Merge branch 'main' into fix/chat-template-return-compat

Author: zhaoyinglia

.github/configs/ascend.yml

@@ -20,9 +20,9 @@ hardware_name: ascend
 display_name: "Ascend NPU Tests"
 
 # Docker image for this hardware
-ci_image: harbor.baai.ac.cn/flagscale/vllm-plugin-fl:v0.1.0-ascend-ci
-ci_train_image: harbor.baai.ac.cn/flagscale/vllm-plugin-fl:v0.1.0-ascend-ci
-ci_inference_image: harbor.baai.ac.cn/flagscale/vllm-plugin-fl:v0.1.0-ascend-ci
+ci_image: harbor.baai.ac.cn/flagos-dev/flagscale:manual-20260812-ascend-dev-inference
+ci_train_image: harbor.baai.ac.cn/flagos-dev/flagscale:manual-20260812-ascend-dev-train
+ci_inference_image: harbor.baai.ac.cn/flagos-dev/flagscale:manual-20260812-ascend-dev-inference
 
 # Runner labels for this hardware
 runner_labels: ["flagscale-ascend-ascend910-gpu2-32c-128g"]
@@ -35,6 +35,35 @@ container_volumes:
   - /usr/local/Ascend/add-ons:/usr/local/Ascend/add-ons:ro
   - /usr/local/sbin/npu-smi:/usr/local/sbin/npu-smi:ro
 
+# Image build contract consumed by the generic build workflow.
+image_build:
+  enabled: true
+  registry: harbor.baai.ac.cn
+  target: dev
+  network: host
+  max_parallel: 1
+  registry_login: true
+  validation_script: tools/install/ascend/validate_image_build.sh
+  tasks:
+    train:
+      dockerfile: docker/ascend/Dockerfile.train
+      image: flagos-dev/flagscale-train
+      test_roles: [train]
+      base_image: harbor.baai.ac.cn/flagos-dev/megatron-lm-fl:manual-20260717-ascend-dev
+      build_args:
+        PKG_MGR: pip
+      source_refs:
+        FLAGSCALE_TE_REF: transformer_engine_fl
+        FLAGSCALE_MEGATRON_REF: megatron_lm_fl
+    inference:
+      dockerfile: docker/ascend/Dockerfile.inference
+      image: flagos-dev/flagscale-inference
+      test_roles: [inference]
+      base_image: harbor.baai.ac.cn/flagos-dev/flagscale:manual-20260812-ascend-dev-inference
+      build_args: {}
+      source_refs:
+        VLLM_PLUGIN_FL_REF: vllm_plugin_fl
+        FLAGGEMS_REF: flaggems
 # Container options for Ascend NPU (no --gpus, use --device for NPU cards)
 container_options: >-
   --device /dev/davinci_manager
@@ -55,6 +84,12 @@ container_options: >-
   --security-opt apparmor=unconfined
   --privileged
 
+# Runtime allowances for the currently validated Ascend images. Keep these
+# values in platform configuration so common workflows remain chip-agnostic.
+timeouts:
+  inference_minutes: 15
+  serve_ready_seconds: 600
+
 # =============================================================================
 # Package Manager Configuration
 # =============================================================================

.github/configs/cuda.yml

@@ -24,12 +24,76 @@ ci_image: harbor.baai.ac.cn/flagos/flagscale:cuda12.8.1-cudnn9.7.1-python3.12-to
 ci_train_image: harbor.baai.ac.cn/flagscale/flagscale-train:dev-cu128-py3.12-20260228210721
 ci_inference_image: harbor.baai.ac.cn/flagscale/flagscale-inference:dev-cu128-py3.12-20260302102033
 
-# Directory to store image tar files (shared between build and push workflows)
-tar_dir: /mnt/airs-business/cicd/image_tar
-
 # Runner labels for this hardware
 runner_labels: ['flagscale-nvidia-a100-gpu2-32c-128g']
 
+# Image build contract consumed by the generic image-build workflow.
+image_build:
+  enabled: true
+  registry: harbor.baai.ac.cn
+  target: dev
+  network: default
+  max_parallel: 1
+  registry_login: false
+  runner_labels: [self-hosted, Linux, X64, nvidia-0, gpus-8]
+  validation_script: tools/install/cuda/validate_image_build.sh
+  tasks:
+    train:
+      dockerfile: docker/cuda/Dockerfile.train
+      image: flagos-dev/flagscale-train
+      base_image: nvidia/cuda:12.8.1-devel-ubuntu22.04
+      test_roles: [train]
+      build_args:
+        CUDA_VERSION: '12.8.1'
+        UBUNTU_VERSION: '22.04'
+        PYTHON_VERSION: '3.12'
+        UV_VERSION: '0.7.2'
+        PKG_MGR: conda
+        PYTORCH_INDEX: https://download.pytorch.org/whl/cu128
+      source_refs:
+        FLAGSCALE_MEGATRON_REF: megatron_lm_fl
+        FLAGSCALE_TE_REF: transformer_engine_fl
+    inference:
+      dockerfile: docker/cuda/Dockerfile.inference
+      image: flagos-dev/flagscale-inference
+      base_image: nvidia/cuda:12.8.1-devel-ubuntu22.04
+      test_roles: [inference]
+      build_args:
+        CUDA_VERSION: '12.8.1'
+        UBUNTU_VERSION: '22.04'
+        PYTHON_VERSION: '3.12'
+        UV_VERSION: '0.7.2'
+        PKG_MGR: conda
+        PYTORCH_INDEX: https://download.pytorch.org/whl/cu128
+    all:
+      dockerfile: docker/cuda/Dockerfile.all
+      image: flagos-dev/flagscale-all
+      base_image: nvidia/cuda:12.8.1-devel-ubuntu22.04
+      test_roles: [train, inference]
+      test_environments:
+        train:
+          pkg_mgr: conda
+          env_name: flagscale-all
+          env_path: /opt/flagscale/miniconda3
+        inference:
+          pkg_mgr: conda
+          env_name: flagscale-all
+          env_path: /opt/flagscale/miniconda3
+        serve:
+          pkg_mgr: conda
+          env_name: flagscale-all
+          env_path: /opt/flagscale/miniconda3
+      build_args:
+        CUDA_VERSION: '12.8.1'
+        UBUNTU_VERSION: '22.04'
+        PYTHON_VERSION: '3.12'
+        UV_VERSION: '0.7.2'
+        PKG_MGR: conda
+        PYTORCH_INDEX: https://download.pytorch.org/whl/cu128
+      source_refs:
+        FLAGSCALE_MEGATRON_REF: megatron_lm_fl
+        FLAGSCALE_TE_REF: transformer_engine_fl
+
 # Container volumes (hardware-specific paths)
 container_volumes:
   - /mnt/airs-business/cicd/baai_datasets:/home/gitlab-runner/data
@@ -76,4 +140,5 @@ env_names:
   train: "flagscale-train"
   hetero_train: "flagscale-train"
   inference: "flagscale-inference"
+  serve: "flagscale-inference"
   rl: "flagscale-rl"

.github/configs/image_sources.yml

@@ -0,0 +1,32 @@
+# 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.
+
+# Source overlays shared by platform image builds. Moving refs are resolved once
+# per workflow; native components owned by a vendor base remain digest-pinned.
+sources:
+  megatron_lm_fl:
+    repository: https://github.com/flagos-ai/Megatron-LM-FL.git
+    policy: branch
+    branch: main
+  transformer_engine_fl:
+    repository: https://github.com/flagos-ai/TransformerEngine-FL.git
+    policy: branch
+    branch: main
+  vllm_plugin_fl:
+    repository: https://github.com/flagos-ai/vllm-plugin-FL.git
+    policy: branch
+    branch: main
+  flaggems:
+    repository: https://github.com/flagos-ai/FlagGems.git
+    policy: latest_release

.github/configs/metax.yml

@@ -21,8 +21,8 @@ display_name: "MetaX C550 Tests"
 
 # Docker images for MetaX C550 (MACA-based)
 ci_image: harbor.baai.ac.cn/flagscale/megatron-lm-with-te:202603231839
-ci_train_image: harbor.baai.ac.cn/flagscale/megatron-lm-with-te:202603231839
-ci_inference_image: harbor.baai.ac.cn/flagscale/megatron-lm-with-te:202603231839
+ci_train_image: harbor.baai.ac.cn/flagos-dev/flagscale:0a6cc4b-metax-train-dev
+ci_inference_image: harbor.baai.ac.cn/flagos-dev/flagscale-inference:86c8e54-metax-dev
 
 # Runner labels for MetaX C550 hardware
 runner_labels: ["flagscale-metax-c550-gpu2-8c-256g"]
@@ -32,6 +32,37 @@ container_volumes:
   - /mnt/airs-business/cicd/baai_datasets:/home/gitlab-runner/data
   - /mnt/airs-business/cicd/baai_tokenizers:/home/gitlab-runner/tokenizers
 
+# Image build contract consumed by the generic build workflow.
+image_build:
+  enabled: true
+  registry: harbor.baai.ac.cn
+  target: dev
+  network: default
+  max_parallel: 1
+  # The C550 runner exposes an eight-device host. Individual image smoke
+  # tests still use runtime_smoke_nproc=2 below because the smoke collective
+  # only needs two processes.
+  runtime_device_count: '8'
+  runtime_smoke_nproc: '2'
+  registry_login: true
+  validation_script: tools/install/metax/validate_image_build.sh
+  tasks:
+    train:
+      dockerfile: docker/metax/Dockerfile.train
+      image: flagos-dev/flagscale-train
+      test_roles: [train]
+      base_image: harbor.baai.ac.cn/flagscale/megatron-lm-with-te:202603231839
+      build_args: {}
+      source_refs:
+        FLAGSCALE_MEGATRON_REF: megatron_lm_fl
+        FLAGSCALE_TE_REF: transformer_engine_fl
+    inference:
+      dockerfile: docker/metax/Dockerfile.inference
+      image: flagos-dev/flagscale-inference
+      test_roles: [inference]
+      base_image: harbor.baai.ac.cn/flagos-dev/vllm-plugin-fl:v0.20.2-metax-ci
+      source_refs:
+        VLLM_PLUGIN_FL_REF: vllm_plugin_fl
 # Container options for MetaX C550 (MACA runtime)
 container_options: >-
   --hostname=flagscale-ci
@@ -52,6 +83,13 @@ container_options: >-
   --cap-add NET_ADMIN
   --security-opt seccomp=unconfined
 
+# MetaX vLLM may spend about three minutes loading the first model. Keep this
+# platform allowance in configuration so the common serve workflow stays
+# hardware-neutral.
+timeouts:
+  inference_minutes: 15
+  serve_ready_seconds: 600
+
 # =============================================================================
 # Package Manager Configuration
 # =============================================================================
@@ -65,4 +103,5 @@ env_names:
   train: "base"
   hetero_train: ""
   inference: ""
+  serve: ""
   rl: ""

.github/configs/musa.yml

@@ -0,0 +1,101 @@
+# 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.
+
+# MThreads MUSA S5000 Hardware Configuration
+# This file defines CI/CD settings for MUSA-based testing.
+
+hardware_name: musa
+display_name: "MThreads MUSA Tests"
+
+# Latest train/inference images validated end-to-end on S5000.
+# The platform currently does not publish an all-in-one image.
+ci_image: harbor.baai.ac.cn/flagscale/flagscale-train:dev-musa-s5000-py3.10-8aa5314e94bc
+ci_train_image: harbor.baai.ac.cn/flagscale/flagscale-train:dev-musa-s5000-py3.10-8aa5314e94bc
+ci_inference_image: harbor.baai.ac.cn/flagscale/flagscale-inference:dev-musa-s5000-py3.10-8aa5314e94bc
+
+# Runner labels for MThreads MUSA hardware
+runner_labels: ["flagscale-mthreads-s5000-gpu8"]
+
+# Image build contract consumed by the generic image-build workflow.
+# max_parallel keeps the three image tasks serialized on each S5000 runner.
+image_build:
+  enabled: true
+  registry: harbor.baai.ac.cn
+  target: dev
+  network: default
+  max_parallel: 1
+  runtime_device_count: '8'
+  registry_login: true
+  validation_script: tools/install/musa/validate_image_build.sh
+  tasks:
+    train:
+      dockerfile: docker/musa/Dockerfile.train
+      image: flagos-dev/flagscale-train
+      base_image: harbor.baai.ac.cn/flagos-dev/transformerengine-fl:manual-20260802-musa-torch2.7.1-e73781e-vllm29ceebd
+      runtime_smoke_nproc: '2'
+      test_roles: [train]
+      build_args:
+        PKG_MGR: pip
+        PIP_INDEX_URL: https://pypi.org/simple
+        PIP_DEFAULT_TIMEOUT: '300'
+        FLAGSCALE_TE_VERSION: 2.0.0+e73781e
+      source_refs:
+        FLAGSCALE_MEGATRON_REF: megatron_lm_fl
+    inference:
+      dockerfile: docker/musa/Dockerfile.inference
+      image: flagos-dev/flagscale-inference
+      base_image: harbor.baai.ac.cn/flagos-dev/transformerengine-fl:manual-20260802-musa-torch2.7.1-e73781e-vllm29ceebd
+      runtime_smoke_nproc: '2'
+      test_roles: [inference]
+      build_args:
+        PKG_MGR: pip
+        PIP_INDEX_URL: https://pypi.org/simple
+        PIP_DEFAULT_TIMEOUT: '300'
+# Container volumes (hardware-specific paths)
+container_volumes:
+  - /mnt/airs-business/cicd/baai_datasets:/home/gitlab-runner/data
+  - /mnt/airs-business/cicd/baai_tokenizers:/home/gitlab-runner/tokenizers
+
+# Container options for MThreads MUSA runtime
+container_options: >-
+  --hostname=flagscale-ci
+  --privileged
+  --ipc=host
+  --shm-size=100g
+  --ulimit memlock=-1
+  --env MTHREADS_VISIBLE_DEVICES=all
+  --env MTHREADS_DRIVER_CAPABILITIES=all
+  --user root
+  --ulimit nofile=65535:65535
+
+# Runtime allowances for models loaded from the shared network volume.
+timeouts:
+  inference_minutes: 45
+  serve_ready_seconds: 600
+
+# =============================================================================
+# Package Manager Configuration
+# =============================================================================
+pkg_mgr: "pip"
+
+# Environment path (conda installation path)
+env_path: ""
+
+# Conda environment names
+env_names:
+  train: ""
+  hetero_train: ""
+  inference: ""
+  serve: ""
+  rl: ""