A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 9, 18:15 UTC
[torch.compile][CI] Add back attn fusion on hopper/ada (#32940)
Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Luka Govedič committed
Jan 23, 2026 at 11:49 UTC
bbbd696af9ebfadf38553e2361f3a31ed759e0ed
1 file changed
+4
-5
tests/compile/test_fusion_attn.py
+4
-5
@@ -18,7 +18,7 @@ from tests.compile.fusion_test_utils import (
18
is_blackwell,
19
run_model,
20
)
21
-from tests.utils import cuda_device_count_stateless, flat_product
21
+from tests.utils import flat_product
22
from tests.v1.attention.utils import BatchSpec, create_common_attn_metadata
23
from vllm._custom_ops import cutlass_scaled_fp4_mm, scaled_fp4_quant
24
from vllm.attention.layer import Attention
@@ -265,13 +265,13 @@ if current_platform.is_cuda():
265
HEADS = [(64, 8), (40, 8)]
266
PATTERN_TEST_MODELS_FP8 = [
267
(
268
- "nvidia/Llama-4-Scout-17B-16E-Instruct-FP8",
268
+ "RedHatAI/Meta-Llama-3.1-8B-FP8",
269
TestAttentionFp8StaticQuantPatternModel,
270
)
271
]
272
PATTERN_TEST_MODELS_FP4 = [
273
(
274
- "nvidia/Llama-4-Scout-17B-16E-Instruct-FP4",
274
+ "nvidia/Llama-3.1-8B-Instruct-NVFP4",
275
TestAttentionNvfp4QuantPatternModel,
276
)
277
]
@@ -331,9 +331,8 @@ def test_attention_quant_pattern(
331
if backend == AttentionBackendEnum.FLASHINFER and (
332
not current_platform.is_device_capability((10, 0)) or not has_flashinfer()
333
):
334
+ # This also captures the FP4 case
335
pytest.skip("FlashInfer attn fusion requires Blackwell and flashinfer")
335
- if "Llama-4-Scout" in model_name and cuda_device_count_stateless() < 2:
336
- pytest.skip("Llama-4-Scout requires at least 2 GPUs")
336
337
custom_ops_list = custom_ops.split(",") if custom_ops else []
338