| 1234567891011121314151617181920212223242526272829303132333435363738394041424344454647 |
- # Version profile: DeepSeek-V4.1-Flash r36 (jovian-judgement, qualified 2026-09-12).
- #
- # The lab's FIRST published DS4.1 lock: rtx6kpro models/deepseek-v4.1-flash.md
- # ("Status: qualified" for TP4/DCP1 startup, seeded sampling, text/Vision
- # smoke, bounded throughput; image localinferencelab/vllm:
- # jovian-judgement-community-20260912-r36 @ sha256:23ab683d...). The DS4.1
- # serving profile is native DSpark K7 (weights inside the target checkpoint,
- # greedy draft, standard rejection, adaptive verification), B12X
- # attention/MoE/linear, 131K context, Engram ngram storage RAM/disk.
- #
- # The image is two-layer: the GLM entrypoint (our r35 lane) + the DS4.1
- # entrypoint (serve-ds41-jovian.sh over the bundled serve-ds41-flash.sh).
- # We build the same source composition here for GB10 sm_121a (the published
- # image's native extension is compiled for sm_120 / CUDA 13.3 x86).
- #
- # Pinned by commit + tree hash (gated per repo in the Dockerfile). What moved
- # over the r35 profile (r35.env):
- # vLLM de982a50 -> 202a11a98eb (JJ b40673cd + #740 startup correction;
- # 0.26.1rc0+glm53.r36.vllm202a11a9; adds the DS4.1 DSpark/Engram
- # serving path, greedy-Markov sampler fusion #738)
- # B12X 98086604 -> 323107ff (clean master; includes the DS4.1 native
- # kernels)
- # LMCache 29bc5a2e — UNCHANGED (same as r35)
- # FlashKDA 3b225bf2 + in-tree patch — UNCHANGED
- # DeepGEMM 8b1392b9 (nv_dev) — UNCHANGED (verified in the r36 tree's
- # tools/install_deepgemm.sh + cmake/external_projects/deepgemm.cmake)
- # requirements/cuda.txt (r36 tree) — torch 2.13.0, flashinfer 0.6.17:
- # SAME as our r35 GB10 pins (CUDA 13.2 aarch64 / cu130 base, documented
- # deviation from their cu133 x86 runtime).
- # NEW: liburing 2.5 (apt, locks/apt-packages.txt) — io_uring reads for the
- # DS4.1 Engram ngram tables in disk mode (their runtime.liburing pin).
- VLLM_PIN=202a11a98eb360b73552586f2b5c5880cbd48c8c
- VLLM_TREE=e20d152fa6972a4a7e624e74d198403fccb7af73
- VLLM_PRETEND_VERSION=0.26.1rc0+glm53.r36.vllm202a11a9
- B12X_PIN=323107ff948ca532f1f7c793b4b550c30ba5212b
- B12X_TREE=18cb59ec640a5cb8d624bebfc81b919fee88340b
- B12X_PRETEND_VERSION=0.1.dev0+glm53b12x.r36.323107ff
- LMCACHE_PIN=29bc5a2efde737c436b04499eb62cd1776cebeec
- LMCACHE_TREE=5a88a1ea9d2627c76288d056e7193f7454669b64
- LMCACHE_VERSION=0.5.5.dev0+glm53checkpoints.29bc5a2e
- # Same r35 behavior: empty on purpose — the r36 tree gates FlashKDA on the
- # packed-checkpoint API and applies its own in-tree patch (same pin 3b225bf2
- # as r35; our r27 unpatched checkout would fail the gate).
- FLASH_KDA_SRC_DIR=
|
