From d223c900d85224c02f2162ee2c757a769e99f519 Mon Sep 17 00:00:00 2001 From: Nick Hill Date: Mon, 27 Jul 2026 21:36:45 -0700 Subject: [PATCH] [Bugfix] Only pad transformers backend `value` when it is narrower (#50060) Signed-off-by: Nick Hill --- vllm/model_executor/models/transformers/__init__.py | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/vllm/model_executor/models/transformers/__init__.py b/vllm/model_executor/models/transformers/__init__.py index 9dbfe4b503..ff4bb6cd43 100644 --- a/vllm/model_executor/models/transformers/__init__.py +++ b/vllm/model_executor/models/transformers/__init__.py @@ -64,12 +64,15 @@ def vllm_attention_forward( head_dim_v = value.shape[-1] query, key, value = (x.transpose(1, 2) for x in (query, key, value)) query, key, value = (x.reshape(hidden, -1) for x in (query, key, value)) - # Pad `value` up to the query/key head size when they differ (expanded MLA). - if head_dim_v != head_dim_qk: + # Pad `value` up to the query/key head size when it is smaller (expanded + # MLA). A larger last dim just means `value` isn't split per head, e.g. + # packed grouped/multi-query projections, and needs no padding. + pad_value = head_dim_v < head_dim_qk + if pad_value: value = F.pad(value.view(-1, head_dim_v), (0, head_dim_qk - head_dim_v)) value = value.reshape(hidden, -1) attn_output = self_attn.forward(query, key, value) - if head_dim_v != head_dim_qk: + if pad_value: attn_output = attn_output.view(-1, head_dim_qk)[..., :head_dim_v] attn_output = attn_output.reshape(hidden, -1) return attn_output, None