| 12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152 |
- # Version profile: GLM-5.3-Flash r38 (jovian-judgement tip, 2026-09-19 cut).
- #
- # Operator direction (2026-09-19): use the LATEST jovian-judgement + the
- # current lab quants, not the r12 legacy rebuild. This profile pins the lab
- # mainline at the day-19 tip instead of the r37 DS4.1-optimized release.
- #
- # What this is (verified 2026-09-19 against the public repos):
- # vLLM 8e1f1e587f8 = local-inference-lab/vllm `dev/jovian-judgement` tip
- # (2026-09-16, "Fix deferred grammar rejection in v2 speculative
- # sampling"). 25+ commits over the r37 pin: B12X MXFP8 linear kernel
- # (mxfp8/b12x.py — the small-GDN-GEMM path that flashinfer mm_mxfp8
- # rejects at n/k < 128 on the Spark quant), per-projection layer-max
- # activation scales for b12x NVFP4 MoE, GLM B12X startup contracts
- # for target+draft (#772), sparse-MLA module owner (#771), KV
- # admission fix (#770), KDA __init__ override for native-FP8
- # checkpoints, RoCE/b12x tuning. Model code keeps the SPLIT GDN
- # naming (q_proj/k_proj/g_a/g_b/f_a/f_b/b_proj) that both current
- # lab quants (GLM-5.3-Flash-NVFP4 ~200 GB, -NVFP4-Spark 154 GB)
- # carry — the fused in_proj_qkvbfg naming is the Aiden fork only.
- # B12X 0f3a8cbfd1c = local-inference-lab/b12x `master` tip.
- #
- # Delta vs the r37 profile (jovian-r37-ds41.env):
- # vLLM c687594b8a8 (release/ds41-optimized-r37) -> 8e1f1e587f8 (JJ tip)
- # B12X f1c4e9dd5b1 -> 0f3a8cbfd1c
- # LMCache 29bc5a2e — UNCHANGED
- # FlashKDA 3b225bf2 — UNCHANGED (JJ tip cmake gate verified 2026-09-19)
- # DeepGEMM 8b1392b9 (nv_dev) — UNCHANGED
- # requirements/cuda.txt (JJ tip) — torch 2.13.0, flashinfer 0.6.17:
- # SAME as our r27/r35 GB10 pins. cmake/external_projects pins (cutlass
- # v4.4.2, flashkda 3b225bf, flashmla a8f794d, MSA 087c161, qutlass
- # e74319e, tml-fa4 b206834, flash-attention f3e1a4f, triton v3.5.1,
- # deepgemm 8b1392b9) verified identical to the r37 tree's — fetch-deps.sh
- # needs no resync.
- #
- # Build: clone-at-pin (docker-images/glm53-jovian-r27 Dockerfile), in-build
- # git only (no pre-staged tarballs — operator direction 2026-09-19).
- VLLM_PIN=8e1f1e587f8d24faf606f334a1c4bdaaa6bd4368
- VLLM_TREE=d9f35de1ed5e47dd006f6d420dd25707fb284eae
- VLLM_PRETEND_VERSION=0.26.1rc0+glm53.r38.vllm8e1f1e58
- B12X_PIN=0f3a8cbfd1c11d27f04e3ab37a802d522f4f1c68
- B12X_TREE=ff2218d710ca9cae71459e1b9eb7ca3fe10ae08e
- B12X_PRETEND_VERSION=0.1.dev0+glm53b12x.r38.0f3a8cbf
- LMCACHE_PIN=29bc5a2efde737c436b04499eb62cd1776cebeec
- LMCACHE_TREE=5a88a1ea9d2627c76288d056e7193f7454669b64
- LMCACHE_VERSION=0.5.5.dev0+glm53checkpoints.29bc5a2e
- # Same r35/r36/r37 behavior: empty on purpose — the tree gates FlashKDA on
- # the packed-checkpoint API and applies its own in-tree patch (same pin
- # 3b225bf2; our unpatched checkout would fail the gate).
- FLASH_KDA_SRC_DIR=
|
