feat(logprob): add deterministic ROCm vocab-parallel path - #328

Merged

Flink-ddd merged 60 commits into

RL-Align:testfrom

hihaluemen:work/ws2-logprob-rocm

Sep 3, 2026

Commits on Aug 2, 2026

feat(ws2): add TP-aware logprob contract and dispatch metadata
ryankert01
committed
b7ffb3a
fix(ws2): address CodeRabbit review on logprob contract PR
ryankert01
committed
cdc11ba
fix(ws2): address second CodeRabbit round on logprob contract
ryankert01
committed
6455715
feat(ws2): make determinism scope and invocation surface part of the typed contract
ryankert01
committed
3b4eaef
docs(ws2): drop standalone design doc per review
ryankert01
committed
e6dbeef
style(ws2): align comment density with sibling kernel modules
ryankert01
committed
878ba88

Commits on Aug 4, 2026

feat: add single-gpu logprob comparison harness
hihaluemen
committed
934bc5b
fix: keep logprob CLI stdout machine readable
hihaluemen
committed
b69426d
refactor: simplify logprob comparison harness
hihaluemen
committed
0efcfe1
docs: document SM90 logprob validation
hihaluemen
committed
115d86c
fix: address logprob harness lint and provenance
hihaluemen
committed
c028b5b
fix: type heterogeneous logprob backends
hihaluemen
committed
7ba09b5

Commits on Aug 5, 2026

init vocab parallel logp
KJLdefeated
committed
a63bea2
init vocab parallel logp
KJLdefeated
committed
4fcdc30
adding cross tp testing
KJLdefeated
committed
6ffade4

Commits on Aug 6, 2026

fix comment
KJLdefeated
committed
4231625

Commits on Aug 7, 2026

test(ws2): align dispatch tests with the auto+TP>1 unsafe-dispatch guard

KJLdefeated
and
claude
committed
2360a71

Commits on Aug 8, 2026

Merge branch 'RL-Align:main' into feat/ws2-logprob-single-gpu-harness-pr2
hihaluemen
authored
99e59f8
refactor: colocate logprob harness tooling and docs
hihaluemen
committed
4eebb3b
test: resolve logprob CLI path reliably
hihaluemen
committed
19488cc
Merge PR1 logprob contract into PR2 integration base
hihaluemen
committed
6e2a79e
init vocab parallel logp

KJLdefeated
authored and
hihaluemen
committed
b7d9d89
init vocab parallel logp

KJLdefeated
authored and
hihaluemen
committed
3866d3c
adding cross tp testing

KJLdefeated
authored and
hihaluemen
committed
65f3c6f
test: align PR3 dispatch with latest PR1 guard
hihaluemen
committed
05d19eb

Commits on Aug 10, 2026

Merge branch 'RL-Align:main' into feat/ws2-logprob-single-gpu-harness-pr2
hihaluemen
authored
a46891a

Commits on Aug 11, 2026

Commits on Aug 12, 2026

ws2 deterministic grpo loss PR5
KJLdefeated
committed
3627fcd
Merge branch 'main' into feat/ws2-logprob-tp-contract-pr1
KJLdefeated
authored
e668339

Commits on Aug 19, 2026

Merge issue 241 PR1
hihaluemen
committed
d407561
Merge issue 241 PR2
hihaluemen
committed
e3be4f9
Merge issue 241 PR3
hihaluemen
committed
628a9f6
Merge issue 241 PR4
hihaluemen
committed
28dde33
Merge issue 241 PR5
hihaluemen
committed
b52fa19
feat(ws2): record cross-topology logprob fingerprints
hihaluemen
committed
c2c99c1
Merge PR4 cross-topology fingerprints
hihaluemen
committed
9aca059

Commits on Aug 21, 2026

Commits on Aug 22, 2026

test: isolate ROCm vocab logprob dispatch cases
hihaluemen
committed
9a05c1e
style: format ROCm dispatch test with black
hihaluemen
committed
1c73a42

Commits on Aug 23, 2026

logp benchmark
KJLdefeated
committed
8e793c9
modify report
KJLdefeated
committed
e9f1d2a
feat(logprob): add Triton WS2 vocab-parallel backend, shared fused kernel path, report update
KJLdefeated
committed
dd5fe05

Commits on Aug 24, 2026

Update CPU and Cuda golden baseline
frank-2077
committed
46ec8d7

Commits on Aug 31, 2026

Commits on Sep 3, 2026

Merge latest test into PR #328
Flink-ddd
committed
52234b5