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 5, 23:30 UTC
[Revert] Remove DeepGEMM availability check in DeepseekV32IndexerMetadataBuilder (#38076)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Chauncey committed
Mar 26, 2026 at 09:43 UTC
87f05d6880d0795d658fd4a5dacc0109704a8c4c
1 file changed
-7
vllm/v1/attention/backends/mla/indexer.py
-7
@@ -10,7 +10,6 @@ from vllm.platforms import current_platform
10
from vllm.utils.deep_gemm import (
11
get_paged_mqa_logits_metadata,
12
has_deep_gemm,
13
- is_deep_gemm_supported,
13
)
14
from vllm.utils.math_utils import cdiv
15
from vllm.utils.platform_utils import num_compute_units
@@ -216,12 +215,6 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
215
vllm_config: VllmConfig,
216
kv_cache_spec: AttentionSpec,
217
) -> AttentionCGSupport:
219
- if not is_deep_gemm_supported():
220
- logger.warning_once(
221
- "DeepGEMM is not available. Disabling CUDA graph support "
222
- "for sparse attention indexer. This may reduce performance.",
223
- )
224
- return AttentionCGSupport.NEVER
218
return AttentionCGSupport.UNIFORM_BATCH
219
220
def __init__(self, *args, **kwargs):