Merge branch 'main' into qwen3vl-cp-vision-grad-fix

Author: going-song

docker/Dockerfile.ci

@@ -95,6 +95,8 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
         CXXFLAGS="-I/usr/local/cuda/include/cccl${CXXFLAGS:+ $CXXFLAGS} -DNDEBUG" && \
     # Reinstall nvidia-cutlass-dsl in system site packages
     # uv install can run into non-deterministic install issues
+    # Keep this pin compatible with cuDNN Frontend; mismatched versions can break
+    # its CUTLASS-backed kernels.
     pip uninstall -y nvidia-cutlass-dsl nvidia-cutlass-dsl-libs-base nvidia-cutlass-dsl-libs-cu13 && \
     pip install nvidia-cutlass-dsl[cu13]==4.5.0 && \
     if [ "$MCORE_TRIGGERED_TESTING" = "true" ]; then \

docker/common/fw_pyproject.toml

@@ -103,7 +103,7 @@ override-dependencies = [
     "nvidia-cusparse-cu12; sys_platform == 'never'",
     "nvidia-cusparselt-cu12; sys_platform == 'never'",
     "nvidia-nccl-cu12; sys_platform == 'never'",
-    "nvidia-cudnn-frontend==1.24.0",
+    "nvidia-cudnn-frontend==1.26.0",
     "nvidia-cutlass-dsl[cu13]; sys_platform == 'never'", # Rely on system site packages install
     "tensorrt-llm; sys_platform == 'never'",
     "tensorrt; sys_platform == 'never'",

pyproject.toml

@@ -140,7 +140,7 @@ override-dependencies = [
     "onnx>=1.21.0; python_version < '3.13'",                # onnx<1.21 has missing lower bound on ml-dtypes, uses float4_e2m1fn (requires ml-dtypes>=0.5.0)
     # TODO: Remove this cap once huggingface/transformers#46381 is merged and included in a Transformers release.
     "tokenizers>=0.22.0,<0.23",                             # tokenizers 0.23 renames RobertaProcessing's 'cls' kwarg to 'cls_token'; transformers still passes cls= (breaks CLIPTokenizer)
-    "nvidia-cudnn-frontend==1.24.0",
+    "nvidia-cudnn-frontend==1.26.0",
     "nvidia-cutlass-dsl[cu13]; sys_platform == 'never'",    # Rely on system site packages install
     "fast-hadamard-transform @ git+https://github.com/Dao-AILab/fast-hadamard-transform.git@v1.1.0"  # pin 1.1.0; PyPI sdist is incomplete, install from git tag (see DSv4/GLM-5 READMEs)
 ]