root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-24 10:24:35 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

lora : fix llama conversion script with ROPE_FREQS (#9117)

Browse Source
This commit is contained in:
Xuan Son Nguyen 2024-08-23 12:58:53 +02:00 committed by GitHub
parent a07c32ea54
commit 3ba780e2a8
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 12 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
convert_hf_to_gguf.py
View File

@ -63,6 +63,7 @@ class Model:
model_name: str | None model_name: str | None
metadata_override: Path | None metadata_override: Path | None
dir_model_card: Path dir_model_card: Path
is_lora: bool
# subclasses should define this! # subclasses should define this!
model_arch: gguf.MODEL_ARCH model_arch: gguf.MODEL_ARCH
@ -70,7 +71,7 @@ class Model:
def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False, def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False,
use_temp_file: bool = False, eager: bool = False, use_temp_file: bool = False, eager: bool = False,
metadata_override: Path | None = None, model_name: str | None = None, metadata_override: Path | None = None, model_name: str | None = None,
split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False): split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False, is_lora: bool = False):
if type(self) is Model: if type(self) is Model:
raise TypeError(f"{type(self).__name__!r} should not be directly instantiated") raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
@ -92,6 +93,7 @@ class Model:
self.metadata_override = metadata_override self.metadata_override = metadata_override
self.model_name = model_name self.model_name = model_name
self.dir_model_card = dir_model # overridden in convert_lora_to_gguf.py self.dir_model_card = dir_model # overridden in convert_lora_to_gguf.py
self.is_lora = is_lora # true if model is used inside convert_lora_to_gguf.py
# Apply heuristics to figure out typical tensor encoding based on first layer tensor encoding type # Apply heuristics to figure out typical tensor encoding based on first layer tensor encoding type
if self.ftype == gguf.LlamaFileType.GUESSED: if self.ftype == gguf.LlamaFileType.GUESSED:
@ -1593,7 +1595,8 @@ class LlamaModel(Model):
smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor) smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
rope_factors.append(1 / ((1 - smooth) / factor + smooth)) rope_factors.append(1 / ((1 - smooth) / factor + smooth))
self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32)) if not self.is_lora:
self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
super().prepare_tensors() super().prepare_tensors()
@ -2140,8 +2143,9 @@ class Phi3MiniModel(Model):
if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2: if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}') raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32)) if not self.is_lora:
self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32)) self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32))
self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
@Model.register("PlamoForCausalLM") @Model.register("PlamoForCausalLM")
@ -3839,7 +3843,8 @@ class ExaoneModel(Model):
smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor) smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
rope_factors.append(1 / ((1 - smooth) / factor + smooth)) rope_factors.append(1 / ((1 - smooth) / factor + smooth))
self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32)) if not self.is_lora:
self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
super().prepare_tensors() super().prepare_tensors()

1
convert_lora_to_gguf.py
View File

@ -386,6 +386,7 @@ if __name__ == '__main__':
dry_run=args.dry_run, dry_run=args.dry_run,
dir_lora_model=dir_lora, dir_lora_model=dir_lora,
lora_alpha=alpha, lora_alpha=alpha,
is_lora=True,
) )
logger.info("Exporting model...") logger.info("Exporting model...")

2
tests/test-lora-conversion-inference.sh
View File

@ -14,7 +14,7 @@ MODELS_REPO_URL=https://huggingface.co/ggml-org/$MODELS_REPO
# Clone the Hugging Face repository if the directory does not exist # Clone the Hugging Face repository if the directory does not exist
if [ ! -d "$MODELS_REPO" ]; then if [ ! -d "$MODELS_REPO" ]; then
echo "Cloning the Hugging Face repository..." echo "Cloning the Hugging Face repository..."
git clone $MODELS_REPO_URL git clone $MODELS_REPO_URL --depth 1
else else
echo "Repository already exists. Skipping clone." echo "Repository already exists. Skipping clone."
fi fi