[deps] 6/n towards Kimi K2.6: CUDA-13 deploy stack for Blackwell-Ultra (sm103 / B300)

[casper-hansen]

What

Ships the Blackwell-Ultra (sm103 / B300) dependency stack as deploy/cuda13/: a uv-exported pinned requirements file, the mirrored override set for the uv pip interface, and a three-phase installer that handles the sdist builds (transformer-engine-torch needs an in-tree build for its vendored build_tools helper). CUDA-12 defaults (pyproject.toml, uv.lock) are untouched.

Why

The cu12x vLLM wheels ship sm_100a/sm_120a SASS only (no sm103, no PTX), so every compiled vLLM kernel fails with "no kernel image is available" on B300. Under CUDA 13, plain sm_100 cubins are family-compatible with sm103 (verified by driver-loading the PyPI wheel's cubins on a B300), and TE's cu13 core ships explicit sm_103a kernels.

A default-preserving CUDA-variant extra cannot be expressed in one uv project (extra/group markers do not bind in sources or overrides, group-scoped sources force the variant packages into default groups, and a sub-project wrapper inherits the parent's source mappings), hence a deploy manifest.

Notable pins: torch 2.11.0+cu130 (PyPI default), vllm 0.23.0 (PyPI wheel is the CUDA-13 build), TE core cu13, flashinfer-jit-cache cu130, causal-conv1d / mamba-ssm from sdist, flash-attn removed (no torch-2.11/cu13 wheels; a CUDA-less stub crashes TE -- TE uses cuDNN fused attention and HF models fall back to SDPA).

Verified on B300: a from-scratch install resolves to a package set identical to the uv.lock-managed venv that completed multi-step GRPO training; Kimi bridge GPU tests pass 3/3 against it.

Part of the Kimi K2.x series (follow-up to #1862). Independent of the other PRs in the series.

Made with Cursor

[gemini-code-assist]

Code Review

This pull request introduces a CUDA-13 stack deployment configuration tailored for Blackwell-Ultra GPUs (sm103), including installation scripts, dependency overrides, and pinned requirements. The review feedback focuses on improving the robustness of the deployment scripts: specifically, checking for an empty 'TE_TORCH_VERSION' and using the '-f' flag with 'curl' in 'install.sh' to handle network failures gracefully, as well as explicitly specifying 'utf-8' encoding when opening files in the Python block of 'regenerate.sh' to ensure cross-platform compatibility.

[gemini-code-assist]

To improve robustness and error diagnostics:

1. Check if TE_TORCH_VERSION is empty: If transformer-engine-torch is missing or formatted differently in requirements-megatron.txt, TE_TORCH_VERSION will be empty, leading to an invalid PyPI API request.
2. Use curl -sfSL instead of curl -sL: By default, curl exits with status 0 even on HTTP errors (like 404 or 500). Adding the -f/--fail flag ensures curl exits with a non-zero status code on failure, which triggers pipefail and set -e to halt the script immediately with a clear exit code rather than passing empty/HTML content to Python or tar.

Suggested change

	TE_TORCH_VERSION="$(sed -n 's/^transformer-engine-torch==\([0-9.]*\).*/\1/p' "$REQ" | head -1)"
	SDIST_URL="$(curl -sL "https://pypi.org/pypi/transformer-engine-torch/$TE_TORCH_VERSION/json" \
	| "$PY" -c "import json,sys; print([u['url'] for u in json.load(sys.stdin)['urls'] if u['packagetype']=='sdist'][0])")"
	curl -sL "$SDIST_URL" | tar xz
	TE_TORCH_VERSION="$(sed -n 's/^transformer-engine-torch==\\([0-9.]*\\).*/\\1/p' "$REQ" | head -1)"
	if [ -z "$TE_TORCH_VERSION" ]; then
	echo "Error: Could not extract transformer-engine-torch version from $REQ" >&2
	exit 1
	fi
	SDIST_URL="$(curl -sfSL "https://pypi.org/pypi/transformer-engine-torch/$TE_TORCH_VERSION/json" \\
	| "$PY" -c "import json,sys; print([u['url'] for u in json.load(sys.stdin)['urls'] if u['packagetype']=='sdist'][0])")"
	curl -sfSL "$SDIST_URL" | tar xz

[casper-hansen]

Done in 60a9c6d — empty-pin check with a clear error, and curl -sfSL on both fetches so HTTP errors stop the script instead of piping error pages into python/tar.

[gemini-code-assist]

Specify encoding="utf-8" when opening files in Python. This ensures cross-platform consistency and prevents potential UnicodeDecodeError on environments where the default system locale encoding is not UTF-8 (e.g., some minimal Docker containers or Windows environments).

Suggested change

	path = sys.argv[1]
	lines = open(path).read().splitlines(True)
	out, skip_via = [], False
	for line in lines:
	# Only the flashinfer cu130 index is used by these pins; other emitted
	# extra indexes (cu128/cu129/...) would reintroduce ambiguity.
	if line.startswith("--extra-index-url") and "flashinfer.ai/whl/cu130" not in line:
	continue
	# The editable gym path is cwd-sensitive; install.sh installs it (and
	# skyrl itself) explicitly with absolute paths instead.
	if line.startswith("-e ./skyrl-gym"):
	skip_via = True
	continue
	if skip_via and line.strip().startswith("# via"):
	skip_via = False
	continue
	skip_via = False
	out.append(line)
	open(path, "w").writelines(out)
	path = sys.argv[1]
	lines = open(path, encoding="utf-8").read().splitlines(True)
	out, skip_via = [], False
	for line in lines:
	# Only the flashinfer cu130 index is used by these pins; other emitted
	# extra indexes (cu128/cu129/...) would reintroduce ambiguity.
	if line.startswith("--extra-index-url") and "flashinfer.ai/whl/cu130" not in line:
	continue
	# The editable gym path is cwd-sensitive; install.sh installs it (and
	# skyrl itself) explicitly with absolute paths instead.
	if line.startswith("-e ./skyrl-gym"):
	skip_via = True
	continue
	if skip_via and line.strip().startswith("# via"):
	skip_via = False
	continue
	skip_via = False
	out.append(line)
	open(path, "w", encoding="utf-8").writelines(out)

[casper-hansen]

Done in 60a9c6d.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{Reviewed by Cursor Bugbot for commit 5902652. Configure here.}