tree 3af06b274b6acd884c636e61c4727ce5e94875c5
parent 3d0c17b9a5ae682ff71cf6ec35d4b39a11c9321e
author WMC <tnwilly@gmail.com> 1788168281 +0800
committer GitHub <noreply@github.com> 1788168281 +0200
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJqlUhZCRC1aQ7uu5UhlAAA3FkQADh8xsVpjnnPgvoNtFa9qjWV
 fu15Nkxd/XNnTOvzeKDeGtA0F2+2OVb8nHg2/ftx8t/sWotyT2wJycSDgzwKgrL8
 UEt437BSPGpu0+zR9R3WChQIeEJe6phxCGmrSM36Bz/s6F/0Ve1aPZbqbefdtVDx
 zCFzAI7/6zvpQHzlJZm+Do8+QhwQPQUbeGSjwbVV7eU1RZzc6hW0HeGNsl1PLjeW
 0sZpMImeA7o8pKawwVjuckcP3nxLOYh0AcIfIOh6HcBY+dv5s0bayAInh7mh2hBk
 Me7h7Cw4EkMxbVXwD9J3TdAYtWXGTinPy3du/nBvwjYnjq34/88+rcrcwBKUFcsK
 KK03Iih2RU1Hjrv9PPS4MlgPSInfw9Rx9/xJJ4w2tqznKyChw4b5R/X10sgNxm8o
 R6ioixTtMFVtE7ymI7w2WWlq+LjLC41cUSHgdFAbMbQfbN/vG2fL8Wyt/i5U5/9v
 61H+Ee0mg5BKZJhlAk2V8yRFNmOuYl8wbkdRJKtvhv1NKQUfJrYMm4qM0boasfC1
 RcfwSUtcHtJX/T3M4TP18hEobwzUvUUErOG3pe/nqK/MNaoyNSq3YAFeCSvekgVT
 LI9nTaW2fNVQm+fEkVipnxgJszF40bPzT/i6v9SqYeHqdM8zBttCxKuGxxqMOIz/
 4HKn/U7PHJNZUZj6/D0k
 =a7YQ
 -----END PGP SIGNATURE-----
 

[LLVMGPU] Add dense and paged attention runtime correctness tests (#24829)

## Summary

This PR adds a CUDA end-to-end runtime correctness test for paged-KV
attention as Stage 2 of
[#24766](https://github.com/iree-org/iree/issues/24766).

This change also fixes and enables the shared host-side attention
checker used by the existing dense-attention e2e tests. Before this
change, the reference softmax was numerically incorrect and the checker
generated expected results without comparing them against device output.

## Changes

- Adds CUDA paged-KV attention coverage with non-contiguous, distinct
key and value page tables, `iree_linalg_ext.gather`, and
`iree_linalg_ext.online_attention`.
- Makes the page-table dimension dynamic (`tensor<4x?xi64>`), matching
the runtime `NUM_PAGES` use case.
- Fixes the shared stable-softmax reference and makes it compare device
output with expected results.
- Refactors the paged checker to gather paged K/V into dense tensors and
reuse the shared attention reference and checker.
- Passes the f16-rounded `1 / sqrt(head_dim)` scale to both the
operation and host reference.

## Testing
Validated on NVIDIA RTX PRO 6000 Blackwell Server Edition

- All existing CPU dense-attention configurations.
- CUDA paged attention with 4 pages.
- The same compiled paged-attention VMFB with 2, 4, and 6 runtime pages.

Attention e2e tests use:

```text
--require_exact_results=false
--acceptable_fp_delta=0.01
```

---------

Signed-off-by: weimin023 <tnwilly@gmail.com>