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 14, 05:30 UTC
[gpt-oss] fix model config with hf_config (#22401)
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Yongye Zhu committed
Aug 6, 2025 at 18:04 UTC
5c7cc33f4dafd4949a3f4bda815fa980d71ba45f
1 file changed
+3
-3
vllm/model_executor/models/gpt_oss.py
+3
-3
@@ -61,9 +61,9 @@ class OAIAttention(nn.Module):
61
"original_max_position_embeddings":
62
config.rope_scaling["original_max_position_embeddings"],
63
"beta_fast":
64
- config.rope_ntk_beta,
64
+ config.rope_scaling["beta_fast"],
65
"beta_slow":
66
- config.rope_ntk_alpha,
66
+ config.rope_scaling["beta_slow"],
67
},
68
is_neox_style=True,
69
)
@@ -154,7 +154,7 @@ class MLPBlock(torch.nn.Module):
154
dtype=torch.bfloat16)
155
assert config.intermediate_size % self.world_size == 0
156
self.experts = FusedMoE(num_experts=config.num_local_experts,
157
- top_k=config.num_experts_per_token,
157
+ top_k=config.num_experts_per_tok,
158
hidden_size=config.hidden_size,
159
intermediate_size=config.intermediate_size,
160
reduce_results=True,