Fix FP16 overflow in GQA attention and concat_past_present buffer overflow#4677

Merged

causten merged 8 commits intoROCm:developfrom

aditya-dl:fix-fp16-overflow-gqa-attention

Apr 14, 2026

Commits on Mar 31, 2026

Fix FP16 overflow: insert convert(f16->f32) after dot output
Aditya Lohia
committed

Commits on Apr 1, 2026

Add test for kv_cache_attention fusion after FP32 softmax upcast
Aditya Lohia
committed
Add test for mul without upstream dot (develop behavior preserved)
Aditya Lohia
committed
Add regression test for FP16 overflow in Q*K attention dot product
Aditya Lohia
committed

Commits on Apr 10, 2026

Merge branch 'develop' into fix-fp16-overflow-gqa-attention
causten
authored

Commits on Apr 11, 2026

Separate dot-softmax FP32 upcast into its own matcher
Aditya Lohia
committed

Commits on Apr 13, 2026

Fix CI formatting: use 'or' instead of '||', apply clang-format
Aditya Lohia
committed
Use can_eval to identify data path in backward walk
Aditya Lohia
committed