Star181
852
代码
代码
Issues161
Pull Requests325
流水线
项目讨论
Wiki
分析
Star181
852
已合并
[fix][ops]MatmulReduceScatter算子修复string_view悬空指针 #5383
chenjl创建于 7月6日
[fix][ops]MatmulReduceScatter算子修复string_view悬空指针 #5383
已合并
chenjl创建于 7月6日
从chenjiale56/op-plugin26.1.0合入到Ascend/op-plugin26.1.0
共 2 个文件变更+13-7
Mop_plugin/ops/opapi/MatmulReduceScatterBaseKernelNpuOpApi.cpp+7-4
@@ -32,8 +32,11 @@ at::Tensor npu_mm_reduce_scatter_base(const at::Tensor & self, const at::Tensor
32 self.size(1), " and the K-axis of x2 is ", x2.size(0), "." + OPS_ERROR(ErrCode::PARAM));32 self.size(1), " and the K-axis of x2 is ", x2.size(0), "." + OPS_ERROR(ErrCode::PARAM));
33 TORCH_CHECK(self.size(0) % world_size == 0, "The M-axis in input of Matmul should be be divisible by world_size."33 TORCH_CHECK(self.size(0) % world_size == 0, "The M-axis in input of Matmul should be be divisible by world_size."
34 + OPS_ERROR(ErrCode::PARAM));34 + OPS_ERROR(ErrCode::PARAM));
35+ std::string hcom_str = std::string(hcom);
36+ std::string reduce_op_str = std::string(reduce_op);
35 bool isSocBelowAscend950 = (c10_npu::GetSocVersion() < c10_npu::SocVersion::Ascend950);37 bool isSocBelowAscend950 = (c10_npu::GetSocVersion() < c10_npu::SocVersion::Ascend950);
36 c10::string_view comm_mode_value = comm_mode.value_or("ai_cpu");38 c10::string_view comm_mode_value = comm_mode.value_or("ai_cpu");
39+ std::string comm_mode_str = std::string(comm_mode_value);
37 auto output_size = {self.size(0) / world_size, x2.size(1)};40 auto output_size = {self.size(0) / world_size, x2.size(1)};
38 auto result_dtype = self.scalar_type();41 auto result_dtype = self.scalar_type();
39 bool has_quant = x2_scale.has_value();42 bool has_quant = x2_scale.has_value();
@@ -41,9 +44,9 @@ at::Tensor npu_mm_reduce_scatter_base(const at::Tensor & self, const at::Tensor
41 result_dtype = x2_scale.value().scalar_type() == at::kLong ? at::kHalf: output_dtype.value_or(at::kBFloat16);44 result_dtype = x2_scale.value().scalar_type() == at::kLong ? at::kHalf: output_dtype.value_or(at::kBFloat16);
42 }45 }
43 auto result = at_npu::native::OpPreparation::apply_tensor_without_format(output_size, self.options().dtype(result_dtype));46 auto result = at_npu::native::OpPreparation::apply_tensor_without_format(output_size, self.options().dtype(result_dtype));
44- char *reduce_op_ptr = const_cast<char *>(reduce_op.data());47+ char *reduce_op_ptr = const_cast<char*>(reduce_op_str.data());
45- char *hcom_ptr = const_cast<char *>(hcom.data());48+ char *hcom_ptr = const_cast<char*>(hcom_str.data());
46- char *comm_mode_ptr = const_cast<char *>(comm_mode_value.data());49+ char *comm_mode_ptr = const_cast<char*>(comm_mode_str.data());
47 const at::Tensor &bias_real = bias.value_or(at::Tensor());50 const at::Tensor &bias_real = bias.value_or(at::Tensor());
48 int64_t stream_mode = ACL_STOP_ON_FAILURE;51 int64_t stream_mode = ACL_STOP_ON_FAILURE;
49 int64_t block_size = 0;52 int64_t block_size = 0;
@@ -51,7 +54,7 @@ at::Tensor npu_mm_reduce_scatter_base(const at::Tensor & self, const at::Tensor
51 at::Tensor quant_scale;54 at::Tensor quant_scale;
52 at::Tensor amax_out;55 at::Tensor amax_out;
53 if (isSocBelowAscend950) {56 if (isSocBelowAscend950) {
54- if (comm_mode_value == "ai_cpu") {57+ if (comm_mode_str == "ai_cpu") {
55 TORCH_CHECK(!has_quant, "When comm_mode is ai_cpu, quantization not supported." + OPS_ERROR(ErrCode::PARAM));58 TORCH_CHECK(!has_quant, "When comm_mode is ai_cpu, quantization not supported." + OPS_ERROR(ErrCode::PARAM));
56 EXEC_NPU_CMD(aclnnMatmulReduceScatter, self, x2, bias_real, hcom_ptr, reduce_op_ptr, comm_turn, stream_mode, result);59 EXEC_NPU_CMD(aclnnMatmulReduceScatter, self, x2, bias_real, hcom_ptr, reduce_op_ptr, comm_turn, stream_mode, result);
57 } else {60 } else {
Mop_plugin/ops/opapi/QuantMatmulReduceScatterKernelNpuOpApi.cpp+6-3
@@ -45,6 +45,9 @@ std::tuple<at::Tensor, at::Tensor> npu_quant_mm_reduce_scatter(
45 bool isSocBelowAscend950 = (c10_npu::GetSocVersion() < c10_npu::SocVersion::Ascend950);45 bool isSocBelowAscend950 = (c10_npu::GetSocVersion() < c10_npu::SocVersion::Ascend950);
46 c10::string_view default_comm_mode = isSocBelowAscend950 ? "aiv" : "ai_cpu";46 c10::string_view default_comm_mode = isSocBelowAscend950 ? "aiv" : "ai_cpu";
47 c10::string_view comm_mode_value = comm_mode.value_or(default_comm_mode);47 c10::string_view comm_mode_value = comm_mode.value_or(default_comm_mode);
48+ std::string hcom_str = std::string(hcom);
49+ std::string reduce_op_str = std::string(reduce_op);
50+ std::string comm_mode_str = std::string(comm_mode_value);
48 at::IntArrayRef group_size_list = group_sizes.value_or(at::IntArrayRef{});51 at::IntArrayRef group_size_list = group_sizes.value_or(at::IntArrayRef{});
49 int64_t group_size = op_plugin::utils::check_and_get_group_size(group_size_list);52 int64_t group_size = op_plugin::utils::check_and_get_group_size(group_size_list);
50 TORCH_CHECK(group_size != -1, "Invalid group_sizes.", OPS_ERROR(ErrCode::PARAM));53 TORCH_CHECK(group_size != -1, "Invalid group_sizes.", OPS_ERROR(ErrCode::PARAM));
@@ -69,9 +72,9 @@ std::tuple<at::Tensor, at::Tensor> npu_quant_mm_reduce_scatter(
69 }72 }
70 c10::TensorOptions options = self.options().dtype(output_scalar_type);73 c10::TensorOptions options = self.options().dtype(output_scalar_type);
71 auto result = npu_preparation::apply_tensor_without_format(output_size, options);74 auto result = npu_preparation::apply_tensor_without_format(output_size, options);
72- char* reduce_op_ptr = const_cast<char*>(reduce_op.data());75+ char *reduce_op_ptr = const_cast<char*>(reduce_op_str.data());
73- char* hcom_ptr = const_cast<char*>(hcom.data());76+ char *hcom_ptr = const_cast<char*>(hcom_str.data());
74- char *comm_mode_ptr = const_cast<char *>(comm_mode_value.data());77+ char *comm_mode_ptr = const_cast<char*>(comm_mode_str.data());
75 const at::Tensor& bias_real = bias.value_or(at::Tensor());78 const at::Tensor& bias_real = bias.value_or(at::Tensor());
76 const at::Tensor& quant_scale_real = quant_scale.value_or(at::Tensor());79 const at::Tensor& quant_scale_real = quant_scale.value_or(at::Tensor());
77 int64_t stream_mode = ACL_STOP_ON_FAILURE;80 int64_t stream_mode = ACL_STOP_ON_FAILURE;