Remove router weight upcast for DSv2-related models (#47970)
Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Thien Tran committed
Jul 8, 2026 at 19:19 UTC
e7b3853bac9de3d0b6fd35619f5ca1dc32853e07
1 file changed
-2
vllm/model_executor/models/deepseek_v2.py
-2
@@ -307,9 +307,7 @@ class DeepseekV2MoE(nn.Module):
307
self.gate = GateLinear(
308
config.hidden_size,
309
config.n_routed_experts,
310
- params_dtype=self.router_dtype,
310
out_dtype=self.router_dtype,
312
- force_fp32_compute=self.router_dtype == torch.float32,
311
prefix=f"{prefix}.gate",
312
)
313
if getattr(config, "topk_method", None) == "noaux_tc":