| 12345678910111213141516171819202122232425262728 |
- # DeepSeek-V4.1-Flash (552B MoE, MXFP4 experts, 1M ctx) — DGX Spark image pins.
- #
- # Provenance: tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark (2026-09-10,
- # MIT) — their overlay1/3/5 chain, re-expressed as one clone-at-pin Dockerfile:
- # base vllm/vllm-openai nightly image (digest-pinned, arm64)
- # vLLM vllm-project/vllm @ dsv41-feat (day-0 DSV4.1 support, commit-pinned)
- # FI FlashInfer 0.7.0rc1 from source (pinned commit + submodules)
- # ext _C_stable_libtorch rebuilt for sm_121a (their build_stable_ext.sh)
- # patches the 7 boot-10 file replacements, baked + md5-gated (their manifest)
- #
- # versions.env convention: KEY=VAL, no spaces.
- VLLM_BASE_IMAGE=docker.io/vllm/vllm-openai@sha256:a551e05307cd2e0092139d84db32af9c97e67d2eeeff072d21e429131d8c23f0
- VLLM_REPO=https://github.com/vllm-project/vllm.git
- VLLM_REF=e47aa780bccf59f59dfa2cbb18e17a10b4fe69ba
- VLLM_REF_TREE=8bf2b4034505699b829f9fcc1a8cc8e041c49384
- VLLM_PRETEND_VERSION=0.1.dev0+dsv41.e47aa780
- # _C_stable_libtorch stable-only build (external_projects includes commented out)
- CUTLASS_TAG=v4.7.1
- CUTLASS_URL=https://codeload.github.com/NVIDIA/cutlass/tar.gz/refs/tags/v4.7.1
- # FlashInfer 0.7.0rc1 from source (0.6.18 in the base has no SM120 sparse-MLA
- # decode kernel for page_block_size=32 — their boot-1 death)
- FI_SHA=07869c61ba581e6d6b8ad8d142f4a6c89b707cc1
- FI_CUTLASS_SHA=b46b16d003484063bca4ed365e44095c4c6ed633
- FI_CCCL_SHA=16bd510c9b712e82b0ab6cbb630d8e29ba1f7116
- FI_SPDLOG_SHA=c3aed4b68373955e1cc94307683d44dca1515d2b
|
