R1-Style distributed GRPO (#2326)

Co-authored-by: Felipe Mello <fmellomascarenhas@gmail.com>
Co-authored-by: ebsmothers <ebs@meta.com>
Co-authored-by: salman <salman.mohammadi@outlook.com>

recipes/configs/dev/3B_full_grpo.yaml

@@ -0,0 +1,140 @@
+# Config for multi-node GRPO in dev/grpo_full_finetune_distributed.py
+# using a Llama3.2 3B Base model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download meta-llama/Llama-3.2-3B --output-dir /tmp/Llama-3.2-3B --ignore-patterns "original/consolidated.00.pth"
+#
+# It can be beneficial to first train the base model with SFT using the 3B_sft recipe.
+#
+# To launch on 4 devices, run the following command from root:
+# tune run --nproc_per_node 4 dev/grpo_full_finetune_distributed --config dev/3B_full_grpo
+#
+# You can add specific overrides through the command line. For example
+# to override the checkpointer directory while launching training
+# you can run:
+#   tune run --nproc_per_node 4 dev/grpo_full_finetune_distributed --config dev/grpo/3B_full_rl checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works best when the model is being fine-tuned on 2+ GPUs.
+#
+# Furthermore, you can launch it on multiple nodes by going to recipes/dev/ and using
+#  sbatch multinode_grpo.sbatch
+
+name: grpo_llama3b
+
+output_dir: /tmp/checkpoints/${name}
+base_model_path: /tmp/llama3B_gsm8k_sft_part0/epoch_0  # Use this to train from the slightly trained SFT model
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.llama3.llama3_tokenizer
+  path: /tmp/Llama-3.2-3B/original/tokenizer.model
+  max_seq_len: null
+
+# Dataset
+dataset:
+  _component_: torchtune.dev.grpo.gsm8k.gsm8k_dataset
+  partition: 1-9/10
+seed: null
+shuffle: False
+
+# Model Arguments
+model:
+  _component_: torchtune.models.llama3_2.llama3_2_3b
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: ${base_model_path}
+  checkpoint_files: [
+    model-00001-of-00002.safetensors,
+    model-00002-of-00002.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: LLAMA3
+
+
+ref_checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: ${base_model_path}
+  checkpoint_files: [
+    model-00001-of-00002.safetensors,
+    model-00002-of-00002.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: ${output_dir}/ref  # shouldn't be used?
+  model_type: LLAMA3
+
+
+resume_from_checkpoint: False
+save_every_n_epochs: 1
+
+# Fine-tuning arguments
+batch_size: 1
+grpo_samples: 16
+forward_batch_size: 1
+max_generated_tokens: 512
+top_k: null
+temperature: 1.0
+
+ppo_epochs: 1
+
+num_steps: 200
+
+clip_grad_norm: 1.0
+
+epochs: 10
+optimizer:
+  _component_: torch.optim.AdamW
+  lr: 1e-5
+  fused: True
+lr_scheduler:
+  _component_: torchtune.training.lr_schedulers.get_cosine_schedule_with_warmup
+  num_warmup_steps: 50
+loss:
+  _component_: torchtune.dev.grpo.loss.GRPOSimpleLoss
+  kl_coeff: 0.01
+  epsilon: 0.2
+
+# Training env
+device: cuda
+
+# Memory management
+enable_activation_checkpointing: True  # True reduces memory
+compile: False  # pytorch compile, set to true for better perf/memory
+
+# Reduced precision
+dtype: bf16
+
+
+# Logging
+metric_logger:
+  _component_: torchtune.training.metric_logging.DiskLogger
+  log_dir: ${output_dir}/logs
+log_every_n_steps: 1
+log_peak_memory_stats: True
+
+# Profiler (disabled)
+profiler:
+  _component_: torchtune.training.setup_torch_profiler
+  enabled: True
+
+  #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: True
+  with_stack: True
+  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

recipes/configs/dev/3B_sft_for_grpo.yaml

@@ -0,0 +1,109 @@
+# Config for multi-device SFT for reasoning in full_finetune_distributed.py
+# using a Llama3.2 3B Base model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download meta-llama/Meta-Llama-3.2-3B --output-dir /tmp/Meta-Llama-3.2-3B-Instruct --ignore-patterns "original/consolidated.00.pth"
+#
+# To launch on 4 devices, run the following command from root:
+#   tune run --nproc_per_node 4 full_finetune_distributed --config dev/3B_grpo_sft
+#
+# You can add specific overrides through the command line. For example
+# to override the checkpointer directory while launching training
+# you can run:
+#   tune run --nproc_per_node 4 full_finetune_distributed --config dev/grpo/3B_sft checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works best when the model is being fine-tuned on 2+ GPUs.
+
+
+name: llama3B_gsm8k_sft_part0
+
+output_dir: /tmp/${name}
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.llama3.llama3_tokenizer
+  path: /tmp/Llama-3.2-3B/original/tokenizer.model
+  max_seq_len: null
+
+# Dataset
+dataset:
+  _component_: torchtune.dev.grpo.gsm8k.gsm8k_sft
+  partition: 0-0/10
+seed: null
+shuffle: True
+
+# Model Arguments
+model:
+  _component_: torchtune.models.llama3_2.llama3_2_3b
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Llama-3.2-3B/
+  checkpoint_files: [
+    model-00001-of-00002.safetensors,
+    model-00002-of-00002.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: LLAMA3
+resume_from_checkpoint: False
+
+# Fine-tuning arguments
+batch_size: 2
+epochs: 1
+
+optimizer:
+  _component_: torch.optim.AdamW
+  lr: 1e-5
+  fused: True
+loss:
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
+max_steps_per_epoch: null
+clip_grad_norm: null
+compile: False  # torch.compile the model + loss, True increases speed + decreases memory
+optimizer_in_bwd: False  # True saves memory. Requires gradient_accumulation_steps=1
+gradient_accumulation_steps: 1  # Use to increase effective batch size
+
+# Training env
+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
+log_every_n_steps: 1
+log_peak_memory_stats: True
+
+
+# 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