From 2db39c7049ae7830682150a30f3faf3ac7f8029c Mon Sep 17 00:00:00 2001 From: qizixi <22851944+zixi-qi@users.noreply.github.com> Date: Wed, 15 Jul 2026 20:34:41 -0700 Subject: [PATCH] [Bugfix][Spec Decode] Fix eagle3 first-layer qkv_proj prefix for quantized drafts (#48068) Signed-off-by: zixi-qi Co-authored-by: Claude Opus 4.8 (1M context) --- vllm/model_executor/models/llama_eagle3.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm/model_executor/models/llama_eagle3.py b/vllm/model_executor/models/llama_eagle3.py index 2d859bd491..549e8b7bf6 100644 --- a/vllm/model_executor/models/llama_eagle3.py +++ b/vllm/model_executor/models/llama_eagle3.py @@ -60,7 +60,7 @@ class LlamaDecoderLayer(LlamaDecoderLayer): self.self_attn.total_num_kv_heads, bias=qkv_bias, quant_config=quant_config, - prefix=maybe_prefix(prefix, "qkv_proj"), + prefix=maybe_prefix(prefix, "self_attn.qkv_proj"), ) self.hidden_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)