r18: add PR #646 (fine-grained SWA prefix hits + decoupled GLM53 target/recurrent blocks, 5.4x KV capacity); advance bases to new tips (vllm a8c796f3 w/ #649, b12x b58f34ea w/ #315)

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

@@ -24,36 +24,42 @@ FLASHINFER_VERSION=0.6.17
 
 # vLLM — local-inference-lab fork, dev/jovian-judgement base + PR heads.
 #
-# 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-# base 9c4dd054 (which itself superseded r12's 54f6e9826c + PR set 550-557;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,
+# 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+# base 9c4dd054, which superseded r12's 54f6e9826c + PR set 550-557; do NOT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. 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).
+# 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 @ 343661fe63dd64fe4125cfe1897c0c6c21673bb2
+# (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=f564dffe9a6625ae3d87706530ce35974f1d2989
-VLLM_PR_HEADS="582:4df4cf5e123b55aab2fbcb310468b076995e9410"
+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 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).
+# 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=613cfd9bbc542e2a7a9e5e216d1f62a8becda87f
+B12X_BASE_REF=b58f34eaf978277621efced6678e6713fd7122e4
 B12X_PRS="284"
 
 # DeepGEMM — fetched by vllm's cmake; pre-cloned (raw-SHA fetches of a