Repository navigation

feat(serve): logprobs on the chat route, and the V3 serving-shape probe that reads them (PMAT-4971) - #5188

Open

noahgift wants to merge 35 commits into

mainfrom

la-711a/4971-v3a-margin-m4

Commits on Oct 12, 2026
- feat(serve): top-k logprobs with the raw logit, for the V3 top-2 margin (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  5436bf3
- feat(serve): Session records each step's logprobs; a greedy step keeps the device argmax (PMAT-4971)
  noahgift
  committed
  d65af56
- feat(serve): chat logprobs on the quantized CPU path; every other backend refuses 501 (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  ef4d871
- feat(serve): streamed chat logprobs on the quantized CPU path (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  a7f578b
- feat(serve): chat logprobs on the Qwen3.5 hybrid, streamed and not (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  b8d2fa2
- fix(serve): the chat request names TopLogprobs in scope; DeviceGreedy test init gets its new fields (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  b833380
- refactor(serve): split the CUDA batch scheduler loop and process_cuda_batch into named steps (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  3c71ad1
- feat(serve): chat logprobs on the dense CUDA backend, direct and through the scheduler's m=1 path (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  cb7b27e
- feat(serve): chat logprobs on a batched CUDA turn, m > 1 (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  42d739e
- refactor(serve): split process_iteration_batch into named steps (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  4948474
- feat(serve): chat logprobs on the iteration scheduler's batched path (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  f4f7e02
- feat(serve): chat routes outside realizar's router refuse logprobs by name (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  7416d81
- feat(perf-gate): V3 shape check in Rust, and the top-2 margin on the PP-26 witness (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  a2ec4bb
- feat(apr-cli): `apr test llm shape-check`, the CLI over the V3 shape check (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  2f6bc5f
- fix(apr-cli): a RED V3 witness exits 4 as InvalidInput, not "Invalid APR format" (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  718ade7
- feat(perf-gate): the PP-26 witness and its top-2 margin from OpenAI logprob entries (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  2ff74d8
- feat(perf-gate): the perf041 probe's decisions in Rust (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  3de50dd
- feat(test-lib): the perf041 probe's requests over LlmClient (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  4789f16
- feat(apr-cli): `apr test llm parity`, the perf041 probe's front door (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  712b4f2
- fix(apr-cli): `apr test llm parity` writes its witness to --witness, not --json (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  a9a6409
- fix(apr-cli): parse the whole `apr` Cli on a 16 MB stack; quote apr's in the PERF041 row (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  134ed00
- feat(contracts): serving-shape-parity-v1, the V3 bar and its admission rule (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  0cc45a1
- feat(perf_gate): V3 reads S7, the near-tie cap and the declared_min floor (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  09af4ba
- docs(contracts): serving-shape-parity-v1 records the S7 run on gx10 (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  e43f6d0
- test(perf_gate): kill the three g16 survivors in the V3 probe (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  0a09a76
- test(perf_gate): pin check_reference at exactly declared_min (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  4911d7d
- docs(contracts): serving-shape-parity-v1 records the Rust mutation runs (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  ca9701a
- test(serve): logprobs_4971 asks ignore_eos, so the fixture generates tokens (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  6f97a10
- fix(apr-cli): the model-path gate lists `llm ttft-verdict` beside shape-check and parity (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  6099a21
- chore(contracts): ONT witness for serving-shape-parity-v1 on the re-cut (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  e102b99
- fix(serve): the APR CUDA chat route refuses logprobs instead of dropping them (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  fd6f25f
- fix(serve): a logprobs step stays graphed under BATCHED_GRAPH=1 (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  23e42db
- docs(contracts): serving-shape-parity-v1 cites no internal paths or names (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  fa0af16
- docs(audits): re-derive the chat SSE row after the logprobs change (PMAT-4971)
  
  noahgift
  and
  claude
  committed
  ed43b1e
- style: format the Qwen3.5 chat backend fragment (#4971)
  
  noahgift
  and
  claude
  committed
  db2f729

Navigation Menu

Repository navigation

feat(serve): logprobs on the chat route, and the V3 serving-shape probe that reads them (PMAT-4971) - #5188

feat(serve): logprobs on the chat route, and the V3 serving-shape probe that reads them (PMAT-4971)#5188
noahgift wants to merge 35 commits into
mainfrom
la-711a/4971-v3a-margin-m4

Commits on Oct 12, 2026

feat(serve): top-k logprobs with the raw logit, for the V3 top-2 margin (PMAT-4971)

feat(serve): Session records each step's logprobs; a greedy step keeps the device argmax (PMAT-4971)

feat(serve): chat logprobs on the quantized CPU path; every other backend refuses 501 (PMAT-4971)

feat(serve): streamed chat logprobs on the quantized CPU path (PMAT-4971)

feat(serve): chat logprobs on the Qwen3.5 hybrid, streamed and not (PMAT-4971)

fix(serve): the chat request names TopLogprobs in scope; DeviceGreedy test init gets its new fields (PMAT-4971)

refactor(serve): split the CUDA batch scheduler loop and process_cuda_batch into named steps (PMAT-4971)

feat(serve): chat logprobs on the dense CUDA backend, direct and through the scheduler's m=1 path (PMAT-4971)

feat(serve): chat logprobs on a batched CUDA turn, m > 1 (PMAT-4971)

refactor(serve): split process_iteration_batch into named steps (PMAT-4971)

feat(serve): chat logprobs on the iteration scheduler's batched path (PMAT-4971)

feat(serve): chat routes outside realizar's router refuse logprobs by name (PMAT-4971)

feat(perf-gate): V3 shape check in Rust, and the top-2 margin on the PP-26 witness (PMAT-4971)

feat(apr-cli): `apr test llm shape-check`, the CLI over the V3 shape check (PMAT-4971)

fix(apr-cli): a RED V3 witness exits 4 as InvalidInput, not "Invalid APR format" (PMAT-4971)

feat(perf-gate): the PP-26 witness and its top-2 margin from OpenAI logprob entries (PMAT-4971)

feat(perf-gate): the perf041 probe's decisions in Rust (PMAT-4971)

feat(test-lib): the perf041 probe's requests over LlmClient (PMAT-4971)

feat(apr-cli): `apr test llm parity`, the perf041 probe's front door (PMAT-4971)

fix(apr-cli): `apr test llm parity` writes its witness to --witness, not --json (PMAT-4971)

fix(apr-cli): parse the whole `apr` Cli on a 16 MB stack; quote apr's in the PERF041 row (PMAT-4971)

feat(contracts): serving-shape-parity-v1, the V3 bar and its admission rule (PMAT-4971)

feat(perf_gate): V3 reads S7, the near-tie cap and the declared_min floor (PMAT-4971)

docs(contracts): serving-shape-parity-v1 records the S7 run on gx10 (PMAT-4971)

test(perf_gate): kill the three g16 survivors in the V3 probe (PMAT-4971)

test(perf_gate): pin check_reference at exactly declared_min (PMAT-4971)

docs(contracts): serving-shape-parity-v1 records the Rust mutation runs (PMAT-4971)

test(serve): logprobs_4971 asks ignore_eos, so the fixture generates tokens (PMAT-4971)

fix(apr-cli): the model-path gate lists `llm ttft-verdict` beside shape-check and parity (PMAT-4971)

chore(contracts): ONT witness for serving-shape-parity-v1 on the re-cut (PMAT-4971)

fix(serve): the APR CUDA chat route refuses logprobs instead of dropping them (PMAT-4971)

fix(serve): a logprobs step stays graphed under BATCHED_GRAPH=1 (PMAT-4971)

docs(contracts): serving-shape-parity-v1 cites no internal paths or names (PMAT-4971)

docs(audits): re-derive the chat SSE row after the logprobs change (PMAT-4971)

style: format the Qwen3.5 chat backend fragment (#4971)