[torch.compile][CI] Add back attn fusion on hopper/ada (vllm-project#32940)

Signed-off-by: Luka Govedič <[email protected]>

Author: ProExpertProg

tests/compile/test_fusion_attn.py

@@ -18,7 +18,7 @@
     is_blackwell,
     run_model,
 )
-from tests.utils import cuda_device_count_stateless, flat_product
+from tests.utils import flat_product
 from tests.v1.attention.utils import BatchSpec, create_common_attn_metadata
 from vllm._custom_ops import cutlass_scaled_fp4_mm, scaled_fp4_quant
 from vllm.attention.layer import Attention
@@ -265,13 +265,13 @@ def forward(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):
     HEADS = [(64, 8), (40, 8)]
     PATTERN_TEST_MODELS_FP8 = [
         (
-            "nvidia/Llama-4-Scout-17B-16E-Instruct-FP8",
+            "RedHatAI/Meta-Llama-3.1-8B-FP8",
             TestAttentionFp8StaticQuantPatternModel,
         )
     ]
     PATTERN_TEST_MODELS_FP4 = [
         (
-            "nvidia/Llama-4-Scout-17B-16E-Instruct-FP4",
+            "nvidia/Llama-3.1-8B-Instruct-NVFP4",
             TestAttentionNvfp4QuantPatternModel,
         )
     ]
@@ -331,9 +331,8 @@ def test_attention_quant_pattern(
     if backend == AttentionBackendEnum.FLASHINFER and (
         not current_platform.is_device_capability((10, 0)) or not has_flashinfer()
     ):
+        # This also captures the FP4 case
         pytest.skip("FlashInfer attn fusion requires Blackwell and flashinfer")
-    if "Llama-4-Scout" in model_name and cuda_device_count_stateless() < 2:
-        pytest.skip("Llama-4-Scout requires at least 2 GPUs")
 
     custom_ops_list = custom_ops.split(",") if custom_ops else []