]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commitdiff
convert.py: Support models which are stored in a single pytorch_model.bin (#1469)
authorTom Jobbins <redacted>
Tue, 16 May 2023 22:04:35 +0000 (23:04 +0100)
committerGitHub <redacted>
Tue, 16 May 2023 22:04:35 +0000 (00:04 +0200)
* Support models in a single pytorch_model.bin

* Remove spurious line with typo

convert.py

index 8f4f0399e1c52dd1875441229dbc80b7ff783a92..ece5a0266836529444a845ecfd2ef64413bd4a7e 100644 (file)
@@ -121,7 +121,6 @@ def make_tensors_list() -> List[str]:
             f'layers.{i}.feed_forward.w1.weight',
             f'layers.{i}.feed_forward.w2.weight',
             f'layers.{i}.feed_forward.w3.weight',
-            f'layers.{i}.atttention_norm.weight',
             f'layers.{i}.ffn_norm.weight',
         ]
     return ret
@@ -1055,7 +1054,7 @@ def load_some_model(path: Path) -> ModelPlus:
         files = list(path.glob("model-00001-of-*.safetensors"))
         if not files:
             # Try the PyTorch patterns too, with lower priority
-            globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt"]
+            globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin" ]
             files = [file for glob in globs for file in path.glob(glob)]
         if not files:
             # Try GGML too, but with lower priority, since if both a non-GGML