[ws1]: WS1 Full Qwen3-8B Dense Train-Inference Closeout by maxiaosong1124 · Pull Request #315 · RL-Align/RL-Kernel

Commits on Aug 11, 2026

feat(ws1): land C1 four-judgment numerical contract (#267)
maxiaosong1124
committed
af4d9c2
docs(ws1): add #267 C1 closeout evidence map
maxiaosong1124
committed
70286fb
docs(ws1): streamline C1 gtest documentation
maxiaosong1124
committed
087156e
fix(ws1): address tolerance contract review
maxiaosong1124
committed
81ddd65

Commits on Aug 12, 2026

feat(ws1): land C2 canonical workload identity (#268)
maxiaosong1124
committed
e857084
fix(ws1): address PR 292 review feedback
maxiaosong1124
committed
fca1656
style(testing): apply isort export ordering
maxiaosong1124
committed
362562e
fix(ws1): satisfy mypy workload validation
maxiaosong1124
committed
b41c6f8
feat(ws1): land C3 forward config-invariance harness (#269)
maxiaosong1124
committed
f1bfbc5
fix ws1 provenance review issues
maxiaosong1124
committed
2e4a30a
feat(ws1): land C4 gradient-invariance harness and adapters (#270)

maxiaosong1124
and
claude
committed
596feb0

Commits on Aug 13, 2026

Commits on Aug 15, 2026

feat(ws1): complete C6-C11 closeout gates
maxiaosong1124
committed
c386d47

Commits on Aug 16, 2026

chain test
maxiaosong1124
committed
85e1373
fix(ws1): accumulate Triton LM-head hidden gradients in fp32
maxiaosong1124
committed
af07ca2
fix(ws1): avoid retained transposed weights in chain gate
maxiaosong1124
committed
e4bf38e
fix(ws1): canonicalize chunked training backward replay
maxiaosong1124
committed
3b2e640
fix(ws1): close chunked dual-backend gradient paths
maxiaosong1124
committed
fdf5bcc
docs(ws1): record H20 C8 and C10 evidence
maxiaosong1124
committed
10ccc85
fix(ci): satisfy lint and secure fork GPU gate
maxiaosong1124
committed
27f5e01
fix(ci): allow CPU collection without Triton
maxiaosong1124
committed
be609fa

Navigation Menu

Uh oh!

[ws1]: WS1 Full Qwen3-8B Dense Train-Inference Closeout - #315

[ws1]: WS1 Full Qwen3-8B Dense Train-Inference Closeout#315
frank-2077 merged 31 commits into
RL-Align:testfrom
maxiaosong1124:feat/ws1-c6-c11-closeout-266

Commits on Aug 11, 2026

feat(ws1): land C1 four-judgment numerical contract (#267)

docs(ws1): add #267 C1 closeout evidence map

docs(ws1): streamline C1 gtest documentation

fix(ws1): address tolerance contract review

Commits on Aug 12, 2026

feat(ws1): land C2 canonical workload identity (#268)

fix(ws1): address PR 292 review feedback

style(testing): apply isort export ordering

fix(ws1): satisfy mypy workload validation

feat(ws1): land C3 forward config-invariance harness (#269)

fix ws1 provenance review issues

feat(ws1): land C4 gradient-invariance harness and adapters (#270)

Commits on Aug 13, 2026

feat(ws1): land C1-C5/C8 gtest framework and single-op gates

feat(ws1): close C8 four-judgment gtest port and evidence gaps

fix(ws1): record launched C2 candidate id on C8 invariance cells

docs(ws1): bind C8 execute evidence to 5c33dcd on H20

test(ws1): fix CPU sampling smoke and add CUDA/Triton C8 CI gate

fix(ws1): keep C8 evidence git-clean and tighten GPU CI safety

chore: ignore local ws1-c8-ci.json dumps

Merge branch 'test' into feat/ws1-c1-c5-c8-gtest

fix(ci): satisfy pre-commit formatting and mypy checks

fix(ws1): address PR #305 review feedback

style(ws1): apply black formatting for CI lint

Commits on Aug 15, 2026

feat(ws1): complete C6-C11 closeout gates

Commits on Aug 16, 2026

chain test

fix(ws1): accumulate Triton LM-head hidden gradients in fp32

fix(ws1): avoid retained transposed weights in chain gate

fix(ws1): canonicalize chunked training backward replay

fix(ws1): close chunked dual-backend gradient paths

docs(ws1): record H20 C8 and C10 evidence

fix(ci): satisfy lint and secure fork GPU gate

fix(ci): allow CPU collection without Triton