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

Enable modelopt gemma3 nvfp4/fp8, make workflow more robust (#22771)

Browse Source 
Signed-off-by: Zhiyu Cheng <zhiyuc@nvidia.com>
Signed-off-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
Zhiyu
2025-09-19 15:40:33 -07:00
committed by GitHub
parent 711e912946
commit 431535b522
7 changed files with 82 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
vllm/model_executor/models/siglip.py
View File

@@ -20,7 +20,8 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.vocab_parallel_embedding import (
VocabParallelEmbedding)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader, maybe_remap_kv_scale_name)
from .vision import VisionEncoderInfo, resolve_visual_encoder_outputs
@@ -506,6 +507,21 @@ class SiglipVisionModel(nn.Module):
if layer_idx >= layer_count:
continue
# Check if this is a scale parameter that needs remapping first
if name.endswith(
(".k_scale", ".v_scale", ".q_scale", ".prob_scale")):
# Try to remap the scale name first
remapped_name = maybe_remap_kv_scale_name(name, params_dict)
if remapped_name is not None and remapped_name in params_dict:
# Successfully remapped, use the remapped name
param = params_dict[remapped_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(remapped_name)
continue
# If remapping failed, continue with normal processing
for (param_name, weight_name, shard_id) in stacked_params_mapping:
if weight_name not in name:
continue