Merge dev-zyh into dev-hr-musa

Author: root

.github/scripts/hygon/setup.sh

@@ -15,19 +15,77 @@ git config --global --add safe.directory "$(pwd)"
 : "${DEVICE_LIB_PATH:?DEVICE_LIB_PATH is not set}"
 : "${LD_LIBRARY_PATH:?LD_LIBRARY_PATH is not set}"
 
+IMAGE_PLUGIN_ROOT="${VLLM_FL_IMAGE_PLUGIN_ROOT:-/opt/vllm-src/vllm-plugin-FL}"
+test -d "${IMAGE_PLUGIN_ROOT}/vllm_fl"
+
+# The Hygon CI image already contains the validated plugin commit. Keep the
+# checkout available for tests and configs, but load vllm_fl from the image.
+HYGON_SITE_DIR="${RUNNER_TEMP:-/tmp}/hygon-python-site"
+mkdir -p "${HYGON_SITE_DIR}"
+cat > "${HYGON_SITE_DIR}/sitecustomize.py" <<'PY'
+import importlib.abc
+import importlib.util
+import os
+import sys
+from pathlib import Path
+
+
+class _ImageVllmFLFinder(importlib.abc.MetaPathFinder):
+    def __init__(self, root):
+        self.package_dir = Path(root) / "vllm_fl"
+
+    def find_spec(self, fullname, path=None, target=None):
+        if fullname != "vllm_fl":
+            return None
+        init_file = self.package_dir / "__init__.py"
+        if not init_file.exists():
+            return None
+        return importlib.util.spec_from_file_location(
+            fullname,
+            init_file,
+            submodule_search_locations=[str(self.package_dir)],
+        )
+
+
+_root = os.environ.get("VLLM_FL_IMAGE_PLUGIN_ROOT")
+if _root:
+    sys.meta_path.insert(0, _ImageVllmFLFinder(_root))
+PY
+
+export VLLM_FL_IMAGE_PLUGIN_ROOT="${IMAGE_PLUGIN_ROOT}"
+export PYTHONPATH="${HYGON_SITE_DIR}${PYTHONPATH:+:${PYTHONPATH}}"
+
+if [[ -n "${GITHUB_ENV:-}" ]]; then
+    {
+        echo "VLLM_FL_IMAGE_PLUGIN_ROOT=${VLLM_FL_IMAGE_PLUGIN_ROOT}"
+        echo "PYTHONPATH=${PYTHONPATH}"
+    } >> "${GITHUB_ENV}"
+fi
+
 echo "DTK_HOME=${DTK_HOME}"
 echo "LD_LIBRARY_PATH=${LD_LIBRARY_PATH}"
+echo "VLLM_FL_IMAGE_PLUGIN_ROOT=${VLLM_FL_IMAGE_PLUGIN_ROOT}"
+echo "PYTHONPATH=${PYTHONPATH}"
 test -e "${HIP_PATH}/lib/libgalaxyhip.so.5"
 test -e "${DTK_HOME}/llvm/lib/libomp.so"
 
-python -m pip install --no-build-isolation --no-deps -e .
-
 python - <<'PY'
+import os
+from pathlib import Path
+
 import flag_gems
 import torch
 import vllm
 import vllm_fl
 
+image_plugin_root = Path(os.environ["VLLM_FL_IMAGE_PLUGIN_ROOT"]).resolve()
+expected_package = image_plugin_root / "vllm_fl"
+plugin_file = Path(vllm_fl.__file__).resolve()
+if plugin_file != expected_package / "__init__.py" and expected_package not in plugin_file.parents:
+    raise RuntimeError(
+        f"Unexpected vllm_fl path: {plugin_file}; expected under {expected_package}"
+    )
+
 print(f"vLLM import ok: {vllm.__version__}")
 print(f"vLLM-FL import ok: {vllm_fl.__file__}")
 print(f"FlagGems import ok: {getattr(flag_gems, '__version__', 'unknown')}")

.github/workflows/_benchmark_test.yml

@@ -68,47 +68,17 @@ jobs:
         run: bash .github/scripts/${{ inputs.platform }}/setup.sh
 
       - name: Run benchmark smoke tests
-        id: benchmark_test
         run: |
           python tests/run.py \
             --platform ${{ inputs.platform }} \
             --scope benchmark
 
-      - name: Generate benchmark report
-        if: always() && steps.benchmark_test.outcome == 'success'
-        run: |
-          python tests/benchmarks/format_serve_report.py \
-            --input benchmark-results/serve_result.json \
-            --output benchmark-results/benchmark_metrics.json
-
-      - name: Install jq
-        if: always() && steps.benchmark_test.outcome == 'success'
-        run: apt-get update && apt-get install -y jq
-
-      - name: Upload benchmark data to backend
-        continue-on-error: true
-        if: always() && steps.benchmark_test.outcome == 'success'
-        uses: flagos-ai/FlagOps/actions/post-benchmark-report@main
-        env:
-          NO_PROXY: "flagcicd-inner.flagos.net"
-        with:
-          backend_url: 'http://flagcicd-inner.flagos.net:8000'
-          user_id: '000000000000000000'
-          report_path: benchmark-results/benchmark_metrics.json
-          list_code: 'benchmark_${{ inputs.platform }}_vllm_serve_smoke'
-          list_name: 'Benchmark-vLLM-Serve-Smoke-${{ inputs.platform }}'
-          header_config: '[{"field":"metric","name":"Benchmark Metric","required":true,"sortable":true,"type":"string"},{"field":"values","name":"Per Step Details","required":false,"sortable":false,"type":"array"},{"field":"avg","name":"Avg","required":true,"sortable":true,"type":"number"},{"field":"p50","name":"P50","required":true,"sortable":true,"type":"number"},{"field":"p99","name":"P99","required":true,"sortable":true,"type":"number"}]'
-          repository_name: '${{ github.repository }}'
-          commit_id: ${{ github.event.pull_request.head.sha || github.sha }}
-          workflow_id: ${{ github.run_id }}
-          fail_on_error: 'false'
-
       - name: Upload test results
         if: always()
         uses: actions/upload-artifact@v4
+        continue-on-error: true
         with:
           name: benchmark-${{ inputs.platform }}
           path: |
             test-results-${{ inputs.platform }}.xml
             test-results-${{ inputs.platform }}.json
-            benchmark-results/**

.github/workflows/_build_wheel.yml

@@ -63,6 +63,7 @@ jobs:
       - name: Upload wheel
         if: inputs.upload-artifact
         uses: actions/upload-artifact@v4
+        continue-on-error: true
         with:
           name: wheel
           path: dist/*.whl

.github/workflows/_functional_test.yml

@@ -76,6 +76,7 @@ jobs:
       - name: Upload test results
         if: always()
         uses: actions/upload-artifact@v4
+        continue-on-error: true
         with:
           name: functional-${{ inputs.platform }}
           path: |

README.md

@@ -138,7 +138,7 @@ import torch
 from vllm.config.compilation import CompilationConfig
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     prompts = [
         "Hello, my name is",
     ]

docker/hygon/Dockerfile

@@ -43,7 +43,7 @@ COPY .hygon-runtime/ /
 RUN python -m pip install --upgrade \
         ${INDEX_URL:+--index-url ${INDEX_URL}} \
         ${EXTRA_INDEX_URL:+--extra-index-url ${EXTRA_INDEX_URL}} \
-        pip setuptools wheel setuptools-scm
+        pip "setuptools<81" wheel setuptools-scm
 
 # Replace the base image's vLLM with official vLLM built in empty mode.
 RUN python -m pip uninstall -y vllm || true \
@@ -95,7 +95,10 @@ RUN rm -rf /opt/vllm-src/FlagGems \
     && git clone https://github.com/flagos-ai/FlagGems /opt/vllm-src/FlagGems \
     && cd /opt/vllm-src/FlagGems \
     && git checkout "${FLAGGEMS_VERSION}" \
-    && python -m pip install --no-build-isolation -e .
+    && python -m pip install \
+        ${INDEX_URL:+--index-url ${INDEX_URL}} \
+        ${EXTRA_INDEX_URL:+--extra-index-url ${EXTRA_INDEX_URL}} \
+        --no-build-isolation -e .
 
 # Install vllm-plugin-FL from the pinned source commit.
 RUN rm -rf /opt/vllm-src/vllm-plugin-FL \
@@ -104,7 +107,10 @@ RUN rm -rf /opt/vllm-src/vllm-plugin-FL \
     && cd /opt/vllm-src/vllm-plugin-FL \
     && git checkout "${VLLM_PLUGIN_FL_VERSION}" \
     && CMAKE_ARGS="-DCMAKE_PREFIX_PATH=/opt/dtk;/opt/dtk/hip;/opt/dtk/hsa;/opt/hyhal;/usr/local/hyhal ${CMAKE_ARGS:-}" \
-        VLLM_VENDOR=cuda python -m pip install --no-build-isolation -e .
+        VLLM_VENDOR=cuda python -m pip install \
+        ${INDEX_URL:+--index-url ${INDEX_URL}} \
+        ${EXTRA_INDEX_URL:+--extra-index-url ${EXTRA_INDEX_URL}} \
+        --no-build-isolation -e .
 
 WORKDIR /workspace
 

tests/README.md

@@ -303,6 +303,7 @@ Unit tests should be fast, isolated, and not require GPU or model weights. Use t
 import pytest
 import torch
 
+
 class TestMyOp:
     def test_basic(self, device):
         """The `device` fixture returns cuda:0 or npu:0 automatically."""
@@ -322,6 +323,7 @@ import torch
 
 pytestmark = pytest.mark.gpu
 
+
 def test_my_kernel(device):
     x = torch.randn(16, 256, device=device)
     result = my_kernel(x)

tests/e2e_tests/serving/test_bge_m3.py

@@ -17,10 +17,14 @@
 
 SENTENCES_1 = ["What is BGE M3?", "Defination of BM25"]
 SENTENCES_2 = [
-    "BGE M3 is an embedding model supporting dense retrieval, "
-    "lexical matching and multi-vector interaction.",
-    "BM25 is a bag-of-words retrieval function that ranks a set "
-    "of documents based on the query terms appearing in each document",
+    (
+        "BGE M3 is an embedding model supporting dense retrieval, "
+        "lexical matching and multi-vector interaction."
+    ),
+    (
+        "BM25 is a bag-of-words retrieval function that ranks a set "
+        "of documents based on the query terms appearing in each document"
+    ),
 ]
 
 SIMILARITY_REFERENCE = [[0.6265, 0.3477], [0.3499, 0.678]]

tests/models/qwen3_6/27b_tp2_eager.yaml

@@ -1,32 +1,43 @@
-# Qwen3.6-27B Hygon smoke configuration (TP=2, eager).
+# Copyright 2026 FlagOS Contributors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Qwen3.6-27B smoke configuration (TP=2, eager).
 
 llm:
   model: "/workspace/Qwen3.6-27B"
   tensor_parallel_size: 2
   pipeline_parallel_size: 1
-  max_model_len: 262144
-  gpu_memory_utilization: 0.90
+  max_model_len: 8192
+  gpu_memory_utilization: 0.95
   enforce_eager: true
-  enable_log_requests: false
-  enable_prefix_caching: false
   trust_remote_code: false
   disable_custom_all_reduce: true
 
 generate:
   prompts:
-    - "介绍一下北京"
+    - "Introduce yourself,please"
   sampling:
     max_tokens: 100
     temperature: 0.0
 
 serve:
   served_model_name: "qwen"
-  startup_retries: 120
   endpoints: ["chat"]
   stream: false
   max_tokens: 256
   chat_messages:
     - role: "user"
-      content: "介绍一下北京"
+      content: "Introduce yourself,please"
   sampling:
-    temperature: 0.7
+    temperature: 0.0

tests/models/qwen3_6/27b_tp2_eager_hygon.yaml

@@ -0,0 +1,49 @@
+# Copyright 2026 FlagOS Contributors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Qwen3.6-27B Hygon smoke configuration (TP=2, eager).
+
+llm:
+  model: "/workspace/Qwen3.6-27B"
+  tensor_parallel_size: 2
+  pipeline_parallel_size: 1
+  max_model_len: 262144
+  gpu_memory_utilization: 0.90
+  enforce_eager: true
+  trust_remote_code: false
+  disable_custom_all_reduce: true
+
+generate:
+  prompts:
+    - "介绍一下北京"
+  sampling:
+    max_tokens: 100
+    temperature: 0.0
+
+serve:
+  served_model_name: "qwen"
+  extra_engine:
+    enforce_eager: false
+    disable_custom_all_reduce: false
+    enable_log_requests: false
+    enable_prefix_caching: false
+  startup_retries: 120
+  endpoints: ["chat"]
+  stream: false
+  max_tokens: 256
+  chat_messages:
+    - role: "user"
+      content: "介绍一下北京"
+  sampling:
+    temperature: 0.7