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 14, 05:30 UTC
[Bug] Fix Long Context OOM Issue (#25290)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Sep 22, 2025 at 22:04 UTC
4741239db7b75949d577f99eb1c3a9e463dc77f3
1 file changed
+1
-1
vllm/v1/attention/backends/mla/common.py
+1
-1
@@ -481,7 +481,7 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
481
# which would result in up-projected context being
482
# 2*(192*128)*(64*1024) = 3gb
483
# (assuming 192 QK head dim, 128 heads, and fp16)
484
- 128 * 1024)
484
+ 64 * 1024)
485
assert self.chunked_prefill_workspace_size >= \
486
scheduler_config.max_num_seqs * cache_config.block_size
487
if self.dcp_world_size > 1: