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, 10:30 UTC
fix: remove unused norm for dpskv4 (#41710)
Signed-off-by: inisis <desmond.yao@buaa.edu.cn> Co-authored-by: Yongye Zhu <zyy1102000@gmail.com>
inisis committed
May 18, 2026 at 18:33 UTC
df852ed503ac1a79e568271cd6f136a7b2698f5e
1 file changed
+1
-2
vllm/model_executor/layers/deepseek_v4_attention.py
+1
-2
@@ -47,7 +47,7 @@ from vllm.logger import init_logger
47
from vllm.model_executor.custom_op import PluggableLayer
48
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
49
from vllm.model_executor.layers.deepseek_compressor import DeepseekCompressor
50
-from vllm.model_executor.layers.layernorm import LayerNorm, RMSNorm
50
+from vllm.model_executor.layers.layernorm import RMSNorm
51
from vllm.model_executor.layers.quantization import QuantizationConfig
52
from vllm.model_executor.layers.quantization.input_quant_fp8 import (
53
QuantFP8,
@@ -1111,7 +1111,6 @@ class DeepseekV4Indexer(nn.Module):
1111
quant_config=None,
1112
prefix=f"{prefix}.weights_proj",
1113
)
1114
- self.k_norm = LayerNorm(self.head_dim, eps=1e-6)
1114
self.softmax_scale = self.head_dim**-0.5
1115
1116
self.scale_fmt = "ue8m0"