| 12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455 |
- # Version profile: DeepSeek-V4.1-Flash r37 (jovian-judgement, qualified 2026-09-13).
- #
- # rtx6kpro models/deepseek-v4.1-flash.md ("R37 dispatch and projection
- # qualification"; image localinferencelab/vllm:
- # jovian-judgement-community-20260913-r37). Two-layer image like r36: the GLM
- # entrypoint (our r35 lane) + the DS4.1 entrypoint. We build the same source
- # composition here for GB10 sm_121a (the published image's native extension is
- # sm_120 / CUDA 13.3 x86).
- #
- # What R37 claims over its public-source control (their RTX quartet, TP4/DCP1,
- # RAM Engram, DSpark K7/adaptive, batch 4096):
- # uncached 32K prefill 14,896 -> 21,399 tok/s (+43.65%)
- # C1 output (budget 50, three-cell median) 238.8 -> 281.1 tok/s (+17.70%)
- # C1 verifier 99.0 -> 110.8 steps/s (+11.92%); Sieve 385.2 -> 416.7 (+8.17%)
- # Cause (their changelog): removes repeated CPU argument-list expansion,
- # mutable-schema inspection, scalar dtype allocations and tensor-equality calls
- # during kernel dispatch — no quantization/sampling approximation — plus
- # measured B12X projection tiles, four FP32 projection partials, visible
- # index-score tiles, compact mHC prefill config, removed attention-output
- # copies, bounded short-index scratch. Also corrects the named reasoning
- # budget `high` from 50 to 75 (a behavior change, not just speed: our dev65
- # benches set reasoning_effort explicitly, so expect no surprise there) and
- # adds the B12X small-tile barrier fix (#363) + TP3 prefill-head support
- # (TP3 not serving-qualified).
- #
- # Pinned by commit + tree hash (gated per repo in the Dockerfile). Delta vs
- # the r36 profile (r36-ds41.env):
- # vLLM 202a11a98eb -> c687594b8a8 (composed "release/ds41-optimized-r37":
- # JJ b40673cd + #740/#743/#742/#744)
- # B12X 323107ff -> f1c4e9dd5b1 (clean master 05d2c43b + #364/#367/#366/#365)
- # LMCache 29bc5a2e — UNCHANGED; FlashKDA 3b225bf2 + in-tree patch — UNCHANGED
- # DeepGEMM 8b1392b9 (nv_dev) — UNCHANGED (verified in the r37 tree's
- # tools/install_deepgemm.sh)
- # requirements/cuda.txt (r37 tree) — torch 2.13.0, flashinfer 0.6.17:
- # SAME as our r35/r36 GB10 pins (CUDA 13.2 aarch64 / cu130 base, the
- # documented deviation from their cu133 x86 runtime). cmake/
- # external_projects listing identical to r36 (no dep add/remove), so
- # fetch-deps.sh needs no resync.
- # liburing 2.5 (+ -dev headers) already in locks/apt-packages.txt since r36.
- VLLM_PIN=c687594b8a8082e18af9fe2f64eb5f9ee442e251
- VLLM_TREE=3fb88fdfd0b5acdd4460076c6443a4299e5edf7d
- VLLM_PRETEND_VERSION=0.26.1rc0+glm53.r37.vllmc687594b
- B12X_PIN=f1c4e9dd5b1d841c10cda70ee49c3190cf3db0be
- B12X_TREE=480ea232d52fff4271b0412470d4ec340196735a
- B12X_PRETEND_VERSION=0.1.dev0+glm53b12x.r37.f1c4e9dd
- LMCACHE_PIN=29bc5a2efde737c436b04499eb62cd1776cebeec
- LMCACHE_TREE=5a88a1ea9d2627c76288d056e7193f7454669b64
- LMCACHE_VERSION=0.5.5.dev0+glm53checkpoints.29bc5a2e
- # Same r35/r36 behavior: empty on purpose — the r37 tree gates FlashKDA on the
- # packed-checkpoint API and applies its own in-tree patch (same pin 3b225bf2;
- # our r27 unpatched checkout would fail the gate).
- FLASH_KDA_SRC_DIR=
|
