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 7, 19:00 UTC
[Bug][DSV3.2] Always prepare metadata for DeepGEMM Sparse Attention (#35075)
Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
Benjamin Chislett committed
Feb 24, 2026 at 17:55 UTC
9fa5b25a238c08fae8acf507e5dbc923f5b2e5cb
1 file changed
+4
-2
vllm/v1/attention/backends/mla/indexer.py
+4
-2
@@ -8,7 +8,7 @@ import torch
8
from vllm.config import VllmConfig
9
from vllm.logger import init_logger
10
from vllm.platforms import current_platform
11
-from vllm.utils.deep_gemm import get_paged_mqa_logits_metadata, is_deep_gemm_supported
11
+from vllm.utils.deep_gemm import get_paged_mqa_logits_metadata, has_deep_gemm
12
from vllm.v1.attention.backend import (
13
AttentionBackend,
14
AttentionCGSupport,
@@ -342,7 +342,9 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
342
offsets = None
343
344
seq_lens = common_attn_metadata.seq_lens[:num_decodes]
345
- if is_deep_gemm_supported():
345
+
346
+ # DeepGEMM is required for the paged MQA logits on CUDA devices
347
+ if current_platform.is_cuda() and has_deep_gemm():
348
self.scheduler_metadata_buffer[:] = get_paged_mqa_logits_metadata(
349
seq_lens, self.kv_cache_spec.block_size, self.num_sms
350
)