fix(attention): avoid FA3 with batch invariance on Hopper

Downgrade FlashAttention 3 to FlashAttention 2 on SM90 when batch invariance is enabled.

Fixes #47069

Co-authored-by: Devin AI <noreply@devin.ai>

Signed-off-by: Sidharth Rajmohan <dumko.raj@gmail.com>
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

Author: dumko2001

vllm/v1/attention/backends/fa_utils.py

@@ -171,6 +171,17 @@ def get_flash_attn_version(
         ):
             fa_version = vllm_config.attention_config.flash_attn_version
 
+        if (
+            envs.VLLM_BATCH_INVARIANT
+            and device_capability.major == 9
+            and fa_version == 3
+        ):
+            logger.warning_once(
+                "Cannot use FlashAttention 3 with batch invariance on "
+                "Hopper; defaulting to FlashAttention 2."
+            )
+            fa_version = 2
+
         # 3. fallback for unsupported combinations
         if device_capability.major >= 10 and fa_version == 3:
             logger.warning_once(