Skip to content

feat(npu): support CANN 9.1 MegaMoe for Qwen3.5 #1989

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

Draft
pjgao wants to merge 7 commits into
base: main
Choose a base branch
from
Draft

feat(npu): support CANN 9.1 MegaMoe for Qwen3.5 #1989

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
43d3b5c
feat(npu): add CANN 9.1 MegaMoe adapter for Qwen3.5
pjgao Jul 19, 2026
d4788d7
fix(npu): validate MegaMoe remote buffer spans
pjgao Jul 19, 2026
a14a79d
fix(npu): allow MegaMoe with attention data parallelism
pjgao Jul 20, 2026
0b50240
fix(npu): stabilize MegaMoe EP8 execution
pjgao Jul 21, 2026
024e1de
test(npu): avoid environment-specific paths
pjgao Jul 21, 2026
0a838b8
build: migrate NPU submodules to GitHub
pjgao Jul 21, 2026
8aeef08
build: drop unnecessary NPU submodule dependencies
pjgao Jul 23, 2026
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
4 changes: 2 additions & 2 deletions .gitmodules
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,7 +20,7 @@
fetchRecurseSubmodules = false
[submodule "third_party/xllm_ops"]
path = third_party/xllm_ops
url = https://gitcode.com/xLLM-AI/xllm_ops.git
url = https://github.com/xLLM-AI/xllm-ops.git
fetchRecurseSubmodules = true
[submodule "third_party/etcd_cpp_apiv3"]
path = third_party/etcd_cpp_apiv3
Expand All @@ -36,7 +36,7 @@
fetchRecurseSubmodules = false
[submodule "third_party/torch_npu_ops"]
path = third_party/torch_npu_ops
url = https://gitcode.com/xLLM-AI/torch_npu_ops.git
url = https://github.com/xLLM-AI/torch-npu-ops.git
fetchRecurseSubmodules = false
[submodule "third_party/cutlass"]
path = third_party/cutlass
Expand Down
34 changes: 34 additions & 0 deletions tests/core/framework/config/config_json_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@ limitations under the License.

#include <gtest/gtest.h>

#include <algorithm>
#include <filesystem>
#include <fstream>
#include <nlohmann/json.hpp>
Expand All @@ -24,6 +25,7 @@ limitations under the License.
#include "core/common/global_flags.h"
#include "core/framework/config/config_utils.h"
#include "core/framework/config/execution_config.h"
#include "core/framework/config/kernel_config.h"
#include "core/framework/config/kv_cache_config.h"
#include "core/framework/config/model_config.h"
#include "core/framework/config/parallel_config.h"
Expand Down Expand Up @@ -236,6 +238,38 @@ TEST(ConfigJsonTest, FromJsonUsesParsedOverrides) {
EXPECT_EQ(scheduler_config.max_decode_token_per_sequence(), 256);
}

#if defined(USE_NPU)
TEST(KernelConfigTest, MegaMoeDefaultsOffAndReadsJsonOverride) {
KernelConfig kernel_config;
EXPECT_EQ(kernel_config.mega_moe_mode(), "off");
EXPECT_EQ(kernel_config.mega_moe_weight_cache_budget_bytes(),
4LL * 1024 * 1024 * 1024);

const JsonReader json = config::parse_json_string(
R"json({"mega_moe_mode": "on", "mega_moe_weight_cache_budget_bytes": 123456})json");
kernel_config.from_json(json);
EXPECT_EQ(kernel_config.mega_moe_mode(), "on");
EXPECT_EQ(kernel_config.mega_moe_weight_cache_budget_bytes(), 123456);

nlohmann::ordered_json dumped_config;
kernel_config.append_config_json(dumped_config);
ASSERT_TRUE(dumped_config.contains("mega_moe_mode"));
EXPECT_EQ(dumped_config.at("mega_moe_mode").get<std::string>(), "on");
ASSERT_TRUE(
dumped_config.contains("mega_moe_weight_cache_budget_bytes"));
EXPECT_EQ(dumped_config.at("mega_moe_weight_cache_budget_bytes")
.get<int64_t>(),
123456);

const auto& names = KernelConfig::option_category().option_names;
EXPECT_NE(std::find(names.begin(), names.end(), "mega_moe_mode"),
names.end());
EXPECT_NE(std::find(names.begin(), names.end(),
"mega_moe_weight_cache_budget_bytes"),
names.end());
}
#endif

TEST(KVCacheConfigValidationTest, AcceptsSupportedIndexerCacheDtypes) {
KVCacheConfig config;

Expand Down
13 changes: 13 additions & 0 deletions tests/core/framework/parallel_state/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,6 +37,19 @@ if(USE_NPU)
hccl
c_sec
nnopbase)

cc_test(
NAME
mega_moe_comm_resource_test
SRCS
mega_moe_comm_resource_test.cpp
DEPS
mega_moe_comm_resource
torch
GTest::gtest_main
hccl
hcomm
)
endif()

if(USE_MLU)
Expand Down
Loading