| 12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758 |
- # Authoritative build inputs — karmic-kraken (local-inference-lab vLLM fork)
- # clean build. Provenance: sparkring dev-20261001-kraken-cuda1342-nccl2323-status034
- # (build-metadata.yaml: 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 — the sparkring image's pin (commit 36ad74ad, version 0.7.1)
- FLASHINFER_VERSION=0.7.0.post1
- FLASHINFER_COMMIT=36ad74addfa1260f8f368d650ade47c856fa90e1
- # 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=ab86b70734009c34e91db456bb96340c3faf3d4e
- VLLM_PRETEND_VERSION=0.1.dev21553+gab86b7073.d20261001
- # B12X — lukealonso/b12x@master (the sparkring image's pin)
- B12X_REPO=https://github.com/lukealonso/b12x.git
- B12X_PIN=e4084d2eef4932e0fa06db3f7a94deb83ad132d7
- B12X_PRETEND_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
|
