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 13, 09:30 UTC
[Bugfix] Fix gpt_oss packed_modules_mapping (#28536)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Nov 12, 2025 at 21:02 UTC
a9d18b51078d29f732e2811e91e8c002b1d1d0c4
1 file changed
+5
-5
vllm/model_executor/models/gpt_oss.py
+5
-5
@@ -92,7 +92,7 @@ class OAIAttention(nn.Module):
92
self.scaling = self.head_dim**-0.5
93
self.rope_theta = config.rope_theta
94
95
- self.qkv = QKVParallelLinear(
95
+ self.qkv_proj = QKVParallelLinear(
96
hidden_size=self.hidden_size,
97
head_size=self.head_dim,
98
total_num_heads=self.num_attention_heads,
@@ -129,7 +129,7 @@ class OAIAttention(nn.Module):
129
def forward(
130
self, hidden_states: torch.Tensor, positions: torch.Tensor
131
) -> torch.Tensor:
132
- qkv, _ = self.qkv(hidden_states)
132
+ qkv, _ = self.qkv_proj(hidden_states)
133
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
134
q, k = self.rotary_emb(positions, q, k)
135
v = v.contiguous()
@@ -606,9 +606,9 @@ class GptOssModel(nn.Module):
606
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
607
stacked_params_mapping = [
608
# (param_name, shard_name, shard_id)
609
- (".qkv", ".q_proj", "q"),
610
- (".qkv", ".k_proj", "k"),
611
- (".qkv", ".v_proj", "v"),
609
+ (".qkv_proj", ".q_proj", "q"),
610
+ (".qkv_proj", ".k_proj", "k"),
611
+ (".qkv_proj", ".v_proj", "v"),
612
]
613
614
tp_rank = get_tensor_model_parallel_rank()