Fix fp16/GQA static-cache export correctness (split from #340) - #351

Merged

titaiwangms merged 18 commits into

mainfrom

fix/gqa-fp16-fold

Jun 12, 2026

Commits on Jun 10, 2026

Fix fp16 GQA export emitting fp32 packed weights in fold passes
titaiwangms
committed
3db2653
Strip dead pre-pack weights in FoldConcatInitializersPass
titaiwangms
committed
87abd27
test(fold-concat): name the live-packed-result invariant in DCE assert
titaiwangms
committed
4392d32
test(fold-concat): assert survived packed-QKV values are exact after DCE
titaiwangms
committed
fbfd4cd
test(fold-concat): gate packed-QKV values through serialize→reload
titaiwangms
committed
d4db4c2
test(fold-concat): add mean|abs| degeneracy assert + poison negative control
titaiwangms
committed
a444699
test: use IR-native ir.save instead of onnx.save in _fold_concat_test
titaiwangms
committed
8d842d0
test(dtype): assert warning fires on declared/const_value dtype disagreement
titaiwangms
committed
c784a3e
test(fp16): add e2e regression test guarding df203cc fold-pass dtype retention
titaiwangms
committed
f18b339
test(fp16): make fp16 fold e2e fixture function-scoped for hermetic, xdist-safe runs
titaiwangms
committed
a079efe
fix(tasks): stamp explicit present KV-cache output shapes for GQA
titaiwangms
committed
fad723f
docs: add fp16/GQA ONNX-export gotchas skill + fp16 GQA fold-fix changelog
titaiwangms
committed
109a7c4
style: apply lintrunner (ruff format + D205) to salvaged tests
titaiwangms
committed
786aab7
docs: re-add GQA present-KV changelog entry + de-anchor SKILL.md from cherry-pick SHAs
titaiwangms
committed
285a246
fix(kv-cache): fail-closed on partial present-shape parameter sets

titaiwangms
and
Copilot
committed
28578a3

Commits on Jun 12, 2026

Fail closed on contradictory initializer dtype (PR #351 review)

titaiwangms
and
Copilot
committed
3b0fa13
Merge branch 'main' into fix/gqa-fp16-fold
titaiwangms
authored
72cee28
Fix D205 docstring lint on test_raises_on_dtype_contradiction

titaiwangms
and
Copilot
committed
3361f5d

Navigation Menu

Uh oh!

Fix fp16/GQA static-cache export correctness (split from #340) - #351

Fix fp16/GQA static-cache export correctness (split from #340)#351
titaiwangms merged 18 commits into
mainfrom
fix/gqa-fp16-fold

Commits on Jun 10, 2026

Fix fp16 GQA export emitting fp32 packed weights in fold passes

Strip dead pre-pack weights in FoldConcatInitializersPass

test(fold-concat): name the live-packed-result invariant in DCE assert

test(fold-concat): assert survived packed-QKV values are exact after DCE

test(fold-concat): gate packed-QKV values through serialize→reload

test(fold-concat): add mean|abs| degeneracy assert + poison negative control

test: use IR-native ir.save instead of onnx.save in _fold_concat_test

test(dtype): assert warning fires on declared/const_value dtype disagreement

test(fp16): add e2e regression test guarding df203cc fold-pass dtype retention

test(fp16): make fp16 fold e2e fixture function-scoped for hermetic, xdist-safe runs

fix(tasks): stamp explicit present KV-cache output shapes for GQA

docs: add fp16/GQA ONNX-export gotchas skill + fp16 GQA fold-fix changelog

style: apply lintrunner (ruff format + D205) to salvaged tests

docs: re-add GQA present-KV changelog entry + de-anchor SKILL.md from cherry-pick SHAs

fix(kv-cache): fail-closed on partial present-shape parameter sets

Commits on Jun 12, 2026

Fail closed on contradictory initializer dtype (PR #351 review)

Merge branch 'main' into fix/gqa-fp16-fold

Fix D205 docstring lint on test_raises_on_dtype_contradiction