Fix llm.export_ckpt

Signed-off-by: Hemil Desai <[email protected]>
NVIDIA · Dec 1, 2024 · 2f68efe · 2f68efe
1 parent 8becd57
commit 2f68efe
Show file tree

Hide file tree

Showing 2 changed files with 3 additions and 3 deletions.
diff --git a/nemo/collections/llm/api.py b/nemo/collections/llm/api.py
@@ -599,7 +599,7 @@ def import_ckpt(
 
 
 def load_connector_from_trainer_ckpt(path: Path, target: str) -> io.ModelConnector:
-    return io.load_context(path).model.exporter(target, path)
+    return io.load_context(path, subpath="model").exporter(target, path)
 
 
 @run.cli.entrypoint(name="export", namespace="llm")

diff --git a/nemo/collections/llm/gpt/model/llama.py b/nemo/collections/llm/gpt/model/llama.py
@@ -366,11 +366,11 @@ def convert_state(self, source, target):
 
     @property
     def tokenizer(self):
-        return io.load_context(str(self)).model.tokenizer.tokenizer
+        return io.load_context(str(self), subpath="model").tokenizer.tokenizer
 
     @property
     def config(self) -> "HFLlamaConfig":
-        source: LlamaConfig = io.load_context(str(self)).model.config
+        source: LlamaConfig = io.load_context(str(self), subpath="model.config")
 
         from transformers import LlamaConfig as HFLlamaConfig