[Bugfix][Spec Decode] Fix eagle3 first-layer qkv_proj prefix for quantized drafts (#48068)
Signed-off-by: zixi-qi <zixi@inferact.ai> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
qizixi committed
Jul 15, 2026 at 20:34 UTC
2db39c7049ae7830682150a30f3faf3ac7f8029c
1 file changed
+1
-1
vllm/model_executor/models/llama_eagle3.py
+1
-1
@@ -60,7 +60,7 @@ class LlamaDecoderLayer(LlamaDecoderLayer):
60
self.self_attn.total_num_kv_heads,
61
bias=qkv_bias,
62
quant_config=quant_config,
63
- prefix=maybe_prefix(prefix, "qkv_proj"),
63
+ prefix=maybe_prefix(prefix, "self_attn.qkv_proj"),
64
)
65
66
self.hidden_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)