Fix/h200 cu130 heyifan

.pre-commit-config.yaml

@@ -1,77 +1,116 @@
+# Offline-friendly config: all hooks are `repo: local` and install via pip
+# (internal PyPI / mirrors). Avoids cloning github.com which is unreachable
+# on many GPU clusters.
 files: >
   (?x)^(
     xtuner/v1/.* |
     autotest/.*
   )$
 repos:
-  - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v5.0.0
+  - repo: local
     hooks:
       - id: check-yaml
+        name: check-yaml
+        entry: check-yaml
+        language: python
+        types: [yaml]
+        additional_dependencies: ["pre-commit-hooks==5.0.0"]
       - id: requirements-txt-fixer
+        name: requirements-txt-fixer
+        entry: requirements-txt-fixer
+        language: python
+        types: [text]
+        files: requirements.*\.txt$
+        additional_dependencies: ["pre-commit-hooks==5.0.0"]
       - id: check-merge-conflict
+        name: check-merge-conflict
+        entry: check-merge-conflict
+        language: python
+        additional_dependencies: ["pre-commit-hooks==5.0.0"]
       - id: mixed-line-ending
+        name: mixed-line-ending
+        entry: mixed-line-ending
+        language: python
         args: ["--fix=lf"]
-  - repo: https://github.com/codespell-project/codespell
-    rev: v2.2.1
-    hooks:
+        additional_dependencies: ["pre-commit-hooks==5.0.0"]
+
       - id: codespell
-  - repo: https://github.com/executablebooks/mdformat
-    rev: 0.7.9
-    hooks:
+        name: codespell
+        entry: codespell
+        language: python
+        types: [text]
+        additional_dependencies: ["codespell==2.2.1"]
+
       - id: mdformat
+        name: mdformat
+        entry: mdformat
+        language: python
+        types: [markdown]
         args: ["--number"]
         additional_dependencies:
+          - mdformat==0.7.9
           - mdformat-openmmlab
           - mdformat_frontmatter
           - linkify-it-py
         exclude: 'docs/zh_cn/user_guides/sequence_parallel.md'
-  - repo: https://github.com/myint/docformatter
-    rev: 06907d0
-    hooks:
+
       - id: docformatter
+        name: docformatter
+        entry: docformatter
+        language: python
+        types: [python]
         args: ["--in-place", "--wrap-descriptions", "119"]
-  - repo: https://github.com/asottile/pyupgrade
-    rev: v3.20.0
-    hooks:
+        additional_dependencies: ["docformatter==1.7.5"]
+
       - id: pyupgrade
+        name: pyupgrade
+        entry: pyupgrade
+        language: python
+        types: [python]
         args: ["--py36-plus"]
+        additional_dependencies: ["pyupgrade==3.20.0"]
 
-  - repo: https://github.com/astral-sh/ruff-pre-commit
-    # Ruff version.
-    rev: v0.9.10
-    hooks:
-      # Run the linter.
-      - id: ruff
-        types_or: [ python, pyi ]
-        args: [ --diff ]
       - id: ruff
-        types_or: [ python, pyi ]
-        args: [ --fix ]
-      # Run the formatter.
-      - id: ruff-format
-        types_or: [ python, pyi ]
-        args: [ --diff ]
+        name: ruff
+        entry: ruff check --force-exclude
+        language: python
+        types_or: [python, pyi]
+        args: [--diff]
+        additional_dependencies: ["ruff==0.9.10"]
+      - id: ruff-fix
+        name: ruff-fix
+        entry: ruff check --force-exclude
+        language: python
+        types_or: [python, pyi]
+        args: [--fix]
+        additional_dependencies: ["ruff==0.9.10"]
+      - id: ruff-format-diff
+        name: ruff-format
+        entry: ruff format --force-exclude
+        language: python
+        types_or: [python, pyi]
+        args: [--diff]
+        additional_dependencies: ["ruff==0.9.10"]
       - id: ruff-format
-        types_or: [ python, pyi ]
-  - repo: local
-  # We do not use pre-commit/mirrors-mypy, 
-  # as it comes with opinionated defaults 
-  # (like --ignore-missing-imports)
-  # and is difficult to configure to run 
-  # with the dependencies correctly installed.
-    hooks:
+        name: ruff-format
+        entry: ruff format --force-exclude
+        language: python
+        types_or: [python, pyi]
+        additional_dependencies: ["ruff==0.9.10"]
+
+      # We do not use pre-commit/mirrors-mypy (opinionated defaults).
+      # Torch is intentionally not pinned here: installing it into the hook
+      # env is heavy; missing third-party imports are ignored via pyproject.
       - id: mypy
         name: mypy
         entry: .dev_scripts/mypy_entrypoint.sh
         language: python
-        # For 3rd party lib, we only do type check for torch
-        additional_dependencies: ["mypy==1.16.1", "torch==2.6.0", "types-requests"]
+        additional_dependencies: ["mypy==1.16.1", "types-requests"]
         require_serial: true
         verbose: true
         types: [python]
       - id: pydantic-extra-check
-        name: pydantic-extra-check 
+        name: pydantic-extra-check
         language: system
         entry: .dev_scripts/test_pydantic.py
         verbose: false

Dockerfile

@@ -1,6 +1,6 @@
 # syntax=docker/dockerfile:1.10.0
 # builder
-ARG BASE_IMAGE=nvcr.io/nvidia/pytorch:25.03-py3
+ARG BASE_IMAGE=nvcr.io/nvidia/pytorch:25.11-py3
 
 ## build base env
 FROM ${BASE_IMAGE} AS setup_env
@@ -21,14 +21,24 @@ RUN if [ -d /etc/pip ] && [ -f /etc/pip/constraint.txt ]; then echo > /etc/pip/c
 RUN pip uninstall flash_attn opencv -y && rm -rf /usr/local/lib/python3.12/dist-packages/cv2
 RUN git config --system --add safe.directory "*"
 
+ENV DEBIAN_FRONTEND=noninteractive
+RUN apt-get update && apt-get install -y \
+    ffmpeg \
+    libavutil-dev \
+    libavcodec-dev \
+    libavformat-dev \
+    libswscale-dev
+
+
 # torch
 ARG TORCH_VERSION
 ARG PYTORCH_WHEELS_URL
 RUN --mount=type=secret,id=HTTPS_PROXY,env=https_proxy \
     --mount=type=secret,id=NO_PROXY,env=no_proxy \
     if [ -n "${TORCH_VERSION}" ]; then \
         pip install torchvision torch==${TORCH_VERSION} \
-        -i ${PYTORCH_WHEELS_URL}/cu128 \
+        -i ${PYTORCH_WHEELS_URL}/cu130 \
+        --extra-index-url ${PYTORCH_WHEELS_URL}/cu128 \
         --extra-index-url ${PYTORCH_WHEELS_URL}/cu126 \
         --no-cache-dir; \
     fi
@@ -58,17 +68,31 @@ ARG FLASH_ATTN_URL
 # force hopper for now, you change it throught build args
 ARG FLASH_ATTN_CUDA_ARCHS="90"
 ARG FLASH_ATTENTION_DISABLE_SM80="TRUE"
+ARG FLASH_ATTENTION_OFFLINE_BUILD="TRUE"
 
 RUN --mount=type=secret,id=HTTPS_PROXY,env=https_proxy \
     git clone $(echo ${FLASH_ATTN_URL} | cut -d '@' -f 1) && \
     cd ${CODESPACE}/flash-attention && \
+    mkdir -p ${CODESPACE}/flash-attention/third_party/nvidia/backend/bin && \
+    ln -snf ${CUDA_HOME:-/usr/local/cuda}/bin/nvcc ${CODESPACE}/flash-attention/third_party/nvidia/backend/bin/nvcc && \
     git checkout $(echo ${FLASH_ATTN_URL} | cut -d '@' -f 2) && \
     git submodule update --init --recursive --force
 
 WORKDIR ${CODESPACE}/flash-attention
 
-RUN cd hopper && FLASH_ATTENTION_FORCE_BUILD=TRUE pip wheel -w ${FLASH_ATTN3_DIR} -v --no-deps .
-RUN FLASH_ATTENTION_FORCE_BUILD=TRUE pip wheel -w ${FLASH_ATTN_DIR} -v --no-deps .
+# CUDA 13 moves CCCL headers into include/cccl, which is missing from default host include paths
+# nvcc 13 falls back to PATH lookup for cicc when invoked from third_party/nvidia/backend
+RUN export PATH="${CUDA_HOME:-/usr/local/cuda}/nvvm/bin:${PATH}" && \
+    CUDA_MAJOR="$(python -c 'import torch; print(torch.version.cuda.split(".", 1)[0])')" && \
+    if [ "${CUDA_MAJOR}" = "13" ]; then \
+        export CPLUS_INCLUDE_PATH="${CUDA_HOME:-/usr/local/cuda}/include/cccl":${CPLUS_INCLUDE_PATH}; \
+    fi && \
+    cd hopper && FLASH_ATTENTION_FORCE_BUILD=TRUE pip wheel --no-build-isolation -w ${FLASH_ATTN3_DIR} -v --no-deps .
+RUN CUDA_MAJOR="$(python -c 'import torch; print(torch.version.cuda.split(".", 1)[0])')" && \
+    if [ "${CUDA_MAJOR}" = "13" ]; then \
+        export CPLUS_INCLUDE_PATH="${CUDA_HOME:-/usr/local/cuda}/include/cccl":${CPLUS_INCLUDE_PATH}; \
+    fi && \
+    FLASH_ATTENTION_FORCE_BUILD=TRUE pip wheel --no-build-isolation -w ${FLASH_ATTN_DIR} -v --no-deps .
 
 # compile adaptive_gemm
 FROM setup_env AS adaptive_gemm
@@ -85,7 +109,7 @@ RUN --mount=type=secret,id=HTTPS_PROXY,env=https_proxy \
 
 WORKDIR ${CODESPACE}/AdaptiveGEMM
 
-RUN pip wheel -w ${ADAPTIVE_GEMM_DIR} -v --no-deps .
+RUN pip wheel --no-build-isolation -w ${ADAPTIVE_GEMM_DIR} -v --no-deps .
 
 # compile grouped_gemm(permute and unpermute)
 FROM setup_env AS grouped_gemm
@@ -102,7 +126,7 @@ RUN --mount=type=secret,id=HTTPS_PROXY,env=https_proxy \
 
 WORKDIR ${CODESPACE}/GroupedGEMM
 
-RUN pip wheel -w ${GROUPED_GEMM_DIR} -v --no-deps .
+RUN pip wheel --no-build-isolation -w ${GROUPED_GEMM_DIR} -v --no-deps .
 
 # compile causal_conv1d
 FROM setup_env AS causal_conv1d
@@ -119,7 +143,7 @@ RUN --mount=type=secret,id=HTTPS_PROXY,env=https_proxy \
 
 WORKDIR ${CODESPACE}/causal-conv1d
 
-RUN CAUSAL_CONV1D_FORCE_BUILD=TRUE pip wheel -w ${CAUSAL_CONV1D_DIR} -v --no-deps --no-build-isolation .
+RUN CAUSAL_CONV1D_FORCE_BUILD=TRUE pip wheel --no-build-isolation -w ${CAUSAL_CONV1D_DIR} -v --no-deps .
 
 # compile nvshmem and deepep
 FROM setup_env AS deep_ep
@@ -156,8 +180,17 @@ RUN --mount=type=secret,id=HTTPS_PROXY,env=https_proxy \
 WORKDIR ${CODESPACE}/DeepEP
 
 RUN --mount=type=secret,id=HTTPS_PROXY,env=https_proxy \
-    (pip show nvidia-nvshmem-cu12 >/dev/null 2>&1 || pip install nvidia-nvshmem-cu12==3.4.5) && \
-    pip wheel -w ${DEEP_EP_DIR} -v --no-deps .
+    CUDA_MAJOR="$(python -c 'import torch; print(torch.version.cuda.split(".", 1)[0])')" && \
+    if [ "${CUDA_MAJOR}" = "13" ]; then \
+        pip show nvidia-nvshmem-cu13 >/dev/null 2>&1 || pip install nvidia-nvshmem-cu13==3.4.5; \
+        export CPLUS_INCLUDE_PATH="${CUDA_HOME:-/usr/local/cuda}/include/cccl":${CPLUS_INCLUDE_PATH}; \
+    elif [ "${CUDA_MAJOR}" = "12" ]; then \
+        pip show nvidia-nvshmem-cu12 >/dev/null 2>&1 || pip install nvidia-nvshmem-cu12==3.4.5; \
+    else \
+        echo "Unsupported CUDA major version: ${CUDA_MAJOR}" >&2; \
+        exit 1; \
+    fi && \
+    pip wheel --no-build-isolation -w ${DEEP_EP_DIR} -v --no-deps .
 
 # compile deep_gemm
 FROM setup_env AS deep_gemm
@@ -174,7 +207,7 @@ RUN --mount=type=secret,id=HTTPS_PROXY,env=https_proxy \
 
 WORKDIR ${CODESPACE}/DeepGEMM
 
-RUN pip wheel -w ${DEEP_GEMM_DIR} -v --no-deps .
+RUN pip wheel --no-build-isolation -w ${DEEP_GEMM_DIR} -v --no-deps .
 
 # integration xtuner
 FROM setup_env AS xtuner_dev
@@ -259,13 +292,15 @@ RUN --mount=type=secret,id=HTTPS_PROXY,env=https_proxy \
 
 # nccl update for torch 2.6.0
 # RUN --mount=type=secret,id=HTTPS_PROXY,env=https_proxy \
-RUN if [ "x${TORCH_VERSION}" = "x2.6.0" ]; then \
+RUN if [ "x${TORCH_VERSION}" = "x2.6.0" ] && \
+       pip show nvidia-nccl-cu12 >/dev/null 2>&1; then \
         pip install nvidia-nccl-cu12==2.25.1 --no-cache-dir -i ${DEFAULT_PYPI_URL}; \
     fi
 
 # cudnn update for torch 2.9.1
 # RUN --mount=type=secret,id=HTTPS_PROXY,env=https_proxy \
-RUN if [ "x${TORCH_VERSION}" = "x2.9.1" ]; then \
+RUN if [ "x${TORCH_VERSION}" = "x2.9.1" ] && \
+       pip show nvidia-cudnn-cu12 >/dev/null 2>&1; then \
         pip install nvidia-cudnn-cu12==9.15.1.9 --no-cache-dir -i ${DEFAULT_PYPI_URL}; \
     fi
 

autotest/config.yaml

@@ -921,7 +921,7 @@ case:
                 output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
             resource:
                 memory_per_task: 1200
-                pip_package: pip install sglang==0.5.10.post1 --timeout 600 --retries 5;pip install transformers==5.2.0 apache-tvm-ffi==0.1.9 nvidia-cudnn-cu12==9.15.1.9;pip uninstall -y nvidia-cutlass-dsl;rm -rf /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl
+                pip_package: pip install sglang==0.5.10.post1 --timeout 600 --retries 5;pip install transformers==5.14.1 apache-tvm-ffi==0.1.9 nvidia-cudnn-cu12==9.15.1.9;pip uninstall -y nvidia-cutlass-dsl;rm -rf /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl
                 envs:
                     - MODEL_PATH=/mnt/shared-storage-user/llmrazor-share/model/Qwen3-30B-A3B
                     - DATA_PATH=/mnt/shared-storage-user/llmrazor-share/data/gsm8k/train-mini.jsonl

image_build.sh

@@ -1,19 +1,19 @@
 export HTTPS_PROXY=$HTTPS_PROXY
 
-export BASE_IMAGE=nvcr.io/nvidia/pytorch:25.03-py3
+export BASE_IMAGE=nvcr.io/nvidia/pytorch:25.11-py3
 export XTUNER_COMMIT=$(git rev-parse HEAD)
 export XTUNER_URL=https://github.com/InternLM/xtuner@${XTUNER_COMMIT}
 export FLASH_ATTN_URL=https://github.com/Dao-AILab/flash-attention@060c9188beec3a8b62b33a3bfa6d5d2d44975fab
-export ADAPTIVE_GEMM_URL=https://github.com/InternLM/AdaptiveGEMM@10411e08b182e853d0f3ecec4c68bf90c90e309f # fix fp8 dw k_grouped_gemm bug
+export ADAPTIVE_GEMM_URL=https://github.com/InternLM/AdaptiveGEMM@c1a0650d3c82a626378fcc76f0022d88b6688245 # fix fp8 dw k_grouped_gemm bug
 export GROUPED_GEMM_URL=https://github.com/InternLM/GroupedGEMM@aa5ffb21cb626d6cd61d99fc42958127b0b99be7
 export DEEP_EP_URL=https://github.com/deepseek-ai/DeepEP@9af0e0d0e74f3577af1979c9b9e1ac2cad0104ee # v1.2.1
 export DEEP_GEMM_URL=https://github.com/deepseek-ai/DeepGEMM@c9f8b34dcdacc20aa746b786f983492c51072870 # v2.1.1.post3
 export CAUSAL_CONV1D_URL=https://github.com/Dao-AILab/causal-conv1d@da6dbaa9fd5a919967f14d3fd031da1288ad5025 # v1.6.0
 export FLA_URL="${FLA_URL-https://github.com/HAOCHENYE/flash-linear-attention@tmp-tensor-cache}"
 
-export TORCH_VERSION=${TORCH_VERSION:-"2.9.1"}
-export LMDEPLOY_URL="${LMDEPLOY_URL:-https://github.com/RunningLeon/lmdeploy@c14f389e9b2346f9835954ed90aefc0a0aac7ad4}"
-# export LMDEPLOY_VERSION="0.15.0"
+export TORCH_VERSION=${TORCH_VERSION:-"2.12.1"}
+export LMDEPLOY_VERSION="0.14.0"
+# export LMDEPLOY_URL=https://github.com/InternLM/lmdeploy@efe3b88607756a7ad9411b89627b5ac6ebaa540e
 export PPA_SOURCE="https://mirrors.aliyun.com"
 export DEFAULT_PYPI_URL=${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_VERSION=$LMDEPLOY_VERSION \
   --progress=plain \
   --label "BASE_IMAGE=$BASE_IMAGE" \
   --label "XTUNER_URL=${XTUNER_URL/@/\/tree\/}" \
@@ -52,8 +52,5 @@ docker build . \
   --label "FLA_URL=${FLA_URL/@/\/tree\/}" \
   --label "DEEP_EP_URL=${DEEP_EP_URL/@/\/tree\/}" \
   --label "DEEP_GEMM_URL=${DEEP_GEMM_URL/@/\/tree\/}" \
-  --label "LMDEPLOY_URL=${LMDEPLOY_URL/@/\/tree\/}"
-
-echo "===== pip list in ${image_name}:${image_tag} ====="
-docker run --rm --entrypoint pip "${image_name}:${image_tag}" list
-
+  --label "LMDEPLOY_VERSION=$LMDEPLOY_VERSION"
+  # --label "LMDEPLOY_URL=${LMDEPLOY_URL/@/\/tree\/}"

pyproject.toml

@@ -24,7 +24,6 @@ classifiers = [
 ]
 dependencies = [
   "astor",
-  "bitsandbytes==0.45.0",
   "datasets<4.0.0",
   "einops",
   "loguru",
@@ -77,11 +76,13 @@ rl = [
   "pylatexenc"
 ]
 video = [
+  "torchcodec==0.15.0",
   "decord",
   "av",
 ]
 all = [
   "jsonlines",
+  "torchcodec==0.15.0",
   "decord",
   "av",
   "ray[default]",

requirements/runtime.txt

@@ -1,4 +1,3 @@
-bitsandbytes==0.45.0
 datasets<4.0.0
 einops
 loguru
@@ -11,7 +10,7 @@ SentencePiece
 tiktoken
 torch>=2.6.0
 torchvision
-transformers==4.56.0
+transformers==5.14.1
 cyclopts
 transformers_stream_generator
 opencv-python-headless