Implement initial_load_path for checkpointer

docs/checkpoint.md

@@ -25,19 +25,19 @@ interval = 500
 
 
 2. SAVE ONLY MODEL WEIGHTS
-By setting `model_weights_only` to `True`, the checkpoint will only contain the model weights and exclude the optimizer state and extra train states, resulting in a smaller checkpoint size.
+By setting `last_save_model_weights_only` to `True`, the checkpoint will only contain the model weights and exclude the optimizer state and extra train states, resulting in a smaller checkpoint size.
 ```
 [checkpoint]
 enable_checkpoint = true
-model_weights_only = true
+last_save_model_weights_only = true
 ```
 
 3. CHOOSE DESIRED EXPORT PRECISION
 The default model states are in `float32`. You can choose to export the checkpoint in a lower precision format such as `bfloat16`.
 ```
 [checkpoint]
 enable_checkpoint = true
-model_weights_only = true
+last_save_model_weights_only = true
 export_dtype = "bfloat16"
 ```
 
@@ -48,7 +48,7 @@ enable_checkpoint = true
 folder = "checkpoint"
 interval = 10
 load_step = 5
-model_weights_only = true
+last_save_model_weights_only = true
 export_dtype = "bfloat16"
 ```
 

tests/integration_tests.py

@@ -122,22 +122,22 @@ def build_test_list():
             [
                 [
                     "--checkpoint.enable_checkpoint",
-                    "--checkpoint.model_weights_only",
+                    "--checkpoint.last_save_model_weights_only",
                 ],
             ],
             "Checkpoint Integration Test - Save Model Weights Only fp32",
-            "model_weights_only_fp32",
+            "last_save_model_weights_only_fp32",
         ),
         OverrideDefinitions(
             [
                 [
                     "--checkpoint.enable_checkpoint",
-                    "--checkpoint.model_weights_only",
+                    "--checkpoint.last_save_model_weights_only",
                     "--checkpoint.export_dtype bfloat16",
                 ],
             ],
             "Checkpoint Integration Test - Save Model Weights Only bf16",
-            "model_weights_only_bf16",
+            "last_save_model_weights_only_bf16",
         ),
         OverrideDefinitions(
             [

tests/unit_tests/test_checkpoint.py

@@ -44,16 +44,22 @@ def fake_get_model_state_dict(model, *args, **kwargs):
     return model.state_dict()
 
 
+# TODO: The unittest is not well structured and does not cover enough paths.
+# It should be refactored.
+
+
 @dataclass
 class DummyCheckpointConfig:
     enable_checkpoint: bool = True
     folder: str = "dummy_folder"
     interval: int = 10
     async_mode: str = "disabled"
     keep_latest_k: int = 0
-    model_weights_only: bool = False
+    last_save_model_weights_only: bool = False
     export_dtype: str = "float32"
     exclude_from_loading = []
+    initial_load_model_weights_only: bool = False
+    initial_load_path: str = ""
 
 
 @dataclass

torchtitan/components/checkpoint.py

@@ -81,6 +81,12 @@ class SaveDone:
     pass
 
 
+# For now, we will manually pop the freqs_cis buffer, as we made this permanent
+# temporarily and we don't want to include it in the exported state_dict.
+# Context: https://github.com/pytorch/torchtitan/blob/main/torchtitan/models/llama3/model.py#L404
+excluded_parameters_for_model_only = {"freqs_cis"}
+
+
 @torch.no_grad()
 def save_with_gc(state, checkpoint_id):
     dcp.save(state, checkpoint_id=checkpoint_id)
@@ -267,6 +273,10 @@ def load_state_dict(state_dict):
         self.staging_stream = torch.cuda.Stream() if self.enable_staging else None
 
         self.folder = os.path.join(job_config.job.dump_folder, ckpt_config.folder)
+        self.initial_load_path = ckpt_config.initial_load_path
+        self.initial_load_model_weights_only = (
+            ckpt_config.initial_load_model_weights_only
+        )
         self.interval = ckpt_config.interval
         async_mode = ckpt_config.async_mode.lower()
         if async_mode == AsyncMode.ASYNC or self.ft_manager:
@@ -287,7 +297,7 @@ def load_state_dict(state_dict):
         else:
             self.purge_thread = None
 
-        self.model_weights_only = ckpt_config.model_weights_only
+        self.last_save_model_weights_only = ckpt_config.last_save_model_weights_only
         self.export_dtype = TORCH_DTYPE_MAP[ckpt_config.export_dtype]
         self.exclude_from_loading = ckpt_config.exclude_from_loading
 
@@ -408,21 +418,38 @@ def load(self, step: int = -1) -> bool:
         if self.ft_manager:
             self._ft_load()
 
-        if not self.enable_checkpoint or not os.path.isdir(self.folder):
+        if not self.enable_checkpoint:
             return False
 
-        if step == -1:
-            step = self._find_load_step()
+        model_only = False
+        if not os.path.exists(self.folder):
+            if self.initial_load_path:
+                checkpoint_id = self.initial_load_path
+                if not os.path.isdir(checkpoint_id):
+                    raise ValueError(
+                        "initial_load_full_checkpoint is specified but the path is not valid."
+                    )
+                model_only = self.initial_load_model_weights_only
+            else:
+                return False
+        else:
+            if self.initial_load_path:
+                logger.info(
+                    "`initial_load_path` is provided but the checkpoint folder exists. "
+                    "Checkpointer will use the checkpoints from the checkpoint folder."
+                )
+            step = self._find_load_step() if step == -1 else step
             if step == -1:
                 return False
+            model_only = step == 0
+            checkpoint_id = self._create_checkpoint_id(step)
 
-        checkpoint_id = self._create_checkpoint_id(step)
-        if not os.path.isdir(checkpoint_id):
-            return False
+            if not os.path.isdir(checkpoint_id):
+                return False
 
-        logger.info(f"Loading the checkpoint at step {step}.")
+        logger.info(f"Loading the checkpoint from {checkpoint_id}.")
         begin = time.monotonic()
-        states = self._states_to_load(step)
+        states = self._states_to_load(model_only)
         dcp.load(states, checkpoint_id=checkpoint_id)
         GarbageCollection.collect("GC collection for checkpoint loading.")
         logger.info(
@@ -521,28 +548,36 @@ def _ft_load(self) -> None:
             f"Finished loading the ft checkpoint in {time.monotonic() - begin:.2f} seconds."
         )
 
-    def _states_to_load(self, step: int) -> dict[str, Any]:
+    def _states_to_load(self, model_only: bool) -> dict[str, Any]:
         """Determines which states to load for the given step.
 
-        When checkpointer determines which step of the checkpoint to load, this API is
-        used to determine which states to load based on the step.
+        This API is used to determine which states to load based on the
+        configurations.
 
         Args:
-            step (int): The step to load the checkpoint for.
+            model_only (bool): Whether to load the model only.
 
         Returns:
             Dict[str, Any]: The states to load for the given step.
         """
         # For the first step, we will only load the model weights.
-        states = {MODEL: self.states[MODEL]} if step == 0 else self.states
-        states_to_load = {
-            k: v for k, v in states.items() if k not in self.exclude_from_loading
-        }
+        if model_only:
+            sd = self.states[MODEL].state_dict()
+            for k in excluded_parameters_for_model_only:
+                sd.pop(k, None)
+            return sd
+
         for exclude_key in self.exclude_from_loading:
-            if exclude_key not in states:
+            if exclude_key not in self.states:
                 raise ValueError(f"{exclude_key} not found in state_dict.")
+
+        states_to_load = {
+            k: v for k, v in self.states.items() if k not in self.exclude_from_loading
+        }
+
         if self.ft_manager:
             states_to_load.pop(DATALOADER)
+
         return states_to_load
 
     def _save_last_step(self, curr_step: int) -> None:
@@ -551,18 +586,16 @@ def _save_last_step(self, curr_step: int) -> None:
         # dtype conversion when we are checkpoint model weights only and the
         # current dtype is not the same as the export dtype at the end of the training.
 
-        if self.model_weights_only:
+        if self.last_save_model_weights_only:
             # We update self.states to keep the model only.
             # After this update, self.states = {
             #      'tok_embeddings.weight':...,
             #      'layers.0.attention.wq.weight': ...
             # }.
             self.states = self.states[MODEL].state_dict()
 
-            # For now, we will manually pop the freqs_cis buffer, as we made this permanent
-            # temporarily and we don't want to include it in the exported state_dict.
-            # Context: https://github.com/pytorch/torchtitan/blob/main/torchtitan/models/llama3/model.py#L404
-            self.states.pop("freqs_cis", None)
+            for k in excluded_parameters_for_model_only:
+                self.states.pop(k, None)
 
             if self.export_dtype != torch.float32:
                 self.states = {

torchtitan/config_manager.py

@@ -373,21 +373,47 @@ class Checkpoint:
     When enable_checkpoint is set to true, checkpoints will be in {--job.dump_folder}/{--checkpoint.folder}.
     """
 
+    initial_load_path: str | None = None
+    """
+    This option specifies the path to the initial checkpoint to load, which is
+    particularly useful for resuming training from a previous run with a
+    different output path or when loading a checkpoint from a pre-trained model.
+    If the checkpoint folder for the current run is not empty,
+    located at {--job.dump_folder}/{--checkpoint.folder}, this option will be ignored.
+    This feature allows users to load an initial checkpoint from a different folder and
+    continue training, saving new checkpoints to the specified folder without affecting
+    the existing ones.
+
+    Note that the path should contain the full path to the checkpoint folder,
+    including the step number, if any; for example,
+    "//pre_train/checkpoints/llama3/llama3_8b/step_10000".
+    """
+
+    initial_load_model_weights_only: bool = True
+    """
+    This option specifies if only the model weights should be loaded during the initial
+    checkpoint load. The option is only used when `initial_load_path` is specified.
+    If False, the checkpoint at `initial_load_path` is treated as a standard training
+    checkpoint, including optimizer and training states.
+    The default setting for this option is True. Note that you will have to use
+    `--checkpoint.no_initial_load_model_weights_only` to override the default setting.
+    """
+
     interval: int = 500
     """Checkpointing interval in steps."""
 
-    model_weights_only: bool = False
+    last_save_model_weights_only: bool = False
     """
-    When model_weights_only=True, only model weights will be saved at the end of training.
-    With this, checkpoints can be loaded using `torch.load(..., weights_only=True)` after conversion.
-    When model_weights_only=False, the full checkpoint will be saved.
+    When last_save_model_weights_only=True, only model weights will be saved at the end of training,
+    the last save.  With this, checkpoints can be loaded using `torch.load(..., weights_only=True)`
+    after conversion.  When last_save_model_weights_only=False, the full checkpoint will be saved.
     A full checkpoint includes model, optimizer and train_state, which can be used to resume training.
     The default value is false.
     """
 
     export_dtype: Literal["float16", "bfloat16", "float32"] = "float32"
     """
-    Converts to the specified precision when training completes and model_weights_only=true.
+    Converts to the specified precision when training completes and last_save_model_weights_only=true.
     """
 
     create_seed_checkpoint: bool = False

torchtitan/experiments/flux/tests/integration_tests.py

@@ -58,11 +58,11 @@ def build_test_list():
             [
                 [
                     "--checkpoint.enable_checkpoint",
-                    "--checkpoint.model_weights_only",
+                    "--checkpoint.last_save_model_weights_only",
                 ],
             ],
             "Checkpoint Integration Test - Save Model Weights Only fp32",
-            "model_weights_only_fp32",
+            "last_save_model_weights_only_fp32",
         ),
         # Parallelism tests.
         OverrideDefinitions(

torchtitan/experiments/flux/train_configs/debug_model.toml

@@ -68,6 +68,6 @@ mode = "full"
 enable_checkpoint = false
 folder = "checkpoint"
 interval = 5
-model_weights_only = false
+last_save_model_weights_only = false
 export_dtype = "float32"
 async_mode = "disabled"  # ["disabled", "async", "async_with_pinned_mem"]

torchtitan/experiments/flux/train_configs/flux_dev_model.toml

@@ -67,6 +67,6 @@ mode = "full"
 enable_checkpoint = false
 folder = "checkpoint"
 interval = 1_000
-model_weights_only = false
+last_save_model_weights_only = false
 export_dtype = "float32"
 async_mode = "disabled"  # ["disabled", "async", "async_with_pinned_mem"]

torchtitan/experiments/flux/train_configs/flux_schnell_model.toml

@@ -67,6 +67,6 @@ mode = "full"
 enable_checkpoint = false
 folder = "checkpoint"
 interval = 1_000
-model_weights_only = false
+last_save_model_weights_only = false
 export_dtype = "float32"
 async_mode = "disabled"  # ["disabled", "async", "async_with_pinned_mem"]

torchtitan/experiments/llama4/scripts/convert_hf_to_dcp_with_gpus.py

@@ -322,7 +322,6 @@ def _reshard_send(
     def _reshard_receive(
         self, assignment: _Assignment, state_dict: dict[str, torch.Tensor]
     ) -> dict[str, torch.Tensor]:
-
         flatten_tensor = torch.empty(
             sum(math.prod(s) for s, d in zip(assignment.shapes, assignment.dtypes)),
             dtype=assignment.dtypes[0],
@@ -535,7 +534,7 @@ def state_dict(self) -> dict[str, torch.Tensor]:
             # oh, this is pretty bad, when can we get rid of the freqs_cis issue?
             state_dict["freqs_cis"] = None
             trainer.checkpointer.states[MODEL] = DummyModel(state_dict)
-            trainer.checkpointer.model_weights_only = True
+            trainer.checkpointer.last_save_model_weights_only = True
             trainer.checkpointer.export_dtype = next(iter(state_dict.values())).dtype
             trainer.checkpointer.save(curr_step=0, force=True)
             time.sleep(2)

torchtitan/experiments/llama4/scripts/convert_meta_to_dcp_with_gpus.py

@@ -529,7 +529,7 @@ def state_dict(self) -> dict[str, torch.Tensor]:
             # oh, this is pretty bad, when can we get rid of the freqs_cis issue?
             state_dict["freqs_cis"] = None
             trainer.checkpointer.states[MODEL] = DummyModel(state_dict)
-            trainer.checkpointer.model_weights_only = True
+            trainer.checkpointer.last_save_model_weights_only = True
             trainer.checkpointer.export_dtype = next(iter(state_dict.values())).dtype
             trainer.checkpointer.save(curr_step=0, force=True)
             time.sleep(2)

torchtitan/experiments/llama4/train_configs/debug_model.toml

@@ -57,7 +57,7 @@ context_parallel_degree = 1
 enable_checkpoint = false
 folder = "checkpoint"
 interval = 10
-model_weights_only = false
+last_save_model_weights_only = false
 export_dtype = "float32"
 async_mode = "disabled"  # ["disabled", "async", "async_with_pinned_mem"]
 

torchtitan/experiments/llama4/train_configs/llama4_17bx128e.toml

@@ -51,7 +51,7 @@ context_parallel_degree = 1
 enable_checkpoint = false
 folder = "checkpoint"
 interval = 500
-model_weights_only = false
+last_save_model_weights_only = false
 export_dtype = "float32"
 async_mode = "disabled" # ["disabled", "async", "async_with_pinned_mem"]
 

torchtitan/experiments/llama4/train_configs/llama4_17bx16e.toml

@@ -49,7 +49,7 @@ context_parallel_degree = 1
 enable_checkpoint = false
 folder = "checkpoint"
 interval = 500
-model_weights_only = false
+last_save_model_weights_only = false
 export_dtype = "float32"
 async_mode = "disabled" # ["disabled", "async", "async_with_pinned_mem"]
 

torchtitan/models/llama3/train_configs/debug_model.toml

@@ -59,7 +59,7 @@ context_parallel_degree = 1
 enable_checkpoint = false
 folder = "checkpoint"
 interval = 10
-model_weights_only = false
+last_save_model_weights_only = false
 export_dtype = "float32"
 async_mode = "disabled"  # ["disabled", "async", "async_with_pinned_mem"]
 

torchtitan/models/llama3/train_configs/llama3_405b.toml

@@ -49,7 +49,7 @@ context_parallel_degree = 1
 enable_checkpoint = false
 folder = "checkpoint"
 interval = 500
-model_weights_only = false
+last_save_model_weights_only = false
 export_dtype = "float32"
 async_mode = "disabled" # ["disabled", "async", "async_with_pinned_mem"]