[ci] change lmdeploy version and related packages

.github/workflows/docker-inner.yml

@@ -3,6 +3,11 @@ name: publish-docker-inner
 on:
   workflow_dispatch:
     inputs:
+      repo_ref:
+        required: false
+        description: 'Set branch or tag or commit id. Default is "main"'
+        type: string
+        default: 'main'
       image_tag:
         required: true
         description: 'Set docker image tag. Default is "latest", you can also set "auto" to generate a formated tag, or input a manual tag'
@@ -39,7 +44,9 @@ jobs:
       SCHEDULE_TAG: 'pt29_latest'
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v3
+        uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event.inputs.repo_ref || 'main' }}
       - name: Parse tag
         run: |
           if [[ "${GITHUB_EVENT_NAME}" == 'schedule' ]]; then

Dockerfile

@@ -222,9 +222,8 @@ RUN --mount=type=secret,id=HTTPS_PROXY,env=https_proxy \
     pip install fastapi fire openai outlines \
         pyzmq aiohttp cloudpickle prometheus_client protobuf numpy pillow einops tiktoken sentencepiece \
         partial_json_parser 'ray[default]<3' shortuuid uvicorn pybase64 \
-        tilelang==0.1.11 \
         'pydantic>2' openai_harmony dlblas --no-cache-dir -i ${DEFAULT_PYPI_URL} && \
-    pip install xgrammar==0.1.32 timm!=1.0.23 --no-cache-dir -i ${DEFAULT_PYPI_URL} --no-deps && \
+    pip install xgrammar>=0.1.33 timm!=1.0.23 --no-cache-dir -i ${DEFAULT_PYPI_URL} --no-deps && \
     if [ -n "${LMDEPLOY_VERSION}" ]; then \
         pip install lmdeploy==${LMDEPLOY_VERSION} --no-deps --no-cache-dir -i ${DEFAULT_PYPI_URL}; \
     else \

autotest/module/train.py

@@ -33,6 +33,9 @@ def get_cmd(config):
             )
             config["work_dir"] = work_dir
 
+            # Print runtime image package list before training for version diagnosis.
+            image_version_check = 'echo "===== runtime pip list ====="; pip list; '
+
             # this patch is for torch 2.9.1 Conv3d memory issue fix
             cudnn_patch = (
                 "TORCH_VERSION=$(python -c 'import torch;print(torch.__version__.split(chr(43))[0])'); "
@@ -41,7 +44,8 @@ def get_cmd(config):
 
             if train_type == "sft":
                 command = (
-                    f"cd {current_dir}; pwd; {pip_package}; export GITHUB_RUN_ID={config.get('run_id')}; export WORK_DIR={work_dir}; "
+                    f"cd {current_dir}; pwd; {pip_package}; {image_version_check}"
+                    f"export GITHUB_RUN_ID={config.get('run_id')}; export WORK_DIR={work_dir}; "
                     + cudnn_patch
                     + f"torchrun --nproc-per-node {nproc_per_node} --master_addr=${{MASTER_ADDR}} --master_port=${{MASTER_PORT}} --nnodes=${{WORLD_SIZE}} --node_rank=${{RANK}} "
                     + f"xtuner/v1/train/cli/{train_type}.py"
@@ -66,7 +70,8 @@ def get_cmd(config):
                 infer_type = config.get("parameters", {}).get("infer_backend", "lmdeploy")
                 accelerator = config.get("parameters", {}).get("accelerator", "GPU")
                 command = (
-                    f"cd {current_dir}; pwd; {pip_package}; export GITHUB_RUN_ID={config.get('run_id')}; export WORK_DIR={work_dir}; "
+                    f"cd {current_dir}; pwd; {pip_package}; {image_version_check}"
+                    f"export GITHUB_RUN_ID={config.get('run_id')}; export WORK_DIR={work_dir}; "
                     + cudnn_patch
                     + f"bash -x autotest/utils/ci_run_rl.sh {accelerator} {infer_type} {config_path} ${{MODEL_PATH}} ${{DATA_PATH}} ${{EVAL_DATA_PATH}}"
                 )

image_build.sh

@@ -12,8 +12,8 @@ export CAUSAL_CONV1D_URL=https://github.com/Dao-AILab/causal-conv1d@da6dbaa9fd5a
 export FLA_URL="${FLA_URL-https://github.com/HAOCHENYE/flash-linear-attention@tmp-tensor-cache}"
 
 export TORCH_VERSION=${TORCH_VERSION:-"2.9.1"}
-# export LMDEPLOY_VERSION="0.13.0dev"
-export LMDEPLOY_URL=https://github.com/InternLM/lmdeploy@efe3b88607756a7ad9411b89627b5ac6ebaa540e
+export LMDEPLOY_URL="${LMDEPLOY_URL:-https://github.com/RunningLeon/lmdeploy@c14f389e9b2346f9835954ed90aefc0a0aac7ad4}"
+# export LMDEPLOY_VERSION="0.15.0"
 export PPA_SOURCE="https://mirrors.aliyun.com"
 export DEFAULT_PYPI_URL=${DEFAULT_PYPI_URL:-"https://mirrors.aliyun.com/pypi/simple"}
 # mirror https://download.pytorch.org/whl
@@ -40,7 +40,7 @@ docker build . \
   --build-arg DEEP_GEMM_URL=$DEEP_GEMM_URL \
   --build-arg XTUNER_URL=$XTUNER_URL \
   --build-arg XTUNER_COMMIT=$XTUNER_COMMIT \
-  --build-arg LMDEPLOY_URL=$LMDEPLOY_URL \
+  --build-arg LMDEPLOY_URL="$LMDEPLOY_URL" \
   --progress=plain \
   --label "BASE_IMAGE=$BASE_IMAGE" \
   --label "XTUNER_URL=${XTUNER_URL/@/\/tree\/}" \
@@ -53,4 +53,7 @@ docker build . \
   --label "DEEP_EP_URL=${DEEP_EP_URL/@/\/tree\/}" \
   --label "DEEP_GEMM_URL=${DEEP_GEMM_URL/@/\/tree\/}" \
   --label "LMDEPLOY_URL=${LMDEPLOY_URL/@/\/tree\/}"
-  # --label "LMDEPLOY_VERSION=$LMDEPLOY_VERSION"
+
+echo "===== pip list in ${image_name}:${image_tag} ====="
+docker run --rm --entrypoint pip "${image_name}:${image_tag}" list
+

pyproject.toml

@@ -36,9 +36,10 @@ dependencies = [
   "SentencePiece",
   "tiktoken",
   "tilelang==0.1.11",
+  "apache-tvm-ffi==0.1.11",
   "torch>=2.6.0",
   "torchvision",
-  "transformers==5.2.0",
+  "transformers==5.14.1",
   "cyclopts",
   "transformers_stream_generator",
   "opencv-python-headless<=4.12.0.88",