A high-throughput and memory-efficient inference and serving engine for LLMs
fix: prefix DeepSeek V4 MTP projections (#44821)
Signed-off-by: Yufeng He <40085740+he-yufeng@users.noreply.github.com>
Yufeng He committed
Jun 10, 2026 at 23:47 UTC
4673ca1d7869cafa469e327dbe079eefebd8e368
2 files changed
+4
vllm/models/deepseek_v4/amd/mtp.py
+2
@@ -86,6 +86,7 @@ class DeepSeekV4MultiTokenPredictorLayer(nn.Module):
86
bias=False,
87
return_bias=False,
88
quant_config=quant_config,
89
+ prefix=f"{prefix}.e_proj",
90
)
91
self.h_proj = ReplicatedLinear(
92
config.hidden_size,
@@ -93,6 +94,7 @@ class DeepSeekV4MultiTokenPredictorLayer(nn.Module):
94
bias=False,
95
return_bias=False,
96
quant_config=quant_config,
97
+ prefix=f"{prefix}.h_proj",
98
)
99
100
self.hc_eps = config.hc_eps
vllm/models/deepseek_v4/nvidia/mtp.py
+2
@@ -92,6 +92,7 @@ class DeepSeekV4MultiTokenPredictorLayer(nn.Module):
92
bias=False,
93
return_bias=False,
94
quant_config=quant_config,
95
+ prefix=f"{prefix}.e_proj",
96
)
97
self.h_proj = ReplicatedLinear(
98
config.hidden_size,
@@ -99,6 +100,7 @@ class DeepSeekV4MultiTokenPredictorLayer(nn.Module):
100
bias=False,
101
return_bias=False,
102
quant_config=quant_config,
103
+ prefix=f"{prefix}.h_proj",
104
)
105
106
self.hc_eps = config.hc_eps