[WIP][CI][Bugfix] Fix `test_run_eagle_dp` (#38584)

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

tests/v1/distributed/test_eagle_dp.py

@@ -69,9 +69,7 @@ async def test_run_eagle_dp(monkeypatch: pytest.MonkeyPatch, attn_backend: str):
     )
 
     prompt = "This is a test of data parallel with eagle"
-    # This test might be flaky, see
-    # https://github.com/vllm-project/vllm/issues/31913
-    num_expected_tokens = 20
+    num_expected_tokens = 100
     sampling_params = SamplingParams(
         max_tokens=num_expected_tokens,
         ignore_eos=True,

vllm/v1/attention/backends/flash_attn.py

@@ -389,8 +389,11 @@ class FlashAttentionMetadataBuilder(AttentionMetadataBuilder[FlashAttentionMetad
         slot_mapping = common_attn_metadata.slot_mapping
         causal = common_attn_metadata.causal
 
-        # the overhead of the aot schedule is not worth it for spec-decode
-        aot_schedule = self.aot_schedule and not fast_build
+        # Disable AOT schedule for spec-decode proposer (not worth the overhead)
+        # and for batch invariance (schedule varies with max_seqlen_q/k).
+        aot_schedule = (
+            self.aot_schedule and not fast_build and not envs.VLLM_BATCH_INVARIANT
+        )
 
         if self.aot_sliding_window is None:
             self.aot_sliding_window = (-1, -1)