CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-30 08:42:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

convert : force patch_embd weights to F16 or F32 to avoid broken GGUFs (#15367)

Browse Source 
* force patch_embd weights to f32

* use MmprojModel base tensor_force_quant instead
This commit is contained in:
Sigbjørn Skjæret
2025-08-17 14:47:42 +02:00
committed by GitHub
parent b143fbc87a
commit 4d196981d4
1 changed files with 12 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
convert_hf_to_gguf.py
View File

@@ -1334,6 +1334,12 @@ class MmprojModel(ModelBase):
return None return None
raise KeyError(f"could not find any of: {keys}") raise KeyError(f"could not find any of: {keys}")
def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, name, n_dims # unused
if ".patch_embd.weight" in new_name:
return gguf.GGMLQuantizationType.F16 if self.ftype == gguf.LlamaFileType.MOSTLY_F16 else gguf.GGMLQuantizationType.F32
return False
@ModelBase.register("GPTNeoXForCausalLM") @ModelBase.register("GPTNeoXForCausalLM")
class GPTNeoXModel(TextModel): class GPTNeoXModel(TextModel):
@@ -2305,10 +2311,9 @@ class SmolVLMModel(MmprojModel):
self.gguf_writer.add_vision_use_gelu(True) self.gguf_writer.add_vision_use_gelu(True)
def tensor_force_quant(self, name, new_name, bid, n_dims): def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, new_name, n_dims # unused
if ".embeddings." in name: if ".embeddings." in name:
return gguf.GGMLQuantizationType.F32 return gguf.GGMLQuantizationType.F32
return False return super().tensor_force_quant(name, new_name, bid, n_dims)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused del bid # unused
@@ -3296,12 +3301,9 @@ class Qwen2VLVisionModel(MmprojModel):
self.gguf_writer.add_vision_attention_layernorm_eps(self.global_config.get("rms_norm_eps", 1e-6)) self.gguf_writer.add_vision_attention_layernorm_eps(self.global_config.get("rms_norm_eps", 1e-6))
def tensor_force_quant(self, name, new_name, bid, n_dims): def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, name, n_dims # unused
if ".patch_embd." in new_name:
return gguf.GGMLQuantizationType.F16
if ".position_embd." in new_name: if ".position_embd." in new_name:
return gguf.GGMLQuantizationType.F32 return gguf.GGMLQuantizationType.F32
return False return super().tensor_force_quant(name, new_name, bid, n_dims)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused del bid # unused
@@ -3374,10 +3376,9 @@ class Qwen25OmniModel(Qwen2VLVisionModel):
yield ("audio_tower.embed_positions.weight", pos_embd) yield ("audio_tower.embed_positions.weight", pos_embd)
def tensor_force_quant(self, name, new_name, bid, n_dims): def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, new_name, n_dims # unused
if ".conv" in name and ".weight" in name: if ".conv" in name and ".weight" in name:
return gguf.GGMLQuantizationType.F16 return gguf.GGMLQuantizationType.F16
return False return super().tensor_force_quant(name, new_name, bid, n_dims)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
if name.startswith("thinker."): if name.startswith("thinker."):
@@ -3423,12 +3424,9 @@ class InternVisionModel(MmprojModel):
self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio)) self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))
def tensor_force_quant(self, name, new_name, bid, n_dims): def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, name, n_dims # unused
if ".patch_embd." in new_name:
return gguf.GGMLQuantizationType.F16
if ".position_embd." in new_name: if ".position_embd." in new_name:
return gguf.GGMLQuantizationType.F32 return gguf.GGMLQuantizationType.F32
return False return super().tensor_force_quant(name, new_name, bid, n_dims)
def _mapping_interns1_name(self, name): def _mapping_interns1_name(self, name):
names_map = { names_map = {
@@ -5062,13 +5060,12 @@ class Gemma3VisionModel(MmprojModel):
self.gguf_writer.add_vision_projector_scale_factor(proj_scale_factor) self.gguf_writer.add_vision_projector_scale_factor(proj_scale_factor)
def tensor_force_quant(self, name, new_name, bid, n_dims): def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, new_name, n_dims # unused
# related to https://github.com/ggml-org/llama.cpp/issues/13025 # related to https://github.com/ggml-org/llama.cpp/issues/13025
if "input_projection" in name: if "input_projection" in name:
return gguf.GGMLQuantizationType.F16 return gguf.GGMLQuantizationType.F16
if ".embeddings." in name: if ".embeddings." in name:
return gguf.GGMLQuantizationType.F32 return gguf.GGMLQuantizationType.F32
return False return super().tensor_force_quant(name, new_name, bid, n_dims)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused del bid # unused
@@ -7727,10 +7724,9 @@ class WhisperEncoderModel(MmprojModel):
self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-5)) self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-5))
def tensor_force_quant(self, name, new_name, bid, n_dims): def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, new_name, n_dims # unused
if ".conv" in name and ".weight" in name: if ".conv" in name and ".weight" in name:
return gguf.GGMLQuantizationType.F16 return gguf.GGMLQuantizationType.F16
return False return super().tensor_force_quant(name, new_name, bid, n_dims)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused del bid # unused