launcher/docker-images/glm53-jovian-r18/versions.env

# Authoritative build inputs — GLM-5.3-Flash r12 (jovian-judgement) clean build.
# Every entry is a tag or exact version — no :latest, no main, no ranges.
# Pattern: solar-open2-nvfp4 (raw CUDA base, staged wheel builds, ccache/uv
# buildx caches) — content r12-specific per the rtx6kpro source contract.

# CUDA base (aarch64; digest-pinned, tested by solar-open2-nvfp4)
CUDA_BASE_IMAGE=nvidia/cuda:13.2.0-devel-ubuntu24.04
CUDA_BASE_DIGEST=sha256:a5b6256e470196fc1d5f8f62139d57d3662867746dfe1cb352d7652024047020

# Python / uv
UV_VERSION=0.11.28
PYPI_INDEX_URL=https://pypi.org/simple
PYTORCH_INDEX_URL=https://download.pytorch.org/whl/cu130

# Torch stack — the fork's requirements/cuda.txt pins (r12 era, cu130 aarch64)
TORCH_VERSION=2.13.0
TORCHVISION_VERSION=0.28.0
TORCHAUDIO_VERSION=2.11.0

# FlashInfer — the fork's cuda.txt uses flashinfer-python==0.6.17 (PyPI);
# flashinfer-cubin is NOT a build dep (setup.py skips it; the runtime
# fetches cubins when the package is absent).
FLASHINFER_VERSION=0.6.17

# vLLM — local-inference-lab fork, dev/jovian-judgement base + PR heads.
#
# r18 composition (2026-09-05): branch tip a8c796f3af74 (supersedes the r17
# base 9c4dd054, which superseded r12's 54f6e9826c + PR set 550-557; do NOT
# re-merge 550-557, they conflict). PR set 579/580/581/586 merged into the
# tip; still open, pinned by head SHA: 582 (KDA gate side-stream) and 646
# (prefix-cache fine-grained sliding-window hits + decoupled GLM-5.3
# target/recurrent blocks — 2048-token target / 256 recurrent /
# prefix_match_unit 256 = 5.4x KV capacity for the same budget, fine
# prefix-hit resume restored; VLLM_GLM53_SPLIT_MAMBA_BLOCK_SIZE may now
# divide the target block). New in window vs r17: RoCEnante one-shot RoCE
# collectives (#597; env VLLM_ENABLE_ROCE_ALLREDUCE, off by default), B12X
# NVFP4 KV cache (fp8_ds_mla / nvfp4_ds_mla), DFlash split-KV paged decode
# attention, L2 weight-prefetch decode, DSpark input-budget split,
# spec-decode KV-connector + prefix-cache-block fixes, GLM tool-delimiter
# parser hardening, prefill fairness engines, KDA prefill backend, FlashKDA
# stable-inverse fix, #649 logits top-k single exchange, GLM5Next decode
# defaults + LM-head/metadata reuse. No requirements/setup.py drift.
# Staged tree on the build host: sm121/r18 @ 67e7e6e2ae204f40d6b60b96326bdcdc50a61b96
# (base + pr582 + pr646 merges + staged-cutlass cmake patch; /tmp/vllm-src.tar.gz).
VLLM_REPO=https://github.com/local-inference-lab/vllm.git
VLLM_BASE_REF=a8c796f3af74106b2d8d441e9ec54588936a5388
VLLM_PR_HEADS="582:4df4cf5e123b55aab2fbcb310468b076995e9410 646:b4f16bd3f34cd90d2de2823dfdb83e7742279fb7"

# B12X — master + PR heads (pure-python wheel; CuTe-DSL JIT kernels).
# r18: master tip b58f34ea (supersedes r17 base 2bbab479; PR 290 bf16 PCIe
# two-shot merged into it; + #315 RoCE gather-geometry). Still open: 284
# (SM120 mhc decode split), pinned by head SHA. New in window: RoCEnante
# one-shot RoCE all-reduce/all-gather (#295, both-rail striping, DSV4 graph
# collectives), GLM_NEXT NVFP4 KV cache (PR #309), KDA prefill extension
# (full op surface), dense-GEMM/MoE A16 activation-precision selection on
# SM121/GB10 (relevant to the linear-b12x quality toxin — re-bisect with
# this build), QSA large-prefill + paged-CuTe specialization, DSA indexer
# plan-owned lifecycle.
# Staged tree on the build host: sm121/r18 @ 1adb8a38af88e4ccc407ec774f165b7dca20f949
# (base + pr284 merge; /tmp/b12x-src.tar.gz).
B12X_REPO=https://github.com/local-inference-lab/b12x.git
B12X_BASE_REF=b58f34eaf978277621efced6678e6713fd7122e4
B12X_PRS="284"

# Deps staging: deps-src.tar.gz (in this dir) carries the pre-fetched build
# sub-deps. 2026-09-06 re-stage (two fixes): (1) flash-attention/csrc/cutlass
# was an empty stub (staging skipped the submodule) — the r18 vllm tree
# dropped the implicit cutlass include wiring for the FA targets, so the
# stub broke the FA3 build (sm90_get_smem_store_op_for_accumulator missing
# in cutlass 4.4.2). Now populated with FA's own pin (vllm pins
# flash-attention @ f3e1a4f, whose csrc/cutlass gitlink = 62750a2b75c802660e4894434dc55e839f322277,
# nvidia/cutlass). (2) Same for MSA: fmha_sm100 install needs
# MSA/python/fmha_sm100/cutlass/include — populated with MSA's own pin
# (vllm pins vllm-project/MSA @ 087c1618, gitlink eb61c911471867a5fd2466bfd8f29306cea6ebf8).
# Also purged 23k macOS AppleDouble ._* metadata files from the staging.

# DeepGEMM — fetched by vllm's cmake; pre-cloned (raw-SHA fetches of a
# non-default-branch tip 404 over GitHub's git protocol)
DEEPGEMM_REPO=https://github.com/deepseek-ai/DeepGEMM.git
DEEPGEMM_BRANCH=nv_dev
DEEPGEMM_REF=8b1392b978f5a03c828dd1711090d7fb50958b8a

# Target GPU architecture (DGX Spark = sm_121a)
TORCH_CUDA_ARCH_LIST=12.1a
BUILD_JOBS=4