From c6668106760e1f8222abd19a6ee9e93fc060cb19 Mon Sep 17 00:00:00 2001 From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com> Date: Sun, 2 Aug 2026 16:30:54 -0400 Subject: [PATCH] [ROCm][Bugfix][Kimi-K3] Preserve MoE correction bias in FP32 (#50761) Signed-off-by: Fangzhou Ai Co-authored-by: OpenAI Codex --- vllm/models/kimi_k3/amd/linear.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/vllm/models/kimi_k3/amd/linear.py b/vllm/models/kimi_k3/amd/linear.py index 4e0088e9e1..181704ebf9 100644 --- a/vllm/models/kimi_k3/amd/linear.py +++ b/vllm/models/kimi_k3/amd/linear.py @@ -210,7 +210,10 @@ class KimiMoE(nn.Module): prefix=f"{prefix}.gate", ) - self.gate.e_score_correction_bias = nn.Parameter(torch.empty(num_experts)) + # Preserve FP32 checkpoint values and match FP32 router logits. + self.gate.e_score_correction_bias = nn.Parameter( + torch.empty(num_experts, dtype=torch.float32) + ) if self.num_shared_experts is not None: shared_intermediate_size = moe_intermediate_size * self.num_shared_experts