commit

llama_stack/providers/inline/post_training/torchtune/common/checkpointer.py

@@ -4,12 +4,14 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
+import json
 import os
 import shutil
 from pathlib import Path
 from typing import Any, Dict, List
 
 import torch
+from safetensors.torch import save_file
 from torchtune import training
 from torchtune.models import convert_weights
 from torchtune.training.checkpointing._utils import (
@@ -103,7 +105,7 @@ class TorchtuneCheckpointer:
             Path(self._output_dir)
             / f"{self._model_id}-{self._training_algorithm}-{epoch}"
         )
-        if format == "meta":
+        if checkpoint_format == "meta":
             model_file_path.mkdir(parents=True, exist_ok=True)
 
             # copy the related files for inference
@@ -176,79 +178,27 @@ class TorchtuneCheckpointer:
                 )
 
             print("model_file_path", str(model_file_path))
-        elif format == "hf":
+        elif checkpoint_format == "hf":
+            # Note: for saving hugging face format checkpoints, we only suppport saving adapter weights now
+
             # the config.json file contains model params needed for state dict conversion
             config = json.loads(
-                Path.joinpath(self._checkpoint_dir, "config.json").read_text()
+                Path.joinpath(self._checkpoint_dir.parent, "config.json").read_text()
             )
-            if not adapter_only:
-                state_dict[training.MODEL_KEY] = convert_weights.tune_to_hf(
-                    state_dict[training.MODEL_KEY],
-                    num_heads=config["num_attention_heads"],
-                    num_kv_heads=config["num_key_value_heads"],
-                    dim=config["hidden_size"],
-                    head_dim=config.get("head_dim", None),
-                )
 
-                # split the state_dict into separate dicts, one for each output checkpoint file
+            # repo_id is necessary for when saving an adapter config, so its compatible with HF.
-                # e.g. split_state_dicts= {
+            # This json file is produced and saved in the download step.
-                #       "0001": {"key1": tensor1, "key2": tensor2},
+            # contents are {"repo_id": "some_model/some_model_version"}
-                #       "0002": {"key3": tensor3}
+            repo_id_path = Path.joinpath(
-                #       }
+                self._checkpoint_dir.parent, REPO_ID_FNAME
-                split_state_dicts: Dict[str, Dict[str, torch.Tensor]] = {}
+            ).with_suffix(".json")
-                total_size = 0
+            self.repo_id = None
-                for key, weight in state_dict[training.MODEL_KEY].items():
+            if repo_id_path.exists():
-                    cpt_idx = self._weight_map[key]
+                with open(repo_id_path, "r") as json_file:
-
+                    data = json.load(json_file)
-                    # initialize dict
+                    self.repo_id = data.get("repo_id")
-                    if cpt_idx not in split_state_dicts:
-                        split_state_dicts[cpt_idx] = {}
-
-                    split_state_dicts[cpt_idx].update({key: weight})
-                    total_size += weight.numel() * weight.element_size()
-
-                # write the partitioned state dicts to the right checkpoint file
-                # e.g. model-00001-of-00004.safetensors, model-00002-of-00004.safetensors, etc
-                num_shards = len(split_state_dicts)
-                map_original_name_to_new_name = {}
-                for cpt_idx, model_state_dict in split_state_dicts.items():
-                    # TODO: We should probably use the original shard name and just add a prefix
-                    # however, having the SHARD_FNAME standardizes our checkpoints
-                    shard_name = SHARD_FNAME.format(
-                        cpt_idx=f"{cpt_idx}".zfill(5),
-                        num_shards=f"{num_shards}".zfill(5),
-                    )
-                    map_original_name_to_new_name[cpt_idx] = shard_name
-                    output_path = Path.joinpath(model_file_path, shard_name)
-                    output_path.parent.mkdir(parents=True, exist_ok=True)
-                    output_path = output_path.with_suffix(".safetensors")
-                    save_file(model_state_dict, output_path, metadata={"format": "pt"})
-
-                    logger.info(
-                        "Model checkpoint of size "
-                        f"{os.path.getsize(output_path) / 1024**3:.2f} GiB "
-                        f"saved to {output_path}"
-                    )
-
-                # Save the appropriate index file based on serialization format
-                # e.g. {metadata: {total_size: 1234}, weight_map: {"key1": "model_0001.safetensors", "key2": "model_0002.safetensors"}}
-                weight_map = {
-                    k: map_original_name_to_new_name[cpt_idx] + ".safetensors"
-                    for k, cpt_idx in self._weight_map.items()
-                }
-                index_file_name = SAFETENSOR_INDEX_FNAME
-
-                index_path = Path.joinpath(model_file_path, index_file_name)
-
-                index_data = {
-                    "metadata": {"total_size": total_size},
-                    "weight_map": weight_map,
-                }
-                with open(index_path, "w") as f:
-                    json.dump(index_data, f, indent=2)
 
             if training.ADAPTER_KEY in state_dict:
-
                 # TODO: saving it "as is" is a requirement because, if we only save with
                 # convert_weights.tune_to_peft_adapter_weights, we do NOT have a fn
                 # convert_weights.peft_to_tune. The .pt format is not needed, but
@@ -273,7 +223,9 @@ class TorchtuneCheckpointer:
                         head_dim=config.get("head_dim", None),
                     )
                 )
-                output_path = Path.joinpath(model_file_path, ADAPTER_MODEL_FNAME)
+                output_path = Path.joinpath(
+                    model_file_path, "adapter", ADAPTER_MODEL_FNAME
+                )
                 output_path.parent.mkdir(parents=True, exist_ok=True)
                 output_path = output_path.with_suffix(".safetensors")
                 save_file(
@@ -300,7 +252,7 @@ class TorchtuneCheckpointer:
                 )
 
                 output_path = Path.joinpath(
-                    model_file_path, ADAPTER_CONFIG_FNAME
+                    model_file_path, "adapter", ADAPTER_CONFIG_FNAME
                 ).with_suffix(".json")
                 with open(output_path, "w") as f:
                     json.dump(state_dict[training.ADAPTER_CONFIG], f)
@@ -313,7 +265,7 @@ class TorchtuneCheckpointer:
             # Save all files in ckpt_dir, except model weights and mapping, to output_dir/epoch_{epoch}
             # So its easy to run inference with the model using this epoch's checkpoint
             copy_files(
-                self._checkpoint_dir,
+                self._checkpoint_dir.parent,
                 model_file_path,
                 ignore_suffixes=SUFFIXES_TO_NOT_COPY,
             )

llama_stack/providers/inline/post_training/torchtune/config.py

@@ -11,4 +11,4 @@ from pydantic import BaseModel
 
 class TorchtunePostTrainingConfig(BaseModel):
     torch_seed: Optional[int] = None
-    checkpoint_format: Optional[str] = "hf"
+    checkpoint_format: Optional[str] = "meta"