Full DPO Distributed

docs/source/recipes/dpo.rst

@@ -13,7 +13,7 @@ To see the best results when using this recipe, it may be helpful to first fine-
 on-distribution for the domain you're interested in. To do this, check out our other fine-tuning recipes in the :ref:`recipe overview <recipes_overview_label>` which
 support a variety of SFT paradigms.
 
-After supervised fine-tuning, here is an example of DPO with Llama 3.1 8B:
+After supervised fine-tuning, here is an example of using either LoRA-based finetuning, or full-finetuning Llama 3.1 8B with DPO:
 
 .. note::
 
@@ -27,12 +27,15 @@ After supervised fine-tuning, here is an example of DPO with Llama 3.1 8B:
     --ignore-patterns "original/consolidated.00.pth"
     --HF_TOKEN <HF_TOKEN>
 
-    # run on a single device
+    # run lora dpo on a single device
     tune run lora_dpo_single_device --config llama3_1/8B_lora_dpo_single_device
 
-    # run on two gpus
+    # run lora dpo on two gpus
     tune run --nproc_per_node 2 lora_dpo_distributed --config llama3_1/8B_lora_dpo
 
+    # run full dpo on four gpus
+    tune run --nproc_per_node 4 full_dpo_distributed --config llama3_1/8B_full_dpo
+
 It's easy to get started with this recipe with your dataset of choice, including custom local datasets,
 and datasets from Hugging Face. Check out our primer on :ref:`preference datasets <preference_dataset_usage_label>` to
 see how to do this.

recipes/configs/llama3_1/8B_full_dpo.yaml

@@ -0,0 +1,98 @@
+# Config for multi-device full DPO alignment in full_dpo_distributed.py
+# using a Llama3.1 8B model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download meta-llama/Meta-Llama-3.1-8B-Instruct --output-dir /tmp/Meta-Llama-3.1-8B-Instruct --ignore-patterns "original/consolidated.00.pth"
+#
+# To launch on 4 devices, run the following command from root:
+#   tune run --nnodes 1 --nproc_per_node 4 full_dpo_distributed --config llama3_1/8B_full_dpo
+#
+# You can add specific overrides through the command line. For example
+# to override the checkpointer directory while launching training
+# you can run:
+#   tune run --nnodes 1 --nproc_per_node 4 full_dpo_distributed --config llama3_1/8B_full_dpo checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+
+output_dir: /tmp/torchtune/llama3_1_8B/full_dpo # /tmp may be deleted by your system. Change it to your preference.
+
+# Model Arguments
+model:
+  _component_: torchtune.models.llama3_1.llama3_1_8b
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.llama3.llama3_tokenizer
+  path: /tmp/Meta-Llama-3.1-8B-Instruct/original/tokenizer.model
+  max_seq_len: 1024 # higher increases memory
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Meta-Llama-3.1-8B-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: LLAMA3
+resume_from_checkpoint: False
+
+# The ref_checkpointer should always point to the original weights.
+ref_checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Meta-Llama-3.1-8B-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: LLAMA3
+
+# Dataset and Sampler
+dataset:
+  _component_: torchtune.datasets.stack_exchange_paired_dataset
+seed: null
+shuffle: True
+batch_size: 4
+
+# Optimizer and Scheduler
+optimizer:
+  _component_: torch.optim.AdamW
+  fused: True
+  weight_decay: 0.05
+  lr: 2e-5
+lr_scheduler:
+  _component_: torchtune.training.lr_schedulers.get_cosine_schedule_with_warmup
+  num_warmup_steps: 20
+
+loss:
+  _component_: torchtune.rlhf.loss.DPOLoss
+  beta: 0.05
+  label_smoothing: 0
+
+# Training
+epochs: 1
+max_steps_per_epoch: 1000
+gradient_accumulation_steps: 8  # Use to increase effective batch size
+compile: False  # torch.compile the model + loss, True increases speed + decreases memory
+
+# Logging
+metric_logger:
+  _component_: torchtune.training.metric_logging.DiskLogger
+  log_dir: ${output_dir}/logs
+log_every_n_steps: 1
+log_peak_memory_stats: True
+
+# Environment
+device: cuda
+dtype: bf16
+
+# Memory management
+enable_activation_checkpointing: True  # True reduces memory
+enable_activation_offloading: False  # True reduces memory