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

convert-hf : save memory with lazy evaluation (#7075)

Browse source 
* convert-hf : begin refactoring write_tensor

* convert : upgrade to sentencepiece v0.2.0

* convert-hf : remove unused n_dims in extra_*_tensors

* convert-hf : simplify MoE weights stacking

* convert-hf : flake8 linter doesn't like semicolons

* convert-hf : allow unusual model part names

For example, loading `model-00001-of-00001.safetensors` now works.

* convert-hf : fix stacking MoE expert tensors

`torch.stack` and `torch.cat` don't do the same thing.

* convert-hf : fix Mamba conversion

Tested to work even with a SentencePiece-based tokenizer.

* convert : use a string for the SentencePiece tokenizer path

* convert-hf : display tensor shape

* convert-hf : convert norms to f32 by default

* convert-hf : sort model part names

`os.listdir` is said to list files in arbitrary order.
Sorting the file names should let "model-00009-of-00042.safetensors"
be loaded before "model-00010-of-00042.safetensors".

* convert-hf : use an ABC for Model again

It seems Protocol can't be used as a statically type-checked ABC,
because its subclasses also can't be instantiated. (why did it seem to work?)

At least there's still a way to throw an error when forgetting to define
the `model_arch` property of any registered Model subclasses.

* convert-hf : use a plain class for Model, and forbid direct instantiation

There are no abstract methods used anyway,
so using ABC isn't really necessary.

* convert-hf : more consistent formatting of cmdline args

* convert-hf : align the message logged for converted tensors

* convert-hf : fix Refact conversion

* convert-hf : save memory with lazy evaluation

* convert-hf : flake8 doesn't like lowercase L as a variable name

* convert-hf : remove einops requirement for InternLM2

* convert-hf : faster model parts loading

Instead of pre-loading them all into a dict, iterate on the tensors
in the model parts progressively as needed in Model.write_tensors

Conversion for some architectures relies on checking for the presence
of specific tensor names, so for multi-part models, the weight map is read
from the relevant json file to quickly get these names up-front.

* convert-hf : minor changes for consistency

* gguf-py : add tqdm as a dependency

It's small, and used for a progress bar
in GGUFWriter.write_tensors_to_file
This commit is contained in:
compilade 2024-05-08 18:16:38 -04:00 • committed by GitHub
parent bc4bba364f
commit f98eb31c51
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
14 changed files with 873 additions and 1285 deletions
Download patch file Download diff file Expand all files Collapse all files

20
convert.py
View file

@ -284,6 +284,7 @@ class Params:
n_experts = None
n_experts_used = None
f_rope_freq_base = None
n_ff = None
# hack to determine LLaMA v1 vs v2 vs CodeLlama
if config.get("moe"):
@ -308,6 +309,8 @@ class Params:
n_experts_used = config["moe"]["num_experts_per_tok"]
f_rope_freq_base = 1e6
assert n_ff is not None
return Params(
n_vocab = model["tok_embeddings.weight"].shape[0],
n_embd = config["dim"],
@ -462,7 +465,8 @@ class SentencePieceVocab(Vocab):
# not found in alternate location either
raise FileNotFoundError('Cannot find tokenizer.model')
self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
self.sentencepiece_tokenizer = SentencePieceProcessor()
self.sentencepiece_tokenizer.LoadFromFile(str(fname_tokenizer))
vocab_size = self.sentencepiece_tokenizer.vocab_size()
new_tokens = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
@ -482,23 +486,23 @@ class SentencePieceVocab(Vocab):
def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
tokenizer = self.sentencepiece_tokenizer
for i in range(tokenizer.vocab_size()):
piece = tokenizer.id_to_piece(i)
piece = tokenizer.IdToPiece(i)
text = piece.encode("utf-8")
score: float = tokenizer.get_score(i)
score: float = tokenizer.GetScore(i)
toktype = gguf.TokenType.NORMAL
if tokenizer.is_unknown(i):
if tokenizer.IsUnknown(i):
toktype = gguf.TokenType.UNKNOWN
if tokenizer.is_control(i):
if tokenizer.IsControl(i):
toktype = gguf.TokenType.CONTROL
# NOTE: I think added_tokens are user defined.
# ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto
# if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED
if tokenizer.is_unused(i):
if tokenizer.IsUnused(i):
toktype = gguf.TokenType.UNUSED
if tokenizer.is_byte(i):
if tokenizer.IsByte(i):
toktype = gguf.TokenType.BYTE
yield text, score, toktype
@ -906,7 +910,7 @@ class LazyUnpickler(pickle.Unpickler):
def rebuild_from_type_v2(func, new_type, args, state):
return func(*args)
CLASSES = {
CLASSES: dict[tuple[str, str], type[LazyTensor] | LazyStorageKind] = {
# getattr used here as a workaround for mypy not being smart enough to determine
# the staticmethods have a __func__ attribute.
('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),