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

4
vllm/model_executor/models/deepseek_v2.py
View File

@@ -85,7 +85,7 @@ from vllm.v1.attention.backends.mla.indexer import (
)
from vllm.v1.kv_cache_interface import KVCacheSpec, MLAAttentionSpec
from .interfaces import MixtureOfExperts, SupportsLoRA, SupportsPP
from .interfaces import MixtureOfExperts, SupportsEagle, SupportsLoRA, SupportsPP
from .utils import (
PPMissingLayer,
is_pp_missing_parameter,
@@ -1311,7 +1311,7 @@ class DeepseekV2MixtureOfExperts(MixtureOfExperts):
class DeepseekV2ForCausalLM(
nn.Module, SupportsPP, DeepseekV2MixtureOfExperts, SupportsLoRA
nn.Module, SupportsPP, DeepseekV2MixtureOfExperts, SupportsLoRA, SupportsEagle
):
packed_modules_mapping = {
"gate_up_proj": ["gate_proj", "up_proj"],