From 4d196981d4db79e0105b939eaa7ecd40385b721c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Sun, 17 Aug 2025 14:47:42 +0200 Subject: [PATCH] convert : force patch_embd weights to F16 or F32 to avoid broken GGUFs (#15367) * force patch_embd weights to f32 * use MmprojModel base tensor_force_quant instead --- convert_hf_to_gguf.py | 28 ++++++++++++---------------- 1 file changed, 12 insertions(+), 16 deletions(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index bd21e55f4..b45c8f1d7 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -1334,6 +1334,12 @@ class MmprojModel(ModelBase): return None 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") class GPTNeoXModel(TextModel): @@ -2305,10 +2311,9 @@ class SmolVLMModel(MmprojModel): self.gguf_writer.add_vision_use_gelu(True) def tensor_force_quant(self, name, new_name, bid, n_dims): - del bid, new_name, n_dims # unused if ".embeddings." in name: 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]]: 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)) 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: 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]]: del bid # unused @@ -3374,10 +3376,9 @@ class Qwen25OmniModel(Qwen2VLVisionModel): yield ("audio_tower.embed_positions.weight", pos_embd) 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: 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]]: if name.startswith("thinker."): @@ -3423,12 +3424,9 @@ class InternVisionModel(MmprojModel): self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio)) 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: return gguf.GGMLQuantizationType.F32 - return False + return super().tensor_force_quant(name, new_name, bid, n_dims) def _mapping_interns1_name(self, name): names_map = { @@ -5062,13 +5060,12 @@ class Gemma3VisionModel(MmprojModel): self.gguf_writer.add_vision_projector_scale_factor(proj_scale_factor) 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 if "input_projection" in name: return gguf.GGMLQuantizationType.F16 if ".embeddings." in name: 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]]: 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)) 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: 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]]: del bid # unused