nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 23 Packages Projects Releases Wiki Activity

convert : fix Pixtral 12B --mistral-format conversion (3 bugs) (#22981)

Browse Source
This commit is contained in:
fredzillman
2026-05-12 21:46:01 +02:00
committed by GitHub
parent dded58b450
commit cce09f0b2b
1 changed files with 11 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
convert_hf_to_gguf.py
+9 -1
View File
@@ -2865,6 +2865,10 @@ class LlamaModel(TextModel):
# fix for SmolVLM2, missing `num_attention_heads` in config.json # fix for SmolVLM2, missing `num_attention_heads` in config.json
if self.hf_arch == "VLlama3ForCausalLM": if self.hf_arch == "VLlama3ForCausalLM":
self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32) self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
# Mistral consolidated format has no config.json; origin_hf_arch is HF-only.
if self.is_mistral_format:
self.origin_hf_arch = None
else:
hparams = ModelBase.load_hparams(self.dir_model, is_mistral_format=False) hparams = ModelBase.load_hparams(self.dir_model, is_mistral_format=False)
self.origin_hf_arch = hparams.get('architectures', [None])[0] self.origin_hf_arch = hparams.get('architectures', [None])[0]
@@ -13409,7 +13413,7 @@ class PixtralModel(LlavaVisionModel):
self.gguf_writer.add_vision_use_silu(True) self.gguf_writer.add_vision_use_silu(True)
# spatial_merge_size # spatial_merge_size
if self.find_vparam(["mm_projector_id"]) == "patch_merge": if self.find_vparam(["mm_projector_id"], optional=True) == "patch_merge":
self.gguf_writer.add_vision_spatial_merge_size( self.gguf_writer.add_vision_spatial_merge_size(
self.find_vparam(["spatial_merge_size"]) self.find_vparam(["spatial_merge_size"])
) )
@@ -13417,8 +13421,12 @@ class PixtralModel(LlavaVisionModel):
def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", ".bias")) -> str: def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", ".bias")) -> str:
if name == "vision_language_adapter.w_in.weight": if name == "vision_language_adapter.w_in.weight":
return "mm.1.weight" return "mm.1.weight"
elif name == "vision_language_adapter.w_in.bias":
return "mm.1.bias"
elif name == "vision_language_adapter.w_out.weight": elif name == "vision_language_adapter.w_out.weight":
return "mm.2.weight" return "mm.2.weight"
elif name == "vision_language_adapter.w_out.bias":
return "mm.2.bias"
return super().map_tensor_name(name, try_suffixes) return super().map_tensor_name(name, try_suffixes)