Enable CUDA graph support for llama 3.2 vision (#14917)

Signed-off-by: Matt Ritter <100659061+mritterfigma@users.noreply.github.com>
This commit is contained in:
Matt Ritter
2025-03-19 23:29:16 -07:00
committed by GitHub
parent 2f726b241e
commit a8652f4f0f
3 changed files with 1 additions and 13 deletions
+1 -1
View File
@@ -1368,7 +1368,7 @@ class MllamaForConditionalGeneration(nn.Module, SupportsMultiModal,
full_text_row_masked_out_mask = (
attn_metadata.encoder_seq_lens_tensor
!= 0).reshape(-1, 1).to(input_ids.device)
skip_cross_attention = max(attn_metadata.encoder_seq_lens) == 0
skip_cross_attention = attn_metadata.max_encoder_seq_len == 0
# For image-present prefill.
else: