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 : avoid dequantizing mxfp4 for GPT-OSS

Browse Source
This commit is contained in:
Francis Couture-Harpin
2025-10-24 07:46:34 -04:00
parent 5a91109a5d
commit d7f794eadb
1 changed files with 7 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
convert_hf_to_gguf.py
View File

@@ -8943,6 +8943,13 @@ class SmolLM3Model(LlamaModel):
class GptOssModel(TextModel): class GptOssModel(TextModel):
model_arch = gguf.MODEL_ARCH.GPT_OSS model_arch = gguf.MODEL_ARCH.GPT_OSS
# TODO: remove once MXFP4 is supported more generally
def dequant_model(self):
quant_config = self.hparams.get("quantization_config")
if quant_config is not None and quant_config.get("quant_method") == "mxfp4":
return
return super().dequant_model()
def transform_nibble_layout(self, tensor): def transform_nibble_layout(self, tensor):
assert tensor.dtype == torch.uint8 assert tensor.dtype == torch.uint8
assert tensor.shape[-1] == 16 assert tensor.shape[-1] == 16