[PyTorch][torch.compile] DotProductAttention: declarative packed QKV/KV inputs - #3200

Merged

pggPL merged 18 commits into

NVIDIA:mainfrom

pggPL:dpa_packed_qkv_api

Jul 23, 2026

Commits on Jul 7, 2026

[PyTorch] DotProductAttention: declarative packed qkv/kv inputs

pggPL
and
claude
committed
a9cd697

Commits on Jul 8, 2026

[PyTorch] MultiheadAttention: pass packed projection output to DPA declaratively

pggPL
and
claude
committed
5f70d12

Commits on Jul 9, 2026

[PyTorch] DotProductAttention: factor packed-input handling into _unpack_packed_qkv
pggPL
committed
5713048
[PyTorch] DotProductAttention: validate packed-input last-dim stride
pggPL
committed
2255bcc
[PyTorch] DotProductAttention: pass packed qkv/kv buffers as separate arguments
pggPL
committed
dc4469b
[PyTorch] Move packed qkv/kv input tests into test_attention.py
pggPL
committed
f82ba70
[PyTorch] Deprecate pointer-based detection of packed qkv layouts
pggPL
committed
5cd0351
[PyTorch] Fix implicit string concatenation lint warning
pggPL
committed
361a2a0
[PyTorch] Trim packed-input test section to equivalence and MHA tests
pggPL
committed
627dd3f

Commits on Jul 10, 2026

Commits on Jul 23, 2026