Re-enable FlashInfer for Llama4 on Blackwell in e2e fusion tests (#28966)

Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: ProExpertProg <11367180+ProExpertProg@users.noreply.github.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>

.buildkite/test-pipeline.yaml

@@ -930,6 +930,8 @@ steps:
   - csrc/quantization/fp4/
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
+  - vllm/v1/worker/
+  - vllm/v1/cudagraph_dispatcher.py
   - vllm/compilation/
   # can affect pattern matching
   - vllm/model_executor/layers/layernorm.py

tests/compile/distributed/test_fusions_e2e.py

@@ -47,12 +47,8 @@ if current_platform.is_cuda():
         ModelBackendTestCase(
             # Use smaller model for L40s in CI
             model_name="RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8",
-            # TODO while llama4 is broken, use FLASHINFER for llama3 on Blackwell
-            #  so FI attention+fp8_quant is at least tested once
             model_kwargs=dict(max_model_len=1024, kv_cache_dtype="fp8"),
-            backend=AttentionBackendEnum.FLASHINFER
-            if is_blackwell()
-            else AttentionBackendEnum.TRITON_ATTN,
+            backend=AttentionBackendEnum.TRITON_ATTN,
             matches=Matches(
                 attention_fusion=32,
                 allreduce_fusion=65,
@@ -65,9 +61,9 @@ if current_platform.is_cuda():
             model_kwargs=dict(max_model_len=1024, kv_cache_dtype="fp8"),
             # TODO FlashInfer attn broken on Hopper with kvcache=fp8:
             # https://github.com/vllm-project/vllm/issues/28568
-            # TODO FlashInfer attn broken on Blackwell for llama4:
-            # https://github.com/vllm-project/vllm/issues/28604
-            backend=AttentionBackendEnum.TRITON_ATTN,
+            backend=AttentionBackendEnum.FLASHINFER
+            if is_blackwell()
+            else AttentionBackendEnum.TRITON_ATTN,
             matches=Matches(
                 attention_fusion=48,
                 allreduce_fusion=96,