xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 03:05:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Model] Gemma3: Fix GGUF loading and quantization (#26189)

Browse Source 
Signed-off-by: Luciano Martins <lucianommartins@users.noreply.github.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Luciano Martins <lucianommartins@users.noreply.github.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Luciano Martins 2025-10-09 04:00:53 -03:00 committed by GitHub
parent 5e49c3e777
commit 1317028aa8
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 14 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/model_executor/model_loader/gguf_loader.py
View File

@ -72,6 +72,10 @@ class GGUFModelLoader(BaseModelLoader):
# hack: ggufs have a different name than transformers
if model_type == "cohere":
model_type = "command-r"
if model_type == "gemma3_text":
# Gemma3 models use "gemma3_text" in HuggingFace but
# "gemma3" in GGUF architecture naming
model_type = "gemma3"
if model_type in ("deepseek_v3", "deepseek_v2"):
model_type = "deepseek2"
# GGUF layer map assumes that we will have a merged expert weights

10
vllm/model_executor/models/gemma3.py
View File

@ -372,6 +372,7 @@ class Gemma3Model(nn.Module):
self.embed_tokens = VocabParallelEmbedding(
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=f"{prefix}.embed_tokens",
)
self.start_layer, self.end_layer, self.layers = make_layers(
@ -442,6 +443,15 @@ class Gemma3Model(nn.Module):
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
for name, loaded_weight in weights:
# Revert +1 during llama.cpp conversion
# see: https://github.com/ggml-org/llama.cpp/blob/be7c3034108473beda214fd1d7c98fd6a7a3bdf5/convert_hf_to_gguf.py#L3397-L3400
if (
self.quant_config
and self.quant_config.get_name() == "gguf"
and name.endswith("norm.weight")
):
loaded_weight -= 1
if self.quant_config is not None and (
scale_name := self.quant_config.get_cache_scale(name)
):