Repository navigation

fix(#4621): apr chat sizes its device KV to the context (D3-SPLIT C, root 3) - #4623

Closed

noahgift wants to merge 53 commits into

mainfrom

ec/4621-chat-session-kv

Closed

fix(#4621): apr chat sizes its device KV to the context (D3-SPLIT C, root 3)#4623
noahgift wants to merge 53 commits into
mainfrom
ec/4621-chat-session-kv

Commits on Sep 28, 2026
- fix(serve): a turn past the device KV cache is a 400 "context exceeds N", not a 500 (D5)
  
  noahgift
  and
  claude
  committed
  bebc388
- roadmap: PMAT-4616 (D5): the ticket #4614 implements, not #3715
  
  noahgift
  and
  claude
  committed
  167bd5e
- evidence(PMAT-4616): pr-review receipt for 167bd5e126 (reviewer sonnet-5, DEGRADED, advisory finding)
  
  noahgift
  and
  claude
  committed
  9243b42
- evidence(PMAT-4616): commit the CI signer's signature for the 167bd5e126 receipt
  
  noahgift
  and
  claude
  committed
  099d07d
- fix(cuda): chunk the dense batched prefill; QK-norm models stop taking the serial prefill (#3715)
  
  noahgift
  and
  claude
  committed
  b96392c
- fix(chat): size the chat device KV to the context when VRAM allows (#3715)
  
  noahgift
  and
  claude
  committed
  1c4b84c
- fix(cuda): Qwen3.5 prefill dequantizes F16 and IQ4_XS on the device (#3715)
  
  noahgift
  and
  claude
  committed
  81a7908
- fix(cuda): a QK-norm model keeps the serial prefill when its FP16 batched prefill does not fit (#3715)
  
  noahgift
  and
  claude
  committed
  fdc165f
Commits on Sep 29, 2026
- evidence(pr-review): #4620 non-author receipt for fdc165f50d (DEGRADED: mutation unreachable)
  
  noahgift
  and
  claude
  committed
  18b6c83
- test(cuda): the fit-check no-op cases assert the path of the profile they called (#4621)
  
  noahgift
  and
  claude
  committed
  65370ad
- fix(PMAT-4616): D5 gate reds — the cap is CUDA-free and tested, try_cuda_backend no longer grows
  
  noahgift
  and
  claude
  committed
  926edb2
- evidence(pr-review): #4620 non-author receipt for 65370adc9a (DEGRADED: mutation sweep unreachable; targeted mutant killed)
  
  noahgift
  and
  claude
  committed
  76555e8
- evidence(pr-review): drop the superseded fdc165f50d receipt from #4620's tree
  
  noahgift
  and
  claude
  committed
  2b8d79a
- docs(audits): quorum PMAT-4621 AGREED 3/3 (sonnet-5, gemini-3.1-pro-high, haiku-4-5) at 2b8d79acd7
  
  noahgift
  and
  claude
  committed
  2d43ad4
- evidence(pr-review): stamp #4620's receipt with the diff patch-id Arm 4 computes (b3f1fd942a)
  
  noahgift
  and
  claude
  committed
  e3a7890
- fix(PMAT-4616): CUDA /generate and /batch/generate map ContextLimitExceeded to 400
  
  noahgift
  and
  claude
  committed
  015b0c9
- merge origin/main — roadmap.yaml regenerated with make roadmap-aggregate
  
  noahgift
  and
  claude
  committed
  1ead14b
- evidence(pr-review): #4620 receipt passes the whole guard, not only the signature arm
  
  noahgift
  and
  claude
  committed
  b953e8f
- fix(PMAT-4616): every serve generate/logprobs/perplexity error routes through generation_error_status
  
  noahgift
  and
  claude
  committed
  7008062
- split(#4621): #4620 carries roots 1+5 only (chunked dense prefill, F16/IQ4_XS device dequant)
  
  noahgift
  and
  claude
  committed
  23f058b
- fix(#4621): QK-norm models take the FP16 batched prefill when it fits (root 2)
  
  noahgift
  and
  claude
  committed
  cd324d0
- fix(#4621): apr chat sizes its device KV to the context (root 3)
  
  noahgift
  and
  claude
  committed
  7da9a62
- evidence(PMAT-4616): pr-review receipt for 7008062734 (reviewer sonnet-5, DEGRADED, advisory findings)
  
  noahgift
  and
  claude
  committed
  b3e1e46
- evidence(PMAT-4616): bind the 7008062734 receipt to diff patch-id 2fe3407e2f
  
  noahgift
  and
  claude
  committed
  595b014
- evidence(PMAT-4616): 7008062734 receipt — crux block and citations made guard-complete
  
  noahgift
  and
  claude
  committed
  bd46408
- fix(#4621): a chunked prefill's later chunks never replay the prefill graph
  
  noahgift
  and
  claude
  committed
  66c07c4
- evidence(pr-review): drop #4620's pre-split receipt (it reviewed roots 2+3, now in the stacked PRs)
  
  noahgift
  and
  claude
  committed
  31cd3f6
- evidence(pr-review): #4620 Arm-4 receipt for split head 31cd3f6c40 (PMAT-4621)
  
  noahgift
  and
  claude
  committed
  6562916
- merge ec/3715-lambda-cells (A @6562916bee: PREFILL_GRAPH chunk fix) into B
  
  noahgift
  and
  claude
  committed
  451bc77
- merge B (ec/4621-fit-and-chat-kv @451bc77c52) into C
  
  noahgift
  and
  claude
  committed
  f61c073
- evidence(pr-review): #4620 post-split quorum (3/3 AGREED, scope A) + S4.2-compliant receipt SARIF (PMAT-4621)
  
  noahgift
  and
  claude
  committed
  91e70ca
- merge ec/3715-lambda-cells (A @91e70ca127: A-scope quorum json) into B
  
  noahgift
  and
  claude
  committed
  1c97bea
- docs(audits): quorum PMAT-4621 AGREED 3/3 (sonnet-5, gemini-3.1-pro-high, haiku-4-5) at 1c97bea00e (#4624, D3-SPLIT B)
  
  noahgift
  and
  claude
  committed
  f3bf4ae
- merge B (ec/4621-fit-and-chat-kv @f3bf4ae869) into C
  
  noahgift
  and
  claude
  committed
  42477d4
- fix(serve): K1 — size the GPU KV to the model's context, not a fixed 4096 (#4603)
  
  noahgift
  and
  claude
  committed
  2cab59f
- test(serve): K1 — the 24 GB 1.7B serve budget includes the batched slots
  
  noahgift
  and
  claude
  committed
  98abc8a
- merge #4614 (2e/d5-serve-kv-4xx) into K1 — R2: use its serving-context 4xx, no second path
  
  noahgift
  and
  claude
  committed
  0c0e824
- refactor(serve): K1 adds serving_kv_len beside #4623's session_kv_len, editing none of its lines
  
  noahgift
  and
  claude
  committed
  bc201cc
- test(serve): K1 — at the limit 200 whole, past it a 400, on the device (#4603)
  
  noahgift
  and
  claude
  committed
  b77bd07
- fix(serve): size the GPU KV from the model's context; refuse past it with a 400 (K1 #4603, D5 #4614)
  
  noahgift
  and
  claude
  committed
  1683fa1
- merge devont/k1-kv-only: #4630 becomes the KV-only vehicle on main (cop 11:03Z)
  
  noahgift
  and
  claude
  committed
  067f906
- ci: first main-based run for #4630 (retarget does not fire pull_request)
  
  noahgift
  and
  claude
  committed
  5d983db
- merge main: pick up #4632 wasmtime 48.0.3 (RUSTSEC-2026-0315/0316) for #4630
  
  noahgift
  and
  claude
  committed
  6a6a1c2
- docs(roadmap): PMAT-4636 K1 entry (#4636); PMAT-4616 AC1 names the serving context N
  
  noahgift
  and
  claude
  committed
  45d616e
- docs(audits): quorum PMAT-4616,PMAT-4636 AGREED 3/3 at 45d616e6e7 (#4630 Q0)
  
  noahgift
  and
  claude
  committed
  8df8d6d
- docs(pr-review): #4630 receipt at 8df8d6d0a7 (DEGRADED, independent sonnet reviewer)
  
  noahgift
  and
  claude
  committed
  44cb53e
Commits on Oct 4, 2026
- merge origin/main into ec/4621-chat-session-kv
  
  noahgift
  and
  claude
  committed
  d593b5c
- merge devont/k1-serve-kv (#4630) onto main (316dee2cd4)
  
  noahgift
  and
  claude
  committed
  da4f4f9
- evidence(pr-review): receipt for #4623 at d593b5c99b
  
  noahgift
  and
  claude
  committed
  56ad4d4
- fold: #4630 into #4623 -- serve sizes its GPU KV to the model context beside chat's session KV
  
  noahgift
  and
  claude
  committed
  3cc3d7a
- evidence(pr-review): receipt for #4623 at 3cc3d7a71b
  
  noahgift
  and
  claude
  committed
  616ff4c
Commits on Oct 6, 2026
- test(serve): K1 fails, never skips, when the device is unreachable (L25)
  
  noahgift
  and
  claude
  committed
  07a5877
- evidence(pr-review): receipt for #4623 at 07a58772fd
  
  noahgift
  and
  claude
  committed
  688a0e2

Commits

Commits on Sep 28, 2026

Commits on Sep 29, 2026

Commits on Oct 4, 2026

Commits on Oct 6, 2026