docker: keep the cu12 dependency markers after checking out sglang-miles

[yueming-yuan]

What breaks

Docker Build & Push on main fails at Install sglang-miles:

#38 [sglang 33/39] RUN cd /sgl-workspace/sglang && git fetch origin sglang-miles && ...
 * branch            sglang-miles -> FETCH_HEAD
error: Your local changes to the following files would be overwritten by checkout:
	python/pyproject.toml
Please commit your changes or stash them before you switch branches.

First failure: run 30491453615 (the first release build after #1795).

Why

sglang's own Dockerfile rewrites the dependency markers in place for CUDA 12 images and never commits the result:

# sgl-project/sglang, docker/Dockerfile @ v0.5.16, L713
cd /sgl-workspace/sglang \
&& if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
       sed -i 's/cuda-python>=13\.0/cuda-python>=12,<13/' python/pyproject.toml && \
       sed -i 's/flashinfer_python\[cu13\]/flashinfer_python[cu12]/' python/pyproject.toml && \
       sed -i 's/nvidia-cutlass-dsl\[cu13\]/nvidia-cutlass-dsl/' python/pyproject.toml; \
   fi \
&& python3 -m pip install --no-deps -e "python[${BUILD_TYPE}]"

So v0.5.16-cu129 ships a dirty working tree and our checkout aborts. The release build is the only path that hits this: it pins SGLANG_IMAGE_TAG=v0.5.16-cu129 in docker/build.py, while PR CI uses the default v0.5.16 (cu130), which builds with CUDA_VERSION=13.0.1 and never runs that rewrite. v0.5.15-cu129 was clean, so this is new in the v0.5.16 CUDA 12 image rather than a regression in our tree.

The fix

Force the checkout, then re-apply the same three rewrites when ENABLE_CUDA_13 != 1.

Forcing on its own is not enough. The tree we check out carries the cu13 markers, pip install -e publishes them as sglang's installed metadata, and every later pip resolve then sees unsatisfied cu13 requirements on a cu12 image — the same mechanism that silently downgraded cuDNN in #1836.

Checks

• The three sed patterns match sglang-miles's python/pyproject.toml (lines 26, 34, 48) and rewrite exactly those three lines.
• humming-kernels[cu13] is left alone, matching upstream — sglang's own build does not rewrite it either.
• sglang-miles differs from the v0.5.16 tag in this file only by adding xxhash, so after the rewrite the tree matches the base image's dependency markers plus that one addition.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[yueming-yuan]

Verified against the real lmsysorg/sglang:v0.5.16-cu129 image on a CPU build box, not just read off upstream's Dockerfile.

The dirty file is exactly the three markers

$ git -C /sgl-workspace/sglang status --porcelain
 M python/pyproject.toml

$ git diff -- python/pyproject.toml
-  "cuda-python>=13.0",
+  "cuda-python>=12,<13",
-  "flashinfer_python[cu13]==0.6.14",
+  "flashinfer_python[cu12]==0.6.14",
-  "nvidia-cutlass-dsl[cu13]==4.6.0",
+  "nvidia-cutlass-dsl==4.6.0",

$ git log --oneline -1
fdebc93 [Cherry-pick to release/v0.5.16] Fix stale flashinfer-MLA fallback poisoning spec verify capture (#32288) (#32346)

The failure reproduces verbatim

$ git fetch origin sglang-miles && git checkout FETCH_HEAD
error: Your local changes to the following files would be overwritten by checkout:
	python/pyproject.toml
Please commit your changes or stash them before you switch branches.
Aborting

This PR's sequence resolves it

Running the new block with ENABLE_CUDA_13=0 inside that image:

checkout+sed OK   HEAD=3003d70 [sglang-miles] ...

26:  "cuda-python>=12,<13",
34:  "flashinfer_python[cu12]==0.6.14",
37:  "humming-kernels[cu13]==0.1.10",     <- left alone, as upstream does
48:  "nvidia-cutlass-dsl==4.6.0",
90:  "xxhash",  # /pull_weights delta checksum   <- sglang-miles' own addition, preserved

The tree ends up at sglang-miles with the cu12 markers restored, so the cu13 specifiers never reach sglang's installed metadata on a cu12 image.

Note this PR's own CI exercises the cu130 path only; the cu129 path is covered by the run above and by the release build after merge.

[guapisolo]

approved to unblock.