A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 23:15 UTC
[Misc] Fix arg names in quantizer script (#5507)
Allen.Dou committed
Jun 14, 2024 at 10:02 UTC
55d6361b13ae6328de809f57a69b719c1600040a
1 file changed
+8
-8
examples/fp8/quantizer/quantize.py
+8
-8
@@ -332,7 +332,7 @@ def main(args):
332
333
if __name__ == "__main__":
334
parser = argparse.ArgumentParser(description=__doc__)
335
- parser.add_argument("--model_dir",
335
+ parser.add_argument("--model-dir",
336
help="Specify where the HuggingFace model is",
337
required=True)
338
parser.add_argument("--device", default="cuda")
@@ -346,19 +346,19 @@ if __name__ == "__main__":
346
"full_prec"
347
],
348
)
349
- parser.add_argument("--batch_size",
349
+ parser.add_argument("--batch-size",
350
help="Batch size for calibration.",
351
type=int,
352
default=1)
353
- parser.add_argument("--calib_size",
353
+ parser.add_argument("--calib-size",
354
help="Number of samples for calibration.",
355
type=int,
356
default=512)
357
- parser.add_argument("--output_dir", default="exported_model")
358
- parser.add_argument("--tp_size", type=int, default=1)
359
- parser.add_argument("--pp_size", type=int, default=1)
360
- parser.add_argument("--awq_block_size", type=int, default=128)
361
- parser.add_argument("--kv_cache_dtype",
357
+ parser.add_argument("--output-dir", default="exported_model")
358
+ parser.add_argument("--tp-size", type=int, default=1)
359
+ parser.add_argument("--pp-size", type=int, default=1)
360
+ parser.add_argument("--awq-block-size", type=int, default=128)
361
+ parser.add_argument("--kv-cache-dtype",
362
help="KV Cache dtype.",
363
default=None,
364
choices=["int8", "fp8", None])