Disable Cascade Attention for Batch Invariance (#32561)

Signed-off-by: frankwang28 <frank.wbb@hotmail.com>
Signed-off-by: Frank Wang <41319051+frankwang28@users.noreply.github.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>

tests/v1/determinism/utils.py

@@ -7,7 +7,6 @@ import pytest
 import torch
 
 from vllm.platforms import current_platform
-from vllm.utils.flashinfer import has_flashinfer
 from vllm.v1.attention.backends.fa_utils import flash_attn_supports_mla
 
 skip_unsupported = pytest.mark.skipif(
@@ -22,8 +21,10 @@ BACKENDS: list[str] = [
     "TRITON_MLA",
 ]
 
-if has_flashinfer():
-    BACKENDS.append("FLASHINFER")
+# FlashInfer temporarily disabled due to invariant CTA sizes.
+# See FlashInfer issue #2424
+# if has_flashinfer():
+#     BACKENDS.append("FLASHINFER")
 
 if flash_attn_supports_mla():
     BACKENDS.append("FLASH_ATTN_MLA")
@@ -78,9 +79,10 @@ def _random_prompt(min_words: int = 1024, max_words: int = 1024 * 2) -> str:
         # For longer prompts, repeat context
         padding_text = (
             " This is an interesting topic that deserves more explanation. "
+            # TODO: Update to * (target_words // 10) to better align with word ratio
             * (target_words // 50)
         )
-        base_prompt = base_prompt + padding_text
+        base_prompt = padding_text + base_prompt
 
     return base_prompt