happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

qwen2.5vl ok

This commit is contained in:
Xuan Son Nguyen 2025-04-30 13:06:01 +02:00
parent 792387b164
commit f7260c2d2b
2 changed files with 6 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
convert_hf_to_gguf.py
View File

@ -1103,7 +1103,7 @@ class VisionModel(ModelBase):
# preprocessor config # preprocessor config
self.gguf_writer.add_vision_image_mean(self.preprocessor_config["image_mean"]) self.gguf_writer.add_vision_image_mean(self.preprocessor_config["image_mean"])
self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_mean"]) self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_std"])
def write_vocab(self): def write_vocab(self):
raise ValueError("VisionModel does not support vocab writing") raise ValueError("VisionModel does not support vocab writing")
@ -2563,8 +2563,9 @@ class Qwen2VLVisionModel(VisionModel):
# rename config.json values # rename config.json values
self.hparams["num_attention_heads"] = self.hparams.get("num_heads") self.hparams["num_attention_heads"] = self.hparams.get("num_heads")
self.hparams["num_hidden_layers"] = self.hparams.get("depth") self.hparams["num_hidden_layers"] = self.hparams.get("depth")
self.hparams["intermediate_size"] = self.hparams.get("hidden_size") if "embed_dim" in self.hparams: # qwen2vl
self.hparams["hidden_size"] = self.hparams.get("embed_dim") self.hparams["intermediate_size"] = self.hparams.get("hidden_size")
self.hparams["hidden_size"] = self.hparams.get("embed_dim")
def set_gguf_parameters(self): def set_gguf_parameters(self):
super().set_gguf_parameters() super().set_gguf_parameters()

4
gguf-py/gguf/tensor_mapping.py
View File

@ -990,7 +990,7 @@ class TensorNameMap:
"model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3 (note: name is swapped) "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3 (note: name is swapped)
"vision_tower.transformer.layers.{bid}.feed_forward.up_proj", # pixtral "vision_tower.transformer.layers.{bid}.feed_forward.up_proj", # pixtral
"visual.blocks.{bid}.mlp.fc2", # qwen2vl "visual.blocks.{bid}.mlp.fc2", # qwen2vl
"visual.blocks.{bid}.mlp.down_proj", # qwen2.5vl "visual.blocks.{bid}.mlp.up_proj", # qwen2.5vl
), ),
MODEL_TENSOR.V_ENC_FFN_GATE: ( MODEL_TENSOR.V_ENC_FFN_GATE: (
@ -1004,7 +1004,7 @@ class TensorNameMap:
"model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3 (note: name is swapped) "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3 (note: name is swapped)
"vision_tower.transformer.layers.{bid}.feed_forward.down_proj", # pixtral "vision_tower.transformer.layers.{bid}.feed_forward.down_proj", # pixtral
"visual.blocks.{bid}.mlp.fc1", # qwen2vl "visual.blocks.{bid}.mlp.fc1", # qwen2vl
"visual.blocks.{bid}.mlp.up_proj", # qwen2.5vl "visual.blocks.{bid}.mlp.down_proj", # qwen2.5vl
), ),
MODEL_TENSOR.V_PRE_NORM: ( MODEL_TENSOR.V_PRE_NORM: (