[Attention] Use FA4 for MLA prefill (#34732)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>

vllm/v1/attention/backends/fa_utils.py

@@ -125,11 +125,14 @@ def get_flash_attn_version(
         # FA4 on SM100 (Blackwell) has TMEM capacity limits that restrict
         # supported head dimensions.
         # See: https://github.com/Dao-AILab/flash-attention/issues/1959
+        # Exception: hdim 192 is supported for MLA's diff-headdim case
+        # (qk=192, v=128), added upstream in commits 1a15733e/1b36ab19.
         if (
             fa_version == 4
             and device_capability.major >= 10
             and head_size is not None
             and head_size > 128
+            and head_size != 192
         ):
             logger.warning_once(
                 "FA4 on Blackwell does not support head_size=%d due to TMEM "