recipes/docker-images/jovian-judgement/profiles/r31.env

# Version profile: GLM-5.3-Flash r31 source lock (2026-09-09 cut, all forks).
#
# Referenced from recipes via `build_profile = "r31"`; 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); branch names are provenance, not the contract:
#   vLLM    voipmonitor/vllm    integration/jovian-warmup-buffer-reuse-20260909
#           (R31 cut: MoE warmup buffer reuse across route-id dtypes on top of
#           the R28.1-R30 line: scheduler hardening, FP8 checkpoint serving,
#           engine-driven async SHM LMCache + atomic recurrent checkpoints,
#           GLM Flash MTP projection loader fix — our upstream issue,
#           FlashInfer GDN prefill on SM12x)
#   b12x    voipmonitor/b12x    integration/glm-fp8-checkpoint-serving-20260908
#           (frozen since the 09-08 cut; unchanged across R28.1-R31)
#   LMCache local-inference-lab/LMCache
#           integration/jovian-write-through-restore-headroom-20260909
#           (write-through L1 headroom reclaim on top of the 09-09 dedup line)
#
# The R31 cut is the latest published lock
# (rtx6kpro models/glm-5.3-flash/validation/warmup-retention-r31.source.lock,
# "qualified" status, published image
# localinferencelab/vllm:jovian-judgement-community-20260909-r31).
#
# Build-dep gates verified against the r31 vLLM tree (55769270):
# deepgemm.cmake still declares 8b1392b9 (= our versions.env pin, no bump),
# requirements/cuda.txt still pins torch 2.13.0 + flashinfer 0.6.17.
VLLM_PIN=5576927057cf71b6ec61d120932338b333efa089
VLLM_TREE=6a2da27b08a1f062b00719c17cffdfa087ae4210
VLLM_PRETEND_VERSION=0.1.dev0+glm53.r31.55769270

B12X_PIN=3edbcbce70f491741b82f5eab9c1b30b39447228
B12X_TREE=c4bfeee9f3c9457400d191c870eb2e44fbcd5c2e
B12X_PRETEND_VERSION=0.1.dev0+glm53b12x.r31.3edbcbce

LMCACHE_PIN=f868580a475b7e6f7fd81e586526481d4dbead0e
LMCACHE_TREE=6e5f080f794962bf8c667ffcdb46b210d7023b15
LMCACHE_VERSION=0.5.5.dev0+glm53cache.engine.f868580a

# The r28.1+ vLLM tree gates FlashKDA on the packed-checkpoint API
# (cmake/external_projects/flashkda.cmake) and applies its own in-tree patch
# (flashkda-packed-checkpoints.patch, locked in the r31 source lock) when the
# shared /deps2 checkout is not injected. Empty here on purpose: the build
# falls back to the tree's pinned 3b225bf2 + patch instead of our r27
# unpatched checkout (which fails the gate).
FLASH_KDA_SRC_DIR=