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 15, 18:30 UTC
[Model] Pixtral: Remove layer instantiation duplication (#15053)
Signed-off-by: Julien Denize <julien.denize@mistral.ai>
Julien Denize committed
Mar 19, 2025 at 03:34 UTC
437f9162d0eacf5b3f179c864c9664dfc8a020ff
1 file changed
-9
vllm/model_executor/models/pixtral.py
-9
@@ -365,16 +365,7 @@ class PixtralForConditionalGeneration(nn.Module, SupportsMultiModal,
365
spatial_merge_size=self.vision_args.spatial_merge_size,
366
use_mlp_bias=False,
367
)
368
- if self.vision_args.add_pre_mm_projector_layer_norm:
369
- self.pre_mm_projector_norm = RMSNorm(self.vision_args.hidden_size,
370
- eps=1e-5)
368
372
- if self.vision_args.mm_projector_id == PATCH_MERGE:
373
- self.patch_merger = PatchMerger(
374
- vision_encoder_dim=self.vision_args.hidden_size,
375
- spatial_merge_size=self.vision_args.spatial_merge_size,
376
- use_mlp_bias=False,
377
- )
369
self.vision_language_adapter = VisionLanguageAdapter(
370
self.vision_args, dim=config.text_config.hidden_size)
371