fix(docker): pin transformer_engine to 2.16.0 for 1.39.0

Temporarily pin for the 1.39.0 release: TE 2.17 has no prebuilt
transformer_engine_torch wheels and fails to compile from source.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
suluyan
2026-07-20 10:11:46 +08:00
parent 5ee9b27220
commit 4f223056cd

View File

@@ -71,7 +71,7 @@ RUN if [ "$IMAGE_TYPE" = "gpu" ]; then \
pip install --no-cache-dir math_verify "gradio<5.33" "deepspeed<0.19" ray -U && \
pip install --no-cache-dir mcore-bridge -i https://pypi.org/simple/ -U && \
pip install --no-cache-dir liger_kernel wandb swanlab nvitop pre-commit "transformers<5.9" "trl<1.0" "peft<0.20" huggingface-hub -U && \
pip install --no-cache-dir --no-build-isolation transformer_engine[pytorch]; \
pip install --no-cache-dir --no-build-isolation "transformer_engine[pytorch]==2.16.0"; \
cd /tmp && GIT_LFS_SKIP_SMUDGE=1 git clone https://github.com/NVIDIA/apex && \
cd apex && pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" ./ && \
cd / && rm -fr /tmp/apex && pip cache purge; \