recipes/docker-images/jovian-judgement-r39/profiles/r39.env

# Version profile: GLM-5.3-Flash r39 (jovian-judgement tip, 2026-09-21 cut).
#
# The lab's continuous beta line after r38 (8e1f1e5, 09-19): two days of
# high-frequency cuts (7 vLLM beta tags on 09-21 alone). This profile pins
# the 20:00 UTC tip pair — vLLM tag vllm-jovian-cu134-beta-9e5d1793fa3 and
# the same-moment b12x master commit (RoCE traffic class for switched Spark
# fabrics — the first b12x commit aimed at our exact topology).
#
# What this is (verified 2026-09-21 against the public repos):
#   vLLM  9e5d1793fa3 = local-inference-lab/vllm tip, tag
#         vllm-jovian-cu134-beta-9e5d1793fa3 (2026-09-21 20:00, "Update
#         Spark launchers for switched RoCE and DeepSeek TP3"). 100+ lab
#         commits over the r38 pin, notably a memory-management series in
#         the b12x preparation area ("release preparation probes",
#         "reclaim temporary startup memory before KV allocation on every
#         rank", "Fix GLM KDA projection scratch overlap") — the same area
#         where r38 deadlocks on 2x GB10 (BUG-2026-09-19) — plus
#         b12x verifier tuning coordination across startup ranks, sparse-MLA
#         owner fixes, DeepGEMM DeepSelect routing.
#   B12X  c2dc1cf022 = local-inference-lab/b12x master tip (same cut,
#         "Configure RoCE native traffic class for switched Spark
#         fabrics") + W4A8/W4A16 perf + DCP program-identity fix.
#   LMCache 29bc5a2e — UNCHANGED (no LMCache code since 09-14, CI only).
#
# External deps MOVED in this cut (verified by reading the r39 tree's
# cmake/external_projects directly — this is why the r39 build lives in its
# own dir docker-images/jovian-judgement-r39/ with its own fetch-deps.sh +
# deps.lock; the r27/r35 dir keeps the r38-era contract for the prod lanes):
#   cutlass       v4.4.2 -> v4.7.1
#   flash-attention f3e1a4f7 -> 506341a1
#   MSA           087c1618 -> f355c37e
#   FlashMLA      a8f794d1 -> 0eee43b1
#   FlashKDA      3b225bf2 -> b59532f1 (vendored cutlass gitlink now 5c149f52)
#   tml-fa4       b2068346 -> 75765e76
#   DeepGEMM      deepseek-ai/DeepGEMM 8b1392b9 -> vllm-project/DeepGEMM
#                 ad1f1726 (the lab fork)
#   DeepSelect    NEW dep d96d33af (csrc/3rdparty/cutlass ae6bccf3 gitlink)
#   qutlass e74319e3, triton_kernels v3.5.1 — UNCHANGED
#   FlashKDA patches moved into the tree (cmake/external_projects/patches/
#   + apply_flashkda_checkpoint_patch.cmake) — the tree applies them itself.
#
# requirements/cuda.txt (r39 tree) — torch 2.13.0, flashinfer 0.6.17:
#   SAME as our r27/r35 GB10 runtime pins (unchanged).
#
# Build: clone-at-pin (docker-images/jovian-judgement-r39 Dockerfile),
# in-build git only (no pre-staged tarballs).
VLLM_PIN=9e5d1793fa34db4e672664711690e8a68d90fcd3
VLLM_TREE=01f937ec5a23953845e74d952bb782741d2b765a
VLLM_PRETEND_VERSION=0.26.1rc0+glm53.r39.vllm9e5d1793

B12X_PIN=c2dc1cf02295b8241fc6a7728be7b6e8c23dda2f
B12X_TREE=1f77ebffba81828c7015c968d7b007f6c0c1f89b
B12X_PRETEND_VERSION=0.1.dev0+glm53b12x.r39.c2dc1cf0

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

# Same r31+ behavior: empty on purpose — the r39 tree gates FlashKDA on the
# packed-checkpoint API and applies its own in-tree patches (cmake
# apply_flashkda_checkpoint_patch.cmake); a pre-staged unpatched checkout
# would fail the gate.
FLASH_KDA_SRC_DIR=