Nano v3 lora

3rdparty/Megatron-Bridge-workspace/setup.py

@@ -44,6 +44,8 @@
     "mamba-ssm",
     "nvidia-resiliency-ext",
     "causal-conv1d",
+    "timm",
+    "open-clip-torch>=3.2.0",
 ]
 
 # If the bridge source exists, compare cached dependencies with the submodule's pyproject

examples/configs/sft_nanov3.yaml

@@ -0,0 +1,216 @@
+# SFT Algorithm Configuration
+sft:
+  ## total number of steps to train will equal
+  ## min((max_num_epochs * len(train_dataloader)), max_num_steps)
+  max_num_epochs: 1
+  max_num_steps: 60
+
+  val_period: 10
+  val_batches: 8
+  val_global_batch_size: 32
+  val_micro_batch_size: 1
+  val_at_start: true
+  seed: 42
+
+checkpointing:
+  enabled: true
+  checkpoint_dir: "results/sft_nanov3_lora_permute_fusion_true"
+  metric_name: "val:val_loss" # one of "val:" or "train:" followed by the metric name
+  higher_is_better: false
+  keep_top_k: 3
+  save_period: 10
+  checkpoint_must_save_by: null
+
+policy:
+  model_name: "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16"
+  tokenizer:
+    name: ${policy.model_name} ## specify if you'd like to use a tokenizer different from the model's default
+    # chat_template can be a Jinja template string or path to a .jinja file
+    chat_template: "{% for message in messages %}{%- if message['role'] == 'system'  %}{{'Context: ' + message['content'].strip()}}{%- elif message['role'] == 'user'  %}{{' Question: ' + message['content'].strip() + ' Answer:'}}{%- elif message['role'] == 'assistant'  %}{{' ' + message['content'].strip()}}{%- endif %}{% endfor %}"
+    chat_template_kwargs: null # can be used to pass kwargs to the chat template, e.g., enable_thinking=true
+  train_global_batch_size: 32
+  train_micro_batch_size: 1
+  max_total_sequence_length: 1024
+  precision: "bfloat16"
+
+  offload_optimizer_for_logprob: false
+
+  dtensor_cfg:
+    enabled: false
+    env_vars: {}
+    cpu_offload: False
+    sequence_parallel: false
+    activation_checkpointing: false
+    tensor_parallel_size: 1
+    context_parallel_size: 1
+    custom_parallel_plan: null
+
+  dynamic_batching:
+    enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    sequence_length_round: 64
+
+  sequence_packing:
+    enabled: False
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
+
+  # makes the training sequence length divisible by the tensor parallel size
+  # this is useful for sequence parallel training
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+
+  optimizer:
+    name: "torch.optim.AdamW"
+    kwargs:
+      lr: 5.0e-6
+      weight_decay: 0.1
+      betas: [0.9, 0.98]
+      eps: 1e-5
+      # when using Dtensor, we need to set foreach
+      # and fused to False
+      foreach: False
+      fused: False
+
+  ## ignored since enabled=false, but needed for testing purposes
+  megatron_cfg:
+    enabled: true
+    env_vars: {}
+    empty_unused_memory_level: 1
+    activation_checkpointing: false
+    tensor_model_parallel_size: 8
+    expert_tensor_parallel_size: 1
+    expert_model_parallel_size: 8
+    pipeline_model_parallel_size: 1
+    context_parallel_size: 1
+    pipeline_dtype: ${policy.precision}
+    num_layers_in_first_pipeline_stage: null
+    num_layers_in_last_pipeline_stage: null
+    sequence_parallel: true
+    freeze_moe_router: false
+    moe_router_dtype: null
+    moe_router_load_balancing_type: "none"
+    moe_router_bias_update_rate: 1e-3
+    moe_permute_fusion: true # try this as true
+    #gives ~20% training perf speedup with sequence packing 
+    apply_rope_fusion: True
+    # gives ~25% training perf speedup with sequence packing and apply_rope_fusion
+    bias_activation_fusion: False
+    defer_fp32_logits: False
+    moe_per_layer_logging: False
+
+    peft:
+      enabled: false
+      target_modules: []
+      dim: 32
+      alpha: 32
+      dropout: 0.0
+      dropout_position: "pre"
+      lora_A_init_method: "xavier"
+      lora_B_init_method: "zero"
+      a2a_experimental: false
+      lora_dtype: None
+
+
+    optimizer:
+      optimizer: "adam"
+      lr: 5.0e-6
+      min_lr: 4.9999e-6
+      weight_decay: 0.1
+      bf16: false
+      fp16: false
+      params_dtype: "float32"
+
+      #adam
+      adam_beta1: 0.9
+      adam_beta2: 0.98
+      adam_eps: 1e-5
+
+      #sgd
+      sgd_momentum: 0.9
+
+      #distributed optimizer
+      use_distributed_optimizer: true
+      use_precision_aware_optimizer: true
+
+      clip_grad: ${policy.max_grad_norm}
+
+      # optimizer cpu offload
+      optimizer_cpu_offload: false
+      optimizer_offload_fraction: 0.0
+
+    scheduler:
+      start_weight_decay: ${policy.megatron_cfg.optimizer.weight_decay}
+      end_weight_decay: ${policy.megatron_cfg.optimizer.weight_decay}
+      weight_decay_incr_style: "constant"
+      lr_decay_style: "constant"
+      lr_decay_iters: 1000
+      lr_warmup_iters: 50
+      lr_warmup_init: 4.9999e-6
+
+    distributed_data_parallel_config:
+      grad_reduce_in_fp32: false
+      overlap_grad_reduce: false
+      overlap_param_gather: true
+      data_parallel_sharding_strategy: "optim_grads_params"
+      use_custom_fsdp: false
+
+data:
+  max_input_seq_length: ${policy.max_total_sequence_length}
+  add_bos: true
+  add_eos: true
+  add_generation_prompt: false
+  shuffle: false
+  num_workers: 1
+
+  dataset_name: "squad"
+  # You can use custom response datasets for training and validation. For example:
+  #   data:
+  #     dataset_name: ResponseDataset
+  #     train_data_path: <PathToTrainingDataset>  # e.g., /path/to/local/dataset.jsonl or hf_org/hf_dataset_name (HuggingFace)
+  #     val_data_path: <PathToValidationDataset>
+  #     input_key: <QuestionKey>, default is "input"
+  #     output_key: <AnswerKey>, default is "output"
+  #     train_split: <TrainSplit>, default is None  # used for HuggingFace datasets
+  #     val_split: <ValSplit>, default is None  # used for HuggingFace datasets
+  # See https://github.com/NVIDIA-NeMo/RL/blob/main/docs/guides/sft.md#datasets for more details.
+
+  ## unused with squad dataset
+  prompt_file: null
+  split: null
+  output_key: null
+  seed: null
+
+
+  ## OpenAI format specific configs
+  # train_data_path: "/path/to/train.jsonl"  # Path to training data
+  # val_data_path: "/path/to/val.jsonl"      # Path to validation data
+  # chat_key: "messages"                     # Key for messages in the data
+  # system_key: null                         # Key for system message (optional)
+  # system_prompt: null                      # Default system prompt (optional)
+  # tool_key: "tools"                        # Key for tools in the data
+  # use_preserving_dataset: false            # If true, uses PreservingDataset to preserve heterogeneous schemas (e.g., tool calls with varying argument structures)
+
+logger:
+  log_dir: "logs"  # Base directory for all logs
+  wandb_enabled: true # Make sure you do a ``wandb login [Your API key]'' before running
+  tensorboard_enabled: true
+  mlflow_enabled: false
+  swanlab_enabled: false # Disable SwanLab logging
+  monitor_gpus: true  # If true, will monitor GPU usage and log to wandb and/or tensorboard
+  wandb:
+    project: "sft-dev"
+    name: "sft-dev-${data.dataset_name}"
+  tensorboard:
+    log_dir: "tb_logs-sft-dev-${data.dataset_name}"
+  mlflow:
+    experiment_name: "sft-dev"
+    run_name: "sft-dev-${data.dataset_name}"
+  gpu_monitoring:
+    collection_interval: 10  # How often to collect GPU usage metrics (in seconds)
+    flush_interval: 10  # How often to flush GPU usage metrics to the loggers (in seconds)
+
+cluster:
+  gpus_per_node: 8
+  num_nodes: 1

nemo_rl/models/policy/workers/megatron_policy_worker.py

@@ -115,7 +115,7 @@ def _apply_transformer_engine_patch():
     except Exception as e:
         print(f"Error checking/patching transformer_engine: {e}")
 
-
+from megatron.bridge.peft.lora import LoRA 
 from megatron.bridge import AutoBridge
 from megatron.bridge.models.model_provider import get_model
 from megatron.bridge.training import fault_tolerance
@@ -143,6 +143,7 @@ def _apply_transformer_engine_patch():
 from megatron.bridge.training.optim import setup_optimizer
 from megatron.bridge.training.setup import (
     _update_model_config_funcs,
+    _create_peft_pre_wrap_hook
 )
 from megatron.bridge.training.state import GlobalState
 from megatron.bridge.training.tokenizers.tokenizer import build_tokenizer
@@ -178,6 +179,7 @@ def _apply_transformer_engine_patch():
 )
 from megatron.core.pipeline_parallel import get_forward_backward_func
 from megatron.core.rerun_state_machine import get_rerun_state_machine
+from megatron.core.transformer import MegatronModule
 from megatron.core.transformer.module import Float16Module
 from megatron.core.transformer.transformer_config import TransformerConfig
 from megatron.training.utils import get_ltor_masks_and_position_ids
@@ -366,6 +368,32 @@ def freeze_moe_router(megatron_model):
 
         mixed_precision_wrapper = CustomFloat16Module
         pre_wrap_hook.extend([freeze_moe_router])
+
+    if policy_cfg["megatron_cfg"].get("peft", {}).get("enabled", False):
+        lora_cfg = policy_cfg["megatron_cfg"].get("peft", {})
+        peft_cfg = LoRA(
+            target_modules=lora_cfg["target_modules"],
+            dim=lora_cfg["dim"],
+            alpha=lora_cfg["alpha"],
+            dropout=lora_cfg["dropout"],
+            dropout_position=lora_cfg["dropout_position"],
+            lora_A_init_method=lora_cfg["lora_A_init_method"],
+            lora_B_init_method=lora_cfg["lora_B_init_method"],
+            a2a_experimental=lora_cfg["a2a_experimental"],
+            lora_dtype=lora_cfg["lora_dtype"])
+    else:
+        peft_cfg = None
+    cfg.lora_cfg = peft_cfg
+
+    if cfg.lora_cfg is not None:
+        pre_peft_hook = _create_peft_pre_wrap_hook(cfg, state)
+        cfg.model.register_pre_wrap_hook(pre_peft_hook)
+        def composed_peft_hook(model: list[MegatronModule]) -> list[MegatronModule]:
+            model = pre_peft_hook(model)
+            return model
+        peft_hook = composed_peft_hook
+    else:
+        peft_hook = []
 
     # Model, optimizer, and learning rate.
     model = get_model(
@@ -374,7 +402,7 @@ def freeze_moe_router(megatron_model):
         use_torch_fsdp2=cfg.dist.use_torch_fsdp2,
         overlap_param_gather_with_optimizer_step=cfg.optimizer.overlap_param_gather_with_optimizer_step,
         data_parallel_random_init=cfg.rng.data_parallel_random_init,
-        pre_wrap_hook=pre_wrap_hook,
+        pre_wrap_hook=peft_hook, # @TODO @adithyare should integrate with pre-wrap-hook thats defined for freezing moe layers
         mixed_precision_wrapper=mixed_precision_wrapper,
     )
     if load_optimizer:
@@ -391,14 +419,16 @@ def freeze_moe_router(megatron_model):
     print("Model, optimizer, and learning rate scheduler built")
     torch.distributed.barrier()
 
-    # Load checkpoint if applicable
-    if (
-        cfg.checkpoint.load is not None
-        or cfg.checkpoint.pretrained_checkpoint is not None
-    ) and (
-        checkpoint_exists(cfg.checkpoint.load)
-        or checkpoint_exists(cfg.checkpoint.pretrained_checkpoint)
-    ):
+    if cfg.lora_cfg is not None:
+        should_load_checkpoint = (cfg.checkpoint.load is not None and checkpoint_exists(cfg.checkpoint.load))
+        if should_load_checkpoint:
+            # The finetune toggle is explicitly set to True in order to avoid loading optimizer and RNG states
+            # This is switched off here in order to load these states from the checkpoint
+            cfg.checkpoint.finetune = False
+    else:
+        should_load_checkpoint = (cfg.checkpoint.load is not None and checkpoint_exists(cfg.checkpoint.load)) or (cfg.checkpoint.pretrained_checkpoint is not None and checkpoint_exists(cfg.checkpoint.pretrained_checkpoint))
+
+    if should_load_checkpoint:
         load_checkpoint(
             state,
             model,

pyproject.toml

@@ -218,6 +218,7 @@ override-dependencies = [
   "transformer-engine[pytorch]==2.8.0",
   "opencv-python-headless>=4.11.0",
   "nvidia-modelopt[torch]>=0.39.0",
+  "timm<=1.0.22",
 ]
 # CVE fxies
 constraint-dependencies = [