| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869 |
- # Authoritative build inputs — GLM-5.3-Flash r27 (jovian-judgement) clean build.
- # 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. Provenance: local-inference-lab/rtx6kpro master
- # models/glm-5.3-flash/build/glm53-jovian-judgement-community-20260906-r27.source.lock
- # (embedded at /opt/glm53-flash/source.lock in their image).
- #
- # Runtime deviation from the RTX PRO 6000 lock (documented, GB10 adaptation):
- # their qualified runtime = CUDA 13.3 / flashinfer 0.6.18+cu133 / NCCL 2.31.2
- # on x86_64 sm_120; ours = CUDA 13.2 devel aarch64 / torch 2.13.0+cu130 /
- # flashinfer 0.6.17 — the SAME pins the r27 tree's requirements/cuda.txt asks
- # for (torch==2.13.0, flashinfer-python==0.6.17), proven on GB10 sm_121a.
- # CUDA base (aarch64; digest-pinned, tested by solar-open2-nvfp4 / r18)
- CUDA_BASE_IMAGE=nvidia/cuda:13.2.0-devel-ubuntu24.04
- CUDA_BASE_DIGEST=sha256:a5b6256e470196fc1d5f8f62139d57d3662867746dfe1cb352d7652024047020
- # Python / uv
- UV_VERSION=0.11.28
- PYPI_INDEX_URL=https://pypi.org/simple
- PYTORCH_INDEX_URL=https://download.pytorch.org/whl/cu130
- # DeepGEMM — cloned in-build; pin gated in fetch-deps.sh against the tree's
- # own deepgemm.cmake declaration (same pin in the r27 tree).
- 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 (same in r27)
- TORCH_VERSION=2.13.0
- TORCHVISION_VERSION=0.28.0
- TORCHAUDIO_VERSION=2.11.0
- # FlashInfer — the fork's cuda.txt pin (0.6.17), unchanged from r18
- FLASHINFER_VERSION=0.6.17
- # vLLM — r27 source lock: pinned by COMMIT + tree hash (branch name
- # integration/glm53-r27-release-20260906 in comments only — branches get
- # renamed, the SHA is the contract; merge order documented in vLLM issue 651).
- VLLM_REPO=https://github.com/voipmonitor/vllm.git
- # branch integration/glm53-r27-release-20260906
- VLLM_PIN=63a82f8d323e8538cbe6f88ae1812a1c01577a0f
- VLLM_TREE=f54cd9ca2b9434727715197d32150b75e82a9ebf
- VLLM_PRETEND_VERSION=0.1.dev0+glm53.r27.63a82f8d
- # B12X — r27 source lock, same SHA+tree pinning (branch
- # integration/glm53-r27-release-20260906).
- B12X_REPO=https://github.com/voipmonitor/b12x.git
- B12X_PIN=e8ad299b174f16e2e8fb5879bea272f4efbb53f2
- B12X_TREE=f3cd8a9eb00d3226a1acbbed1efedf10cc1c3e71
- B12X_PRETEND_VERSION=0.1.dev0+glm53b12x.r27.e8ad299b
- # LMCache — r27 source lock: pinned by COMMIT + tree hash (the lock's
- # integration branch integration/glm53-r26-current-head-20260905 was renamed
- # upstream on 2026-09-06 — the pin is the commit; branch names are not the
- # contract). Qualified tree = 008ac3e0, verified at clone time.
- LMCACHE_REPO=https://github.com/local-inference-lab/LMCache.git
- LMCACHE_PIN=822330db1160d2847538d53f9e6f4eeee2404f4d
- LMCACHE_TREE=008ac3e09ae5917aa0849147480d7bd5b9f8b37a
- LMCACHE_VERSION=0.5.5.dev0+glm53cache.engine.822330db
- # Target GPU architecture (DGX Spark = sm_121a) — also used by the LMCache build.
- TORCH_CUDA_ARCH_LIST=12.1a
- BUILD_JOBS=4
|
