Skip to content

feat: support parsing dtype field from config. #1118

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
DongheJin wants to merge 1 commit into
base: main
Choose a base branch
from
+65 −6
Open

feat: support parsing dtype field from config. #1118

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
28 changes: 22 additions & 6 deletions xllm/core/framework/hf_model_loader.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,17 @@ namespace xllm {

namespace {

JsonReader normalize_config_torch_dtype(const JsonReader& reader) {
auto config = reader.data();
if (!config.contains("torch_dtype") && config.contains("dtype")) {
config["torch_dtype"] = config["dtype"];
}

JsonReader normalized_reader;
normalized_reader.parse_text(config.dump());
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Mar 26, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

The current implementation of normalize_config_torch_dtype creates a new JsonReader by first serializing the nlohmann::json object to a string (config.dump()) and then deserializing it (normalized_reader.parse_text()). This roundtrip serialization and deserialization is inefficient and introduces unnecessary overhead. It would be more performant if the JsonReader class provided a constructor that directly accepts an nlohmann::json object, allowing for direct construction without the intermediate string conversion.

return normalized_reader;
}

bool is_compressed_tensors_fp8_scheme(const nlohmann::json& config) {
auto type_it = config.find("type");
auto num_bits_it = config.find("num_bits");
Expand Down Expand Up @@ -725,7 +736,8 @@ bool HFModelLoader::load_model_args(const std::string& model_weights_path) {
<< model_type;
return false;
}
model_args_loader(reader, &args_);
const JsonReader config_reader = normalize_config_torch_dtype(reader);
model_args_loader(config_reader, &args_);

return true;
}
Expand All @@ -738,16 +750,20 @@ bool HFModelLoader::load_quant_args(const std::string& model_weights_path) {
return false;
}

if (!load_quant_cfg(reader, quant_args_)) {
const JsonReader config_reader = normalize_config_torch_dtype(reader);

if (!load_quant_cfg(config_reader, quant_args_)) {
return false;
}

// load quantization args for npu if exists
if (reader.contains("quantize")) {
quant_args_.quantize_type() = reader.value_or<std::string>("quantize", "");
if (config_reader.contains("quantize")) {
quant_args_.quantize_type() =
config_reader.value_or<std::string>("quantize", "");
}
if (reader.contains("torch_dtype")) {
quant_args_.torch_dtype() = reader.value_or<std::string>("torch_dtype", "");
if (config_reader.contains("torch_dtype")) {
quant_args_.torch_dtype() =
config_reader.value_or<std::string>("torch_dtype", "");
}

// awq quantization args
Expand Down
43 changes: 43 additions & 0 deletions xllm/core/framework/hf_model_loader_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,9 @@ limitations under the License.

#include <gtest/gtest.h>

#include <filesystem>
#include <fstream>

#include "core/platform/device.h"

namespace xllm {
Expand Down Expand Up @@ -71,4 +74,44 @@ TEST(HFModelLoaderTest, KeepLegacyFp8ConfigUnchanged) {
EXPECT_FALSE(quant_args.activation_dynamic());
}

class HFModelLoaderConfigTest : public ::testing::Test {
protected:
void SetUp() override {
test_dir_ = std::filesystem::temp_directory_path() / "hf_model_loader_test";
if (std::filesystem::exists(test_dir_)) {
std::filesystem::remove_all(test_dir_);
}
std::filesystem::create_directories(test_dir_);
}

void TearDown() override {
if (std::filesystem::exists(test_dir_)) {
std::filesystem::remove_all(test_dir_);
}
}

void write_file(const std::filesystem::path& path, const std::string& data) {
std::ofstream file(path);
ASSERT_TRUE(file.is_open());
file << data;
file.close();
}

std::filesystem::path test_dir_;
};

TEST_F(HFModelLoaderConfigTest, TreatsTopLevelDtypeAsTorchDtype) {
write_file(test_dir_ / "config.json", R"json(
{
"model_type": "qwen3",
"dtype": "bfloat16"
}
)json");
write_file(test_dir_ / "model.safetensors", "");

HFModelLoader loader(test_dir_.string());
EXPECT_EQ(loader.model_args().dtype(), "bfloat16");
EXPECT_EQ(loader.quant_args().torch_dtype(), "bfloat16");
}

} // namespace xllm
Loading