From ccd0d1d9067a0bf24330a87044dca272e4a5228c Mon Sep 17 00:00:00 2001 From: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Date: Mon, 13 Apr 2026 10:53:45 -0400 Subject: [PATCH] [Bug] Fix rocm sparse attn indexer issue (#39225) Signed-off-by: yewentao256 Co-authored-by: TJian --- vllm/v1/attention/ops/rocm_aiter_mla_sparse.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py b/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py index 683b33b450..35a23b7a60 100644 --- a/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py +++ b/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py @@ -532,6 +532,11 @@ def rocm_aiter_sparse_attn_indexer( has_prefill = layer_attn_metadata.num_prefills > 0 num_decode_tokens = layer_attn_metadata.num_decode_tokens + # during speculative decoding, k may be padded to the CUDA graph batch + # size while slot_mapping only covers actual tokens. + num_tokens = slot_mapping.shape[0] + k = k[:num_tokens] + ops.indexer_k_quant_and_cache( k, kv_cache,