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 14, 05:30 UTC
[Model] Add missing prefix to glm4_1v (#22716)
Signed-off-by: zRzRzRzRzRzRzR <2448370773@qq.com>
Yuxuan Zhang committed
Aug 13, 2025 at 16:23 UTC
9e7e5baaa83b1e5070a3cf3823c134b28eaa2a1c
1 file changed
+7
-1
vllm/model_executor/models/glm4_1v.py
+7
-1
@@ -453,25 +453,30 @@ class Glm4vPatchMerger(nn.Module):
453
context_dim: int,
454
quant_config: Optional[QuantizationConfig] = None,
455
bias: bool = False,
456
+ prefix: str = "",
457
) -> None:
458
super().__init__()
459
self.hidden_size = d_model
460
self.proj = ColumnParallelLinear(self.hidden_size,
461
self.hidden_size,
462
bias=bias,
462
- gather_output=True)
463
+ gather_output=True,
464
+ quant_config=quant_config,
465
+ prefix=f"{prefix}.proj")
466
self.post_projection_norm = nn.LayerNorm(self.hidden_size)
467
self.gate_up_proj = MergedColumnParallelLinear(
468
input_size=self.hidden_size,
469
output_sizes=[context_dim] * 2,
470
bias=bias,
471
quant_config=quant_config,
472
+ prefix=f"{prefix}.gate_up_proj",
473
)
474
self.down_proj = RowParallelLinear(
475
context_dim,
476
self.hidden_size,
477
bias=bias,
478
quant_config=quant_config,
479
+ prefix=f"{prefix}.down_proj",
480
)
481
self.act_fn = SiluAndMul()
482
self.extra_activation_func = nn.GELU()
@@ -661,6 +666,7 @@ class Glm4vVisionTransformer(nn.Module):
666
context_dim=vision_config.intermediate_size,
667
quant_config=quant_config,
668
bias=False,
669
+ prefix=f"{prefix}.merger",
670
)
671
self.embeddings = Glm4vVisionEmbeddings(vision_config)
672