[CICD] Unify dependency-coupled CI across accelerator platforms

.github/configs/ascend.yml

@@ -106,3 +106,38 @@ env_names:
   inference: ""
   serve: ""
   rl: ""
+
+# Setup script for environment preparation
+setup_script: .github/scripts/set_env_ascend.sh
+
+# Test matrix configuration for unit tests
+test_matrix:
+  unit:
+    nproc_per_node: 2
+
+# =============================================================================
+# Dependencies Configuration (Megatron-LM-FL & TE-FL)
+# =============================================================================
+dependencies:
+  megatron_lm_fl:
+    enabled: true
+    repository: flagos-ai/Megatron-LM-FL
+    ref: main
+    runtime:
+      install_mode: pythonpath
+      environment: {}
+
+  te_fl:
+    enabled: true
+    repository: flagos-ai/TransformerEngine-FL
+    ref: main
+    build_mode: python
+    build_environment:
+      TE_FL_SKIP_CUDA: "1"
+    runtime:
+      environment:
+        TE_FL_SKIP_CUDA: "1"
+        TE_FL_PREFER: flagos
+      install_pip_args: []
+      pip_packages:
+        - "megatron-energon[av_decode]~=7.0"

.github/configs/cuda.yml

@@ -142,3 +142,40 @@ env_names:
   inference: "flagscale-inference"
   serve: "flagscale-inference"
   rl: "flagscale-rl"
+
+# Setup script for environment preparation
+setup_script: .github/scripts/set_env_cuda.sh
+
+# Test matrix configuration for unit tests
+test_matrix:
+  unit:
+    nproc_per_node: 8
+
+# =============================================================================
+# Dependencies Configuration (Megatron-LM-FL & TE-FL)
+# =============================================================================
+dependencies:
+  megatron_lm_fl:
+    enabled: true
+    repository: flagos-ai/Megatron-LM-FL
+    ref: main
+    runtime:
+      install_mode: pythonpath
+      environment: {}
+
+  te_fl:
+    enabled: true
+    repository: flagos-ai/TransformerEngine-FL
+    ref: main
+    build_mode: cuda
+    build_environment:
+      # NCCL-EP requires the internal nccl_device.h header, which is not
+      # shipped by the system NCCL installation in the CUDA CI image.
+      NVTE_WITH_NCCL_EP: "0"
+    runtime:
+      environment:
+        NVTE_WITH_NCCL_EP: "0"
+        TE_FL_PREFER: flagos
+      install_pip_args: []
+      pip_packages:
+        - "megatron-energon[av_decode]~=7.0"

.github/configs/enflame.yml

@@ -48,3 +48,47 @@ pkg_mgr: "pip"
 env_path: ""
 env_names:
   train: ""
+  hetero_train: ""
+  inference: ""
+  serve: ""
+  rl: ""
+
+# Setup script for environment preparation
+setup_script: .github/scripts/set_env_enflame.sh
+
+# Test matrix configuration for unit tests
+test_matrix:
+  unit:
+    nproc_per_node: 2
+
+# =============================================================================
+# Dependencies Configuration (Megatron-LM-FL & TE-FL)
+# =============================================================================
+dependencies:
+  megatron_lm_fl:
+    enabled: true
+    repository: flagos-ai/Megatron-LM-FL
+    ref: main
+    runtime:
+      install_mode: pythonpath
+      environment: {}
+
+  te_fl:
+    enabled: true
+    repository: flagos-ai/TransformerEngine-FL
+    ref: main
+    build_mode: python
+    build_environment:
+      TE_FL_SKIP_CUDA: "1"
+      NVTE_FLASH_ATTN: "0"
+      NVTE_FUSED_ATTN: "0"
+    runtime:
+      environment:
+        TE_FL_SKIP_CUDA: "1"
+        TE_FL_PREFER: flagos
+        NVTE_FLASH_ATTN: "0"
+        NVTE_FUSED_ATTN: "0"
+      install_pip_args: []
+      pip_packages:
+        - onnxscript
+        - megatron-energon[av_decode]~=7.0

.github/configs/hygon.yml

@@ -109,3 +109,38 @@ env_names:
   inference: ""
   serve: ""
   rl: ""
+
+# Setup script for environment preparation
+setup_script: .github/scripts/set_env_hygon.sh
+
+# Test matrix configuration for unit tests
+test_matrix:
+  unit:
+    nproc_per_node: 2
+
+# =============================================================================
+# Dependencies Configuration (Megatron-LM-FL & TE-FL)
+# =============================================================================
+dependencies:
+  megatron_lm_fl:
+    enabled: true
+    repository: flagos-ai/Megatron-LM-FL
+    ref: main
+    runtime:
+      install_mode: pythonpath
+      environment: {}
+
+  te_fl:
+    enabled: true
+    repository: flagos-ai/TransformerEngine-FL
+    ref: main
+    build_mode: python
+    build_environment:
+      TE_FL_SKIP_CUDA: "1"
+    runtime:
+      environment:
+        TE_FL_SKIP_CUDA: "1"
+        TE_FL_PREFER: flagos
+      install_pip_args: []
+      pip_packages:
+        - "megatron-energon[av_decode]~=7.0"

.github/configs/kunlunxin.yml

@@ -60,3 +60,41 @@ pkg_mgr: "conda"
 env_path: "/root/miniconda"
 env_names:
   train: "python310_torch29_cuda"
+  hetero_train: "python310_torch29_cuda"
+  inference: "python310_torch29_cuda"
+  serve: "python310_torch29_cuda"
+  rl: "python310_torch29_cuda"
+
+# Setup script for environment preparation
+setup_script: .github/scripts/set_env_kunlunxin.sh
+
+# Test matrix configuration for unit tests
+test_matrix:
+  unit:
+    nproc_per_node: 2
+
+# =============================================================================
+# Dependencies Configuration (Megatron-LM-FL & TE-FL)
+# =============================================================================
+dependencies:
+  megatron_lm_fl:
+    enabled: true
+    repository: flagos-ai/Megatron-LM-FL
+    ref: main
+    runtime:
+      install_mode: pythonpath
+      environment: {}
+
+  te_fl:
+    enabled: true
+    repository: flagos-ai/TransformerEngine-FL
+    ref: main
+    build_mode: python
+    build_environment:
+      TE_FL_SKIP_CUDA: "1"
+    runtime:
+      environment:
+        TE_FL_SKIP_CUDA: "1"
+        TE_FL_PREFER: flagos
+      install_pip_args: []
+      pip_packages: ["megatron-energon[av_decode]~=7.0"]

.github/configs/metax.yml

@@ -105,3 +105,40 @@ env_names:
   inference: ""
   serve: ""
   rl: ""
+
+# Setup script for environment preparation
+setup_script: .github/scripts/set_env_metax.sh
+
+# Test matrix configuration for unit tests
+test_matrix:
+  unit:
+    nproc_per_node: 2
+
+# =============================================================================
+# Dependencies Configuration (Megatron-LM-FL & TE-FL)
+# =============================================================================
+dependencies:
+  megatron_lm_fl:
+    enabled: true
+    repository: flagos-ai/Megatron-LM-FL
+    ref: main
+    runtime:
+      install_mode: pythonpath
+      environment: {}
+
+  te_fl:
+    enabled: true
+    repository: flagos-ai/TransformerEngine-FL
+    ref: main
+    build_mode: python
+    build_environment:
+      TE_FL_SKIP_CUDA: "1"
+    runtime:
+      environment:
+        TE_FL_SKIP_CUDA: "1"
+        TE_FL_PREFER: flagos
+      install_pip_args: []
+      pip_packages:
+        - flag-gems
+        - sqlalchemy
+        - "megatron-energon[av_decode]~=7.0"

.github/configs/musa.yml

@@ -99,3 +99,43 @@ env_names:
   inference: ""
   serve: ""
   rl: ""
+
+# Setup script for environment preparation
+setup_script: .github/scripts/set_env_musa.sh
+
+# Test matrix configuration for unit tests
+test_matrix:
+  unit:
+    nproc_per_node: 2
+
+# =============================================================================
+# Dependencies Configuration (Megatron-LM-FL & TE-FL)
+# =============================================================================
+dependencies:
+  megatron_lm_fl:
+    enabled: true
+    repository: flagos-ai/Megatron-LM-FL
+    ref: main
+    runtime:
+      install_mode: pythonpath
+      environment: {}
+
+  te_fl:
+    enabled: true
+    repository: flagos-ai/TransformerEngine-FL
+    ref: main
+    build_mode: python
+    build_environment:
+      TE_FL_SKIP_CUDA: "1"
+      NVTE_FLASH_ATTN: "0"
+      NVTE_FUSED_ATTN: "0"
+    runtime:
+      environment:
+        TE_FL_SKIP_CUDA: "1"
+        TE_FL_PREFER: flagos
+        NVTE_FLASH_ATTN: "0"
+        NVTE_FUSED_ATTN: "0"
+      install_pip_args: []
+      pip_packages:
+        - onnxscript
+        - megatron-energon[av_decode]~=7.0

.github/scripts/install_megatron_runtime.sh

@@ -0,0 +1,62 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+: "${MEGATRON_INSTALL_DIR:?MEGATRON_INSTALL_DIR is required}"
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "$SCRIPT_DIR/set_env_common.sh"
+
+python_bin="${CI_PYTHON_BIN:-$(command -v python3)}"
+if [ ! -x "$python_bin" ]; then
+  echo "::error::Python executable not found: $python_bin" >&2
+  exit 1
+fi
+
+if [ ! -d "$MEGATRON_INSTALL_DIR" ]; then
+  echo "::error::Megatron-LM-FL install directory is missing: $MEGATRON_INSTALL_DIR" >&2
+  exit 1
+fi
+
+# Remove only a site-packages copy.  Keep the prepared target directory intact
+# even when it is already present on PYTHONPATH from an earlier CI step.
+echo "Removing any site-packages megatron-core to ensure PYTHONPATH priority..."
+env -u PYTHONPATH PYTHONNOUSERSITE=1 \
+  "$python_bin" -m pip uninstall -y megatron-core >/dev/null 2>&1 || true
+
+ci_configure_training_pythonpath
+ci_export_env PYTHONNOUSERSITE 1
+
+"$python_bin" - <<'PY'
+import sys
+import os
+
+megatron_path = os.environ.get("MEGATRON_INSTALL_DIR")
+if megatron_path and megatron_path not in sys.path:
+    sys.path.insert(0, megatron_path)
+
+try:
+    import megatron
+    import megatron.core
+    print(f"Megatron-LM-FL Python: {sys.executable}")
+    print(f"Megatron-LM-FL import passed: {megatron.__file__}")
+    print(f"Megatron-Core version: {getattr(megatron.core, '__version__', 'unknown')}")
+    print(f"Megatron-Core path: {megatron.core.__file__}")
+
+    expected_root = os.path.realpath(megatron_path)
+    actual_path = os.path.realpath(megatron.core.__file__)
+    if os.path.commonpath([expected_root, actual_path]) != expected_root:
+        raise ImportError(
+            f"megatron.core resolved outside the prepared runtime: {actual_path}"
+        )
+
+    # Verify critical modules exist
+    try:
+        from megatron.core.transformer.cuda_graph_config import CudaGraphConfig
+        print(f"✓ cuda_graph_config module available")
+    except (ImportError, ModuleNotFoundError) as e:
+        print(f"⚠ Warning: cuda_graph_config not available (expected for Megatron-Core < 0.18): {e}")
+except ImportError as e:
+    print(f"::error::Failed to import Megatron-LM-FL: {e}", file=sys.stderr)
+    sys.exit(1)
+PY