CompressedGemma commited on
Commit
e6d5776
·
verified ·
1 Parent(s): 2790fea

Upload slimshard.py

Browse files
Files changed (1) hide show
  1. slimshard.py +4 -0
slimshard.py CHANGED
@@ -2297,6 +2297,8 @@ def main() -> None:
2297
  rec.add_argument("--output", required=True, help="full model .safetensors")
2298
  rec.add_argument("--config", default=None, help="config.json (defaults to index/shard dir)")
2299
  rec.add_argument("--export-hf-dir", default=None, help="optional ready-to-load HF model directory")
 
 
2300
  rec.add_argument("--device", default="auto", choices=("auto", "cuda", "cpu"))
2301
 
2302
  args = ap.parse_args()
@@ -2325,6 +2327,8 @@ def main() -> None:
2325
  device=args.device,
2326
  verbose=True,
2327
  export_hf_dir=args.export_hf_dir,
 
 
2328
  )
2329
  else:
2330
  print(json.dumps(slimshard_info(args.input), indent=2))
 
2297
  rec.add_argument("--output", required=True, help="full model .safetensors")
2298
  rec.add_argument("--config", default=None, help="config.json (defaults to index/shard dir)")
2299
  rec.add_argument("--export-hf-dir", default=None, help="optional ready-to-load HF model directory")
2300
+ rec.add_argument("--target-params-b", type=float, default=None, help="optional target parameter budget in billions (e.g. 30.0)")
2301
+ rec.add_argument("--max-experts", type=int, default=None, help="optional max routed experts per layer to keep")
2302
  rec.add_argument("--device", default="auto", choices=("auto", "cuda", "cpu"))
2303
 
2304
  args = ap.parse_args()
 
2327
  device=args.device,
2328
  verbose=True,
2329
  export_hf_dir=args.export_hf_dir,
2330
+ target_params_b=args.target_params_b,
2331
+ max_experts=args.max_experts,
2332
  )
2333
  else:
2334
  print(json.dumps(slimshard_info(args.input), indent=2))