From a966aaed30b9132191e8ba88d4f61e76657b690d Mon Sep 17 00:00:00 2001 From: Uranus <109661872+UranusSeven@users.noreply.github.com> Date: Thu, 30 Apr 2026 07:14:50 +0800 Subject: [PATCH] [Bugfix][MLA] Size arange_buffer to max_num_batched_tokens to prevent CUDA IMA (#39277) Signed-off-by: UranusSeven <109661872+UranusSeven@users.noreply.github.com> Signed-off-by: Woosuk Kwon Co-authored-by: Claude Opus 4.6 Co-authored-by: Woosuk Kwon --- vllm/v1/attention/backends/mla/indexer.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/vllm/v1/attention/backends/mla/indexer.py b/vllm/v1/attention/backends/mla/indexer.py index ded3218346..5d12d27e76 100644 --- a/vllm/v1/attention/backends/mla/indexer.py +++ b/vllm/v1/attention/backends/mla/indexer.py @@ -303,7 +303,10 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder): device=self.device, ) self.arange_buffer = torch.arange( - scheduler_config.max_num_seqs * next_n, + max( + scheduler_config.max_num_seqs * next_n, + scheduler_config.max_num_batched_tokens, + ), dtype=torch.int32, device=self.device, )