r18: bump jovian-judgement pins to branch tip (RoCEnante #597 + NVFP4 KV + DFlash split-KV + SM121 A16 precision); fix stale r12 pins; Dockerfile header -> versions.env is source of truth

docker-images/glm53-jovian-r12/Dockerfile

@@ -5,17 +5,13 @@
 # pattern), NOT the day-0 vllm image. No baked-vllm contamination: torch,
 # flashinfer, vLLM and b12x all come from pinned sources in this file.
 #
-# Stack (rtx6kpro models/glm-5.3-flash.md source contract):
-#   vLLM  = local-inference-lab/vllm @ 54f6e9826c (dev/jovian-judgement base)
-#           + sequential merges of PR heads 550,552,553,554,555,557
-#           (the doc's "composition commit" 7d8a09c… is not a public ref)
-#   b12x  = local-inference-lab/b12x @ 139e04048b (master) + PR heads
-#           260,267,268,269,270. b12x is pure-Python: the kernels are
-#           CuTe-DSL / tvm-ffi JIT (nvidia-cutlass-dsl 4.6.2), no CUDA
-#           compile at install time.
-#   torch = 2.13.0 cu130 (the fork's requirements/cuda.txt pin)
-#   flashinfer = 0.6.17 (day-0-era; not a pip dep of the fork since 0.6.14,
-#           so installed from the flashinfer.ai wheel index)
+# Stack: vLLM/b12x pins live in versions.env (single source of truth — the
+# staged vllm-src.tar.gz / b12x-src.tar.gz on the build host are built from
+# those pins, composition notes included). torch = 2.13.0 cu130 (the fork's
+# requirements/cuda.txt pin); flashinfer = 0.6.17 (not a pip dep of the fork
+# since 0.6.14, installed from the flashinfer.ai wheel index). b12x is
+# pure-Python: CuTe-DSL / tvm-ffi JIT (nvidia-cutlass-dsl 4.6.2), no CUDA
+# compile at install time.
 #
 # GB10 sm_121 fix baked in (patch_cuda_sm12x.py): the sm12x auto-priority
 # list in vllm/platforms/cuda.py offers only TRITON_MLA (no sparse) and

docker-images/glm53-jovian-r12/versions.env

@@ -23,16 +23,38 @@ TORCHAUDIO_VERSION=2.11.0
 FLASHINFER_VERSION=0.6.17
 
 # vLLM — local-inference-lab fork, dev/jovian-judgement base + PR heads.
-# PR 550 closed (unmerged), which deleted its refs/pull/550/head pseudo-ref,
-# so pin the PR head commits by SHA (GitHub keeps the objects).
+#
+# r18 composition (2026-09-05): branch tip f564dffe9a66 — supersedes the r17
+# base 9c4dd054 (which itself superseded r12's 54f6e9826c + PR set 550-557;
+# do NOT re-merge 550-557, they conflict). The r17 PR set 579/580/581/586 is
+# now merged into the tip; only 582 is still open, pinned by head SHA.
+# New in this window (9c4dd054..f564dffe + b12x below): RoCEnante one-shot
+# RoCE collectives (PR #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. No requirements/setup.py drift vs r17.
+# Staged tree on the build host: sm121/r18 @ 725e36d3189cc7cbd8c27a16323c4b9643fd8cf2
+# (base + pr582 merge + staged-cutlass cmake patch; see /tmp/vllm-src.tar.gz).
 VLLM_REPO=https://github.com/local-inference-lab/vllm.git
-VLLM_BASE_REF=54f6e9826c20ef06ed65d838c0ad497ad0abdecf
-VLLM_PR_HEADS="550:52006e858766911cebccc8a7b2b21a260f5c97f2 552:65ef1cd4d1d595a6b54698e3c2983ed4af421445 553:c6b1530e895e5188b6f57109db1f609f5615ce19 554:88cff3d5a98b95df9ceeff77aa676ef2f18a03b6 555:78bfa29002d4d7a2d6c16877074af449ba1b0ce7 557:3ef77b3125253d6f797f1524f9a52ae478dad4f2"
+VLLM_BASE_REF=f564dffe9a6625ae3d87706530ce35974f1d2989
+VLLM_PR_HEADS="582:4df4cf5e123b55aab2fbcb310468b076995e9410"
 
-# B12X — master + PR heads (pure-python wheel; CuTe-DSL JIT kernels)
+# B12X — master + PR heads (pure-python wheel; CuTe-DSL JIT kernels).
+# r18: master tip 613cfd9b (supersedes r17 base 2bbab479; PR 290 bf16 PCIe
+# two-shot now merged into it). 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 @ c594d447d3129e717f8aa0a7c56d7b777912bec3
+# (base + pr284 merge; see /tmp/b12x-src.tar.gz).
 B12X_REPO=https://github.com/local-inference-lab/b12x.git
-B12X_BASE_REF=139e04048bc3bb4f7210c99e7184d8d2f0e345e7
-B12X_PRS="260 267 268 269 270"
+B12X_BASE_REF=613cfd9bbc542e2a7a9e5e216d1f62a8becda87f
+B12X_PRS="284"
 
 # DeepGEMM — fetched by vllm's cmake; pre-cloned (raw-SHA fetches of a
 # non-default-branch tip 404 over GitHub's git protocol)