Skip to content

Defer resolution of the default value of arguments used by quantize #2738

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 2 commits into from
Closed

Defer resolution of the default value of arguments used by quantize #2738

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d91861e
Plumb group_size to 4b quant (#2734)
Jack-Khuu Mar 28, 2024
0bdf70f
Defer resolution of the default value of arguments used by quantize (…
Jack-Khuu Mar 28, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
40 changes: 29 additions & 11 deletions examples/models/llama2/export_llama_lib.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -218,8 +218,8 @@ def quantize(
group_size: int = 128,
# following arguments only used for GPTQ
calibration_tasks: Optional[list] = None,
calibration_limit: int = 5,
calibration_seq_length: int = 100,
calibration_limit: int = 100,
calibration_seq_length: int = 2048,
pad_calibration_inputs: bool = False,
percdamp: float = 0.01,
blocksize: int = 128,
Expand Down Expand Up @@ -254,7 +254,9 @@ def quantize(
elif qmode == "8da4w":
from torchao.quantization.quant_api import Int8DynActInt4WeightQuantizer

model = Int8DynActInt4WeightQuantizer(precision=torch_dtype).quantize(model)
model = Int8DynActInt4WeightQuantizer(
precision=torch_dtype, group_size=group_size
).quantize(model)
if verbose_export():
print("quantized model:", model)
return model
Expand Down Expand Up @@ -340,19 +342,19 @@ def build_args_parser() -> argparse.ArgumentParser:
"--calibration_tasks",
nargs="+",
type=str,
default=[],
default=None,
help="Tasks for GPTQ calibration",
)
parser.add_argument(
"--calibration_limit",
type=int,
default=5,
default=None,
help="number of samples used for calibration",
)
parser.add_argument(
"--calibration_seq_length",
type=int,
default=2048,
default=None,
help="Sequence length for GPTQ calibration",
)
parser.add_argument(
Expand Down Expand Up @@ -406,7 +408,11 @@ def build_args_parser() -> argparse.ArgumentParser:
help="Use cProfile to profile model export. Results saved to profile_path as a html file.",
)
parser.add_argument(
"-G", "--group_size", default=None, help="group_size for weight quantization"
"-G",
"--group_size",
type=int,
default=None,
help="group_size for weight quantization",
)

parser.add_argument(
Expand Down Expand Up @@ -525,9 +531,25 @@ def _prepare_for_llama_export(modelname: str, args) -> LlamaEdgeManager:
transforms = []
if args.quantization_mode:
modelname = f"{modelname}_q"

# If these optional args are None, don't provide them to quantize()
quant_args_str = [
"group_size",
"calibration_tasks",
"calibration_limit",
"calibration_seq_length",
]
arg_dict = vars(args)
quant_args = {
param: val
for param in quant_args_str
if (val := arg_dict.get(param)) is not None
}

transforms.append(
partial(
quantize,
**quant_args,
qmode=args.quantization_mode,
activation_dtype=dtype_override,
checkpoint_path=(
Expand All @@ -536,10 +558,6 @@ def _prepare_for_llama_export(modelname: str, args) -> LlamaEdgeManager:
tokenizer_path=(
Path(path) if (path := args.tokenizer_path) is not None else None
),
group_size=args.group_size,
calibration_tasks=args.calibration_tasks,
calibration_limit=args.calibration_limit,
calibration_seq_length=args.calibration_seq_length,
)
)

Expand Down