ci: use a single environment for CUDA all images

Author: AlexMa616

.github/configs/ascend.yml

@@ -42,8 +42,6 @@ image_build:
   target: dev
   network: host
   max_parallel: 1
-  runtime_device_count: '2'
-  runtime_smoke_nproc: '2'
   registry_login: true
   validation_script: tools/install/ascend/validate_image_build.sh
   tasks:
@@ -66,30 +64,6 @@ image_build:
       source_refs:
         VLLM_PLUGIN_FL_REF: vllm_plugin_fl
         FLAGGEMS_REF: flaggems
-    all:
-      dockerfile: docker/ascend/Dockerfile.all
-      image: flagos-dev/flagscale-all
-      test_roles: [train, inference]
-      base_image: harbor.baai.ac.cn/flagos-dev/flagscale:manual-20260812-ascend-dev-inference
-      build_args:
-        TRAIN_BASE_IMAGE: harbor.baai.ac.cn/flagos-dev/flagscale:manual-20260812-ascend-dev-train
-        INFERENCE_BASE_IMAGE: harbor.baai.ac.cn/flagos-dev/flagscale:manual-20260812-ascend-dev-inference
-      source_refs:
-        FLAGSCALE_TE_REF: transformer_engine_fl
-        FLAGSCALE_MEGATRON_REF: megatron_lm_fl
-        VLLM_PLUGIN_FL_REF: vllm_plugin_fl
-        FLAGGEMS_REF: flaggems
-      test_environments:
-        train:
-          pkg_mgr: runtime
-          env_path: /opt/flagscale/runtimes/train
-        inference:
-          pkg_mgr: runtime
-          env_path: /opt/flagscale/runtimes/inference
-        serve:
-          pkg_mgr: runtime
-          env_path: /opt/flagscale/runtimes/serve
-
 # Container options for Ascend NPU (no --gpus, use --device for NPU cards)
 container_options: >-
   --device /dev/davinci_manager

.github/configs/cuda.yml

@@ -72,14 +72,17 @@ image_build:
       test_roles: [train, inference]
       test_environments:
         train:
-          pkg_mgr: runtime
-          env_path: /opt/flagscale/runtimes/train
+          pkg_mgr: conda
+          env_name: flagscale-all
+          env_path: /opt/flagscale/miniconda3
         inference:
-          pkg_mgr: runtime
-          env_path: /opt/flagscale/runtimes/inference
+          pkg_mgr: conda
+          env_name: flagscale-all
+          env_path: /opt/flagscale/miniconda3
         serve:
-          pkg_mgr: runtime
-          env_path: /opt/flagscale/runtimes/serve
+          pkg_mgr: conda
+          env_name: flagscale-all
+          env_path: /opt/flagscale/miniconda3
       build_args:
         CUDA_VERSION: '12.8.1'
         UBUNTU_VERSION: '22.04'

.github/configs/metax.yml

@@ -63,29 +63,6 @@ image_build:
       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
-    all:
-      dockerfile: docker/metax/Dockerfile.all
-      image: flagos-dev/flagscale-all
-      test_roles: [train, inference]
-      base_image: harbor.baai.ac.cn/flagos-dev/flagscale-inference:86c8e54-metax-dev
-      build_args:
-        TRAIN_BASE_IMAGE: harbor.baai.ac.cn/flagos-dev/flagscale:0a6cc4b-metax-train-dev
-        INFERENCE_BASE_IMAGE: harbor.baai.ac.cn/flagos-dev/flagscale-inference:86c8e54-metax-dev
-      source_refs:
-        FLAGSCALE_MEGATRON_REF: megatron_lm_fl
-        FLAGSCALE_TE_REF: transformer_engine_fl
-        VLLM_PLUGIN_FL_REF: vllm_plugin_fl
-      test_environments:
-        train:
-          pkg_mgr: runtime
-          env_path: /opt/flagscale/runtimes/train
-        inference:
-          pkg_mgr: runtime
-          env_path: /opt/flagscale/runtimes/inference
-        serve:
-          pkg_mgr: runtime
-          env_path: /opt/flagscale/runtimes/serve
-
 # Container options for MetaX C550 (MACA runtime)
 container_options: >-
   --hostname=flagscale-ci

.github/configs/musa.yml

@@ -18,10 +18,9 @@
 hardware_name: musa
 display_name: "MThreads MUSA Tests"
 
-# Latest train/inference/all images validated end-to-end on S5000.
-# Keep all tasks on the same FlagScale revision so common tests use a
-# consistent software stack without rebuilding images for every PR.
-ci_image: harbor.baai.ac.cn/flagscale/flagscale-all:dev-musa-s5000-py3.10-8aa5314e94bc
+# 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
 
@@ -63,30 +62,6 @@ image_build:
         PKG_MGR: pip
         PIP_INDEX_URL: https://pypi.org/simple
         PIP_DEFAULT_TIMEOUT: '300'
-    all:
-      dockerfile: docker/musa/Dockerfile.all
-      image: flagos-dev/flagscale-all
-      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, inference]
-      test_environments:
-        train:
-          pkg_mgr: runtime
-          env_path: /opt/flagscale/runtimes/train
-        inference:
-          pkg_mgr: runtime
-          env_path: /opt/flagscale/runtimes/inference
-        serve:
-          pkg_mgr: runtime
-          env_path: /opt/flagscale/runtimes/serve
-      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
-
 # Container volumes (hardware-specific paths)
 container_volumes:
   - /mnt/airs-business/cicd/baai_datasets:/home/gitlab-runner/data

.github/workflows/all_tests_common.yml

@@ -45,6 +45,10 @@ on:
         required: false
         type: string
         default: ""
+      train_env_name:
+        required: false
+        type: string
+        default: ""
       train_env_path:
         required: false
         type: string
@@ -53,6 +57,10 @@ on:
         required: false
         type: string
         default: ""
+      inference_env_name:
+        required: false
+        type: string
+        default: ""
       inference_env_path:
         required: false
         type: string
@@ -61,6 +69,10 @@ on:
         required: false
         type: string
         default: ""
+      serve_env_name:
+        required: false
+        type: string
+        default: ""
       serve_env_path:
         required: false
         type: string
@@ -190,18 +202,27 @@ jobs:
           if [ -n "${{ inputs.train_pkg_mgr }}" ]; then
             echo "pkg_mgr_train=${{ inputs.train_pkg_mgr }}" >> "$GITHUB_OUTPUT"
           fi
+          if [ -n "${{ inputs.train_env_name }}" ]; then
+            echo "env_name_train=${{ inputs.train_env_name }}" >> "$GITHUB_OUTPUT"
+          fi
           if [ -n "${{ inputs.train_env_path }}" ]; then
             echo "env_path_train=${{ inputs.train_env_path }}" >> "$GITHUB_OUTPUT"
           fi
           if [ -n "${{ inputs.inference_pkg_mgr }}" ]; then
             echo "pkg_mgr_inference=${{ inputs.inference_pkg_mgr }}" >> "$GITHUB_OUTPUT"
           fi
+          if [ -n "${{ inputs.inference_env_name }}" ]; then
+            echo "env_name_inference=${{ inputs.inference_env_name }}" >> "$GITHUB_OUTPUT"
+          fi
           if [ -n "${{ inputs.inference_env_path }}" ]; then
             echo "env_path_inference=${{ inputs.inference_env_path }}" >> "$GITHUB_OUTPUT"
           fi
           if [ -n "${{ inputs.serve_pkg_mgr }}" ]; then
             echo "pkg_mgr_serve=${{ inputs.serve_pkg_mgr }}" >> "$GITHUB_OUTPUT"
           fi
+          if [ -n "${{ inputs.serve_env_name }}" ]; then
+            echo "env_name_serve=${{ inputs.serve_env_name }}" >> "$GITHUB_OUTPUT"
+          fi
           if [ -n "${{ inputs.serve_env_path }}" ]; then
             echo "env_path_serve=${{ inputs.serve_env_path }}" >> "$GITHUB_OUTPUT"
           fi

.github/workflows/build_image_common.yml

@@ -72,10 +72,13 @@ jobs:
       train_image: ${{ steps.contract.outputs.train_image }}
       inference_image: ${{ steps.contract.outputs.inference_image }}
       train_pkg_mgr: ${{ steps.contract.outputs.train_pkg_mgr }}
+      train_env_name: ${{ steps.contract.outputs.train_env_name }}
       train_env_path: ${{ steps.contract.outputs.train_env_path }}
       inference_pkg_mgr: ${{ steps.contract.outputs.inference_pkg_mgr }}
+      inference_env_name: ${{ steps.contract.outputs.inference_env_name }}
       inference_env_path: ${{ steps.contract.outputs.inference_env_path }}
       serve_pkg_mgr: ${{ steps.contract.outputs.serve_pkg_mgr }}
+      serve_env_name: ${{ steps.contract.outputs.serve_env_name }}
       serve_env_path: ${{ steps.contract.outputs.serve_env_path }}
     steps:
       - name: Checkout source
@@ -169,10 +172,13 @@ jobs:
           inference_image=$(/usr/local/bin/yq -r \
             '.ci_inference_image // .ci_image // ""' "$config_file")
           train_pkg_mgr=""
+          train_env_name=""
           train_env_path=""
           inference_pkg_mgr=""
+          inference_env_name=""
           inference_env_path=""
           serve_pkg_mgr=""
+          serve_env_name=""
           serve_env_path=""
 
           for raw_task in "${task_names[@]}"; do
@@ -265,15 +271,36 @@ jobs:
               if jq -e 'index("inference") != null' <<< "$test_roles" >/dev/null; then
                 required_roles+=(inference serve)
               fi
+              reference_role="${required_roles[0]}"
               for role in "${required_roles[@]}"; do
                 pkg_mgr=$(jq -r --arg role "$role" \
                   '.test_environments[$role].pkg_mgr // ""' <<< "$entry")
+                env_name=$(jq -r --arg role "$role" \
+                  '.test_environments[$role].env_name // ""' <<< "$entry")
                 env_path=$(jq -r --arg role "$role" \
                   '.test_environments[$role].env_path // ""' <<< "$entry")
                 if [ -z "$pkg_mgr" ] || [ -z "$env_path" ]; then
                   echo "::error::Missing test_environments.${role} for ${PLATFORM}/${task}"
                   exit 1
                 fi
+                case "$pkg_mgr" in
+                  conda|pip|uv) ;;
+                  *)
+                    echo "::error::Unsupported test environment package manager: ${PLATFORM}/${task}/${role}: $pkg_mgr"
+                    exit 1
+                    ;;
+                esac
+                if [ "$pkg_mgr" = conda ] && [ -z "$env_name" ]; then
+                  echo "::error::Missing test_environments.${role}.env_name for ${PLATFORM}/${task}"
+                  exit 1
+                fi
+                if ! jq -e --arg role "$role" \
+                  --arg reference_role "$reference_role" \
+                  '.test_environments[$role] == .test_environments[$reference_role]' \
+                  <<< "$entry" >/dev/null; then
+                  echo "::error::All image test roles must share one environment: ${PLATFORM}/${task}"
+                  exit 1
+                fi
               done
             fi
 
@@ -345,6 +372,7 @@ jobs:
                     if [ "$all_task_selected" != true ] || [ "$task" = all ]; then
                       train_image="$candidate"
                       train_pkg_mgr=$(jq -r '.test_environments.train.pkg_mgr // ""' <<< "$entry")
+                      train_env_name=$(jq -r '.test_environments.train.env_name // ""' <<< "$entry")
                       train_env_path=$(jq -r '.test_environments.train.env_path // ""' <<< "$entry")
                     fi
                     candidate_test_roles=$((candidate_test_roles + 1))
@@ -353,8 +381,10 @@ jobs:
                     if [ "$all_task_selected" != true ] || [ "$task" = all ]; then
                       inference_image="$candidate"
                       inference_pkg_mgr=$(jq -r '.test_environments.inference.pkg_mgr // ""' <<< "$entry")
+                      inference_env_name=$(jq -r '.test_environments.inference.env_name // ""' <<< "$entry")
                       inference_env_path=$(jq -r '.test_environments.inference.env_path // ""' <<< "$entry")
                       serve_pkg_mgr=$(jq -r '.test_environments.serve.pkg_mgr // ""' <<< "$entry")
+                      serve_env_name=$(jq -r '.test_environments.serve.env_name // ""' <<< "$entry")
                       serve_env_path=$(jq -r '.test_environments.serve.env_path // ""' <<< "$entry")
                     fi
                     candidate_test_roles=$((candidate_test_roles + 1))
@@ -381,10 +411,13 @@ jobs:
             echo "train_image=$train_image"
             echo "inference_image=$inference_image"
             echo "train_pkg_mgr=$train_pkg_mgr"
+            echo "train_env_name=$train_env_name"
             echo "train_env_path=$train_env_path"
             echo "inference_pkg_mgr=$inference_pkg_mgr"
+            echo "inference_env_name=$inference_env_name"
             echo "inference_env_path=$inference_env_path"
             echo "serve_pkg_mgr=$serve_pkg_mgr"
+            echo "serve_env_name=$serve_env_name"
             echo "serve_env_path=$serve_env_path"
           } >> "$GITHUB_OUTPUT"
 
@@ -622,8 +655,11 @@ jobs:
       ci_train_image: ${{ needs.configure.outputs.train_image }}
       ci_inference_image: ${{ needs.configure.outputs.inference_image }}
       train_pkg_mgr: ${{ needs.configure.outputs.train_pkg_mgr }}
+      train_env_name: ${{ needs.configure.outputs.train_env_name }}
       train_env_path: ${{ needs.configure.outputs.train_env_path }}
       inference_pkg_mgr: ${{ needs.configure.outputs.inference_pkg_mgr }}
+      inference_env_name: ${{ needs.configure.outputs.inference_env_name }}
       inference_env_path: ${{ needs.configure.outputs.inference_env_path }}
       serve_pkg_mgr: ${{ needs.configure.outputs.serve_pkg_mgr }}
+      serve_env_name: ${{ needs.configure.outputs.serve_env_name }}
       serve_env_path: ${{ needs.configure.outputs.serve_env_path }}

.github/workflows/functional_tests_benchmark.yml

@@ -192,9 +192,6 @@ jobs:
                 activate_uv_env "$ENV_PATH" || echo "UV activation failed"
               fi
               ;;
-            runtime)
-              activate_runtime_env "$ENV_PATH" || { echo "Runtime activation failed"; exit 1; }
-              ;;
             pip)
               echo "Running tests with pip/system Python"
               ;;
@@ -256,9 +253,6 @@ jobs:
                 activate_uv_env "$ENV_PATH" || { echo "UV activation failed"; exit 1; }
               fi
               ;;
-            runtime)
-              activate_runtime_env "$ENV_PATH" || { echo "Runtime activation failed"; exit 1; }
-              ;;
             pip)
               echo "Using system Python with pip"
               ;;

.github/workflows/functional_tests_cli.yml

@@ -159,9 +159,6 @@ jobs:
                 activate_uv_env "$ENV_PATH" || { echo "❌ UV activation failed"; exit 1; }
               fi
               ;;
-            runtime)
-              activate_runtime_env "$ENV_PATH" || { echo "Runtime activation failed"; exit 1; }
-              ;;
             pip)
               echo "Using system Python with pip"
               ;;
@@ -204,9 +201,6 @@ jobs:
                 activate_uv_env "$ENV_PATH" || { echo "❌ UV activation failed"; exit 1; }
               fi
               ;;
-            runtime)
-              activate_runtime_env "$ENV_PATH" || { echo "Runtime activation failed"; exit 1; }
-              ;;
             pip)
               echo "Using system Python with pip"
               ;;

.github/workflows/functional_tests_hetero_train.yml

@@ -192,9 +192,6 @@ jobs:
                 activate_uv_env "$ENV_PATH" || echo "⚠️  UV activation failed"
               fi
               ;;
-            runtime)
-              activate_runtime_env "$ENV_PATH" || { echo "Runtime activation failed"; exit 1; }
-              ;;
             pip)
               echo "ℹ️  Running tests with pip/system Python"
               ;;

.github/workflows/functional_tests_inference.yml

@@ -167,9 +167,6 @@ jobs:
                 activate_uv_env "$ENV_PATH"
               fi
               ;;
-            runtime)
-              activate_runtime_env "$ENV_PATH"
-              ;;
             pip)
               ;;
             *)
@@ -225,9 +222,6 @@ jobs:
                 activate_uv_env "$ENV_PATH" || echo "⚠️  UV activation failed"
               fi
               ;;
-            runtime)
-              activate_runtime_env "$ENV_PATH" || { echo "Runtime activation failed"; exit 1; }
-              ;;
             pip)
               echo "ℹ️  Running tests with pip/system Python"
               PLATFORM_ENV_SCRIPT="./tools/install/$PLATFORM/env.sh"