zhangyiss/mlx-examples
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx-examples.git synced 2025-09-04 07:44:34 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

fix load model in lora/lora.py

Browse Source
This commit is contained in:
pjh6818
2023-12-27 22:33:09 +09:00
parent 50fceb1a28
commit 77ab988a72
1 changed files with 14 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
lora/lora.py
View File

@@ -334,16 +334,26 @@ def generate(model, prompt, tokenizer, args):
def load_model(folder: str, dtype=mx.float16): def load_model(folder: str, dtype=mx.float16):
model_path = Path(folder) model_path = Path(folder)
weights = mx.load(str(model_path / "weights.npz"))
tokenizer = Tokenizer(str(model_path / "tokenizer.model")) tokenizer = Tokenizer(str(model_path / "tokenizer.model"))
with open(model_path / "params.json", "r") as f: with open(model_path / "params.json", "r") as f:
config = json.loads(f.read()) config = json.loads(f.read())
n_heads = config["n_heads"]
if "n_kv_heads" not in config:
config["n_kv_heads"] = n_heads
if "head_dim" not in config:
config["head_dim"] = config["dim"] // n_heads
if "hidden_dim" not in config:
config["hidden_dim"] = weights["layers.0.feed_forward.w1.weight"].shape[0]
if config.get("vocab_size", -1) < 0: if config.get("vocab_size", -1) < 0:
config["vocab_size"] = tokenizer.vocab_size config["vocab_size"] = weights["output.weight"].shape[-1]
model_args = ModelArgs(**config) unused = ["multiple_of", "ffn_dim_multiplier"]
weights = mx.load(str(model_path / "weights.npz")) for k in unused:
if k in config:
config.pop(k)
model = Model(ModelArgs(**config))
weights = tree_unflatten(list(weights.items())) weights = tree_unflatten(list(weights.items()))
weights = tree_map(lambda p: p.astype(dtype), weights) weights = tree_map(lambda p: p.astype(dtype), weights)
model = Model(model_args)
model.update(weights) model.update(weights)
return model, tokenizer return model, tokenizer