A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 3, 14:04 UTC
gemma3 multi-gpu bug-fix (#42630)
Signed-off-by: Philip Maybank <pmaybank@amd.com> Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
pmaybank committed
May 15, 2026 at 05:32 UTC
1dc3fe08ea12f2e285e3e4e5124745e6c6ab1ee2
1 file changed
+1
vllm/model_executor/models/gemma3n_mm.py
+1
@@ -421,6 +421,7 @@ class Gemma3nMultimodalEmbedder(nn.Module):
421
self.multimodal_hidden_size,
422
self.text_hidden_size,
423
bias=False,
424
+ input_is_parallel=False, # scatter the full-width input internally
425
)
426
427
self.embedding_post_projection_norm = RMSNorm(