Add torchdata Parallel Packer for faster startup

recipes/configs/qwen2_5/32B_full_single_device_mi300.yaml

@@ -0,0 +1,123 @@
+# Config for single device full finetuning in full_finetune_single_device.py
+# using a Qwen2.5 7B
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download Qwen/Qwen2.5-7B-Instruct --output-dir /tmp/Qwen2_5-7B-Instruct
+#
+# The default config uses an optimizer from bitsandbytes. If you do not have it installed,
+# you can install it with
+#   pip install bitsandbytes
+#
+# To launch on a single device, run the following command from root:
+#   tune run full_finetune_single_device --config qwen2_5/7B_full_single_device
+#
+# You can add specific overrides through the command line. For example
+# to override the checkpointer directory while launching training
+# you can run:
+#   tune run full_finetune_single_device --config qwen2_5/7B_full_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works only for training on single device.
+
+output_dir: /tmp/torchtune/qwen2_5_7B/full_single_device # /tmp may be deleted by your system. Change it to your preference.
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.qwen2_5.qwen2_5_tokenizer
+  path: /tmp/Qwen2_5-7B-Instruct/vocab.json
+  merges_file: /tmp/Qwen2_5-7B-Instruct/merges.txt
+  max_seq_len: 2048
+
+# Dataset
+# dataset:
+#   _component_: torchtune.datasets.alpaca_cleaned_dataset
+#   packed: False  # True increases speed
+# seed: null
+# shuffle: True
+dataset:
+  _component_: torchtune.datasets.chat_dataset
+  source: json
+  data_files: /home/andrewkh/local/test.jsonl
+  split: train
+  conversation_column: conversations
+  conversation_style: sharegpt
+  packed: True  # True increases speed
+  packer_num_workers: 12
+
+seed: null
+shuffle: True
+
+# Model Arguments
+model:
+  _component_: torchtune.models.qwen2_5.qwen2_5_7b_instruct
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Qwen2_5-7B-Instruct
+  checkpoint_files: [
+    model-00001-of-00004.safetensors,
+    model-00002-of-00004.safetensors,
+    model-00003-of-00004.safetensors,
+    model-00004-of-00004.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: QWEN2
+resume_from_checkpoint: False
+
+# Fine-tuning arguments
+batch_size: 2
+epochs: 1
+optimizer:
+  _component_: bitsandbytes.optim.PagedAdamW
+  lr: 5e-6
+optimizer_in_bwd: True  # True saves memory. Requires gradient_accumulation_steps=1
+loss:
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
+max_steps_per_epoch: null
+gradient_accumulation_steps: 1  # Use to increase effective batch size
+compile: False  # torch.compile the model + loss, True increases speed + decreases memory
+
+# Training environment
+device: cuda
+
+# Memory management
+enable_activation_checkpointing: True  # True reduces memory
+enable_activation_offloading: False  # True reduces memory
+
+# Reduced precision
+dtype: bf16
+
+# Logging
+metric_logger:
+  # _component_: torchtune.training.metric_logging.DiskLogger
+  # log_dir: ${output_dir}/logs
+  _component_: torchtune.training.metric_logging.StdoutLogger
+log_every_n_steps: 1
+log_peak_memory_stats: False
+
+
+# Profiler (disabled)
+profiler:
+  _component_: torchtune.training.setup_torch_profiler
+  enabled: False
+
+  #Output directory of trace artifacts
+  output_dir: ${output_dir}/profiling_outputs
+
+  #`torch.profiler.ProfilerActivity` types to trace
+  cpu: True
+  cuda: True
+
+  #trace options passed to `torch.profiler.profile`
+  profile_memory: False
+  with_stack: False
+  record_shapes: True
+  with_flops: False
+
+  # `torch.profiler.schedule` options:
+  # wait_steps -> wait, warmup_steps -> warmup, active_steps -> active, num_cycles -> repeat
+  wait_steps: 5
+  warmup_steps: 3
+  active_steps: 2
+  num_cycles: 1

torchtune/datasets/_chat.py

@@ -23,6 +23,7 @@ def chat_dataset(
     packed: bool = False,
     filter_fn: Optional[Callable] = None,
     split: str = "train",
+    packer_num_workers: Optional[int] = None,
     **load_dataset_kwargs: Dict[str, Any],
 ) -> Union[SFTDataset, PackedDataset]:
     """
@@ -183,5 +184,9 @@ def chat_dataset(
             raise ValueError(
                 "PackedDataset requires a max_seq_len to be set on the tokenizer."
             )
-        return PackedDataset(ds, max_seq_len=tokenizer.max_seq_len)
+        return PackedDataset(
+            ds,
+            max_seq_len=tokenizer.max_seq_len,
+            num_workers=packer_num_workers,
+        )
     return ds