Merge PR MUSA S5000 support into dev-zyh

# Conflicts:
#	.github/configs/platforms.yml
#	.github/workflows/ci.yml
#	docker/build.sh

Author: HermiaHuan

.github/configs/musa.yml

@@ -0,0 +1,35 @@
+# 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.
+
+# MUSA Hardware Configuration
+# This file defines CI/CD settings for MUSA testing.
+
+platform: musa
+
+ci_image: harbor.baai.ac.cn/flagos-dev/vllm-plugin-fl:29ceebd-musa-ci
+
+runner_labels:
+  - mt-cicd-vllm-plugin
+
+container_volumes:
+  - /data:/data
+
+container_options: >-
+  --hostname vllm-plugin-fl
+  --privileged
+  --ipc=host
+  --shm-size=64g
+  --env GEMS_VENDOR=mthreads
+  --env VLLM_PLUGINS=fl
+  --env MTHREADS_VISIBLE_DEVICES=all

.github/configs/platforms.yml

@@ -28,3 +28,5 @@ platforms:
     enabled: true
   metax:
     enabled: true
+  musa:
+    enabled: true

.github/scripts/musa/check.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+# Copyright (c) 2026 BAAI. All rights reserved.
+# Check Moore Threads MUSA availability.
+set -euo pipefail
+
+echo "Current time: $(date '+%Y-%m-%d %H:%M:%S')"
+echo "=== Checking Moore Threads MUSA availability ==="
+
+if command -v mthreads-gmi >/dev/null 2>&1; then
+  mthreads-gmi
+else
+  echo "::warning::mthreads-gmi not found; checking through torch_musa."
+fi
+
+python - <<'PY'
+import torch
+import torch_musa
+
+assert torch.musa.is_available(), "MUSA accelerator is unavailable"
+count = torch.musa.device_count()
+assert count > 0, "No MUSA devices detected"
+
+tensor = torch.ones((32, 32), device="musa:0")
+torch.musa.synchronize()
+
+print(f"MUSA devices: {count}")
+print(f"Tensor smoke: {tensor.device} {tuple(tensor.shape)}")
+PY

.github/scripts/musa/setup.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+# Copyright (c) 2026 BAAI. All rights reserved.
+# Setup script for Moore Threads MUSA CI environment.
+set -euo pipefail
+
+git config --global --add safe.directory "$(pwd)"
+
+: "${GEMS_VENDOR:?GEMS_VENDOR is not set}"
+: "${VLLM_PLUGINS:?VLLM_PLUGINS is not set}"
+: "${MTHREADS_VISIBLE_DEVICES:?MTHREADS_VISIBLE_DEVICES is not set}"
+
+python -m pip install --no-build-isolation --no-deps -e .
+
+python - <<'PY'
+import flag_gems
+import torch
+import torch_musa
+import vllm
+import vllm_fl
+from vllm.platforms import current_platform
+
+assert torch.musa.is_available(), "MUSA accelerator is unavailable"
+assert torch.musa.device_count() > 0, "No MUSA devices detected"
+assert current_platform.device_type == "musa", current_platform.device_type
+
+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"Torch import ok: {torch.__version__}")
+print(f"MUSA available: {torch.musa.is_available()}")
+print(f"MUSA devices: {torch.musa.device_count()}")
+print(f"Platform: {current_platform}")
+PY

.github/workflows/ci.yml

@@ -144,3 +144,14 @@ jobs:
     with:
       platform: metax
     secrets: inherit
+
+  # ============================================================
+  # Job 4e: MUSA platform testing
+  # ============================================================
+  test-musa:
+    needs: discover
+    if: contains(fromJson(needs.discover.outputs.platforms), 'musa')
+    uses: ./.github/workflows/_platform_test.yml
+    with:
+      platform: musa
+    secrets: inherit

docker/build.sh

@@ -30,6 +30,12 @@ UBUNTU_VERSION="${UBUNTU_VERSION:-22.04}"
 VLLM_VERSION="${VLLM_VERSION:-0.19.0}"
 CANN_VERSION="${CANN_VERSION:-8.5.1}"
 CANN_CHIP="${CANN_CHIP:-910b}"
+MUSA_BASE_IMAGE="${MUSA_BASE_IMAGE:-registry.mthreads.com/mcconline/inference/vllm:v0.20.2-ph1-4.3.5-torch2.7.1-v1.1.0}"
+MUSA_VERSION="${MUSA_VERSION:-4.3.5}"
+MUSA_VLLM_VERSION="${MUSA_VLLM_VERSION:-0.20.2}"
+MUSA_PYTHON_VERSION="${MUSA_PYTHON_VERSION:-3.10}"
+MUSA_TORCH_VERSION="${MUSA_TORCH_VERSION:-2.7.1}"
+MUSA_FLAGGEMS_VERSION="${MUSA_FLAGGEMS_VERSION:-5.0.0}"
 HYGON_BASE_IMAGE="${HYGON_BASE_IMAGE:-harbor.sourcefind.cn:5443/dcu/admin/base/custom:vllm0.20.0-ubuntu22.04-dtk26.04-py3.10-MiniCPM-V-4.6}"
 HYGON_VLLM_VERSION="${HYGON_VLLM_VERSION:-0.20.2}"
 HYGON_DTK_VERSION="${HYGON_DTK_VERSION:-26.04}"
@@ -141,7 +147,7 @@ Usage: $(basename "$0") [OPTIONS]
 Build the vllm-plugin-FL Docker image.
 
 OPTIONS:
-    --platform PLATFORM    Platform to build: cuda, ascend, hygon, metax (default: ${PLATFORM})
+    --platform PLATFORM    Platform to build: cuda, ascend, hygon, metax, musa (default: ${PLATFORM})
     --target TARGET        Build target: dev, ci, release (default: ${TARGET})
     --image-name NAME      Image name (default: ${IMAGE_NAME})
     --image-tag TAG        Image tag (default: auto-generated)
@@ -161,6 +167,13 @@ VERSIONS (override via environment variables):
   Ascend:
     CANN_VERSION         CANN version (default: ${CANN_VERSION})
     CANN_CHIP            CANN chip: 910b, a3 (default: ${CANN_CHIP})
+  MUSA:
+    MUSA_BASE_IMAGE      Moore Threads base image (default: ${MUSA_BASE_IMAGE})
+    MUSA_VERSION         MUSA version used in image tag (default: ${MUSA_VERSION})
+    MUSA_VLLM_VERSION    vLLM empty-mode version (default: ${MUSA_VLLM_VERSION})
+    MUSA_PYTHON_VERSION  Python version in base image (default: ${MUSA_PYTHON_VERSION})
+    MUSA_TORCH_VERSION   PyTorch version in base image (default: ${MUSA_TORCH_VERSION})
+    MUSA_FLAGGEMS_VERSION FlagGems version in base image (default: ${MUSA_FLAGGEMS_VERSION})
   Hygon:
     HYGON_BASE_IMAGE     Base image (default: ${HYGON_BASE_IMAGE})
     HYGON_VLLM_VERSION   vLLM version installed in empty mode (default: ${HYGON_VLLM_VERSION})
@@ -193,6 +206,9 @@ EXAMPLES:
     # Build MetaX CI image
     ./build.sh --platform metax --target ci --image-name harbor.baai.ac.cn/flagos-dev/vllm-plugin-fl
 
+    # Build Moore Threads MUSA dev image
+    ./build.sh --platform musa --target dev
+
     # Build with custom PyPI mirror
     ./build.sh --target dev --index-url https://pypi.tuna.tsinghua.edu.cn/simple
 
@@ -310,8 +326,18 @@ elif [[ "${PLATFORM}" == "metax" ]]; then
     if [[ -z "${IMAGE_TAG}" ]]; then
         IMAGE_TAG="vllm-metax-${METAX_VLLM_VERSION}-maca.ai${METAX_MACA_VERSION}-torch2.8-${METAX_PYTHON_TAG}-ubuntu22.04-amd64-ci-git"
     fi
+elif [[ "${PLATFORM}" == "musa" ]]; then
+    PYTHON_VERSION="${MUSA_PYTHON_VERSION}"
+    VLLM_VERSION="${MUSA_VLLM_VERSION}"
+    BUILD_ARGS+=(
+        --build-arg "MUSA_BASE_IMAGE=${MUSA_BASE_IMAGE}"
+        --build-arg "FLAGGEMS_VERSION=${MUSA_FLAGGEMS_VERSION}"
+    )
+    if [[ -z "${IMAGE_TAG}" ]]; then
+        IMAGE_TAG="musa${MUSA_VERSION}-vllm${VLLM_VERSION}-torch${MUSA_TORCH_VERSION}-py${MUSA_PYTHON_VERSION}-${TARGET}"
+    fi
 else
-    err "Unknown platform '${PLATFORM}'. Must be 'cuda', 'ascend', 'hygon', or 'metax'."
+    err "Unknown platform '${PLATFORM}'. Must be 'cuda', 'ascend', 'hygon', 'metax', or 'musa'."
 fi
 
 FULL_IMAGE="${IMAGE_NAME}:${IMAGE_TAG}"
@@ -335,6 +361,11 @@ elif [[ "${PLATFORM}" == "metax" ]]; then
     msg "  MACA:           ${METAX_MACA_VERSION}"
     msg "  MetaX Python:   ${METAX_PYTHON_VERSION}"
     msg "  Base image:     ${METAX_BASE_IMAGE}"
+elif [[ "${PLATFORM}" == "musa" ]]; then
+    msg "  MUSA:           ${MUSA_VERSION}"
+    msg "  MUSA base:      ${MUSA_BASE_IMAGE}"
+    msg "  MUSA PyTorch:   ${MUSA_TORCH_VERSION}"
+    msg "  FlagGems:       ${MUSA_FLAGGEMS_VERSION}"
 fi
 msg "  Ubuntu:         ${UBUNTU_VERSION}"
 msg "  Python:         ${PYTHON_VERSION}"

docker/musa/Dockerfile

@@ -0,0 +1,67 @@
+ARG MUSA_BASE_IMAGE=registry.mthreads.com/mcconline/inference/vllm:v0.20.2-ph1-4.3.5-torch2.7.1-v1.1.0
+
+# ---------- base stage ----------
+FROM ${MUSA_BASE_IMAGE} AS base
+
+ARG PYTHON_VERSION=3.10
+ARG VLLM_VERSION=0.20.2
+ARG FLAGGEMS_VERSION=5.0.0
+
+ENV DEBIAN_FRONTEND=noninteractive \
+    VLLM_PLUGINS=fl
+
+# Keep the vendor runtime from the base image and add only the tools required
+# to build the checked-out plugin in CI.
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends \
+        build-essential \
+        ca-certificates \
+        curl \
+        git \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN python -m pip install --no-cache-dir \
+        "setuptools>=77.0.3,<81.0.0" \
+        wheel \
+        setuptools-scm \
+        scikit-build-core==0.11 \
+        pybind11 \
+        ninja \
+        cmake
+
+# Fail early when a different vendor base image is supplied accidentally.
+RUN python -c 'from importlib.metadata import version; \
+vllm = version("vllm"); \
+flag_gems = version("flag-gems"); \
+assert "empty" in vllm, vllm; \
+assert flag_gems == "5.0.0", flag_gems; \
+print("vLLM:", vllm); \
+print("FlagGems:", flag_gems)'
+
+WORKDIR /workspace
+
+# ---------- dev stage ----------
+FROM base AS dev
+
+RUN python -m pip install --no-cache-dir \
+        pytest==8.4.1 \
+        pytest-cov \
+        pytest-json-report \
+        ruff \
+        pre-commit
+
+WORKDIR /workspace
+
+# ---------- ci stage ----------
+FROM dev AS ci
+
+RUN python -m pip install --no-cache-dir \
+        pytest-timeout \
+        pytest-metadata
+
+WORKDIR /workspace
+
+# ---------- release stage ----------
+FROM base AS release
+
+WORKDIR /workspace

tests/models/qwen3/4b_tp2_eager_musa.yaml

@@ -0,0 +1,35 @@
+# Qwen3-4B model configuration (TP=2)
+
+llm:
+  model: "/data/models/Qwen/Qwen3-4B"
+  tensor_parallel_size: 2
+  max_model_len: 1024
+  gpu_memory_utilization: 0.8
+  trust_remote_code: true
+  enforce_eager: true
+  disable_custom_all_reduce: true
+
+generate:
+  prompts:
+    - "Hello, my name is"
+    - "The capital of France is"
+  sampling:
+    max_tokens: 10
+    temperature: 0.7
+
+serve:
+  served_model_name: "qwen"
+  endpoints: ["chat"]
+  stream: true
+  max_tokens: 256
+  chat_messages:
+    - role: "user"
+      content: "Give me a short introduction to large language models."
+  sampling:
+    temperature: 0.7
+    top_p: 0.8
+    presence_penalty: 1.5
+  extra_body:
+    top_k: 20
+    chat_template_kwargs:
+      enable_thinking: false

tests/models/qwen3_6/27b_tp4_eager_musa.yaml

@@ -0,0 +1,31 @@
+# Qwen3.6-27B MUSA smoke configuration (TP=4, eager).
+
+llm:
+  model: "/data/models/Qwen/Qwen3.6-27B"
+  tensor_parallel_size: 4
+  pipeline_parallel_size: 1
+  max_model_len: 8192
+  gpu_memory_utilization: 0.90
+  enforce_eager: true
+  enable_prefix_caching: false
+  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"
+  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

tests/models/qwen3_6/35b_a3b_tp4_eager_musa.yaml

@@ -0,0 +1,31 @@
+# Qwen3.6-35B-A3B MUSA smoke configuration (TP=4, eager).
+
+llm:
+  model: "/data/models/Qwen/Qwen3.6-35B-A3B"
+  tensor_parallel_size: 4
+  pipeline_parallel_size: 1
+  max_model_len: 8192
+  gpu_memory_utilization: 0.90
+  enforce_eager: true
+  enable_prefix_caching: false
+  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"
+  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