Merge branch 'main' into support_kunlunxin_backend_2

Author: luoyc123

.github/configs/enflame.yml

@@ -0,0 +1,37 @@
+# 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.
+
+# Enflame S60 hardware configuration.
+platform: enflame
+
+ci_image: harbor.baai.ac.cn/flagos-dev/vllm-plugin-fl:v0.20.2-enflame-ci
+
+# Replace this with the exact label configured on the Enflame S60 runner.
+runner_labels:
+  - enflame-s60
+
+# Model files are persistent host data and use the same path in the container.
+container_volumes:
+  - /data:/data
+
+container_options: >-
+  --hostname vllm-plugin-fl
+  --ipc=host
+  --privileged
+  --env VLLM_PLUGINS=fl
+  --env TOPS_VISIBLE_DEVICES=2,3
+  --env VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=99999
+  --env TORCH_GCU_ENABLE_INT64_AND_UINT64=1
+  --env ENABLE_I64_CHECK=0
+  --env TORCHGCU_INDUCTOR_ENABLE=1

.github/configs/platforms.yml

@@ -23,10 +23,12 @@ platforms:
   cuda:
     enabled: true
   ascend:
-    enabled: false
+    enabled: true
   hygon:
     enabled: true
   metax:
     enabled: true
   musa:
     enabled: true
+  enflame:
+    enabled: true

.github/scripts/enflame/check.sh

@@ -0,0 +1,27 @@
+#!/bin/bash
+# Copyright (c) 2025 BAAI. All rights reserved.
+# Check Enflame S60 availability.
+set -euo pipefail
+
+echo "Current time: $(date '+%Y-%m-%d %H:%M:%S')"
+echo "=== Checking Enflame S60 availability ==="
+
+if ! command -v efsmi >/dev/null 2>&1; then
+  echo "::error::efsmi is not available in the CI container."
+  exit 1
+fi
+
+efsmi
+
+python - <<'PY'
+import torch
+import torch_gcu  # noqa: F401
+
+if not hasattr(torch, "gcu") or not torch.gcu.is_available():
+    raise RuntimeError("Enflame GCU is not available")
+
+count = torch.gcu.device_count()
+print(f"Enflame GCU count: {count}")
+if count < 2:
+    raise RuntimeError(f"At least 2 GCUs are required, found {count}")
+PY

.github/scripts/enflame/setup.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+# Copyright (c) 2025 BAAI. All rights reserved.
+# Setup script for Enflame S60 CI.
+set -euo pipefail
+
+: "${VLLM_PLUGINS:?VLLM_PLUGINS is not set}"
+: "${TOPS_VISIBLE_DEVICES:?TOPS_VISIBLE_DEVICES is not set}"
+
+git config --global --add safe.directory "$(pwd)"
+
+if [[ -n "${GITHUB_ENV:-}" ]]; then
+  for name in \
+    VLLM_PLUGINS \
+    TOPS_VISIBLE_DEVICES \
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS \
+    TORCH_GCU_ENABLE_INT64_AND_UINT64 \
+    ENABLE_I64_CHECK \
+    TORCHGCU_INDUCTOR_ENABLE; do
+    echo "${name}=${!name}" >> "${GITHUB_ENV}"
+  done
+fi
+
+# The vendor runtime, vLLM, and FlagGems come from the pinned base image.
+python -m pip install --no-build-isolation --no-deps -e .
+
+python - <<'PY'
+import flag_gems
+import torch
+import torch_gcu  # noqa: F401
+import vllm
+import vllm_fl
+
+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')}")
+print(f"FlagGems vendor: {getattr(flag_gems, 'vendor_name', 'auto-detected')}")
+print(f"Torch import ok: {torch.__version__}")
+print(f"Accelerator available: {torch.gcu.is_available()}")
+print(f"Accelerator count: {torch.gcu.device_count()}")
+PY

.github/scripts/generate_matrix.py

@@ -30,6 +30,7 @@
 
 REPO_ROOT = Path(__file__).resolve().parents[2]
 PLATFORMS_DIR = REPO_ROOT / "tests" / "platforms"
+sys.path.insert(0, str(REPO_ROOT))
 
 # YAML task key → test directory name
 TASK_DIR_MAP: dict[str, str] = {
@@ -130,6 +131,70 @@ def build_e2e_matrix(
     return entries
 
 
+def check_model_paths(
+    platform: str,
+    device: str | None = None,
+    cases: str | None = None,
+) -> int:
+    """Validate that the selected CI matrix model directories exist."""
+    from tests.utils.model_config import ModelConfig
+
+    if not cases:
+        print("[check-models] No E2E model cases selected.")
+        return 0
+
+    selected_cases = json.loads(cases)
+    if not isinstance(selected_cases, list):
+        print(
+            "::error title=Invalid cases::--cases must be a JSON array", file=sys.stderr
+        )
+        return 1
+
+    missing = 0
+    print(f"[check-models] Platform: {platform}")
+    print(f"[check-models] Device:   {device or ''}")
+    print(f"[check-models] Cases:    {len(selected_cases)}")
+
+    for case_entry in selected_cases:
+        model = str(case_entry["model"])
+        case = str(case_entry["case"])
+        try:
+            model_cfg = ModelConfig.load(
+                model,
+                case,
+                platform=platform,
+                device=device or None,
+            )
+        except Exception as exc:
+            message = f"{model}/{case}: {exc}"
+            print(
+                f"::error title=Invalid model config::{message}",
+                file=sys.stderr,
+            )
+            missing += 1
+            continue
+
+        model_path = model_cfg.model
+        exists = bool(model_path) and Path(model_path).exists()
+        status = "OK" if exists else "MISSING"
+        print(f"[check-models] {status}: {model}/{case} -> {model_path}")
+        if not exists:
+            message = f"{model}/{case} requires model path: {model_path}"
+            print(
+                f"::error file=tests/models/{model}/{case}.yaml,"
+                f"title=Missing test model::{message}",
+                file=sys.stderr,
+            )
+            missing += 1
+
+    if missing:
+        print(f"[check-models] Missing or invalid model cases: {missing}")
+        return 1
+
+    print("[check-models] All selected model paths exist.")
+    return 0
+
+
 # ---------------------------------------------------------------------------
 # Changed-file filtering for PR-only smart skip
 # ---------------------------------------------------------------------------
@@ -244,8 +309,35 @@ def main(argv: list[str] | None = None) -> int:
         "When provided and non-empty, the e2e matrix is filtered to only "
         "include tests affected by those changes (PR smart-skip).",
     )
+    parser.add_argument(
+        "--check-models",
+        action="store_true",
+        help=(
+            "Validate that selected E2E model paths exist instead of "
+            "generating matrices."
+        ),
+    )
+    parser.add_argument(
+        "--device",
+        default=None,
+        help="Device name to check when --check-models is set.",
+    )
+    parser.add_argument(
+        "--cases",
+        default=None,
+        help=(
+            "JSON array of {model, case} entries to check when --check-models is set."
+        ),
+    )
     args = parser.parse_args(argv)
 
+    if args.check_models:
+        return check_model_paths(
+            args.platform,
+            device=args.device,
+            cases=args.cases,
+        )
+
     config = load_platform(args.platform)
     devices = get_device_sections(config)
     unsupported = config.get("unsupported_features", [])

.github/scripts/hygon/setup.sh

@@ -15,92 +15,22 @@ 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}"
 
-USE_IMAGE_PLUGIN="${HYGON_USE_IMAGE_PLUGIN:-1}"
-
-if [[ "${USE_IMAGE_PLUGIN}" == "1" ]]; then
-    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}"
-        } >> "${GITHUB_ENV}"
-    fi
-else
-    unset VLLM_FL_IMAGE_PLUGIN_ROOT
-fi
-
-export HYGON_USE_IMAGE_PLUGIN="${USE_IMAGE_PLUGIN}"
+unset VLLM_FL_IMAGE_PLUGIN_ROOT
+unset HYGON_USE_IMAGE_PLUGIN
 
 echo "DTK_HOME=${DTK_HOME}"
 echo "LD_LIBRARY_PATH=${LD_LIBRARY_PATH}"
-echo "HYGON_USE_IMAGE_PLUGIN=${HYGON_USE_IMAGE_PLUGIN}"
-if [[ "${USE_IMAGE_PLUGIN}" == "1" ]]; then
-    echo "VLLM_FL_IMAGE_PLUGIN_ROOT=${VLLM_FL_IMAGE_PLUGIN_ROOT}"
-    echo "PYTHONPATH=${PYTHONPATH}"
-fi
 test -e "${HIP_PATH}/lib/libgalaxyhip.so.5"
 test -e "${DTK_HOME}/llvm/lib/libomp.so"
 
-python - <<'PY'
-import os
-from pathlib import Path
+python -m pip install --no-build-isolation --no-deps -e .
 
+python - <<'PY'
 import flag_gems
 import torch
 import vllm
 import vllm_fl
 
-if os.environ.get("HYGON_USE_IMAGE_PLUGIN", "1") == "1":
-    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/_e2e_test.yml

@@ -93,16 +93,17 @@ jobs:
       - name: Check device availability
         run: bash .github/scripts/${{ inputs.platform }}/check.sh
 
-      - name: Prepare host models
-        run: |
-          model_script=".github/scripts/${{ inputs.platform }}/download_models.sh"
-          if [ -f "${model_script}" ]; then
-            bash "${model_script}"
-          fi
-
       - name: Install project
         run: bash .github/scripts/${{ inputs.platform }}/setup.sh
 
+      - name: Check test models
+        run: |
+          python .github/scripts/generate_matrix.py \
+            --platform ${{ inputs.platform }} \
+            --device "${{ inputs.device }}" \
+            --cases '${{ inputs.cases }}' \
+            --check-models
+
       - name: Run E2E test
         run: |
           python tests/run.py \

.github/workflows/ci.yml

@@ -155,3 +155,14 @@ jobs:
     with:
       platform: musa
     secrets: inherit
+
+  # ============================================================
+  # Job 4f: Enflame platform testing
+  # ============================================================
+  test-enflame:
+    needs: discover
+    if: contains(fromJson(needs.discover.outputs.platforms), 'enflame')
+    uses: ./.github/workflows/_platform_test.yml
+    with:
+      platform: enflame
+    secrets: inherit

docker/ascend/README.md

@@ -27,9 +27,10 @@ The common convention is:
     └── Qwen3-0.6B/
 ```
 
-The E2E workflow runs `.github/scripts/ascend/download_models.sh`
-idempotently. It downloads `Qwen/Qwen3-0.6B` through
-`https://hf-mirror.com` only when the model is absent.
+E2E jobs expect these model directories to be provisioned before CI starts.
+The workflow validates the selected model paths with the shared
+`.github/scripts/generate_matrix.py --check-models` check before running
+`tests/run.py`.
 
 To avoid occupying a scarce NPU runner during development, validate changes
 on the host with the same image, setup script, and `tests/run.py` command