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

gguf : eliminate MODEL_TENSOR_NAMES

Browse source
This commit is contained in:
Cebtenzzre 2023-09-30 19:55:24 -04:00
parent fd5e2268a5
commit 3724ad695d
4 changed files with 13 additions and 22 deletions
Download patch file Download diff file Expand all files Collapse all files

3
convert.py
View file

@ -42,7 +42,6 @@ if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'):
NDArray: TypeAlias = 'np.ndarray[Any, Any]' NDArray: TypeAlias = 'np.ndarray[Any, Any]'
ARCH = gguf.MODEL_ARCH.LLAMA ARCH = gguf.MODEL_ARCH.LLAMA
NAMES=gguf.MODEL_TENSOR_NAMES[ARCH]
DEFAULT_CONCURRENCY = 8 DEFAULT_CONCURRENCY = 8
# #
@ -953,7 +952,7 @@ class OutputFile:
of.close() of.close()
def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType: def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
wq_type = model[NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0)+".weight"].data_type wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0)+".weight"].data_type
if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32): if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
return GGMLFileType.AllF32 return GGMLFileType.AllF32

2
examples/finetune/convert-finetune-checkpoint-to-gguf.py
View file

@ -313,7 +313,7 @@ class ModelParams:
gguf_writer.add_feed_forward_length(self.get_n_ff()) gguf_writer.add_feed_forward_length(self.get_n_ff())
def tensor_name(key, bid=None, suffix=".weight"): def tensor_name(key, bid=None, suffix=".weight"):
return gguf.MODEL_TENSOR_NAMES[gguf.MODEL_ARCH.LLAMA][key].format(bid=bid) + suffix return gguf.TENSOR_NAMES[key].format(bid=bid) + suffix
class Layer: class Layer:
def __init__(self, params, lora_params, bid): def __init__(self, params, lora_params, bid):

2
examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py
View file

@ -364,7 +364,7 @@ class ModelParams:
gguf_writer.add_feed_forward_length(self.get_n_ff()) gguf_writer.add_feed_forward_length(self.get_n_ff())
def tensor_name(key, bid=None): def tensor_name(key, bid=None):
return gguf.MODEL_TENSOR_NAMES[gguf.MODEL_ARCH.LLAMA][key].format(bid=bid) + ".weight" return gguf.TENSOR_NAMES[key].format(bid=bid) + ".weight"
class Layer: class Layer:
def __init__(self, params, bid): def __init__(self, params, bid):

26
gguf-py/gguf/gguf.py
View file

@ -212,10 +212,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
# TODO # TODO
} }
MODEL_TENSOR_NAMES: dict[MODEL_ARCH, dict[MODEL_TENSOR, str]] = {
m: {t: TENSOR_NAMES[t] for t in ts} for m, ts in MODEL_TENSORS.items()
}
# tensors that will not be serialized # tensors that will not be serialized
MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_ARCH.LLAMA: [ MODEL_ARCH.LLAMA: [
@ -363,28 +359,24 @@ class TensorNameMap:
mapping: dict[str, tuple[MODEL_TENSOR, str]] mapping: dict[str, tuple[MODEL_TENSOR, str]]
tensor_names: dict[MODEL_TENSOR, str]
def __init__(self, arch: MODEL_ARCH, n_blocks: int): def __init__(self, arch: MODEL_ARCH, n_blocks: int):
mapping = self.mapping = {} self.mapping = {}
tensor_names = self.tensor_names = MODEL_TENSOR_NAMES[arch]
for tensor, keys in self.mappings_cfg.items(): for tensor, keys in self.mappings_cfg.items():
tensor_name = tensor_names.get(tensor) if tensor not in MODEL_TENSORS[ARCH]:
if tensor_name is None:
continue continue
mapping[tensor_name] = (tensor, tensor_name) tensor_name = TENSOR_NAMES[tensor]
self.mapping[tensor_name] = (tensor, tensor_name)
for key in keys: for key in keys:
mapping[key] = (tensor, tensor_name) self.mapping[key] = (tensor, tensor_name)
for bid in range(n_blocks): for bid in range(n_blocks):
for tensor, keys in self.block_mappings_cfg.items(): for tensor, keys in self.block_mappings_cfg.items():
tensor_name = tensor_names.get(tensor) if tensor not in MODEL_TENSORS[ARCH]:
if tensor_name is None:
continue continue
tensor_name = tensor_name.format(bid = bid) tensor_name = TENSOR_NAMES[tensor].format(bid = bid)
mapping[tensor_name] = (tensor, tensor_name) self.mapping[tensor_name] = (tensor, tensor_name)
for key in keys: for key in keys:
key = key.format(bid = bid) key = key.format(bid = bid)
mapping[key] = (tensor, tensor_name) self.mapping[key] = (tensor, tensor_name)
def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None: def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None:
result = self.mapping.get(key) result = self.mapping.get(key)