Lora and Dora finetuning produces identical results

[AndrewMead10]

I was trying to compare lora, dora, and full finetuning on llama 1B, but i found that lora and dora finetuning produced identical results. I am using the orca 10k dataset, like they did in the answer.ai post comparing the 2 methods.

here is the wandb report for the runs

here is my config file, the only thing that i changed between runs was the use_dora field from true to false. The command i ran was tune run lora_finetune_single_device --config benchmark_methods/llama_3_2_1b_lora_adam.yaml

I am using a 3090 gpu.

# Config for single device LoRA finetuning in lora_finetune_single_device.py
# using a Llama3.2 1B Instruct model
#
# This config assumes that you've run the following command before launching
# this run:
#   tune download meta-llama/Llama-3.2-1B-Instruct --output-dir ./tmp/Llama-3.2-1B-Instruct --ignore-patterns "original/consolidated.00.pth"
#
# To launch on a single device, run the following command from root:
#   tune run lora_finetune_single_device --config llama3_2/1B_lora_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 lora_finetune_single_device --config llama3_2/1B_lora_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
#
# This config works only for training on single device.

output_dir: ./tmp/torchtune/llama3_2_1B/lora_single_device # /tmp may be deleted by your system. Change it to your preference.

# Model Arguments
model:
  _component_: torchtune.models.llama3_2.lora_llama3_2_1b
  lora_attn_modules: ['q_proj', 'v_proj', 'output_proj']
  apply_lora_to_mlp: True
  lora_rank: 32  # higher increases accuracy and memory
  lora_alpha: 64  # usually alpha=2*rank
  lora_dropout: 0.0
  use_dora: True

# Tokenizer
tokenizer:
  _component_: torchtune.models.llama3.llama3_tokenizer
  path: ./tmp/Llama-3.2-1B/original/tokenizer.model
  max_seq_len: 2048

checkpointer:
  _component_: torchtune.training.FullModelHFCheckpointer
  checkpoint_dir: ./tmp/Llama-3.2-1B/
  checkpoint_files: [
     model.safetensors
  ]
  recipe_checkpoint: null
  output_dir: ${output_dir}
  model_type: LLAMA3_2
resume_from_checkpoint: False
save_adapter_weights_only: False

# Dataset and Sampler
dataset:
  _component_: torchtune.datasets.chat_dataset
  packed: True  # True increases speed
  source: qnguyen3/orca_math_10k
  conversation_column: conversations
  conversation_style: sharegpt
  split: train
seed: 42
shuffle: True
batch_size: 4

# Optimizer and Scheduler
optimizer:
  _component_: torch.optim.AdamW
  fused: True
  weight_decay: 0.01
  lr: 1e-4
lr_scheduler:
  _component_: torchtune.training.lr_schedulers.get_cosine_schedule_with_warmup
  num_warmup_steps: 100

loss:
  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss

# Training
epochs: 1
max_steps_per_epoch: null
gradient_accumulation_steps: 2  # Use to increase effective batch size
compile: True  # torch.compile the model + loss, True increases speed + decreases memory

# Logging
metric_logger:
  _component_: torchtune.training.metric_logging.WandBLogger
  project: benchmark_torchtune
log_every_n_steps: 1
log_peak_memory_stats: True

# Environment
device: cuda
dtype: bf16

# Activations Memory
enable_activation_checkpointing: False  # True reduces memory
enable_activation_offloading: False  # True reduces memory


# 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

[ebsmothers]

Hi @AndrewMead10 thanks for creating the issue. I am able to repro this, dug in a bit and can see that the DoRA magnitudes are not being updated across iterations. I need to do some further investigation, will keep you posted on my findings. But I think this should be pretty high-priority for us to figure out. Also tagging @SLR722 and @calvinpelletier who are familiar with this code to possibly take a look.