biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Add support for Eagle with separate lm-head and embed_tokens layers (#28549)

Browse Source 
Signed-off-by: Eldar Kurtic <8884008+eldarkurtic@users.noreply.github.com>
This commit is contained in:
Eldar Kurtić
2025-11-15 15:12:02 +01:00
committed by GitHub
parent 085a525332
commit e439c784fa
12 changed files with 205 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/models/llama.py
View File

@@ -58,7 +58,7 @@ from vllm.model_executor.model_loader.weight_utils import (
)
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP
from .interfaces import SupportsEagle, SupportsEagle3, SupportsLoRA, SupportsPP
from .utils import (
AutoWeightsLoader,
PPMissingLayer,
@@ -529,7 +529,9 @@ class LlamaModel(nn.Module):
return loaded_params
class LlamaForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
class LlamaForCausalLM(
nn.Module, SupportsLoRA, SupportsPP, SupportsEagle, SupportsEagle3
):
packed_modules_mapping = {
"qkv_proj": ["q_proj", "k_proj", "v_proj"],
"gate_up_proj": ["gate_proj", "up_proj"],