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 11, 03:45 UTC
[Misc] Fix spelling typos in model comments (#31117)
Signed-off-by: c0de128 <kevin.mckay@outlook.com>
Kevin McKay committed
Dec 21, 2025 at 23:14 UTC
14c3e6ade3605852282ec895397e6403b609847a
2 files changed
+2
-2
vllm/model_executor/models/config.py
+1
-1
@@ -401,7 +401,7 @@ class HybridAttentionMambaModelConfig(VerifyAndUpdateConfig):
401
# of attention tokens that would fit mamba_page_size:
402
# e.g. for mamba page size = 788kB
403
# attn_1_token = 2kB -> fits ~394 tokens
404
- # then round up to a mulitple of 256 -> 512 tokens
404
+ # then round up to a multiple of 256 -> 512 tokens
405
# End result:
406
# attn_block_size = 512
407
# mamba_block_size = 512 (aligned to a multiple of chunk_size)
vllm/model_executor/models/qwen3_omni_moe_thinker.py
+1
-1
@@ -323,7 +323,7 @@ class Qwen3Omni_VisionTransformer(nn.Module):
323
hidden_size=self.hidden_size,
324
)
325
326
- # vit pos embeding, TODO: spatial_patch_size vs patch_size
326
+ # vit pos embedding, TODO: spatial_patch_size vs patch_size
327
if self.apply_vit_abs_pos_embed:
328
self.pos_embed = nn.Embedding(self.num_grid_per_side**2, self.hidden_size)
329
else: