recipes/docker-images/jovian-judgement/profiles/r35.env

# Version profile: GLM-5.3-Flash r35 source lock (2026-09-11 cut, all forks).
#
# Referenced from recipes via `build_profile = "r35"`; merged over this
# image dir's versions.env (profile wins, then recipe [build.versions] wins
# last). Only the keys that differ from versions.env are listed — the rest
# (CUDA base, torch/flashinfer, apt) stay at the image dir's pins.
#
# Pinned by commit + tree hash (the Dockerfile gates HEAD + HEAD^{tree} per
# repo). SHA-only cut (no named branches at these tips; branch names on the
# forks lag at r28). r35 is the latest published lock:
#   rtx6kpro models/glm-5.3-flash/validation/swiglu-reviewed-composition-r35
#   .source.lock — "R35 qualifies the public source composition with GLM
#   DFlash2 TP4/DCP1 serving, C1/C8, cold 32K prefill, prefix restoration and
#   focused GPU correctness tests" (qualified 2026-09-11, published image
#   localinferencelab/vllm:jovian-judgement-community-20260911-r35).
#
# What moved over r31 (per the lab changelog in models/glm-5.3-flash.md):
#   R33: shared-input NVFP4 split MoE prefill (eligible prefills share input
#        quantization across routed experts; +8.6-8.9% 32K prefill on their
#        RTX quartet).
#   R35: MERGED B12X SWIGLU-CLAMP CORRECTION — R33/R34's split path omitted
#        the model's SwiGLU clamp (runaway-generation risk); R35 fixes it
#        without disabling the fast split path. We adopt r35 directly
#        (never r33/r34).
#   plus: VLLM_DEFAULT_MOE_BACKEND=b12x default, LMCache fs-eviction
#   accounting (retired objects leave byte accounting, pending writes
#   protected), DFlash2 K7 qualification.
#
# Build-dep gates verified against the r35 vLLM tree (de982a50, 2026-09-12):
# cmake/external_projects listing identical to r31 (no dep add/remove),
# FlashKDA still 3b225bf2 + in-tree packed-checkpoints patch, DeepGEMM tag
# still 8b1392b9 (= our versions.env pin), requirements/cuda.txt still
# torch 2.13.0 + flashinfer 0.6.17. fetch-deps.sh needs no resync.
VLLM_PIN=de982a50c6a3e4718e5cf9f00423a92192718da1
VLLM_TREE=6d5cc3490735a53e849a4f0909183cb5e3a22e8b
VLLM_PRETEND_VERSION=0.1.dev0+glm53.r35.de982a50

B12X_PIN=98086604c86ec1e78977e5023ce282ecb97ab8a7
B12X_TREE=d963f030c72a426753d2afa412b35d29eeb045f7
B12X_PRETEND_VERSION=0.1.dev0+glm53b12x.r35.98086604

LMCACHE_PIN=29bc5a2efde737c436b04499eb62cd1776cebeec
LMCACHE_TREE=5a88a1ea9d2627c76288d056e7193f7454669b64
LMCACHE_VERSION=0.5.5.dev0+glm53checkpoints.29bc5a2e

# Same r31+ behavior: empty on purpose — the r35 tree gates FlashKDA on the
# packed-checkpoint API and applies its own in-tree patch (same pin + patch
# as r31); our r27 unpatched checkout would fail the gate.
FLASH_KDA_SRC_DIR=