From 49abadaedba2d88c6f35216ebef235edac8c1e82 Mon Sep 17 00:00:00 2001 From: djramic Date: Wed, 8 Jul 2026 23:58:36 +0200 Subject: [PATCH] [ROCm][Bugfix] Fix empty-tensor .max() crash in AITER FA (#47894) Signed-off-by: Djordje Ramic --- vllm/v1/attention/backends/rocm_aiter_fa.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/vllm/v1/attention/backends/rocm_aiter_fa.py b/vllm/v1/attention/backends/rocm_aiter_fa.py index 7e850af3e7..0f289649b1 100644 --- a/vllm/v1/attention/backends/rocm_aiter_fa.py +++ b/vllm/v1/attention/backends/rocm_aiter_fa.py @@ -602,7 +602,11 @@ class AiterFlashAttentionMetadataBuilder( torch.cumsum( chunk_seq_lens, dim=1, out=cu_seq_lens_cpu[:, 1:], dtype=torch.int32 ) - max_cum_tokens = cu_seq_lens_cpu[:, -1].max().item() + # Avoid .max() on an empty tensor when there is no context + # (num_chunks == 0, e.g. Whisper encoder's first pass). + max_cum_tokens = ( + cu_seq_lens_cpu[:, -1].max().item() if num_chunks > 0 else 0 + ) range_idx = torch.arange(max_cum_tokens, dtype=torch.int32)[None, None, :] idx_to_batch_tensor = range_idx == cu_seq_lens_cpu[:, 1:][:, :, None]