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 8, 08:45 UTC
[Bugfix] Enforce DeepGEMM when using sparse_attn_indexer on CUDA (#34374)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Feb 12, 2026 at 15:08 UTC
6d4e27ce29bac0e4cd4975cddf5b0dacc6cb727a
1 file changed
+5
vllm/model_executor/layers/sparse_attn_indexer.py
+5
@@ -10,6 +10,7 @@ from vllm.logger import init_logger
10
from vllm.model_executor.custom_op import CustomOp
11
from vllm.platforms import current_platform
12
from vllm.utils.deep_gemm import fp8_mqa_logits, fp8_paged_mqa_logits
13
+from vllm.utils.import_utils import has_deep_gemm
14
from vllm.utils.torch_utils import direct_register_custom_op
15
from vllm.v1.attention.backends.mla.indexer import (
16
DeepseekV32IndexerMetadata,
@@ -277,6 +278,10 @@ class SparseAttnIndexer(CustomOp):
278
self.max_model_len = max_model_len
279
self.max_total_seq_len = max_total_seq_len
280
self.topk_indices_buffer = topk_indices_buffer
281
+ if current_platform.is_cuda() and not has_deep_gemm():
282
+ raise RuntimeError(
283
+ "Sparse Attention Indexer CUDA op requires DeepGEMM to be installed."
284
+ )
285
286
def forward_native(
287
self,