| 1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071 |
- # Authoritative build inputs — karmic-kraken (local-inference-lab vLLM fork)
- # clean build. Provenance: dev/karmic-kraken tip 636afde9 (2026-10-08), which
- # takes B12X from FlashInfer 0.7.1 (the standalone b12x repo is archived).
- # Supersedes the sparkring dev-20261001-kraken provenance (vllm ab86b7073,
- # b12x e4084d2e, flashinfer 36ad74ad).
- #
- # Every entry is a tag, exact version, branch-name at a pinned commit, or a
- # verified tree hash — no :latest, no ranges, no floating tips.
- #
- # Reproducibility contract (clone-at-pin, NO staged tarballs): the Dockerfile
- # clones every source at the pins below inside the build and GATES each one
- # (HEAD commit + tree hash must match, pinned deps must appear verbatim in the
- # tree's cmake). A fresh clone of this repository + `bash build.sh` reproduces
- # the image.
- # CUDA base (aarch64; the sparkring image uses CUDA 13.4.2)
- CUDA_BASE_IMAGE=nvidia/cuda:13.4.2-devel-ubuntu24.04
- CUDA_BASE_DIGEST=sha256:d44d6dc249c2c8330d0b380c595a514f0e86137fcc717fe7a4389a25720ade37
- # Python / uv
- UV_VERSION=0.11.28
- PYPI_INDEX_URL=https://pypi.org/simple
- PYTORCH_INDEX_URL=https://download.pytorch.org/whl/cu130
- # apt mirror for the apt-base stage: current = live ports.ubuntu.com
- APT_SOURCE=current
- # DeepGEMM — cloned in-build; pin gated in fetch-deps.sh against the tree's
- # own deepgemm.cmake declaration.
- DEEPGEMM_REPO=https://github.com/deepseek-ai/DeepGEMM.git
- DEEPGEMM_BRANCH=nv_dev
- DEEPGEMM_REF=8b1392b978f5a03c828dd1711090d7fb50958b8a
- # Torch stack — the fork's requirements/cuda.txt pins
- TORCH_VERSION=2.13.0
- TORCHVISION_VERSION=0.28.0
- TORCHAUDIO_VERSION=2.11.0
- # FlashInfer — 0.7.1 line (the karmic-kraken tip's pin; B12X now ships inside it).
- # No FINAL 0.7.1 wheel exists yet (the line is at 0.7.1rc5, and 0.7.2rc1 is
- # already out) — so we pin the latest 0.7.1-line wheel, 0.7.1rc5, from the
- # flashinfer.ai/whl index. FlashInfer is NOT a core vllm-wheel dependency
- # (only the `vllm[b12x]` extra pulls it), and we install it separately in
- # torch-base, so the rc pin is safe. Its commit is v0.7.1rc5 = 9a39cddc.
- FLASHINFER_VERSION=0.7.1rc5
- FLASHINFER_COMMIT=9a39cddccee9854a00b2824e64c937d37ad316d7
- # vLLM — karmic-kraken fork (local-inference-lab/vllm@dev/karmic-kraken)
- VLLM_REPO=https://github.com/local-inference-lab/vllm.git
- VLLM_BRANCH=dev/karmic-kraken
- VLLM_PIN=636afde98300e1492d7bfc5665f8dcfaec8b706e
- VLLM_PRETEND_VERSION=0.1.dev21808+g636afde98.d20261008
- # B12X — separate top-level `b12x` package (PyPI, from the archived
- # local-inference-lab/b12x repo). The karmic-kraken tip's requirements/cuda.txt
- # *assumes* b12x ships inside flashinfer, but the latest flashinfer wheel on
- # PyPI (0.7.1rc5) does NOT bundle it (verified: no b12x/ dir, no
- # flashinfer/experimental/b12x) — that bundling lands in a not-yet-released
- # flashinfer. So we install the standalone `b12x` wheel next to flashinfer
- # 0.7.1rc5 (no namespace conflict: rc5 carries no b12x). vLLM imports it as a
- # top-level package (`from b12x.comm.pcie import ...`); 1.5.0 provides
- # b12x/comm/pcie + b12x/preparation.
- B12X_VERSION=1.5.0
- # NCCL — the sparkring image's pin (2.32.3, rev 12df1a11)
- NCCL_VERSION=2.32.3
- NCCL_REVISION=12df1a11afad322be5a204a2db890161cbf8131d
- # Target GPU architecture (DGX Spark = sm_121a)
- TORCH_CUDA_ARCH_LIST=12.1a
- BUILD_JOBS=16
|
