From a1ab51afb61e8120d0f36a730eac45969ea81503 Mon Sep 17 00:00:00 2001 From: aoright <102943475+aoright@users.noreply.github.com> Date: Wed, 8 Jul 2026 19:25:53 +0800 Subject: [PATCH] [Bugfix] Allocate HY V3 expert_bias in float32 to prevent silent downcasting (#47797) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Signed-off-by: 辰言 Signed-off-by: Jee Jee Li Co-authored-by: 辰言 Co-authored-by: Jee Jee Li Co-authored-by: Jee Jee Li --- vllm/model_executor/models/hy_v3.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/vllm/model_executor/models/hy_v3.py b/vllm/model_executor/models/hy_v3.py index cb2c96fa96..62bdd028bb 100644 --- a/vllm/model_executor/models/hy_v3.py +++ b/vllm/model_executor/models/hy_v3.py @@ -177,7 +177,9 @@ class HYV3MoEFused(nn.Module): else: self.shared_mlp = None - self.expert_bias = nn.Parameter(torch.empty(config.num_experts)) + self.expert_bias = nn.Parameter( + torch.empty(config.num_experts, dtype=torch.float32) + ) scoring_func = "sigmoid" e_score_correction_bias = self.expert_bias