# Version profile: GLM-5.3-Flash r28 source lock (2026-09-06 cut, all forks). # # Referenced from recipes via `build_profile = "r28"`; 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/glm53-r28-cache-fixes-20260906 # b12x voipmonitor/b12x integration/glm53-r28-source-qualification-20260906 # (21 commits over the r27 pin: GLM/DeepSeek loader fixes, W4A4 FC1 # fusion, M8 route prep, M1 FC2 overlap, GDN/MHC perf) # LMCache local-inference-lab/LMCache # integration/glm53-lmcache-boundary-qualification-20260906 # (3 MP-transfer robustness fixes over the r26 pin) # # The three cuts landed within minutes of each other (09-06 18:30/18:31 UTC) # — they form ONE stack; never advance only one fork (the 2026-09-08 half-cut # lesson: mixed r28-vLLM + r27-b12x benched fine but skipped loader fixes). # vLLM — r28: prefill/decode compute-time sharing, interleaved bounded # parallel prefills (TTFT -3.6x), chat-template rendering parity, GLM # boundary restore, partial prefix-cache hits. No requirements/cmake changes # vs r27 (verified via the GitHub compare API: 13 commits / 29 files), so the # fetch-deps.sh gates hold. VLLM_PIN=62f6385c2ab633041347925b174e463345bf39b1 VLLM_TREE=6361a741337f60bed8a0431d655c78b90f658210 VLLM_PRETEND_VERSION=0.1.dev0+glm53.r28.62f6385c # B12X — r28 qualification line (strictly ahead of the r27 pin, 0 behind) B12X_PIN=7216f6b97dc8d371bb8c66c07e82a5e3fe88b76c B12X_TREE=be0572c5b2823811817b528d70109315f714a58d B12X_PRETEND_VERSION=0.1.dev0+glm53b12x.r28.7216f6b9 # LMCache — boundary qualification line (strictly ahead of the r26 pin) LMCACHE_PIN=934c287a248e372fc4676232ce271e8a216aa367 LMCACHE_TREE=a860684a8e7da831a984adf6d2b4e659b79d5811 LMCACHE_VERSION=0.5.5.dev0+glm53cache.engine.934c287a