modularization + better logic

pytorch · Dec 2, 2024 · 01bc4be · 01bc4be
1 parent fc6cfbe
commit 01bc4be
Show file tree

Hide file tree

Showing 4 changed files with 233 additions and 177 deletions.
diff --git a/recipes/configs/llama3_2/1B_full_single_device.yaml b/recipes/configs/llama3_2/1B_full_single_device.yaml
@@ -19,12 +19,16 @@
 #
 # This config works only for training on single device.
 
-output_dir: /tmp/felipe/llama_3_2_1b/full_single_device
+output_dir: /tmp/llama_3_2_1b/full_single_device
+
+# Model Arguments
+model:
+  _component_: torchtune.models.llama3_2.llama3_2_1b
 
 # Tokenizer
 tokenizer:
   _component_: torchtune.models.llama3.llama3_tokenizer
-  path:  ${output_dir}/original/tokenizer.model
+  path:  /tmp/Llama-3.2-1B-Instruct/original/tokenizer.model
   max_seq_len: null
 
 # Dataset
@@ -34,9 +38,6 @@ dataset:
 seed: null
 shuffle: True
 
-# Model Arguments
-model:
-  _component_: torchtune.models.llama3_2.llama3_2_1b
 
 checkpointer:
   _component_: torchtune.training.FullModelHFCheckpointer

diff --git a/recipes/configs/llama3_2/1B_lora_single_device.yaml b/recipes/configs/llama3_2/1B_lora_single_device.yaml
@@ -15,7 +15,7 @@
 #
 # This config works only for training on single device.
 
-output_dir: /tmp/felipe/llama_3_2_1b/lora_single_device
+output_dir: /tmp/llama_3_2_1b/lora_single_device
 
 # Model Arguments
 model:
@@ -29,14 +29,14 @@ model:
 # Tokenizer
 tokenizer:
   _component_: torchtune.models.llama3.llama3_tokenizer
-  path: ${output_dir}/original/tokenizer.model
+  path: /tmp/Llama-3.2-1B-Instruct/original/tokenizer.model
   max_seq_len: null
 
 checkpointer:
   _component_: torchtune.training.FullModelHFCheckpointer
   checkpoint_dir: /tmp/Llama-3.2-1B-Instruct/
   checkpoint_files: [
-    model.safetensors
+     epoch_2/ft-model-00001-of-00001.safetensors
   ]
   recipe_checkpoint: null
   output_dir: ${output_dir}