vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

convert-hf : add --outtype auto-f16

Browse source 
A reason for this to exist is for model quantizers who want an initial
GGUF with the most fidelity to the original model while still using
a 16-bit float type instead of 32-bit floats.
This commit is contained in:
Francis Couture-Harpin 2024-05-09 15:16:09 -04:00
parent 95930da30e
commit 58b515cad6
2 changed files with 14 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

14
convert-hf-to-gguf.py
View file

@ -83,6 +83,15 @@ class Model:
self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"]) self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
self.tensor_names = None self.tensor_names = None
if self.ftype == gguf.LlamaFileType.GUESSED:
# NOTE: can't use field "torch_dtype" in config.json, because some finetunes lie.
_, first_tensor = next(self.get_tensors())
if first_tensor.dtype == torch.float16:
logger.info(f"choosing --outtype f16 from first tensor type ({first_tensor.dtype})")
self.ftype = gguf.LlamaFileType.MOSTLY_F16
else:
logger.info(f"choosing --outtype bf16 from first tensor type ({first_tensor.dtype})")
self.ftype = gguf.LlamaFileType.MOSTLY_BF16
@classmethod @classmethod
def __init_subclass__(cls): def __init_subclass__(cls):
@ -2394,8 +2403,8 @@ def parse_args() -> argparse.Namespace:
help="path to write to; default: based on input", help="path to write to; default: based on input",
) )
parser.add_argument( parser.add_argument(
"--outtype", type=str, choices=["f32", "f16", "bf16"], default="f16", "--outtype", type=str, choices=["f32", "f16", "bf16", "auto-f16"], default="f16",
help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16", help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, auto-f16 for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
) )
parser.add_argument( parser.add_argument(
"--bigendian", action="store_true", "--bigendian", action="store_true",
@ -2453,6 +2462,7 @@ def main() -> None:
"f32": gguf.LlamaFileType.ALL_F32, "f32": gguf.LlamaFileType.ALL_F32,
"f16": gguf.LlamaFileType.MOSTLY_F16, "f16": gguf.LlamaFileType.MOSTLY_F16,
"bf16": gguf.LlamaFileType.MOSTLY_BF16, "bf16": gguf.LlamaFileType.MOSTLY_BF16,
"auto-f16": gguf.LlamaFileType.GUESSED, # TODO: use a more appropriate "auto" type
} }
if args.outfile is not None: if args.outfile is not None:

2
gguf-py/gguf/constants.py
View file

@ -861,6 +861,8 @@ class LlamaFileType(IntEnum):
MOSTLY_IQ1_M = 31 # except 1d tensors MOSTLY_IQ1_M = 31 # except 1d tensors
MOSTLY_BF16 = 32 # except 1d tensors MOSTLY_BF16 = 32 # except 1d tensors
GUESSED = 1024 # not specified in the model file
class GGUFEndian(IntEnum): class GGUFEndian(IntEnum):
LITTLE = 0 LITTLE = 0