126: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu62, dtype="float16") + lv156 = R.call_tir(cls.NT_matmul8, (astype126, gpt_neox_layers_14_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add375: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv156, gpt_neox_layers_14_mlp_dense_4h_to_h_bias3) + astype127: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add375, dtype="float16") + add376: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype127, add373) + add377: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add376, add371) + layer_norm128: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add377, gpt_neox_layers_15_input_layernorm_weight3, gpt_neox_layers_15_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv157 = R.call_tir(cls.NT_matmul5, (layer_norm128, gpt_neox_layers_15_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add378: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv157, gpt_neox_layers_15_attention_query_key_value_bias3) + reshape252: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add378, R.shape([1, seq_len, 48, 128])) + reshape253: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape252, R.shape([seq_len, 48, 128])) + 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((seq_len, 16, 128), dtype="float16")) + reshape254: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv64, R.shape([1, seq_len, 16, 128])) + reshape255: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape254, R.shape([1, seq_len, 2048])) + lv158 = R.call_tir(cls.NT_matmul6, (reshape255, gpt_neox_layers_15_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add379: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv158, gpt_neox_layers_15_attention_dense_bias3) + layer_norm129: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add377, 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) + lv159 = R.call_tir(cls.NT_matmul7, (layer_norm129, gpt_neox_layers_15_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add380: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv159, gpt_neox_layers_15_mlp_dense_h_to_4h_bias3) + gelu63: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add380) + astype128: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu63, dtype="float16") + lv160 = R.call_tir(cls.NT_matmul8, (astype128, gpt_neox_layers_15_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add381: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv160, gpt_neox_layers_15_mlp_dense_4h_to_h_bias3) + astype129: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add381, dtype="float16") + add382: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype129, add379) + add383: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add382, add377) + layer_norm130: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add383, gpt_neox_layers_16_input_layernorm_weight3, gpt_neox_layers_16_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv161 = R.call_tir(cls.NT_matmul5, (layer_norm130, gpt_neox_layers_16_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add384: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv161, gpt_neox_layers_16_attention_query_key_value_bias3) + reshape256: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add384, R.shape([1, seq_len, 48, 128])) + reshape257: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape256, R.shape([seq_len, 48, 128])) + lv65 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), reshape257), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape258: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv65, R.shape([1, seq_len, 16, 128])) + reshape259: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape258, R.shape([1, seq_len, 2048])) + lv162 = R.call_tir(cls.NT_matmul6, (reshape259, gpt_neox_layers_16_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add385: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv162, gpt_neox_layers_16_attention_dense_bias3) + layer_norm131: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add383, gpt_neox_layers_16_post_attention_layernorm_weight3, gpt_neox_layers_16_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv163 = R.call_tir(cls.NT_matmul7, (layer_norm131, gpt_neox_layers_16_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add386: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv163, gpt_neox_layers_16_mlp_dense_h_to_4h_bias3) + gelu64: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add386) + astype130: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu64, dtype="float16") + lv164 = R.call_tir(cls.NT_matmul8, (astype130, gpt_neox_layers_16_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add387: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv164, gpt_neox_layers_16_mlp_dense_4h_to_h_bias3) + astype131: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add387, dtype="float16") + add388: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype131, add385) + add389: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add388, add383) + layer_norm132: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add389, gpt_neox_layers_17_input_layernorm_weight3, gpt_neox_layers_17_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv165 = R.call_tir(cls.NT_matmul5, (layer_norm132, gpt_neox_layers_17_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add390: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv165, gpt_neox_layers_17_attention_query_key_value_bias3) + reshape260: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add390, R.shape([1, seq_len, 48, 128])) + reshape261: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape260, R.shape([seq_len, 48, 128])) + lv66 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), reshape261), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape262: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv66, R.shape([1, seq_len, 16, 128])) + reshape263: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape262, R.shape([1, seq_len, 2048])) + lv166 = R.call_tir(cls.NT_matmul6, (reshape263, gpt_neox_layers_17_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add391: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv166, gpt_neox_layers_17_attention_dense_bias3) + layer_norm133: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add389, gpt_neox_layers_17_post_attention_layernorm_weight3, gpt_neox_layers_17_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv167 = R.call_tir(cls.NT_matmul7, (layer_norm133, gpt_neox_layers_17_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add392: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv167, gpt_neox_layers_17_mlp_dense_h_to_4h_bias3) + gelu65: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add392) + astype132: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu65, dtype="float16") + lv168 = R.call_tir(cls.NT_matmul8, (astype132, gpt_neox_layers_17_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add393: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv168, gpt_neox_layers_17_mlp_dense_4h_to_h_bias3) + astype133: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add393, dtype="float16") + add394: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype133, add391) + add395: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add394, add389) + layer_norm134: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add395, gpt_neox_layers_18_input_layernorm_weight3, gpt_neox_layers_18_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv169 = R.call_tir(cls.NT_matmul5, (layer_norm134, gpt_neox_layers_18_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add396: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv169, gpt_neox_layers_18_attention_query_key_value_bias3) + reshape264: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add396, R.shape([1, seq_len, 48, 128])) + reshape265: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape264, R.shape([seq_len, 48, 128])) + lv67 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), reshape265), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape266: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv67, R.shape([1, seq_len, 16, 128])) + reshape267: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape266, R.shape([1, seq_len, 2048])) + lv170 = R.call_tir(cls.NT_matmul6, (reshape267, gpt_neox_layers_18_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add397: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv170, gpt_neox_layers_18_attention_dense_bias3) + layer_norm135: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add395, gpt_neox_layers_18_post_attention_layernorm_weight3, gpt_neox_layers_18_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv171 = R.call_tir(cls.NT_matmul7, (layer_norm135, gpt_neox_layers_18_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add398: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv171, gpt_neox_layers_18_mlp_dense_h_to_4h_bias3) + gelu66: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add398) + astype134: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu66, dtype="float16") + lv172 = R.call_tir(cls.NT_matmul8, (astype134, gpt_neox_layers_18_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add399: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv172, gpt_neox_layers_18_mlp_dense_4h_to_h_bias3) + astype135: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add399, dtype="float16") + add400: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype135, add397) + add401: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add400, add395) + layer_norm136: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add401, gpt_neox_layers_19_input_layernorm_weight3, gpt_neox_layers_19_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv173 = R.call_tir(cls.NT_matmul5, (layer_norm136, gpt_neox_layers_19_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add402: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv173, gpt_neox_layers_19_attention_query_key_value_bias3) + reshape268: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add402, R.shape([1, seq_len, 48, 128])) + reshape269: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape268, R.shape([seq_len, 48, 128])) + lv68 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), reshape269), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape270: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv68, R.shape([1, seq_len, 16, 128])) + reshape271: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape270, R.shape([1, seq_len, 2048])) + lv174 = R.call_tir(cls.NT_matmul6, (reshape271, gpt_neox_layers_19_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add403: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv174, gpt_neox_layers_19_attention_dense_bias3) + layer_norm137: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add401, gpt_neox_layers_19_post_attention_layernorm_weight3, gpt_neox_layers_19_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv175 = R.call_tir(cls.NT_matmul7, (layer_norm137, gpt_neox_layers_19_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add404: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv175, gpt_neox_layers_19_mlp_dense_h_to_4h_bias3) + gelu67: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add404) + astype136: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu67, dtype="float16") + lv176 = R.call_tir(cls.NT_matmul8, (astype136, gpt_neox_layers_19_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add405: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv176, gpt_neox_layers_19_mlp_dense_4h_to_h_bias3) + astype137: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add405, dtype="float16") + add406: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype137, add403) + add407: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add406, add401) + layer_norm138: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add407, gpt_neox_layers_20_input_layernorm_weight3, gpt_neox_layers_20_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv177 = R.call_tir(cls.NT_matmul5, (layer_norm138, gpt_neox_layers_20_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add408: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv177, gpt_neox_layers_20_attention_query_key_value_bias3) + reshape272: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add408, R.shape([1, seq_len, 48, 128])) + reshape273: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape272, R.shape([seq_len, 48, 128])) + lv69 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), reshape273), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape274: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv69, R.shape([1, seq_len, 16, 128])) + reshape275: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape274, R.shape([1, seq_len, 2048])) + lv178 = R.call_tir(cls.NT_matmul6, (reshape275, gpt_neox_layers_20_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add409: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv178, gpt_neox_layers_20_attention_dense_bias3) + layer_norm139: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add407, gpt_neox_layers_20_post_attention_layernorm_weight3, gpt_neox_layers_20_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv179 = R.call_tir(cls.NT_matmul7, (layer_norm139, gpt_neox_layers_20_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add410: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv179, gpt_neox_layers_20_mlp_dense_h_to_4h_bias3) + gelu68: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add410) + astype138: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu68, dtype="float16") + lv180 = R.call_tir(cls.NT_matmul8, (astype138, gpt_neox_layers_20_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add411: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv180, gpt_neox_layers_20_mlp_dense_4h_to_h_bias3) + astype139: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add411, dtype="float16") + add412: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype139, add409) + add413: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add412, add407) + layer_norm140: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add413, gpt_neox_layers_21_input_layernorm_weight3, gpt_neox_layers_21_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv181 = R.call_tir(cls.NT_matmul5, (layer_norm140, gpt_neox_layers_21_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add414: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv181, gpt_neox_layers_21_attention_query_key_value_bias3) + reshape276: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add414, R.shape([1, seq_len, 48, 128])) + reshape277: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape276, R.shape([seq_len, 48, 128])) + lv70 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), reshape277), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape278: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv70, R.shape([1, seq_len, 16, 128])) + reshape279: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape278, R.shape([1, seq_len, 2048])) + lv182 = R.call_tir(cls.NT_matmul6, (reshape279, gpt_neox_layers_21_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add415: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv182, gpt_neox_layers_21_attention_dense_bias3) + layer_norm141: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add413, gpt_neox_layers_21_post_attention_layernorm_weight3, gpt_neox_layers_21_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv183 = R.call_tir(cls.NT_matmul7, (layer_norm141, gpt_neox_layers_21_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add416: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv183, gpt_neox_layers_21_mlp_dense_h_to_4h_bias3) + gelu69: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add416) + astype140: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu69, dtype="float16") + lv184 = R.call_tir(cls.NT_matmul8, (astype140, gpt_neox_layers_21_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add417: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv184, gpt_neox_layers_21_mlp_dense_4h_to_h_bias3) + astype141: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add417, dtype="float16") + add418: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype141, add415) + add419: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add418, add413) + layer_norm142: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add419, gpt_neox_layers_22_input_layernorm_weight3, gpt_neox_layers_22_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv185 = R.call_tir(cls.NT_matmul5, (layer_norm142, gpt_neox_layers_22_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add420: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv185, gpt_neox_layers_22_attention_query_key_value_bias3) + reshape280: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add420, R.shape([1, seq_len, 48, 128])) + reshape281: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape280, R.shape([seq_len, 48, 128])) + lv71 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), reshape281), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape282: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv71, R.shape([1, seq_len, 16, 128])) + reshape283: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape282, R.shape([1, seq_len, 2048])) + lv186 = R.call_tir(cls.NT_matmul6, (reshape283, gpt_neox_layers_22_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add421: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv186, gpt_neox_layers_22_attention_dense_bias3) + layer_norm143: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add419, gpt_neox_layers_22_post_attention_layernorm_weight3, gpt_neox_layers_22_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv187 = R.call_tir(cls.NT_matmul7, (layer_norm143, gpt_neox_layers_22_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add422: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv187, gpt_neox_layers_22_mlp_dense_h_to_4h_bias3) + gelu70: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add422) + astype142: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu70, dtype="float16") + lv188 = R.call_tir(cls.NT_matmul8, (astype142, gpt_neox_layers_22_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add423: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv188, gpt_neox_layers_22_mlp_dense_4h_to_h_bias3) + astype143: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add423, dtype="float16") + add424: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype143, add421) + add425: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add424, add419) + layer_norm144: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add425, gpt_neox_layers_23_input_layernorm_weight3, gpt_neox_layers_23_input_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv189 = R.call_tir(cls.NT_matmul5, (layer_norm144, gpt_neox_layers_23_attention_query_key_value_weight3), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add426: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv189, gpt_neox_layers_23_attention_query_key_value_bias3) + reshape284: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add426, R.shape([1, seq_len, 48, 128])) + reshape285: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape284, R.shape([seq_len, 48, 128])) + lv72 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), reshape285), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape286: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv72, R.shape([1, seq_len, 16, 128])) + reshape287: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape286, R.shape([1, seq_len, 2048])) + lv190 = R.call_tir(cls.NT_matmul6, (reshape287, gpt_neox_layers_23_attention_dense_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add427: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv190, gpt_neox_layers_23_attention_dense_bias3) + layer_norm145: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add425, gpt_neox_layers_23_post_attention_layernorm_weight3, gpt_neox_layers_23_post_attention_layernorm_bias3, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv191 = R.call_tir(cls.NT_matmul7, (layer_norm145, gpt_neox_layers_23_mlp_dense_h_to_4h_weight3), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add428: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv191, gpt_neox_layers_23_mlp_dense_h_to_4h_bias3) + gelu71: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add428) + astype144: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu71, dtype="float16") + lv192 = R.call_tir(cls.NT_matmul8, (astype144, gpt_neox_layers_23_mlp_dense_4h_to_h_weight3), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add429: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv192, gpt_neox_layers_23_mlp_dense_4h_to_h_bias3) + astype145: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add429, dtype="float16") + add430: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype145, add427) + add431: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add430, add425) + layer_norm146: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add431, 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_norm146, logit_positions, axis=1) + lv193 = R.call_tir(cls.NT_matmul9, (take1, embed_out_weight3), out_sinfo=R.Tensor((1, batch_size, vocab_size), dtype="float16")) + astype146: R.Tensor((1, batch_size, vocab_size), dtype="float32") = R.astype(lv193, dtype="float32") + gv3: R.Tuple(R.Tensor((1, batch_size, vocab_size), dtype="float32"), R.Object) = astype146, 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((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_layers_16_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[193] + gpt_neox_layers_16_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[194] + gpt_neox_layers_16_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[195] + gpt_neox_layers_16_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[196] + gpt_neox_layers_16_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[197] + gpt_neox_layers_16_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[198] + gpt_neox_layers_16_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[199] + gpt_neox_layers_16_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[200] + gpt_neox_layers_16_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[201] + gpt_neox_layers_16_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[202] + gpt_neox_layers_16_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[203] + gpt_neox_layers_16_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[204] + gpt_neox_layers_17_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[205] + gpt_neox_layers_17_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[206] + gpt_neox_layers_17_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[207] + gpt_neox_layers_17_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[208] + gpt_neox_layers_17_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[209] + gpt_neox_layers_17_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[210] + gpt_neox_layers_17_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[211] + gpt_neox_layers_17_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[212] + gpt_neox_layers_17_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[213] + gpt_neox_layers_17_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[214] + gpt_neox_layers_17_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[215] + gpt_neox_layers_17_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[216] + gpt_neox_layers_18_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[217] + gpt_neox_layers_18_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[218] + gpt_neox_layers_18_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[219] + gpt_neox_layers_18_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[220] + gpt_neox_layers_18_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[221] + gpt_neox_layers_18_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[222] + gpt_neox_layers_18_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[223] + gpt_neox_layers_18_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[224] + gpt_neox_layers_18_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[225] + gpt_neox_layers_18_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[226] + gpt_neox_layers_18_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[227] + gpt_neox_layers_18_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[228] + gpt_neox_layers_19_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[229] + gpt_neox_layers_19_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[230] + gpt_neox_layers_19_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[231] + gpt_neox_layers_19_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[232] + gpt_neox_layers_19_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[233] + gpt_neox_layers_19_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[234] + gpt_neox_layers_19_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[235] + gpt_neox_layers_19_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[236] + gpt_neox_layers_19_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[237] + gpt_neox_layers_19_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[238] + gpt_neox_layers_19_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[239] + gpt_neox_layers_19_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[240] + gpt_neox_layers_20_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[241] + gpt_neox_layers_20_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[242] + gpt_neox_layers_20_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[243] + gpt_neox_layers_20_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[244] + gpt_neox_layers_20_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[245] + gpt_neox_layers_20_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[246] + gpt_neox_layers_20_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[247] + gpt_neox_layers_20_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[248] + gpt_neox_layers_20_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[249] + gpt_neox_layers_20_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[250] + gpt_neox_layers_20_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[251] + gpt_neox_layers_20_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[252] + gpt_neox_layers_21_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[253] + gpt_neox_layers_21_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[254] + gpt_neox_layers_21_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[255] + gpt_neox_layers_21_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[256] + gpt_neox_layers_21_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[257] + gpt_neox_layers_21_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[258] + gpt_neox_layers_21_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[259] + gpt_neox_layers_21_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[260] + gpt_neox_layers_21_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[261] + gpt_neox_layers_21_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[262] + gpt_neox_layers_21_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[263] + gpt_neox_layers_21_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[264] + gpt_neox_layers_22_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[265] + gpt_neox_layers_22_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[266] + gpt_neox_layers_22_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[267] + gpt_neox_layers_22_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[268] + gpt_neox_layers_22_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[269] + gpt_neox_layers_22_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[270] + gpt_neox_layers_22_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[271] + gpt_neox_layers_22_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[272] + gpt_neox_layers_22_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[273] + gpt_neox_layers_22_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[274] + gpt_neox_layers_22_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[275] + gpt_neox_layers_22_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[276] + gpt_neox_layers_23_input_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[277] + gpt_neox_layers_23_input_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[278] + gpt_neox_layers_23_post_attention_layernorm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[279] + gpt_neox_layers_23_post_attention_layernorm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[280] + gpt_neox_layers_23_attention_query_key_value_weight5: R.Tensor((6144, 2048), dtype="float16") = packed_params[281] + gpt_neox_layers_23_attention_query_key_value_bias5: R.Tensor((6144,), dtype="float16") = packed_params[282] + gpt_neox_layers_23_attention_dense_weight5: R.Tensor((2048, 2048), dtype="float16") = packed_params[283] + gpt_neox_layers_23_attention_dense_bias5: R.Tensor((2048,), dtype="float16") = packed_params[284] + gpt_neox_layers_23_mlp_dense_h_to_4h_weight5: R.Tensor((8192, 2048), dtype="float16") = packed_params[285] + gpt_neox_layers_23_mlp_dense_h_to_4h_bias5: R.Tensor((8192,), dtype="float32") = packed_params[286] + gpt_neox_layers_23_mlp_dense_4h_to_h_weight5: R.Tensor((2048, 8192), dtype="float16") = packed_params[287] + gpt_neox_layers_23_mlp_dense_4h_to_h_bias5: R.Tensor((2048,), dtype="float32") = packed_params[288] + gpt_neox_final_layer_norm_weight5: R.Tensor((2048,), dtype="float16") = packed_params[289] + gpt_neox_final_layer_norm_bias5: R.Tensor((2048,), dtype="float16") = packed_params[290] + embed_out_weight5: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[291] + layer_norm196: 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) + lv194 = R.call_tir(cls.NT_matmul5, (layer_norm196, gpt_neox_layers_0_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add576: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv194, gpt_neox_layers_0_attention_query_key_value_bias5) + reshape384: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add576, R.shape([1, seq_len, 48, 128])) + reshape385: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape384, R.shape([seq_len, 48, 128])) + lv97 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), reshape385), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape386: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv97, R.shape([1, seq_len, 16, 128])) + reshape387: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape386, R.shape([1, seq_len, 2048])) + lv195 = R.call_tir(cls.NT_matmul6, (reshape387, gpt_neox_layers_0_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add577: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv195, gpt_neox_layers_0_attention_dense_bias5) + layer_norm197: 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) + lv196 = R.call_tir(cls.NT_matmul7, (layer_norm197, gpt_neox_layers_0_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add578: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv196, gpt_neox_layers_0_mlp_dense_h_to_4h_bias5) + gelu96: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add578) + astype196: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu96, dtype="float16") + lv197 = R.call_tir(cls.NT_matmul8, (astype196, gpt_neox_layers_0_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add579: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv197, gpt_neox_layers_0_mlp_dense_4h_to_h_bias5) + astype197: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add579, dtype="float16") + add580: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype197, add577) + add581: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add580, input_embeds) + layer_norm198: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add581, gpt_neox_layers_1_input_layernorm_weight5, gpt_neox_layers_1_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv198 = R.call_tir(cls.NT_matmul5, (layer_norm198, gpt_neox_layers_1_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add582: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv198, gpt_neox_layers_1_attention_query_key_value_bias5) + reshape388: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add582, R.shape([1, seq_len, 48, 128])) + reshape389: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape388, R.shape([seq_len, 48, 128])) + lv98 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), reshape389), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape390: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv98, R.shape([1, seq_len, 16, 128])) + reshape391: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape390, R.shape([1, seq_len, 2048])) + lv199 = R.call_tir(cls.NT_matmul6, (reshape391, gpt_neox_layers_1_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add583: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv199, gpt_neox_layers_1_attention_dense_bias5) + layer_norm199: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add581, 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) + lv200 = R.call_tir(cls.NT_matmul7, (layer_norm199, gpt_neox_layers_1_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add584: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv200, gpt_neox_layers_1_mlp_dense_h_to_4h_bias5) + gelu97: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add584) + astype198: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu97, dtype="float16") + lv201 = R.call_tir(cls.NT_matmul8, (astype198, gpt_neox_layers_1_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add585: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv201, gpt_neox_layers_1_mlp_dense_4h_to_h_bias5) + astype199: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add585, dtype="float16") + add586: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype199, add583) + add587: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add586, add581) + layer_norm200: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add587, gpt_neox_layers_2_input_layernorm_weight5, gpt_neox_layers_2_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv202 = R.call_tir(cls.NT_matmul5, (layer_norm200, gpt_neox_layers_2_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add588: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv202, gpt_neox_layers_2_attention_query_key_value_bias5) + reshape392: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add588, R.shape([1, seq_len, 48, 128])) + reshape393: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape392, R.shape([seq_len, 48, 128])) + lv99 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), reshape393), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape394: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv99, R.shape([1, seq_len, 16, 128])) + reshape395: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape394, R.shape([1, seq_len, 2048])) + lv203 = R.call_tir(cls.NT_matmul6, (reshape395, gpt_neox_layers_2_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add589: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv203, gpt_neox_layers_2_attention_dense_bias5) + layer_norm201: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add587, 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) + lv204 = R.call_tir(cls.NT_matmul7, (layer_norm201, gpt_neox_layers_2_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add590: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv204, gpt_neox_layers_2_mlp_dense_h_to_4h_bias5) + gelu98: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add590) + astype200: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu98, dtype="float16") + lv205 = R.call_tir(cls.NT_matmul8, (astype200, gpt_neox_layers_2_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add591: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv205, gpt_neox_layers_2_mlp_dense_4h_to_h_bias5) + astype201: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add591, dtype="float16") + add592: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype201, add589) + add593: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add592, add587) + layer_norm202: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add593, gpt_neox_layers_3_input_layernorm_weight5, gpt_neox_layers_3_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv206 = R.call_tir(cls.NT_matmul5, (layer_norm202, gpt_neox_layers_3_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add594: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv206, gpt_neox_layers_3_attention_query_key_value_bias5) + reshape396: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add594, R.shape([1, seq_len, 48, 128])) + reshape397: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape396, R.shape([seq_len, 48, 128])) + lv100 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), reshape397), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape398: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv100, R.shape([1, seq_len, 16, 128])) + reshape399: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape398, R.shape([1, seq_len, 2048])) + lv207 = R.call_tir(cls.NT_matmul6, (reshape399, gpt_neox_layers_3_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add595: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv207, gpt_neox_layers_3_attention_dense_bias5) + layer_norm203: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add593, 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) + lv208 = R.call_tir(cls.NT_matmul7, (layer_norm203, gpt_neox_layers_3_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add596: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv208, gpt_neox_layers_3_mlp_dense_h_to_4h_bias5) + gelu99: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add596) + astype202: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu99, dtype="float16") + lv209 = R.call_tir(cls.NT_matmul8, (astype202, gpt_neox_layers_3_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add597: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv209, gpt_neox_layers_3_mlp_dense_4h_to_h_bias5) + astype203: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add597, dtype="float16") + add598: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype203, add595) + add599: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add598, add593) + layer_norm204: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add599, gpt_neox_layers_4_input_layernorm_weight5, gpt_neox_layers_4_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv210 = R.call_tir(cls.NT_matmul5, (layer_norm204, gpt_neox_layers_4_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add600: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv210, gpt_neox_layers_4_attention_query_key_value_bias5) + reshape400: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add600, R.shape([1, seq_len, 48, 128])) + reshape401: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape400, R.shape([seq_len, 48, 128])) + lv101 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), reshape401), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape402: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv101, R.shape([1, seq_len, 16, 128])) + reshape403: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape402, R.shape([1, seq_len, 2048])) + lv211 = R.call_tir(cls.NT_matmul6, (reshape403, gpt_neox_layers_4_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add601: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv211, gpt_neox_layers_4_attention_dense_bias5) + layer_norm205: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add599, 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) + lv212 = R.call_tir(cls.NT_matmul7, (layer_norm205, gpt_neox_layers_4_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add602: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv212, gpt_neox_layers_4_mlp_dense_h_to_4h_bias5) + gelu100: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add602) + astype204: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu100, dtype="float16") + lv213 = R.call_tir(cls.NT_matmul8, (astype204, gpt_neox_layers_4_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add603: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv213, gpt_neox_layers_4_mlp_dense_4h_to_h_bias5) + astype205: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add603, dtype="float16") + add604: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype205, add601) + add605: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add604, add599) + layer_norm206: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add605, gpt_neox_layers_5_input_layernorm_weight5, gpt_neox_layers_5_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv214 = R.call_tir(cls.NT_matmul5, (layer_norm206, gpt_neox_layers_5_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add606: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv214, gpt_neox_layers_5_attention_query_key_value_bias5) + reshape404: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add606, R.shape([1, seq_len, 48, 128])) + reshape405: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape404, R.shape([seq_len, 48, 128])) + lv102 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), reshape405), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape406: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv102, R.shape([1, seq_len, 16, 128])) + reshape407: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape406, R.shape([1, seq_len, 2048])) + lv215 = R.call_tir(cls.NT_matmul6, (reshape407, gpt_neox_layers_5_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add607: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv215, gpt_neox_layers_5_attention_dense_bias5) + layer_norm207: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add605, 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) + lv216 = R.call_tir(cls.NT_matmul7, (layer_norm207, gpt_neox_layers_5_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add608: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv216, gpt_neox_layers_5_mlp_dense_h_to_4h_bias5) + gelu101: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add608) + astype206: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu101, dtype="float16") + lv217 = R.call_tir(cls.NT_matmul8, (astype206, gpt_neox_layers_5_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add609: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv217, gpt_neox_layers_5_mlp_dense_4h_to_h_bias5) + astype207: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add609, dtype="float16") + add610: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype207, add607) + add611: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add610, add605) + layer_norm208: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add611, gpt_neox_layers_6_input_layernorm_weight5, gpt_neox_layers_6_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv218 = R.call_tir(cls.NT_matmul5, (layer_norm208, gpt_neox_layers_6_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add612: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv218, gpt_neox_layers_6_attention_query_key_value_bias5) + reshape408: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add612, R.shape([1, seq_len, 48, 128])) + reshape409: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape408, R.shape([seq_len, 48, 128])) + lv103 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(6), R.prim_value(T.float32(1.0)), reshape409), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape410: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv103, R.shape([1, seq_len, 16, 128])) + reshape411: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape410, R.shape([1, seq_len, 2048])) + lv219 = R.call_tir(cls.NT_matmul6, (reshape411, gpt_neox_layers_6_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add613: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv219, gpt_neox_layers_6_attention_dense_bias5) + layer_norm209: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add611, 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) + lv220 = R.call_tir(cls.NT_matmul7, (layer_norm209, gpt_neox_layers_6_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add614: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv220, gpt_neox_layers_6_mlp_dense_h_to_4h_bias5) + gelu102: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add614) + astype208: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu102, dtype="float16") + lv221 = R.call_tir(cls.NT_matmul8, (astype208, gpt_neox_layers_6_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add615: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv221, gpt_neox_layers_6_mlp_dense_4h_to_h_bias5) + astype209: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add615, dtype="float16") + add616: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype209, add613) + add617: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add616, add611) + layer_norm210: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add617, gpt_neox_layers_7_input_layernorm_weight5, gpt_neox_layers_7_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv222 = R.call_tir(cls.NT_matmul5, (layer_norm210, gpt_neox_layers_7_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add618: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv222, gpt_neox_layers_7_attention_query_key_value_bias5) + reshape412: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add618, R.shape([1, seq_len, 48, 128])) + reshape413: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape412, R.shape([seq_len, 48, 128])) + lv104 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(7), R.prim_value(T.float32(1.0)), reshape413), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape414: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv104, R.shape([1, seq_len, 16, 128])) + reshape415: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape414, R.shape([1, seq_len, 2048])) + lv223 = R.call_tir(cls.NT_matmul6, (reshape415, gpt_neox_layers_7_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add619: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv223, gpt_neox_layers_7_attention_dense_bias5) + layer_norm211: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add617, 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) + lv224 = R.call_tir(cls.NT_matmul7, (layer_norm211, gpt_neox_layers_7_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add620: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv224, gpt_neox_layers_7_mlp_dense_h_to_4h_bias5) + gelu103: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add620) + astype210: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu103, dtype="float16") + lv225 = R.call_tir(cls.NT_matmul8, (astype210, gpt_neox_layers_7_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add621: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv225, gpt_neox_layers_7_mlp_dense_4h_to_h_bias5) + astype211: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add621, dtype="float16") + add622: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype211, add619) + add623: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add622, add617) + layer_norm212: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add623, gpt_neox_layers_8_input_layernorm_weight5, gpt_neox_layers_8_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv226 = R.call_tir(cls.NT_matmul5, (layer_norm212, gpt_neox_layers_8_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add624: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv226, gpt_neox_layers_8_attention_query_key_value_bias5) + reshape416: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add624, R.shape([1, seq_len, 48, 128])) + reshape417: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape416, R.shape([seq_len, 48, 128])) + lv105 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(8), R.prim_value(T.float32(1.0)), reshape417), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape418: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv105, R.shape([1, seq_len, 16, 128])) + reshape419: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape418, R.shape([1, seq_len, 2048])) + lv227 = R.call_tir(cls.NT_matmul6, (reshape419, gpt_neox_layers_8_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add625: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv227, gpt_neox_layers_8_attention_dense_bias5) + layer_norm213: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add623, 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) + lv228 = R.call_tir(cls.NT_matmul7, (layer_norm213, gpt_neox_layers_8_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add626: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv228, gpt_neox_layers_8_mlp_dense_h_to_4h_bias5) + gelu104: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add626) + astype212: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu104, dtype="float16") + lv229 = R.call_tir(cls.NT_matmul8, (astype212, gpt_neox_layers_8_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add627: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv229, gpt_neox_layers_8_mlp_dense_4h_to_h_bias5) + astype213: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add627, dtype="float16") + add628: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype213, add625) + add629: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add628, add623) + layer_norm214: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add629, gpt_neox_layers_9_input_layernorm_weight5, gpt_neox_layers_9_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv230 = R.call_tir(cls.NT_matmul5, (layer_norm214, gpt_neox_layers_9_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add630: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv230, gpt_neox_layers_9_attention_query_key_value_bias5) + reshape420: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add630, R.shape([1, seq_len, 48, 128])) + reshape421: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape420, R.shape([seq_len, 48, 128])) + lv106 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(9), R.prim_value(T.float32(1.0)), reshape421), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape422: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv106, R.shape([1, seq_len, 16, 128])) + reshape423: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape422, R.shape([1, seq_len, 2048])) + lv231 = R.call_tir(cls.NT_matmul6, (reshape423, gpt_neox_layers_9_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add631: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv231, gpt_neox_layers_9_attention_dense_bias5) + layer_norm215: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add629, 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) + lv232 = R.call_tir(cls.NT_matmul7, (layer_norm215, gpt_neox_layers_9_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add632: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv232, gpt_neox_layers_9_mlp_dense_h_to_4h_bias5) + gelu105: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add632) + astype214: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu105, dtype="float16") + lv233 = R.call_tir(cls.NT_matmul8, (astype214, gpt_neox_layers_9_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add633: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv233, gpt_neox_layers_9_mlp_dense_4h_to_h_bias5) + astype215: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add633, dtype="float16") + add634: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype215, add631) + add635: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add634, add629) + layer_norm216: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add635, gpt_neox_layers_10_input_layernorm_weight5, gpt_neox_layers_10_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv234 = R.call_tir(cls.NT_matmul5, (layer_norm216, gpt_neox_layers_10_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add636: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv234, gpt_neox_layers_10_attention_query_key_value_bias5) + reshape424: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add636, R.shape([1, seq_len, 48, 128])) + reshape425: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape424, R.shape([seq_len, 48, 128])) + lv107 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(10), R.prim_value(T.float32(1.0)), reshape425), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape426: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv107, R.shape([1, seq_len, 16, 128])) + reshape427: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape426, R.shape([1, seq_len, 2048])) + lv235 = R.call_tir(cls.NT_matmul6, (reshape427, gpt_neox_layers_10_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add637: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv235, gpt_neox_layers_10_attention_dense_bias5) + layer_norm217: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add635, 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) + lv236 = R.call_tir(cls.NT_matmul7, (layer_norm217, gpt_neox_layers_10_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add638: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv236, gpt_neox_layers_10_mlp_dense_h_to_4h_bias5) + gelu106: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add638) + astype216: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu106, dtype="float16") + lv237 = R.call_tir(cls.NT_matmul8, (astype216, gpt_neox_layers_10_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add639: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv237, gpt_neox_layers_10_mlp_dense_4h_to_h_bias5) + astype217: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add639, dtype="float16") + add640: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype217, add637) + add641: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add640, add635) + layer_norm218: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add641, gpt_neox_layers_11_input_layernorm_weight5, gpt_neox_layers_11_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv238 = R.call_tir(cls.NT_matmul5, (layer_norm218, gpt_neox_layers_11_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add642: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv238, gpt_neox_layers_11_attention_query_key_value_bias5) + reshape428: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add642, R.shape([1, seq_len, 48, 128])) + reshape429: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape428, R.shape([seq_len, 48, 128])) + lv108 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(11), R.prim_value(T.float32(1.0)), reshape429), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape430: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv108, R.shape([1, seq_len, 16, 128])) + reshape431: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape430, R.shape([1, seq_len, 2048])) + lv239 = R.call_tir(cls.NT_matmul6, (reshape431, gpt_neox_layers_11_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add643: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv239, gpt_neox_layers_11_attention_dense_bias5) + layer_norm219: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add641, 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) + lv240 = R.call_tir(cls.NT_matmul7, (layer_norm219, gpt_neox_layers_11_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add644: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv240, gpt_neox_layers_11_mlp_dense_h_to_4h_bias5) + gelu107: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add644) + astype218: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu107, dtype="float16") + lv241 = R.call_tir(cls.NT_matmul8, (astype218, gpt_neox_layers_11_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add645: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv241, gpt_neox_layers_11_mlp_dense_4h_to_h_bias5) + astype219: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add645, dtype="float16") + add646: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype219, add643) + add647: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add646, add641) + layer_norm220: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add647, gpt_neox_layers_12_input_layernorm_weight5, gpt_neox_layers_12_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv242 = R.call_tir(cls.NT_matmul5, (layer_norm220, gpt_neox_layers_12_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add648: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv242, gpt_neox_layers_12_attention_query_key_value_bias5) + reshape432: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add648, R.shape([1, seq_len, 48, 128])) + reshape433: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape432, R.shape([seq_len, 48, 128])) + lv109 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(12), R.prim_value(T.float32(1.0)), reshape433), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape434: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv109, R.shape([1, seq_len, 16, 128])) + reshape435: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape434, R.shape([1, seq_len, 2048])) + lv243 = R.call_tir(cls.NT_matmul6, (reshape435, gpt_neox_layers_12_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add649: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv243, gpt_neox_layers_12_attention_dense_bias5) + layer_norm221: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add647, 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) + lv244 = R.call_tir(cls.NT_matmul7, (layer_norm221, gpt_neox_layers_12_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add650: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv244, gpt_neox_layers_12_mlp_dense_h_to_4h_bias5) + gelu108: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add650) + astype220: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu108, dtype="float16") + lv245 = R.call_tir(cls.NT_matmul8, (astype220, gpt_neox_layers_12_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add651: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv245, gpt_neox_layers_12_mlp_dense_4h_to_h_bias5) + astype221: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add651, dtype="float16") + add652: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype221, add649) + add653: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add652, add647) + layer_norm222: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add653, gpt_neox_layers_13_input_layernorm_weight5, gpt_neox_layers_13_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv246 = R.call_tir(cls.NT_matmul5, (layer_norm222, gpt_neox_layers_13_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add654: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv246, gpt_neox_layers_13_attention_query_key_value_bias5) + reshape436: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add654, R.shape([1, seq_len, 48, 128])) + reshape437: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape436, R.shape([seq_len, 48, 128])) + lv110 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(13), R.prim_value(T.float32(1.0)), reshape437), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape438: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv110, R.shape([1, seq_len, 16, 128])) + reshape439: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape438, R.shape([1, seq_len, 2048])) + lv247 = R.call_tir(cls.NT_matmul6, (reshape439, gpt_neox_layers_13_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add655: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv247, gpt_neox_layers_13_attention_dense_bias5) + layer_norm223: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add653, 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) + lv248 = R.call_tir(cls.NT_matmul7, (layer_norm223, gpt_neox_layers_13_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add656: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv248, gpt_neox_layers_13_mlp_dense_h_to_4h_bias5) + gelu109: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add656) + astype222: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu109, dtype="float16") + lv249 = R.call_tir(cls.NT_matmul8, (astype222, gpt_neox_layers_13_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add657: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv249, gpt_neox_layers_13_mlp_dense_4h_to_h_bias5) + astype223: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add657, dtype="float16") + add658: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype223, add655) + add659: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add658, add653) + layer_norm224: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add659, gpt_neox_layers_14_input_layernorm_weight5, gpt_neox_layers_14_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv250 = R.call_tir(cls.NT_matmul5, (layer_norm224, gpt_neox_layers_14_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add660: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv250, gpt_neox_layers_14_attention_query_key_value_bias5) + reshape440: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add660, R.shape([1, seq_len, 48, 128])) + reshape441: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape440, R.shape([seq_len, 48, 128])) + lv111 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(14), R.prim_value(T.float32(1.0)), reshape441), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape442: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv111, R.shape([1, seq_len, 16, 128])) + reshape443: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape442, R.shape([1, seq_len, 2048])) + lv251 = R.call_tir(cls.NT_matmul6, (reshape443, gpt_neox_layers_14_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add661: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv251, gpt_neox_layers_14_attention_dense_bias5) + layer_norm225: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add659, 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) + lv252 = R.call_tir(cls.NT_matmul7, (layer_norm225, gpt_neox_layers_14_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add662: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv252, gpt_neox_layers_14_mlp_dense_h_to_4h_bias5) + gelu110: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add662) + astype224: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu110, dtype="float16") + lv253 = R.call_tir(cls.NT_matmul8, (astype224, gpt_neox_layers_14_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add663: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv253, gpt_neox_layers_14_mlp_dense_4h_to_h_bias5) + astype225: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add663, dtype="float16") + add664: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype225, add661) + add665: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add664, add659) + layer_norm226: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add665, gpt_neox_layers_15_input_layernorm_weight5, gpt_neox_layers_15_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv254 = R.call_tir(cls.NT_matmul5, (layer_norm226, gpt_neox_layers_15_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add666: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv254, gpt_neox_layers_15_attention_query_key_value_bias5) + reshape444: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add666, R.shape([1, seq_len, 48, 128])) + reshape445: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape444, R.shape([seq_len, 48, 128])) + lv112 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(15), R.prim_value(T.float32(1.0)), reshape445), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape446: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv112, R.shape([1, seq_len, 16, 128])) + reshape447: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape446, R.shape([1, seq_len, 2048])) + lv255 = R.call_tir(cls.NT_matmul6, (reshape447, gpt_neox_layers_15_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add667: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv255, gpt_neox_layers_15_attention_dense_bias5) + layer_norm227: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add665, 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) + lv256 = R.call_tir(cls.NT_matmul7, (layer_norm227, gpt_neox_layers_15_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add668: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv256, gpt_neox_layers_15_mlp_dense_h_to_4h_bias5) + gelu111: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add668) + astype226: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu111, dtype="float16") + lv257 = R.call_tir(cls.NT_matmul8, (astype226, gpt_neox_layers_15_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add669: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv257, gpt_neox_layers_15_mlp_dense_4h_to_h_bias5) + astype227: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add669, dtype="float16") + add670: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype227, add667) + add671: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add670, add665) + layer_norm228: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add671, gpt_neox_layers_16_input_layernorm_weight5, gpt_neox_layers_16_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv258 = R.call_tir(cls.NT_matmul5, (layer_norm228, gpt_neox_layers_16_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add672: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv258, gpt_neox_layers_16_attention_query_key_value_bias5) + reshape448: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add672, R.shape([1, seq_len, 48, 128])) + reshape449: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape448, R.shape([seq_len, 48, 128])) + lv113 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), reshape449), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape450: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv113, R.shape([1, seq_len, 16, 128])) + reshape451: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape450, R.shape([1, seq_len, 2048])) + lv259 = R.call_tir(cls.NT_matmul6, (reshape451, gpt_neox_layers_16_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add673: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv259, gpt_neox_layers_16_attention_dense_bias5) + layer_norm229: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add671, gpt_neox_layers_16_post_attention_layernorm_weight5, gpt_neox_layers_16_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv260 = R.call_tir(cls.NT_matmul7, (layer_norm229, gpt_neox_layers_16_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add674: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv260, gpt_neox_layers_16_mlp_dense_h_to_4h_bias5) + gelu112: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add674) + astype228: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu112, dtype="float16") + lv261 = R.call_tir(cls.NT_matmul8, (astype228, gpt_neox_layers_16_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add675: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv261, gpt_neox_layers_16_mlp_dense_4h_to_h_bias5) + astype229: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add675, dtype="float16") + add676: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype229, add673) + add677: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add676, add671) + layer_norm230: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add677, gpt_neox_layers_17_input_layernorm_weight5, gpt_neox_layers_17_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv262 = R.call_tir(cls.NT_matmul5, (layer_norm230, gpt_neox_layers_17_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add678: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv262, gpt_neox_layers_17_attention_query_key_value_bias5) + reshape452: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add678, R.shape([1, seq_len, 48, 128])) + reshape453: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape452, R.shape([seq_len, 48, 128])) + lv114 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), reshape453), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape454: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv114, R.shape([1, seq_len, 16, 128])) + reshape455: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape454, R.shape([1, seq_len, 2048])) + lv263 = R.call_tir(cls.NT_matmul6, (reshape455, gpt_neox_layers_17_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add679: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv263, gpt_neox_layers_17_attention_dense_bias5) + layer_norm231: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add677, gpt_neox_layers_17_post_attention_layernorm_weight5, gpt_neox_layers_17_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv264 = R.call_tir(cls.NT_matmul7, (layer_norm231, gpt_neox_layers_17_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add680: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv264, gpt_neox_layers_17_mlp_dense_h_to_4h_bias5) + gelu113: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add680) + astype230: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu113, dtype="float16") + lv265 = R.call_tir(cls.NT_matmul8, (astype230, gpt_neox_layers_17_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add681: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv265, gpt_neox_layers_17_mlp_dense_4h_to_h_bias5) + astype231: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add681, dtype="float16") + add682: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype231, add679) + add683: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add682, add677) + layer_norm232: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add683, gpt_neox_layers_18_input_layernorm_weight5, gpt_neox_layers_18_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv266 = R.call_tir(cls.NT_matmul5, (layer_norm232, gpt_neox_layers_18_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add684: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv266, gpt_neox_layers_18_attention_query_key_value_bias5) + reshape456: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add684, R.shape([1, seq_len, 48, 128])) + reshape457: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape456, R.shape([seq_len, 48, 128])) + lv115 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), reshape457), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape458: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv115, R.shape([1, seq_len, 16, 128])) + reshape459: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape458, R.shape([1, seq_len, 2048])) + lv267 = R.call_tir(cls.NT_matmul6, (reshape459, gpt_neox_layers_18_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add685: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv267, gpt_neox_layers_18_attention_dense_bias5) + layer_norm233: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add683, gpt_neox_layers_18_post_attention_layernorm_weight5, gpt_neox_layers_18_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv268 = R.call_tir(cls.NT_matmul7, (layer_norm233, gpt_neox_layers_18_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add686: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv268, gpt_neox_layers_18_mlp_dense_h_to_4h_bias5) + gelu114: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add686) + astype232: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu114, dtype="float16") + lv269 = R.call_tir(cls.NT_matmul8, (astype232, gpt_neox_layers_18_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add687: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv269, gpt_neox_layers_18_mlp_dense_4h_to_h_bias5) + astype233: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add687, dtype="float16") + add688: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype233, add685) + add689: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add688, add683) + layer_norm234: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add689, gpt_neox_layers_19_input_layernorm_weight5, gpt_neox_layers_19_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv270 = R.call_tir(cls.NT_matmul5, (layer_norm234, gpt_neox_layers_19_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add690: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv270, gpt_neox_layers_19_attention_query_key_value_bias5) + reshape460: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add690, R.shape([1, seq_len, 48, 128])) + reshape461: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape460, R.shape([seq_len, 48, 128])) + lv116 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), reshape461), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape462: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv116, R.shape([1, seq_len, 16, 128])) + reshape463: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape462, R.shape([1, seq_len, 2048])) + lv271 = R.call_tir(cls.NT_matmul6, (reshape463, gpt_neox_layers_19_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add691: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv271, gpt_neox_layers_19_attention_dense_bias5) + layer_norm235: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add689, gpt_neox_layers_19_post_attention_layernorm_weight5, gpt_neox_layers_19_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv272 = R.call_tir(cls.NT_matmul7, (layer_norm235, gpt_neox_layers_19_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add692: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv272, gpt_neox_layers_19_mlp_dense_h_to_4h_bias5) + gelu115: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add692) + astype234: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu115, dtype="float16") + lv273 = R.call_tir(cls.NT_matmul8, (astype234, gpt_neox_layers_19_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add693: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv273, gpt_neox_layers_19_mlp_dense_4h_to_h_bias5) + astype235: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add693, dtype="float16") + add694: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype235, add691) + add695: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add694, add689) + layer_norm236: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add695, gpt_neox_layers_20_input_layernorm_weight5, gpt_neox_layers_20_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv274 = R.call_tir(cls.NT_matmul5, (layer_norm236, gpt_neox_layers_20_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add696: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv274, gpt_neox_layers_20_attention_query_key_value_bias5) + reshape464: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add696, R.shape([1, seq_len, 48, 128])) + reshape465: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape464, R.shape([seq_len, 48, 128])) + lv117 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), reshape465), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape466: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv117, R.shape([1, seq_len, 16, 128])) + reshape467: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape466, R.shape([1, seq_len, 2048])) + lv275 = R.call_tir(cls.NT_matmul6, (reshape467, gpt_neox_layers_20_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add697: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv275, gpt_neox_layers_20_attention_dense_bias5) + layer_norm237: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add695, gpt_neox_layers_20_post_attention_layernorm_weight5, gpt_neox_layers_20_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv276 = R.call_tir(cls.NT_matmul7, (layer_norm237, gpt_neox_layers_20_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add698: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv276, gpt_neox_layers_20_mlp_dense_h_to_4h_bias5) + gelu116: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add698) + astype236: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu116, dtype="float16") + lv277 = R.call_tir(cls.NT_matmul8, (astype236, gpt_neox_layers_20_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add699: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv277, gpt_neox_layers_20_mlp_dense_4h_to_h_bias5) + astype237: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add699, dtype="float16") + add700: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype237, add697) + add701: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add700, add695) + layer_norm238: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add701, gpt_neox_layers_21_input_layernorm_weight5, gpt_neox_layers_21_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv278 = R.call_tir(cls.NT_matmul5, (layer_norm238, gpt_neox_layers_21_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add702: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv278, gpt_neox_layers_21_attention_query_key_value_bias5) + reshape468: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add702, R.shape([1, seq_len, 48, 128])) + reshape469: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape468, R.shape([seq_len, 48, 128])) + lv118 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), reshape469), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape470: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv118, R.shape([1, seq_len, 16, 128])) + reshape471: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape470, R.shape([1, seq_len, 2048])) + lv279 = R.call_tir(cls.NT_matmul6, (reshape471, gpt_neox_layers_21_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add703: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv279, gpt_neox_layers_21_attention_dense_bias5) + layer_norm239: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add701, gpt_neox_layers_21_post_attention_layernorm_weight5, gpt_neox_layers_21_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv280 = R.call_tir(cls.NT_matmul7, (layer_norm239, gpt_neox_layers_21_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add704: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv280, gpt_neox_layers_21_mlp_dense_h_to_4h_bias5) + gelu117: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add704) + astype238: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu117, dtype="float16") + lv281 = R.call_tir(cls.NT_matmul8, (astype238, gpt_neox_layers_21_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add705: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv281, gpt_neox_layers_21_mlp_dense_4h_to_h_bias5) + astype239: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add705, dtype="float16") + add706: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype239, add703) + add707: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add706, add701) + layer_norm240: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add707, gpt_neox_layers_22_input_layernorm_weight5, gpt_neox_layers_22_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv282 = R.call_tir(cls.NT_matmul5, (layer_norm240, gpt_neox_layers_22_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add708: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv282, gpt_neox_layers_22_attention_query_key_value_bias5) + reshape472: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add708, R.shape([1, seq_len, 48, 128])) + reshape473: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape472, R.shape([seq_len, 48, 128])) + lv119 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), reshape473), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape474: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv119, R.shape([1, seq_len, 16, 128])) + reshape475: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape474, R.shape([1, seq_len, 2048])) + lv283 = R.call_tir(cls.NT_matmul6, (reshape475, gpt_neox_layers_22_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add709: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv283, gpt_neox_layers_22_attention_dense_bias5) + layer_norm241: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add707, gpt_neox_layers_22_post_attention_layernorm_weight5, gpt_neox_layers_22_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv284 = R.call_tir(cls.NT_matmul7, (layer_norm241, gpt_neox_layers_22_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add710: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv284, gpt_neox_layers_22_mlp_dense_h_to_4h_bias5) + gelu118: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add710) + astype240: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu118, dtype="float16") + lv285 = R.call_tir(cls.NT_matmul8, (astype240, gpt_neox_layers_22_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add711: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv285, gpt_neox_layers_22_mlp_dense_4h_to_h_bias5) + astype241: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add711, dtype="float16") + add712: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype241, add709) + add713: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add712, add707) + layer_norm242: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add713, gpt_neox_layers_23_input_layernorm_weight5, gpt_neox_layers_23_input_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv286 = R.call_tir(cls.NT_matmul5, (layer_norm242, gpt_neox_layers_23_attention_query_key_value_weight5), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add714: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv286, gpt_neox_layers_23_attention_query_key_value_bias5) + reshape476: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add714, R.shape([1, seq_len, 48, 128])) + reshape477: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape476, R.shape([seq_len, 48, 128])) + lv120 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), reshape477), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape478: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv120, R.shape([1, seq_len, 16, 128])) + reshape479: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape478, R.shape([1, seq_len, 2048])) + lv287 = R.call_tir(cls.NT_matmul6, (reshape479, gpt_neox_layers_23_attention_dense_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add715: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv287, gpt_neox_layers_23_attention_dense_bias5) + layer_norm243: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add713, gpt_neox_layers_23_post_attention_layernorm_weight5, gpt_neox_layers_23_post_attention_layernorm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv288 = R.call_tir(cls.NT_matmul7, (layer_norm243, gpt_neox_layers_23_mlp_dense_h_to_4h_weight5), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add716: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv288, gpt_neox_layers_23_mlp_dense_h_to_4h_bias5) + gelu119: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add716) + astype242: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu119, dtype="float16") + lv289 = R.call_tir(cls.NT_matmul8, (astype242, gpt_neox_layers_23_mlp_dense_4h_to_h_weight5), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add717: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv289, gpt_neox_layers_23_mlp_dense_4h_to_h_bias5) + astype243: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add717, dtype="float16") + add718: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype243, add715) + add719: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add718, add713) + layer_norm244: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add719, gpt_neox_final_layer_norm_weight5, gpt_neox_final_layer_norm_bias5, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv290 = R.call_tir(cls.NT_matmul9, (layer_norm244, embed_out_weight5), out_sinfo=R.Tensor((1, seq_len, vocab_size), dtype="float16")) + astype244: R.Tensor((1, seq_len, vocab_size), dtype="float32") = R.astype(lv290, dtype="float32") + gv5: R.Tuple(R.Tensor((1, seq_len, vocab_size), dtype="float32"), R.Object) = astype244, 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, 24]), R.prim_value(16), R.prim_value(16), R.prim_value(128), 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((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_layers_16_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[193] + gpt_neox_layers_16_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[194] + gpt_neox_layers_16_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[195] + gpt_neox_layers_16_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[196] + gpt_neox_layers_16_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[197] + gpt_neox_layers_16_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[198] + gpt_neox_layers_16_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[199] + gpt_neox_layers_16_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[200] + gpt_neox_layers_16_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[201] + gpt_neox_layers_16_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[202] + gpt_neox_layers_16_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[203] + gpt_neox_layers_16_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[204] + gpt_neox_layers_17_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[205] + gpt_neox_layers_17_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[206] + gpt_neox_layers_17_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[207] + gpt_neox_layers_17_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[208] + gpt_neox_layers_17_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[209] + gpt_neox_layers_17_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[210] + gpt_neox_layers_17_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[211] + gpt_neox_layers_17_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[212] + gpt_neox_layers_17_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[213] + gpt_neox_layers_17_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[214] + gpt_neox_layers_17_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[215] + gpt_neox_layers_17_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[216] + gpt_neox_layers_18_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[217] + gpt_neox_layers_18_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[218] + gpt_neox_layers_18_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[219] + gpt_neox_layers_18_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[220] + gpt_neox_layers_18_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[221] + gpt_neox_layers_18_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[222] + gpt_neox_layers_18_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[223] + gpt_neox_layers_18_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[224] + gpt_neox_layers_18_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[225] + gpt_neox_layers_18_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[226] + gpt_neox_layers_18_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[227] + gpt_neox_layers_18_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[228] + gpt_neox_layers_19_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[229] + gpt_neox_layers_19_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[230] + gpt_neox_layers_19_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[231] + gpt_neox_layers_19_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[232] + gpt_neox_layers_19_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[233] + gpt_neox_layers_19_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[234] + gpt_neox_layers_19_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[235] + gpt_neox_layers_19_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[236] + gpt_neox_layers_19_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[237] + gpt_neox_layers_19_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[238] + gpt_neox_layers_19_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[239] + gpt_neox_layers_19_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[240] + gpt_neox_layers_20_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[241] + gpt_neox_layers_20_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[242] + gpt_neox_layers_20_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[243] + gpt_neox_layers_20_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[244] + gpt_neox_layers_20_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[245] + gpt_neox_layers_20_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[246] + gpt_neox_layers_20_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[247] + gpt_neox_layers_20_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[248] + gpt_neox_layers_20_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[249] + gpt_neox_layers_20_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[250] + gpt_neox_layers_20_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[251] + gpt_neox_layers_20_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[252] + gpt_neox_layers_21_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[253] + gpt_neox_layers_21_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[254] + gpt_neox_layers_21_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[255] + gpt_neox_layers_21_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[256] + gpt_neox_layers_21_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[257] + gpt_neox_layers_21_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[258] + gpt_neox_layers_21_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[259] + gpt_neox_layers_21_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[260] + gpt_neox_layers_21_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[261] + gpt_neox_layers_21_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[262] + gpt_neox_layers_21_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[263] + gpt_neox_layers_21_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[264] + gpt_neox_layers_22_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[265] + gpt_neox_layers_22_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[266] + gpt_neox_layers_22_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[267] + gpt_neox_layers_22_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[268] + gpt_neox_layers_22_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[269] + gpt_neox_layers_22_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[270] + gpt_neox_layers_22_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[271] + gpt_neox_layers_22_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[272] + gpt_neox_layers_22_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[273] + gpt_neox_layers_22_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[274] + gpt_neox_layers_22_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[275] + gpt_neox_layers_22_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[276] + gpt_neox_layers_23_input_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[277] + gpt_neox_layers_23_input_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[278] + gpt_neox_layers_23_post_attention_layernorm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[279] + gpt_neox_layers_23_post_attention_layernorm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[280] + gpt_neox_layers_23_attention_query_key_value_weight2: R.Tensor((6144, 2048), dtype="float16") = packed_params[281] + gpt_neox_layers_23_attention_query_key_value_bias2: R.Tensor((6144,), dtype="float16") = packed_params[282] + gpt_neox_layers_23_attention_dense_weight2: R.Tensor((2048, 2048), dtype="float16") = packed_params[283] + gpt_neox_layers_23_attention_dense_bias2: R.Tensor((2048,), dtype="float16") = packed_params[284] + gpt_neox_layers_23_mlp_dense_h_to_4h_weight2: R.Tensor((8192, 2048), dtype="float16") = packed_params[285] + gpt_neox_layers_23_mlp_dense_h_to_4h_bias2: R.Tensor((8192,), dtype="float32") = packed_params[286] + gpt_neox_layers_23_mlp_dense_4h_to_h_weight2: R.Tensor((2048, 8192), dtype="float16") = packed_params[287] + gpt_neox_layers_23_mlp_dense_4h_to_h_bias2: R.Tensor((2048,), dtype="float32") = packed_params[288] + gpt_neox_final_layer_norm_weight2: R.Tensor((2048,), dtype="float16") = packed_params[289] + gpt_neox_final_layer_norm_bias2: R.Tensor((2048,), dtype="float16") = packed_params[290] + embed_out_weight2: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[291] + layer_norm49: 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) + lv291 = R.call_tir(cls.NT_matmul10, (layer_norm49, gpt_neox_layers_0_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add144: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv291, gpt_neox_layers_0_attention_query_key_value_bias2) + reshape96: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add144, R.shape([1, 1, 48, 128])) + reshape97: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape96, R.shape([1, 48, 128])) + lv25 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(0), R.prim_value(T.float32(1.0)), reshape97), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape98: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv25, R.shape([1, 1, 16, 128])) + reshape99: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape98, R.shape([1, 1, 2048])) + lv292 = R.call_tir(cls.NT_matmul11, (reshape99, gpt_neox_layers_0_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add145: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv292, gpt_neox_layers_0_attention_dense_bias2) + layer_norm50: 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) + lv293 = R.call_tir(cls.NT_matmul12, (layer_norm50, gpt_neox_layers_0_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(lv293, gpt_neox_layers_0_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") + lv294 = R.call_tir(cls.NT_matmul13, (astype49, gpt_neox_layers_0_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(lv294, gpt_neox_layers_0_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, input_embed) + layer_norm51: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add149, gpt_neox_layers_1_input_layernorm_weight2, gpt_neox_layers_1_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv295 = R.call_tir(cls.NT_matmul10, (layer_norm51, gpt_neox_layers_1_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add150: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv295, gpt_neox_layers_1_attention_query_key_value_bias2) + reshape100: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add150, R.shape([1, 1, 48, 128])) + reshape101: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape100, R.shape([1, 48, 128])) + lv26 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(1), R.prim_value(T.float32(1.0)), reshape101), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape102: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv26, R.shape([1, 1, 16, 128])) + reshape103: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape102, R.shape([1, 1, 2048])) + lv296 = R.call_tir(cls.NT_matmul11, (reshape103, gpt_neox_layers_1_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add151: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv296, gpt_neox_layers_1_attention_dense_bias2) + layer_norm52: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add149, 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) + lv297 = R.call_tir(cls.NT_matmul12, (layer_norm52, gpt_neox_layers_1_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(lv297, gpt_neox_layers_1_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") + lv298 = R.call_tir(cls.NT_matmul13, (astype51, gpt_neox_layers_1_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(lv298, gpt_neox_layers_1_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_2_input_layernorm_weight2, gpt_neox_layers_2_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv299 = R.call_tir(cls.NT_matmul10, (layer_norm53, gpt_neox_layers_2_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add156: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv299, gpt_neox_layers_2_attention_query_key_value_bias2) + reshape104: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add156, R.shape([1, 1, 48, 128])) + reshape105: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape104, R.shape([1, 48, 128])) + lv27 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(2), R.prim_value(T.float32(1.0)), reshape105), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape106: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv27, R.shape([1, 1, 16, 128])) + reshape107: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape106, R.shape([1, 1, 2048])) + lv300 = R.call_tir(cls.NT_matmul11, (reshape107, gpt_neox_layers_2_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add157: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv300, gpt_neox_layers_2_attention_dense_bias2) + layer_norm54: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add155, 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) + lv301 = R.call_tir(cls.NT_matmul12, (layer_norm54, gpt_neox_layers_2_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(lv301, gpt_neox_layers_2_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") + lv302 = R.call_tir(cls.NT_matmul13, (astype53, gpt_neox_layers_2_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(lv302, gpt_neox_layers_2_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_3_input_layernorm_weight2, gpt_neox_layers_3_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv303 = R.call_tir(cls.NT_matmul10, (layer_norm55, gpt_neox_layers_3_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add162: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv303, gpt_neox_layers_3_attention_query_key_value_bias2) + reshape108: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add162, R.shape([1, 1, 48, 128])) + reshape109: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape108, R.shape([1, 48, 128])) + lv28 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(3), R.prim_value(T.float32(1.0)), reshape109), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape110: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv28, R.shape([1, 1, 16, 128])) + reshape111: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape110, R.shape([1, 1, 2048])) + lv304 = R.call_tir(cls.NT_matmul11, (reshape111, gpt_neox_layers_3_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add163: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv304, gpt_neox_layers_3_attention_dense_bias2) + layer_norm56: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add161, 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) + lv305 = R.call_tir(cls.NT_matmul12, (layer_norm56, gpt_neox_layers_3_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(lv305, gpt_neox_layers_3_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") + lv306 = R.call_tir(cls.NT_matmul13, (astype55, gpt_neox_layers_3_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(lv306, gpt_neox_layers_3_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_4_input_layernorm_weight2, gpt_neox_layers_4_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv307 = R.call_tir(cls.NT_matmul10, (layer_norm57, gpt_neox_layers_4_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add168: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv307, gpt_neox_layers_4_attention_query_key_value_bias2) + reshape112: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add168, R.shape([1, 1, 48, 128])) + reshape113: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape112, R.shape([1, 48, 128])) + lv29 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(4), R.prim_value(T.float32(1.0)), reshape113), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape114: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv29, R.shape([1, 1, 16, 128])) + reshape115: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape114, R.shape([1, 1, 2048])) + lv308 = R.call_tir(cls.NT_matmul11, (reshape115, gpt_neox_layers_4_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add169: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv308, gpt_neox_layers_4_attention_dense_bias2) + layer_norm58: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add167, 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) + lv309 = R.call_tir(cls.NT_matmul12, (layer_norm58, gpt_neox_layers_4_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(lv309, gpt_neox_layers_4_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") + lv310 = R.call_tir(cls.NT_matmul13, (astype57, gpt_neox_layers_4_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(lv310, gpt_neox_layers_4_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_5_input_layernorm_weight2, gpt_neox_layers_5_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv311 = R.call_tir(cls.NT_matmul10, (layer_norm59, gpt_neox_layers_5_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add174: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv311, gpt_neox_layers_5_attention_query_key_value_bias2) + reshape116: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add174, R.shape([1, 1, 48, 128])) + reshape117: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape116, R.shape([1, 48, 128])) + lv30 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(5), R.prim_value(T.float32(1.0)), reshape117), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape118: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv30, R.shape([1, 1, 16, 128])) + reshape119: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape118, R.shape([1, 1, 2048])) + lv312 = R.call_tir(cls.NT_matmul11, (reshape119, gpt_neox_layers_5_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add175: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv312, gpt_neox_layers_5_attention_dense_bias2) + layer_norm60: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add173, 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) + lv313 = R.call_tir(cls.NT_matmul12, (layer_norm60, gpt_neox_layers_5_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(lv313, gpt_neox_layers_5_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") + lv314 = R.call_tir(cls.NT_matmul13, (astype59, gpt_neox_layers_5_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(lv314, gpt_neox_layers_5_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_6_input_layernorm_weight2, gpt_neox_layers_6_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv315 = R.call_tir(cls.NT_matmul10, (layer_norm61, gpt_neox_layers_6_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add180: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv315, gpt_neox_layers_6_attention_query_key_value_bias2) + reshape120: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add180, R.shape([1, 1, 48, 128])) + reshape121: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape120, R.shape([1, 48, 128])) + lv31 = 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)), reshape121), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape122: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv31, R.shape([1, 1, 16, 128])) + reshape123: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape122, R.shape([1, 1, 2048])) + lv316 = R.call_tir(cls.NT_matmul11, (reshape123, gpt_neox_layers_6_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add181: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv316, gpt_neox_layers_6_attention_dense_bias2) + layer_norm62: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add179, 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) + lv317 = R.call_tir(cls.NT_matmul12, (layer_norm62, gpt_neox_layers_6_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(lv317, gpt_neox_layers_6_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") + lv318 = R.call_tir(cls.NT_matmul13, (astype61, gpt_neox_layers_6_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(lv318, gpt_neox_layers_6_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_7_input_layernorm_weight2, gpt_neox_layers_7_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv319 = R.call_tir(cls.NT_matmul10, (layer_norm63, gpt_neox_layers_7_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add186: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv319, gpt_neox_layers_7_attention_query_key_value_bias2) + reshape124: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add186, R.shape([1, 1, 48, 128])) + reshape125: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape124, R.shape([1, 48, 128])) + lv32 = 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)), reshape125), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape126: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv32, R.shape([1, 1, 16, 128])) + reshape127: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape126, R.shape([1, 1, 2048])) + lv320 = R.call_tir(cls.NT_matmul11, (reshape127, gpt_neox_layers_7_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add187: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv320, gpt_neox_layers_7_attention_dense_bias2) + layer_norm64: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add185, 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) + lv321 = R.call_tir(cls.NT_matmul12, (layer_norm64, gpt_neox_layers_7_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(lv321, gpt_neox_layers_7_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") + lv322 = R.call_tir(cls.NT_matmul13, (astype63, gpt_neox_layers_7_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(lv322, gpt_neox_layers_7_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_layers_8_input_layernorm_weight2, gpt_neox_layers_8_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv323 = R.call_tir(cls.NT_matmul10, (layer_norm65, gpt_neox_layers_8_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add192: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv323, gpt_neox_layers_8_attention_query_key_value_bias2) + reshape128: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add192, R.shape([1, 1, 48, 128])) + reshape129: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape128, R.shape([1, 48, 128])) + lv33 = 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)), reshape129), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape130: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv33, R.shape([1, 1, 16, 128])) + reshape131: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape130, R.shape([1, 1, 2048])) + lv324 = R.call_tir(cls.NT_matmul11, (reshape131, gpt_neox_layers_8_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add193: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv324, gpt_neox_layers_8_attention_dense_bias2) + layer_norm66: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add191, 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) + lv325 = R.call_tir(cls.NT_matmul12, (layer_norm66, gpt_neox_layers_8_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add194: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv325, gpt_neox_layers_8_mlp_dense_h_to_4h_bias2) + gelu32: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add194) + astype65: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu32, dtype="float16") + lv326 = R.call_tir(cls.NT_matmul13, (astype65, gpt_neox_layers_8_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add195: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv326, gpt_neox_layers_8_mlp_dense_4h_to_h_bias2) + astype66: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add195, dtype="float16") + add196: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype66, add193) + add197: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add196, add191) + layer_norm67: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add197, gpt_neox_layers_9_input_layernorm_weight2, gpt_neox_layers_9_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv327 = R.call_tir(cls.NT_matmul10, (layer_norm67, gpt_neox_layers_9_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add198: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv327, gpt_neox_layers_9_attention_query_key_value_bias2) + reshape132: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add198, R.shape([1, 1, 48, 128])) + reshape133: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape132, R.shape([1, 48, 128])) + lv34 = 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)), reshape133), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape134: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv34, R.shape([1, 1, 16, 128])) + reshape135: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape134, R.shape([1, 1, 2048])) + lv328 = R.call_tir(cls.NT_matmul11, (reshape135, gpt_neox_layers_9_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add199: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv328, gpt_neox_layers_9_attention_dense_bias2) + layer_norm68: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add197, 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) + lv329 = R.call_tir(cls.NT_matmul12, (layer_norm68, gpt_neox_layers_9_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add200: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv329, gpt_neox_layers_9_mlp_dense_h_to_4h_bias2) + gelu33: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add200) + astype67: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu33, dtype="float16") + lv330 = R.call_tir(cls.NT_matmul13, (astype67, gpt_neox_layers_9_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add201: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv330, gpt_neox_layers_9_mlp_dense_4h_to_h_bias2) + astype68: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add201, dtype="float16") + add202: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype68, add199) + add203: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add202, add197) + layer_norm69: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add203, gpt_neox_layers_10_input_layernorm_weight2, gpt_neox_layers_10_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv331 = R.call_tir(cls.NT_matmul10, (layer_norm69, gpt_neox_layers_10_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add204: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv331, gpt_neox_layers_10_attention_query_key_value_bias2) + reshape136: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add204, R.shape([1, 1, 48, 128])) + reshape137: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape136, R.shape([1, 48, 128])) + lv35 = 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)), reshape137), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape138: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv35, R.shape([1, 1, 16, 128])) + reshape139: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape138, R.shape([1, 1, 2048])) + lv332 = R.call_tir(cls.NT_matmul11, (reshape139, gpt_neox_layers_10_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add205: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv332, gpt_neox_layers_10_attention_dense_bias2) + layer_norm70: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add203, 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) + lv333 = R.call_tir(cls.NT_matmul12, (layer_norm70, gpt_neox_layers_10_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add206: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv333, gpt_neox_layers_10_mlp_dense_h_to_4h_bias2) + gelu34: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add206) + astype69: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu34, dtype="float16") + lv334 = R.call_tir(cls.NT_matmul13, (astype69, gpt_neox_layers_10_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add207: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv334, gpt_neox_layers_10_mlp_dense_4h_to_h_bias2) + astype70: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add207, dtype="float16") + add208: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype70, add205) + add209: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add208, add203) + layer_norm71: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add209, gpt_neox_layers_11_input_layernorm_weight2, gpt_neox_layers_11_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv335 = R.call_tir(cls.NT_matmul10, (layer_norm71, gpt_neox_layers_11_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add210: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv335, gpt_neox_layers_11_attention_query_key_value_bias2) + reshape140: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add210, R.shape([1, 1, 48, 128])) + reshape141: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape140, R.shape([1, 48, 128])) + lv36 = 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)), reshape141), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape142: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv36, R.shape([1, 1, 16, 128])) + reshape143: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape142, R.shape([1, 1, 2048])) + lv336 = R.call_tir(cls.NT_matmul11, (reshape143, gpt_neox_layers_11_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add211: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv336, gpt_neox_layers_11_attention_dense_bias2) + layer_norm72: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add209, 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) + lv337 = R.call_tir(cls.NT_matmul12, (layer_norm72, gpt_neox_layers_11_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add212: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv337, gpt_neox_layers_11_mlp_dense_h_to_4h_bias2) + gelu35: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add212) + astype71: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu35, dtype="float16") + lv338 = R.call_tir(cls.NT_matmul13, (astype71, gpt_neox_layers_11_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add213: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv338, gpt_neox_layers_11_mlp_dense_4h_to_h_bias2) + astype72: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add213, dtype="float16") + add214: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype72, add211) + add215: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add214, add209) + layer_norm73: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add215, gpt_neox_layers_12_input_layernorm_weight2, gpt_neox_layers_12_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv339 = R.call_tir(cls.NT_matmul10, (layer_norm73, gpt_neox_layers_12_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add216: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv339, gpt_neox_layers_12_attention_query_key_value_bias2) + reshape144: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add216, R.shape([1, 1, 48, 128])) + reshape145: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape144, R.shape([1, 48, 128])) + lv37 = 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)), reshape145), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape146: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv37, R.shape([1, 1, 16, 128])) + reshape147: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape146, R.shape([1, 1, 2048])) + lv340 = R.call_tir(cls.NT_matmul11, (reshape147, gpt_neox_layers_12_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add217: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv340, gpt_neox_layers_12_attention_dense_bias2) + layer_norm74: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add215, 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) + lv341 = R.call_tir(cls.NT_matmul12, (layer_norm74, gpt_neox_layers_12_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add218: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv341, gpt_neox_layers_12_mlp_dense_h_to_4h_bias2) + gelu36: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add218) + astype73: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu36, dtype="float16") + lv342 = R.call_tir(cls.NT_matmul13, (astype73, gpt_neox_layers_12_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add219: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv342, gpt_neox_layers_12_mlp_dense_4h_to_h_bias2) + astype74: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add219, dtype="float16") + add220: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype74, add217) + add221: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add220, add215) + layer_norm75: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add221, gpt_neox_layers_13_input_layernorm_weight2, gpt_neox_layers_13_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv343 = R.call_tir(cls.NT_matmul10, (layer_norm75, gpt_neox_layers_13_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add222: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv343, gpt_neox_layers_13_attention_query_key_value_bias2) + reshape148: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add222, R.shape([1, 1, 48, 128])) + reshape149: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape148, R.shape([1, 48, 128])) + lv38 = 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)), reshape149), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape150: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv38, R.shape([1, 1, 16, 128])) + reshape151: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape150, R.shape([1, 1, 2048])) + lv344 = R.call_tir(cls.NT_matmul11, (reshape151, gpt_neox_layers_13_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add223: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv344, gpt_neox_layers_13_attention_dense_bias2) + layer_norm76: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add221, 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) + lv345 = R.call_tir(cls.NT_matmul12, (layer_norm76, gpt_neox_layers_13_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add224: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv345, gpt_neox_layers_13_mlp_dense_h_to_4h_bias2) + gelu37: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add224) + astype75: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu37, dtype="float16") + lv346 = R.call_tir(cls.NT_matmul13, (astype75, gpt_neox_layers_13_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add225: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv346, gpt_neox_layers_13_mlp_dense_4h_to_h_bias2) + astype76: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add225, dtype="float16") + add226: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype76, add223) + add227: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add226, add221) + layer_norm77: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add227, gpt_neox_layers_14_input_layernorm_weight2, gpt_neox_layers_14_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv347 = R.call_tir(cls.NT_matmul10, (layer_norm77, gpt_neox_layers_14_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add228: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv347, gpt_neox_layers_14_attention_query_key_value_bias2) + reshape152: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add228, R.shape([1, 1, 48, 128])) + reshape153: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape152, R.shape([1, 48, 128])) + lv39 = 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)), reshape153), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape154: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv39, R.shape([1, 1, 16, 128])) + reshape155: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape154, R.shape([1, 1, 2048])) + lv348 = R.call_tir(cls.NT_matmul11, (reshape155, gpt_neox_layers_14_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add229: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv348, gpt_neox_layers_14_attention_dense_bias2) + layer_norm78: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add227, 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) + lv349 = R.call_tir(cls.NT_matmul12, (layer_norm78, gpt_neox_layers_14_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add230: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv349, gpt_neox_layers_14_mlp_dense_h_to_4h_bias2) + gelu38: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add230) + astype77: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu38, dtype="float16") + lv350 = R.call_tir(cls.NT_matmul13, (astype77, gpt_neox_layers_14_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add231: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv350, gpt_neox_layers_14_mlp_dense_4h_to_h_bias2) + astype78: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add231, dtype="float16") + add232: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype78, add229) + add233: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add232, add227) + layer_norm79: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add233, gpt_neox_layers_15_input_layernorm_weight2, gpt_neox_layers_15_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv351 = R.call_tir(cls.NT_matmul10, (layer_norm79, gpt_neox_layers_15_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add234: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv351, gpt_neox_layers_15_attention_query_key_value_bias2) + reshape156: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add234, R.shape([1, 1, 48, 128])) + reshape157: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape156, R.shape([1, 48, 128])) + lv40 = 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)), reshape157), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape158: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv40, R.shape([1, 1, 16, 128])) + reshape159: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape158, R.shape([1, 1, 2048])) + lv352 = R.call_tir(cls.NT_matmul11, (reshape159, gpt_neox_layers_15_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add235: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv352, gpt_neox_layers_15_attention_dense_bias2) + layer_norm80: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add233, 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) + lv353 = R.call_tir(cls.NT_matmul12, (layer_norm80, gpt_neox_layers_15_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add236: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv353, gpt_neox_layers_15_mlp_dense_h_to_4h_bias2) + gelu39: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add236) + astype79: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu39, dtype="float16") + lv354 = R.call_tir(cls.NT_matmul13, (astype79, gpt_neox_layers_15_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add237: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv354, gpt_neox_layers_15_mlp_dense_4h_to_h_bias2) + astype80: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add237, dtype="float16") + add238: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype80, add235) + add239: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add238, add233) + layer_norm81: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add239, gpt_neox_layers_16_input_layernorm_weight2, gpt_neox_layers_16_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv355 = R.call_tir(cls.NT_matmul10, (layer_norm81, gpt_neox_layers_16_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add240: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv355, gpt_neox_layers_16_attention_query_key_value_bias2) + reshape160: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add240, R.shape([1, 1, 48, 128])) + reshape161: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape160, R.shape([1, 48, 128])) + lv41 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), reshape161), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape162: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv41, R.shape([1, 1, 16, 128])) + reshape163: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape162, R.shape([1, 1, 2048])) + lv356 = R.call_tir(cls.NT_matmul11, (reshape163, gpt_neox_layers_16_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add241: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv356, gpt_neox_layers_16_attention_dense_bias2) + layer_norm82: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add239, gpt_neox_layers_16_post_attention_layernorm_weight2, gpt_neox_layers_16_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv357 = R.call_tir(cls.NT_matmul12, (layer_norm82, gpt_neox_layers_16_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add242: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv357, gpt_neox_layers_16_mlp_dense_h_to_4h_bias2) + gelu40: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add242) + astype81: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu40, dtype="float16") + lv358 = R.call_tir(cls.NT_matmul13, (astype81, gpt_neox_layers_16_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add243: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv358, gpt_neox_layers_16_mlp_dense_4h_to_h_bias2) + astype82: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add243, dtype="float16") + add244: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype82, add241) + add245: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add244, add239) + layer_norm83: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add245, gpt_neox_layers_17_input_layernorm_weight2, gpt_neox_layers_17_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv359 = R.call_tir(cls.NT_matmul10, (layer_norm83, gpt_neox_layers_17_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add246: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv359, gpt_neox_layers_17_attention_query_key_value_bias2) + reshape164: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add246, R.shape([1, 1, 48, 128])) + reshape165: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape164, R.shape([1, 48, 128])) + lv42 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), reshape165), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape166: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv42, R.shape([1, 1, 16, 128])) + reshape167: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape166, R.shape([1, 1, 2048])) + lv360 = R.call_tir(cls.NT_matmul11, (reshape167, gpt_neox_layers_17_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add247: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv360, gpt_neox_layers_17_attention_dense_bias2) + layer_norm84: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add245, gpt_neox_layers_17_post_attention_layernorm_weight2, gpt_neox_layers_17_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv361 = R.call_tir(cls.NT_matmul12, (layer_norm84, gpt_neox_layers_17_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add248: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv361, gpt_neox_layers_17_mlp_dense_h_to_4h_bias2) + gelu41: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add248) + astype83: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu41, dtype="float16") + lv362 = R.call_tir(cls.NT_matmul13, (astype83, gpt_neox_layers_17_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add249: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv362, gpt_neox_layers_17_mlp_dense_4h_to_h_bias2) + astype84: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add249, dtype="float16") + add250: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype84, add247) + add251: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add250, add245) + layer_norm85: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add251, gpt_neox_layers_18_input_layernorm_weight2, gpt_neox_layers_18_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv363 = R.call_tir(cls.NT_matmul10, (layer_norm85, gpt_neox_layers_18_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add252: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv363, gpt_neox_layers_18_attention_query_key_value_bias2) + reshape168: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add252, R.shape([1, 1, 48, 128])) + reshape169: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape168, R.shape([1, 48, 128])) + lv43 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), reshape169), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape170: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv43, R.shape([1, 1, 16, 128])) + reshape171: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape170, R.shape([1, 1, 2048])) + lv364 = R.call_tir(cls.NT_matmul11, (reshape171, gpt_neox_layers_18_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add253: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv364, gpt_neox_layers_18_attention_dense_bias2) + layer_norm86: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add251, gpt_neox_layers_18_post_attention_layernorm_weight2, gpt_neox_layers_18_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv365 = R.call_tir(cls.NT_matmul12, (layer_norm86, gpt_neox_layers_18_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add254: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv365, gpt_neox_layers_18_mlp_dense_h_to_4h_bias2) + gelu42: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add254) + astype85: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu42, dtype="float16") + lv366 = R.call_tir(cls.NT_matmul13, (astype85, gpt_neox_layers_18_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add255: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv366, gpt_neox_layers_18_mlp_dense_4h_to_h_bias2) + astype86: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add255, dtype="float16") + add256: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype86, add253) + add257: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add256, add251) + layer_norm87: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add257, gpt_neox_layers_19_input_layernorm_weight2, gpt_neox_layers_19_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv367 = R.call_tir(cls.NT_matmul10, (layer_norm87, gpt_neox_layers_19_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add258: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv367, gpt_neox_layers_19_attention_query_key_value_bias2) + reshape172: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add258, R.shape([1, 1, 48, 128])) + reshape173: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape172, R.shape([1, 48, 128])) + lv44 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), reshape173), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape174: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv44, R.shape([1, 1, 16, 128])) + reshape175: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape174, R.shape([1, 1, 2048])) + lv368 = R.call_tir(cls.NT_matmul11, (reshape175, gpt_neox_layers_19_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add259: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv368, gpt_neox_layers_19_attention_dense_bias2) + layer_norm88: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add257, gpt_neox_layers_19_post_attention_layernorm_weight2, gpt_neox_layers_19_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv369 = R.call_tir(cls.NT_matmul12, (layer_norm88, gpt_neox_layers_19_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add260: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv369, gpt_neox_layers_19_mlp_dense_h_to_4h_bias2) + gelu43: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add260) + astype87: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu43, dtype="float16") + lv370 = R.call_tir(cls.NT_matmul13, (astype87, gpt_neox_layers_19_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add261: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv370, gpt_neox_layers_19_mlp_dense_4h_to_h_bias2) + astype88: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add261, dtype="float16") + add262: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype88, add259) + add263: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add262, add257) + layer_norm89: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add263, gpt_neox_layers_20_input_layernorm_weight2, gpt_neox_layers_20_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv371 = R.call_tir(cls.NT_matmul10, (layer_norm89, gpt_neox_layers_20_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add264: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv371, gpt_neox_layers_20_attention_query_key_value_bias2) + reshape176: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add264, R.shape([1, 1, 48, 128])) + reshape177: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape176, R.shape([1, 48, 128])) + lv45 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), reshape177), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape178: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv45, R.shape([1, 1, 16, 128])) + reshape179: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape178, R.shape([1, 1, 2048])) + lv372 = R.call_tir(cls.NT_matmul11, (reshape179, gpt_neox_layers_20_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add265: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv372, gpt_neox_layers_20_attention_dense_bias2) + layer_norm90: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add263, gpt_neox_layers_20_post_attention_layernorm_weight2, gpt_neox_layers_20_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv373 = R.call_tir(cls.NT_matmul12, (layer_norm90, gpt_neox_layers_20_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add266: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv373, gpt_neox_layers_20_mlp_dense_h_to_4h_bias2) + gelu44: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add266) + astype89: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu44, dtype="float16") + lv374 = R.call_tir(cls.NT_matmul13, (astype89, gpt_neox_layers_20_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add267: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv374, gpt_neox_layers_20_mlp_dense_4h_to_h_bias2) + astype90: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add267, dtype="float16") + add268: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype90, add265) + add269: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add268, add263) + layer_norm91: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add269, gpt_neox_layers_21_input_layernorm_weight2, gpt_neox_layers_21_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv375 = R.call_tir(cls.NT_matmul10, (layer_norm91, gpt_neox_layers_21_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add270: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv375, gpt_neox_layers_21_attention_query_key_value_bias2) + reshape180: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add270, R.shape([1, 1, 48, 128])) + reshape181: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape180, R.shape([1, 48, 128])) + lv46 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), reshape181), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape182: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv46, R.shape([1, 1, 16, 128])) + reshape183: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape182, R.shape([1, 1, 2048])) + lv376 = R.call_tir(cls.NT_matmul11, (reshape183, gpt_neox_layers_21_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add271: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv376, gpt_neox_layers_21_attention_dense_bias2) + layer_norm92: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add269, gpt_neox_layers_21_post_attention_layernorm_weight2, gpt_neox_layers_21_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv377 = R.call_tir(cls.NT_matmul12, (layer_norm92, gpt_neox_layers_21_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add272: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv377, gpt_neox_layers_21_mlp_dense_h_to_4h_bias2) + gelu45: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add272) + astype91: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu45, dtype="float16") + lv378 = R.call_tir(cls.NT_matmul13, (astype91, gpt_neox_layers_21_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add273: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv378, gpt_neox_layers_21_mlp_dense_4h_to_h_bias2) + astype92: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add273, dtype="float16") + add274: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype92, add271) + add275: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add274, add269) + layer_norm93: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add275, gpt_neox_layers_22_input_layernorm_weight2, gpt_neox_layers_22_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv379 = R.call_tir(cls.NT_matmul10, (layer_norm93, gpt_neox_layers_22_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add276: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv379, gpt_neox_layers_22_attention_query_key_value_bias2) + reshape184: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add276, R.shape([1, 1, 48, 128])) + reshape185: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape184, R.shape([1, 48, 128])) + lv47 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), reshape185), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape186: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv47, R.shape([1, 1, 16, 128])) + reshape187: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape186, R.shape([1, 1, 2048])) + lv380 = R.call_tir(cls.NT_matmul11, (reshape187, gpt_neox_layers_22_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add277: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv380, gpt_neox_layers_22_attention_dense_bias2) + layer_norm94: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add275, gpt_neox_layers_22_post_attention_layernorm_weight2, gpt_neox_layers_22_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv381 = R.call_tir(cls.NT_matmul12, (layer_norm94, gpt_neox_layers_22_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add278: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv381, gpt_neox_layers_22_mlp_dense_h_to_4h_bias2) + gelu46: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add278) + astype93: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu46, dtype="float16") + lv382 = R.call_tir(cls.NT_matmul13, (astype93, gpt_neox_layers_22_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add279: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv382, gpt_neox_layers_22_mlp_dense_4h_to_h_bias2) + astype94: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add279, dtype="float16") + add280: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype94, add277) + add281: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add280, add275) + layer_norm95: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add281, gpt_neox_layers_23_input_layernorm_weight2, gpt_neox_layers_23_input_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv383 = R.call_tir(cls.NT_matmul10, (layer_norm95, gpt_neox_layers_23_attention_query_key_value_weight2), out_sinfo=R.Tensor((1, 1, 6144), dtype="float16")) + add282: R.Tensor((1, 1, 6144), dtype="float16") = R.add(lv383, gpt_neox_layers_23_attention_query_key_value_bias2) + reshape188: R.Tensor((1, 1, 48, 128), dtype="float16") = R.reshape(add282, R.shape([1, 1, 48, 128])) + reshape189: R.Tensor((1, 48, 128), dtype="float16") = R.reshape(reshape188, R.shape([1, 48, 128])) + lv48 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), reshape189), out_sinfo=R.Tensor((1, 16, 128), dtype="float16")) + reshape190: R.Tensor((1, 1, 16, 128), dtype="float16") = R.reshape(lv48, R.shape([1, 1, 16, 128])) + reshape191: R.Tensor((1, 1, 2048), dtype="float16") = R.reshape(reshape190, R.shape([1, 1, 2048])) + lv384 = R.call_tir(cls.NT_matmul11, (reshape191, gpt_neox_layers_23_attention_dense_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + add283: R.Tensor((1, 1, 2048), dtype="float16") = R.add(lv384, gpt_neox_layers_23_attention_dense_bias2) + layer_norm96: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add281, gpt_neox_layers_23_post_attention_layernorm_weight2, gpt_neox_layers_23_post_attention_layernorm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv385 = R.call_tir(cls.NT_matmul12, (layer_norm96, gpt_neox_layers_23_mlp_dense_h_to_4h_weight2), out_sinfo=R.Tensor((1, 1, 8192), dtype="float32")) + add284: R.Tensor((1, 1, 8192), dtype="float32") = R.add(lv385, gpt_neox_layers_23_mlp_dense_h_to_4h_bias2) + gelu47: R.Tensor((1, 1, 8192), dtype="float32") = R.nn.gelu(add284) + astype95: R.Tensor((1, 1, 8192), dtype="float16") = R.astype(gelu47, dtype="float16") + lv386 = R.call_tir(cls.NT_matmul13, (astype95, gpt_neox_layers_23_mlp_dense_4h_to_h_weight2), out_sinfo=R.Tensor((1, 1, 2048), dtype="float32")) + add285: R.Tensor((1, 1, 2048), dtype="float32") = R.add(lv386, gpt_neox_layers_23_mlp_dense_4h_to_h_bias2) + astype96: R.Tensor((1, 1, 2048), dtype="float16") = R.astype(add285, dtype="float16") + add286: R.Tensor((1, 1, 2048), dtype="float16") = R.add(astype96, add283) + add287: R.Tensor((1, 1, 2048), dtype="float16") = R.add(add286, add281) + layer_norm97: R.Tensor((1, 1, 2048), dtype="float16") = R.nn.layer_norm(add287, gpt_neox_final_layer_norm_weight2, gpt_neox_final_layer_norm_bias2, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv387 = R.call_tir(cls.NT_matmul14, (layer_norm97, embed_out_weight2), out_sinfo=R.Tensor((1, 1, vocab_size), dtype="float16")) + astype97: R.Tensor((1, 1, vocab_size), dtype="float32") = R.astype(lv387, dtype="float32") + gv2: R.Tuple(R.Tensor((1, 1, vocab_size), dtype="float32"), R.Object) = astype97, 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((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_norm147: 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_dims291: 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_norm147, permute_dims291, 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"), reshape291: 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_dims292: 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(reshape291, permute_dims292, 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_norm49: 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_dims97: 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_norm49, permute_dims97, 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"), reshape99: 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_dims98: 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(reshape99, permute_dims98, 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_norm50: 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_dims99: 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_norm50, permute_dims99, 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"), astype49: 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_dims100: 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(astype49, permute_dims100, 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_norm97: 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_dims193: 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_norm97, permute_dims193, 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_norm148: 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_dims293: 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_norm148, permute_dims293, 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"), astype147: 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_dims294: 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(astype147, permute_dims294, 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_norm195: 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_dims387: 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_norm195, permute_dims387, 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_norm98: 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_dims194: 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_norm98, permute_dims194, 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"), reshape195: 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_dims195: 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(reshape195, permute_dims195, 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_norm99: 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_dims196: 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_norm99, permute_dims196, 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"), astype98: 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_dims197: 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(astype98, permute_dims197, 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_dims290: 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_dims290, 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((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_layers_16_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[193] + gpt_neox_layers_16_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[194] + gpt_neox_layers_16_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[195] + gpt_neox_layers_16_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[196] + gpt_neox_layers_16_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[197] + gpt_neox_layers_16_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[198] + gpt_neox_layers_16_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[199] + gpt_neox_layers_16_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[200] + gpt_neox_layers_16_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[201] + gpt_neox_layers_16_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[202] + gpt_neox_layers_16_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[203] + gpt_neox_layers_16_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[204] + gpt_neox_layers_17_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[205] + gpt_neox_layers_17_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[206] + gpt_neox_layers_17_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[207] + gpt_neox_layers_17_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[208] + gpt_neox_layers_17_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[209] + gpt_neox_layers_17_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[210] + gpt_neox_layers_17_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[211] + gpt_neox_layers_17_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[212] + gpt_neox_layers_17_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[213] + gpt_neox_layers_17_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[214] + gpt_neox_layers_17_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[215] + gpt_neox_layers_17_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[216] + gpt_neox_layers_18_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[217] + gpt_neox_layers_18_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[218] + gpt_neox_layers_18_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[219] + gpt_neox_layers_18_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[220] + gpt_neox_layers_18_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[221] + gpt_neox_layers_18_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[222] + gpt_neox_layers_18_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[223] + gpt_neox_layers_18_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[224] + gpt_neox_layers_18_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[225] + gpt_neox_layers_18_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[226] + gpt_neox_layers_18_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[227] + gpt_neox_layers_18_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[228] + gpt_neox_layers_19_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[229] + gpt_neox_layers_19_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[230] + gpt_neox_layers_19_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[231] + gpt_neox_layers_19_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[232] + gpt_neox_layers_19_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[233] + gpt_neox_layers_19_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[234] + gpt_neox_layers_19_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[235] + gpt_neox_layers_19_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[236] + gpt_neox_layers_19_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[237] + gpt_neox_layers_19_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[238] + gpt_neox_layers_19_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[239] + gpt_neox_layers_19_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[240] + gpt_neox_layers_20_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[241] + gpt_neox_layers_20_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[242] + gpt_neox_layers_20_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[243] + gpt_neox_layers_20_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[244] + gpt_neox_layers_20_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[245] + gpt_neox_layers_20_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[246] + gpt_neox_layers_20_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[247] + gpt_neox_layers_20_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[248] + gpt_neox_layers_20_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[249] + gpt_neox_layers_20_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[250] + gpt_neox_layers_20_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[251] + gpt_neox_layers_20_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[252] + gpt_neox_layers_21_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[253] + gpt_neox_layers_21_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[254] + gpt_neox_layers_21_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[255] + gpt_neox_layers_21_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[256] + gpt_neox_layers_21_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[257] + gpt_neox_layers_21_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[258] + gpt_neox_layers_21_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[259] + gpt_neox_layers_21_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[260] + gpt_neox_layers_21_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[261] + gpt_neox_layers_21_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[262] + gpt_neox_layers_21_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[263] + gpt_neox_layers_21_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[264] + gpt_neox_layers_22_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[265] + gpt_neox_layers_22_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[266] + gpt_neox_layers_22_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[267] + gpt_neox_layers_22_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[268] + gpt_neox_layers_22_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[269] + gpt_neox_layers_22_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[270] + gpt_neox_layers_22_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[271] + gpt_neox_layers_22_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[272] + gpt_neox_layers_22_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[273] + gpt_neox_layers_22_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[274] + gpt_neox_layers_22_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[275] + gpt_neox_layers_22_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[276] + gpt_neox_layers_23_input_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[277] + gpt_neox_layers_23_input_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[278] + gpt_neox_layers_23_post_attention_layernorm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[279] + gpt_neox_layers_23_post_attention_layernorm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[280] + gpt_neox_layers_23_attention_query_key_value_weight1: R.Tensor((6144, 2048), dtype="float16") = packed_params[281] + gpt_neox_layers_23_attention_query_key_value_bias1: R.Tensor((6144,), dtype="float16") = packed_params[282] + gpt_neox_layers_23_attention_dense_weight1: R.Tensor((2048, 2048), dtype="float16") = packed_params[283] + gpt_neox_layers_23_attention_dense_bias1: R.Tensor((2048,), dtype="float16") = packed_params[284] + gpt_neox_layers_23_mlp_dense_h_to_4h_weight1: R.Tensor((8192, 2048), dtype="float16") = packed_params[285] + gpt_neox_layers_23_mlp_dense_h_to_4h_bias1: R.Tensor((8192,), dtype="float32") = packed_params[286] + gpt_neox_layers_23_mlp_dense_4h_to_h_weight1: R.Tensor((2048, 8192), dtype="float16") = packed_params[287] + gpt_neox_layers_23_mlp_dense_4h_to_h_bias1: R.Tensor((2048,), dtype="float32") = packed_params[288] + gpt_neox_final_layer_norm_weight1: R.Tensor((2048,), dtype="float16") = packed_params[289] + gpt_neox_final_layer_norm_bias1: R.Tensor((2048,), dtype="float16") = packed_params[290] + embed_out_weight1: R.Tensor((vocab_size, 2048), dtype="float16") = packed_params[291] + 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) + lv388 = 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(lv388, gpt_neox_layers_0_attention_query_key_value_bias1) + reshape: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add, R.shape([1, seq_len, 48, 128])) + reshape1: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape2: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv, R.shape([1, seq_len, 16, 128])) + reshape3: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape2, R.shape([1, seq_len, 2048])) + lv389 = 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(lv389, 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) + lv390 = 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(lv390, 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") + lv391 = 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(lv391, 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) + lv392 = 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(lv392, gpt_neox_layers_1_attention_query_key_value_bias1) + reshape4: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add6, R.shape([1, seq_len, 48, 128])) + reshape5: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape4, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape6: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv1, R.shape([1, seq_len, 16, 128])) + reshape7: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape6, R.shape([1, seq_len, 2048])) + lv393 = 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(lv393, 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) + lv394 = 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(lv394, 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") + lv395 = 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(lv395, 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) + lv396 = 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(lv396, gpt_neox_layers_2_attention_query_key_value_bias1) + reshape8: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add12, R.shape([1, seq_len, 48, 128])) + reshape9: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape8, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape10: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv2, R.shape([1, seq_len, 16, 128])) + reshape11: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape10, R.shape([1, seq_len, 2048])) + lv397 = 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(lv397, 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) + lv398 = 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(lv398, 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") + lv399 = 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(lv399, 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) + lv400 = 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(lv400, gpt_neox_layers_3_attention_query_key_value_bias1) + reshape12: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add18, R.shape([1, seq_len, 48, 128])) + reshape13: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape12, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape14: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv3, R.shape([1, seq_len, 16, 128])) + reshape15: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape14, R.shape([1, seq_len, 2048])) + lv401 = 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(lv401, 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) + lv402 = 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(lv402, 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") + lv403 = 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(lv403, 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) + lv404 = 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(lv404, gpt_neox_layers_4_attention_query_key_value_bias1) + reshape16: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add24, R.shape([1, seq_len, 48, 128])) + reshape17: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape16, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape18: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv4, R.shape([1, seq_len, 16, 128])) + reshape19: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape18, R.shape([1, seq_len, 2048])) + lv405 = 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(lv405, 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) + lv406 = 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(lv406, 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") + lv407 = 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(lv407, 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) + lv408 = 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(lv408, gpt_neox_layers_5_attention_query_key_value_bias1) + reshape20: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add30, R.shape([1, seq_len, 48, 128])) + reshape21: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape20, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape22: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv5, R.shape([1, seq_len, 16, 128])) + reshape23: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape22, R.shape([1, seq_len, 2048])) + lv409 = 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(lv409, 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) + lv410 = 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(lv410, 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") + lv411 = 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(lv411, 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) + lv412 = 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(lv412, gpt_neox_layers_6_attention_query_key_value_bias1) + reshape24: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add36, R.shape([1, seq_len, 48, 128])) + reshape25: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape24, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape26: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv6, R.shape([1, seq_len, 16, 128])) + reshape27: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape26, R.shape([1, seq_len, 2048])) + lv413 = 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(lv413, 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) + lv414 = 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(lv414, 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") + lv415 = 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(lv415, 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) + lv416 = 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(lv416, gpt_neox_layers_7_attention_query_key_value_bias1) + reshape28: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add42, R.shape([1, seq_len, 48, 128])) + reshape29: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape28, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape30: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv7, R.shape([1, seq_len, 16, 128])) + reshape31: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape30, R.shape([1, seq_len, 2048])) + lv417 = 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(lv417, 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) + lv418 = 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(lv418, 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") + lv419 = 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(lv419, 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) + lv420 = 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(lv420, gpt_neox_layers_8_attention_query_key_value_bias1) + reshape32: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add48, R.shape([1, seq_len, 48, 128])) + reshape33: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape32, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape34: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv8, R.shape([1, seq_len, 16, 128])) + reshape35: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape34, R.shape([1, seq_len, 2048])) + lv421 = 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(lv421, 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) + lv422 = 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(lv422, 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") + lv423 = 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(lv423, 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) + lv424 = 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(lv424, gpt_neox_layers_9_attention_query_key_value_bias1) + reshape36: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add54, R.shape([1, seq_len, 48, 128])) + reshape37: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape36, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape38: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv9, R.shape([1, seq_len, 16, 128])) + reshape39: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape38, R.shape([1, seq_len, 2048])) + lv425 = 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(lv425, 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) + lv426 = 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(lv426, 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") + lv427 = 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(lv427, 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) + lv428 = 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(lv428, gpt_neox_layers_10_attention_query_key_value_bias1) + reshape40: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add60, R.shape([1, seq_len, 48, 128])) + reshape41: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape40, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape42: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv10, R.shape([1, seq_len, 16, 128])) + reshape43: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape42, R.shape([1, seq_len, 2048])) + lv429 = 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(lv429, 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) + lv430 = 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(lv430, 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") + lv431 = 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(lv431, 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) + lv432 = 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(lv432, gpt_neox_layers_11_attention_query_key_value_bias1) + reshape44: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add66, R.shape([1, seq_len, 48, 128])) + reshape45: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape44, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape46: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv11, R.shape([1, seq_len, 16, 128])) + reshape47: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape46, R.shape([1, seq_len, 2048])) + lv433 = 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(lv433, 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) + lv434 = 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(lv434, 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") + lv435 = 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(lv435, 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) + lv436 = 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(lv436, gpt_neox_layers_12_attention_query_key_value_bias1) + reshape48: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add72, R.shape([1, seq_len, 48, 128])) + reshape49: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape48, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape50: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv12, R.shape([1, seq_len, 16, 128])) + reshape51: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape50, R.shape([1, seq_len, 2048])) + lv437 = 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(lv437, 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) + lv438 = 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(lv438, 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") + lv439 = 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(lv439, 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) + lv440 = 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(lv440, gpt_neox_layers_13_attention_query_key_value_bias1) + reshape52: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add78, R.shape([1, seq_len, 48, 128])) + reshape53: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape52, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape54: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv13, R.shape([1, seq_len, 16, 128])) + reshape55: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape54, R.shape([1, seq_len, 2048])) + lv441 = 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(lv441, 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) + lv442 = 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(lv442, 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") + lv443 = 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(lv443, 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) + lv444 = 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(lv444, gpt_neox_layers_14_attention_query_key_value_bias1) + reshape56: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add84, R.shape([1, seq_len, 48, 128])) + reshape57: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape56, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape58: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv14, R.shape([1, seq_len, 16, 128])) + reshape59: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape58, R.shape([1, seq_len, 2048])) + lv445 = 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(lv445, 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) + lv446 = 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(lv446, 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") + lv447 = 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(lv447, 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) + lv448 = 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(lv448, gpt_neox_layers_15_attention_query_key_value_bias1) + reshape60: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add90, R.shape([1, seq_len, 48, 128])) + reshape61: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape60, R.shape([seq_len, 48, 128])) + 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, 16, 128), dtype="float16")) + reshape62: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv15, R.shape([1, seq_len, 16, 128])) + reshape63: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape62, R.shape([1, seq_len, 2048])) + lv449 = 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(lv449, 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) + lv450 = 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(lv450, 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") + lv451 = 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(lv451, 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_layers_16_input_layernorm_weight1, gpt_neox_layers_16_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv452 = R.call_tir(cls.NT_matmul5, (layer_norm32, gpt_neox_layers_16_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add96: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv452, gpt_neox_layers_16_attention_query_key_value_bias1) + reshape64: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add96, R.shape([1, seq_len, 48, 128])) + reshape65: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape64, R.shape([seq_len, 48, 128])) + lv16 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(16), R.prim_value(T.float32(1.0)), reshape65), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape66: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv16, R.shape([1, seq_len, 16, 128])) + reshape67: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape66, R.shape([1, seq_len, 2048])) + lv453 = R.call_tir(cls.NT_matmul6, (reshape67, gpt_neox_layers_16_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add97: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv453, gpt_neox_layers_16_attention_dense_bias1) + layer_norm33: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add95, gpt_neox_layers_16_post_attention_layernorm_weight1, gpt_neox_layers_16_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv454 = R.call_tir(cls.NT_matmul7, (layer_norm33, gpt_neox_layers_16_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add98: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv454, gpt_neox_layers_16_mlp_dense_h_to_4h_bias1) + gelu16: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add98) + astype32: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu16, dtype="float16") + lv455 = R.call_tir(cls.NT_matmul8, (astype32, gpt_neox_layers_16_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add99: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv455, gpt_neox_layers_16_mlp_dense_4h_to_h_bias1) + astype33: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add99, dtype="float16") + add100: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype33, add97) + add101: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add100, add95) + layer_norm34: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add101, gpt_neox_layers_17_input_layernorm_weight1, gpt_neox_layers_17_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv456 = R.call_tir(cls.NT_matmul5, (layer_norm34, gpt_neox_layers_17_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add102: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv456, gpt_neox_layers_17_attention_query_key_value_bias1) + reshape68: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add102, R.shape([1, seq_len, 48, 128])) + reshape69: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape68, R.shape([seq_len, 48, 128])) + lv17 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(17), R.prim_value(T.float32(1.0)), reshape69), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape70: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv17, R.shape([1, seq_len, 16, 128])) + reshape71: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape70, R.shape([1, seq_len, 2048])) + lv457 = R.call_tir(cls.NT_matmul6, (reshape71, gpt_neox_layers_17_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add103: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv457, gpt_neox_layers_17_attention_dense_bias1) + layer_norm35: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add101, gpt_neox_layers_17_post_attention_layernorm_weight1, gpt_neox_layers_17_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv458 = R.call_tir(cls.NT_matmul7, (layer_norm35, gpt_neox_layers_17_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add104: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv458, gpt_neox_layers_17_mlp_dense_h_to_4h_bias1) + gelu17: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add104) + astype34: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu17, dtype="float16") + lv459 = R.call_tir(cls.NT_matmul8, (astype34, gpt_neox_layers_17_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add105: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv459, gpt_neox_layers_17_mlp_dense_4h_to_h_bias1) + astype35: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add105, dtype="float16") + add106: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype35, add103) + add107: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add106, add101) + layer_norm36: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add107, gpt_neox_layers_18_input_layernorm_weight1, gpt_neox_layers_18_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv460 = R.call_tir(cls.NT_matmul5, (layer_norm36, gpt_neox_layers_18_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add108: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv460, gpt_neox_layers_18_attention_query_key_value_bias1) + reshape72: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add108, R.shape([1, seq_len, 48, 128])) + reshape73: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape72, R.shape([seq_len, 48, 128])) + lv18 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(18), R.prim_value(T.float32(1.0)), reshape73), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape74: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv18, R.shape([1, seq_len, 16, 128])) + reshape75: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape74, R.shape([1, seq_len, 2048])) + lv461 = R.call_tir(cls.NT_matmul6, (reshape75, gpt_neox_layers_18_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add109: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv461, gpt_neox_layers_18_attention_dense_bias1) + layer_norm37: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add107, gpt_neox_layers_18_post_attention_layernorm_weight1, gpt_neox_layers_18_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv462 = R.call_tir(cls.NT_matmul7, (layer_norm37, gpt_neox_layers_18_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add110: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv462, gpt_neox_layers_18_mlp_dense_h_to_4h_bias1) + gelu18: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add110) + astype36: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu18, dtype="float16") + lv463 = R.call_tir(cls.NT_matmul8, (astype36, gpt_neox_layers_18_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add111: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv463, gpt_neox_layers_18_mlp_dense_4h_to_h_bias1) + astype37: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add111, dtype="float16") + add112: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype37, add109) + add113: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add112, add107) + layer_norm38: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add113, gpt_neox_layers_19_input_layernorm_weight1, gpt_neox_layers_19_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv464 = R.call_tir(cls.NT_matmul5, (layer_norm38, gpt_neox_layers_19_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add114: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv464, gpt_neox_layers_19_attention_query_key_value_bias1) + reshape76: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add114, R.shape([1, seq_len, 48, 128])) + reshape77: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape76, R.shape([seq_len, 48, 128])) + lv19 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(19), R.prim_value(T.float32(1.0)), reshape77), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape78: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv19, R.shape([1, seq_len, 16, 128])) + reshape79: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape78, R.shape([1, seq_len, 2048])) + lv465 = R.call_tir(cls.NT_matmul6, (reshape79, gpt_neox_layers_19_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add115: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv465, gpt_neox_layers_19_attention_dense_bias1) + layer_norm39: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add113, gpt_neox_layers_19_post_attention_layernorm_weight1, gpt_neox_layers_19_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv466 = R.call_tir(cls.NT_matmul7, (layer_norm39, gpt_neox_layers_19_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add116: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv466, gpt_neox_layers_19_mlp_dense_h_to_4h_bias1) + gelu19: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add116) + astype38: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu19, dtype="float16") + lv467 = R.call_tir(cls.NT_matmul8, (astype38, gpt_neox_layers_19_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add117: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv467, gpt_neox_layers_19_mlp_dense_4h_to_h_bias1) + astype39: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add117, dtype="float16") + add118: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype39, add115) + add119: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add118, add113) + layer_norm40: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add119, gpt_neox_layers_20_input_layernorm_weight1, gpt_neox_layers_20_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv468 = R.call_tir(cls.NT_matmul5, (layer_norm40, gpt_neox_layers_20_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add120: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv468, gpt_neox_layers_20_attention_query_key_value_bias1) + reshape80: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add120, R.shape([1, seq_len, 48, 128])) + reshape81: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape80, R.shape([seq_len, 48, 128])) + lv20 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(20), R.prim_value(T.float32(1.0)), reshape81), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape82: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv20, R.shape([1, seq_len, 16, 128])) + reshape83: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape82, R.shape([1, seq_len, 2048])) + lv469 = R.call_tir(cls.NT_matmul6, (reshape83, gpt_neox_layers_20_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add121: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv469, gpt_neox_layers_20_attention_dense_bias1) + layer_norm41: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add119, gpt_neox_layers_20_post_attention_layernorm_weight1, gpt_neox_layers_20_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv470 = R.call_tir(cls.NT_matmul7, (layer_norm41, gpt_neox_layers_20_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add122: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv470, gpt_neox_layers_20_mlp_dense_h_to_4h_bias1) + gelu20: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add122) + astype40: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu20, dtype="float16") + lv471 = R.call_tir(cls.NT_matmul8, (astype40, gpt_neox_layers_20_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add123: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv471, gpt_neox_layers_20_mlp_dense_4h_to_h_bias1) + astype41: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add123, dtype="float16") + add124: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype41, add121) + add125: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add124, add119) + layer_norm42: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add125, gpt_neox_layers_21_input_layernorm_weight1, gpt_neox_layers_21_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv472 = R.call_tir(cls.NT_matmul5, (layer_norm42, gpt_neox_layers_21_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add126: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv472, gpt_neox_layers_21_attention_query_key_value_bias1) + reshape84: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add126, R.shape([1, seq_len, 48, 128])) + reshape85: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape84, R.shape([seq_len, 48, 128])) + lv21 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(21), R.prim_value(T.float32(1.0)), reshape85), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape86: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv21, R.shape([1, seq_len, 16, 128])) + reshape87: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape86, R.shape([1, seq_len, 2048])) + lv473 = R.call_tir(cls.NT_matmul6, (reshape87, gpt_neox_layers_21_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add127: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv473, gpt_neox_layers_21_attention_dense_bias1) + layer_norm43: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add125, gpt_neox_layers_21_post_attention_layernorm_weight1, gpt_neox_layers_21_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv474 = R.call_tir(cls.NT_matmul7, (layer_norm43, gpt_neox_layers_21_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add128: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv474, gpt_neox_layers_21_mlp_dense_h_to_4h_bias1) + gelu21: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add128) + astype42: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu21, dtype="float16") + lv475 = R.call_tir(cls.NT_matmul8, (astype42, gpt_neox_layers_21_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add129: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv475, gpt_neox_layers_21_mlp_dense_4h_to_h_bias1) + astype43: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add129, dtype="float16") + add130: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype43, add127) + add131: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add130, add125) + layer_norm44: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add131, gpt_neox_layers_22_input_layernorm_weight1, gpt_neox_layers_22_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv476 = R.call_tir(cls.NT_matmul5, (layer_norm44, gpt_neox_layers_22_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add132: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv476, gpt_neox_layers_22_attention_query_key_value_bias1) + reshape88: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add132, R.shape([1, seq_len, 48, 128])) + reshape89: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape88, R.shape([seq_len, 48, 128])) + lv22 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(22), R.prim_value(T.float32(1.0)), reshape89), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape90: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv22, R.shape([1, seq_len, 16, 128])) + reshape91: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape90, R.shape([1, seq_len, 2048])) + lv477 = R.call_tir(cls.NT_matmul6, (reshape91, gpt_neox_layers_22_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add133: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv477, gpt_neox_layers_22_attention_dense_bias1) + layer_norm45: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add131, gpt_neox_layers_22_post_attention_layernorm_weight1, gpt_neox_layers_22_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv478 = R.call_tir(cls.NT_matmul7, (layer_norm45, gpt_neox_layers_22_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add134: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv478, gpt_neox_layers_22_mlp_dense_h_to_4h_bias1) + gelu22: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add134) + astype44: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu22, dtype="float16") + lv479 = R.call_tir(cls.NT_matmul8, (astype44, gpt_neox_layers_22_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add135: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv479, gpt_neox_layers_22_mlp_dense_4h_to_h_bias1) + astype45: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add135, dtype="float16") + add136: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype45, add133) + add137: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add136, add131) + layer_norm46: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add137, gpt_neox_layers_23_input_layernorm_weight1, gpt_neox_layers_23_input_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv480 = R.call_tir(cls.NT_matmul5, (layer_norm46, gpt_neox_layers_23_attention_query_key_value_weight1), out_sinfo=R.Tensor((1, seq_len, 6144), dtype="float16")) + add138: R.Tensor((1, seq_len, 6144), dtype="float16") = R.add(lv480, gpt_neox_layers_23_attention_query_key_value_bias1) + reshape92: R.Tensor((1, seq_len, 48, 128), dtype="float16") = R.reshape(add138, R.shape([1, seq_len, 48, 128])) + reshape93: R.Tensor((seq_len, 48, 128), dtype="float16") = R.reshape(reshape92, R.shape([seq_len, 48, 128])) + lv23 = R.call_dps_packed("vm.builtin.attention_kv_cache_attention_with_fused_qkv", (paged_kv_cache, R.prim_value(23), R.prim_value(T.float32(1.0)), reshape93), out_sinfo=R.Tensor((seq_len, 16, 128), dtype="float16")) + reshape94: R.Tensor((1, seq_len, 16, 128), dtype="float16") = R.reshape(lv23, R.shape([1, seq_len, 16, 128])) + reshape95: R.Tensor((1, seq_len, 2048), dtype="float16") = R.reshape(reshape94, R.shape([1, seq_len, 2048])) + lv481 = R.call_tir(cls.NT_matmul6, (reshape95, gpt_neox_layers_23_attention_dense_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float16")) + add139: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(lv481, gpt_neox_layers_23_attention_dense_bias1) + layer_norm47: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add137, gpt_neox_layers_23_post_attention_layernorm_weight1, gpt_neox_layers_23_post_attention_layernorm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv482 = R.call_tir(cls.NT_matmul7, (layer_norm47, gpt_neox_layers_23_mlp_dense_h_to_4h_weight1), out_sinfo=R.Tensor((1, seq_len, 8192), dtype="float32")) + add140: R.Tensor((1, seq_len, 8192), dtype="float32") = R.add(lv482, gpt_neox_layers_23_mlp_dense_h_to_4h_bias1) + gelu23: R.Tensor((1, seq_len, 8192), dtype="float32") = R.nn.gelu(add140) + astype46: R.Tensor((1, seq_len, 8192), dtype="float16") = R.astype(gelu23, dtype="float16") + lv483 = R.call_tir(cls.NT_matmul8, (astype46, gpt_neox_layers_23_mlp_dense_4h_to_h_weight1), out_sinfo=R.Tensor((1, seq_len, 2048), dtype="float32")) + add141: R.Tensor((1, seq_len, 2048), dtype="float32") = R.add(lv483, gpt_neox_layers_23_mlp_dense_4h_to_h_bias1) + astype47: R.Tensor((1, seq_len, 2048), dtype="float16") = R.astype(add141, dtype="float16") + add142: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(astype47, add139) + add143: R.Tensor((1, seq_len, 2048), dtype="float16") = R.add(add142, add137) + layer_norm48: R.Tensor((1, seq_len, 2048), dtype="float16") = R.nn.layer_norm(add143, gpt_neox_final_layer_norm_weight1, gpt_neox_final_layer_norm_bias1, axes=[-1], epsilon=1.0000000000000001e-05, center=True, scale=True) + lv24 = R.call_tir(cls.index, (layer_norm48,), out_sinfo=R.Tensor((1, 1, 2048), dtype="float16")) + lv484 = R.call_tir(cls.NT_matmul14, (lv24, embed_out_weight1), out_sinfo=R.Tensor((1, 1, vocab_size), dtype="float16")) + astype48: R.Tensor((1, 1, vocab_size), dtype="float32") = R.astype(lv484, dtype="float32") + gv1: R.Tuple(R.Tensor((1, 1, vocab_size), dtype="float32"), R.Object) = astype48, paged_kv_cache + R.output(gv1) + return gv1 + + @R.function + def softmax_with_temperature(logits: R.Tensor(("batch_size", 1, "vocab_size"), dtype="float32"), temperature: R.Tensor(("batch_size",), dtype="float32")) -> R.Tensor(("batch_size", 1, "vocab_size"), dtype="float32"): + batch_size = T.int64(is_size_var=True) + vocab_size = T.int64(is_size_var=True) + R.func_attr({"relax.memory_plan_dynamic_func_output": True, "tir_non_negative_var": ["vocab_size"], "tir_var_upper_bound": {"batch_size": 128, "seq_len": 2048, "total_seq_len": 2048}}) + cls = Module + with R.dataflow(): + lv: R.Tensor((batch_size, vocab_size), dtype="float32") = R.call_pure_packed("vm.builtin.reshape", logits, R.shape([batch_size, vocab_size]), sinfo_args=(R.Tensor((batch_size, vocab_size), dtype="float32"),)) + lv1 = R.call_tir(cls.chunk_lse, (lv, temperature), out_sinfo=[R.Tensor((batch_size, (vocab_size + 4096 - 1) // 4096), dtype="float32"), R.Tensor((batch_size, (vocab_size + 4096 - 1) // 4096), dtype="float32")]) + lv2: R.Tensor((batch_size, (vocab_size + 4096 - 1) // 4096), dtype="float32") = lv1[0] + lv3: R.Tensor((batch_size, (vocab_size + 4096 - 1) // 4096), dtype="float32") = lv1[1] + lv4 = R.call_tir(cls.softmax_with_chunked_sum, (lv, temperature, lv2, lv3), out_sinfo=R.Tensor((batch_size, vocab_size), dtype="float32")) + gv: R.Tensor((batch_size, 1, vocab_size), dtype="float32") = R.call_pure_packed("vm.builtin.reshape", lv4, R.shape([batch_size, 1, vocab_size]), sinfo_args=(R.Tensor((batch_size, 1, vocab_size), dtype="float32"),)) + R.output(gv) + return gv \ No newline at end of file