NVIDIA NeMo-Aligner 0.6.0rc1.dev0

Prerelease: NVIDIA NeMo-Aligner 0.6.0rc1.dev0 (2024-12-20)'

v0.6.0rc0: fix: fix DPO sequence packing + pipeline parallel (#437)

Signed-off-by: ashors1 <[email protected]>

NVIDIA NeMo-Aligner 0.5.0

New Features and Optimizations

• Implement Kahneman-Tversky Optimization (KTO).
• Sequence packing is now supported when running SFT with SFTChatDataset.

Breaking Changes

Bug Fixes

• Change log_prob_forward_micro_batch_size in DPO to mean the same as the micro_batch_size, which is how many samples(chosen and rejected included) that we process at once.

NVIDIA NeMo-Aligner 0.4.0

• Implement reward-aware preference optimization.
• Added TRT-LLM support in PPO. This can be enabled by doing trainer.ppo.trt_llm.enable=True. There is also a reshard option to reshard out pipeline parallelism during inference for further speedup via trainer.ppo.trt_llm.reshard=True.
• PPO algorithm will now detect if the sample sequence is ended, and if so zero out the gradient of the samples that did not stop properly.
• Added critic warmup to the PPO with the flag trainer.ppo.critic_warmup_steps.

New Features and Optimizations

• Critic and Reward Model server refactored. Now the reward model will have a flag called model.forward_micro_batch_size which determines the micro batch size on which it runs inferences. This can be higher than the training micro batch size since during inference, we have less memory pressure.
• In the critic and reward model server, it is now possible to specify inference_micro_batch_size as a list. This allows us to provide more information to PyTriton regarding the preferred batch sizes for inference.
• It is no longer a requirement to specify num_rollout_samples to be a multiple of inference_micro_batch_size * dp size in PPO.

Breaking Changes

• inference.micro_batch_size is now renamed to inference.inference_micro_batch_size when running reward model inference in inference_rm.yaml. This is to stay consistent with the naming scheme of the PPO critic.
• It is no longer possible to specify add_EOS when running reward model or critic inference.
• NeMo-Aligner now requires Megatron-LM>=0.8.0 for the APIs to calculate the microbatch sizes.

Bug Fixes

• Make num_workers for dataloaders 0 by default. This prevents issues when using MPI (with TRT-LLM) or more sophisticated launchers.

NVIDIA NeMo-Aligner v0.3.1

• SPIN: added rollout_micro_batch_size parameter which allows users to set the batch size for doing generation during SPIN training.
  previously the generation batch size was automatically set to the data parallel size (DP) of the model

New features and optimizations

• Add MoE Support for our reward models.
• SFT/SteerLM: LoRA can now be enabled on all model layers
• DPO: Enable LoRA on all model layers (In this case the actor will be reference model + LoRA weights, we can switch between actor/reference model by enabling/disabling LoRA)
• PPO: Enable LoRA on all model layers (In this case the actor will be init policy + LoRA weights, we can switch between actor/init_policy model by enabling/disabling LoRA)

Breaking changes

Bug Fixes

• Fixed issue where random sampler keeps state when resetting for validation, leading to a different validation batch each validation step. Fixed by using a deterministic sampler
• Fixed crash with float val check interval in DPOTrainer
• Fixed crash with float val check interval when checking progress in DPOTrainer
• Fixed potential crash in SPIN when prompts are longer than encoder_seq_len - generation.max_length
• Fixed crash when calling the generate() method of an SFT model with pipeline parallelism greater than two
• Fixed crash when calling the generate() method of an SFT model with compute_logprob=True and string inputs
• Fixed crash when model.micro_batch_size > 1 in DPO
• Fixed issue when model.encoder_seq_length is mismatched with model.data.train_ds.max_seq_length in SFT and SPIN.
• Delete MegatronPretrainingRandomSampler from Aligner since it has been upstreamed into NeMo

Container

docker pull nvcr.io/nvidia/nemo:24.05

To get access:

1. Sign up to get free and immediate access to NVIDIA NeMo Framework container. If you don’t have an NVIDIA NGC account, you will be prompted to sign up for an account before proceeding.
2. If you don’t have an NVIDIA NGC API key, sign into NVIDIA NGC, selecting organization/team: ea-bignlp/ga-participants and click Generate API key. Save this key for the next step. Else, skip this step.
3. On your machine, docker login to nvcr.io using

docker login nvcr.io
Username: $oauthtoken
Password: <Your Saved NGC API Key>

PyPi

https://pypi.org/project/nemo-aligner/0.3.1/

NVIDIA NeMo-Aligner v0.2.0

New features and optimizations

• Added public-facing official Dockerfile for NeMo-Aligner.
• PPO: memory optimization to help avoid OOM in the actor when sending training data to the critic.
• PPO: it is now possible to use a custom end string in sampling_params.end_strings that is different from <extra_id_1>.
• SFT: added support for custom validation metrics based on model generations.
• Added the ability to do multi-epoch (cfg.max_epochs > 1) training for reward models, DPO, PPO, and SFT
• SFT/SteerLM: added LoRA tuning as an option besides full fine-tuning, only attention_qkv layer is supported

Breaking changes

• We have changed the shuffle logic in the data sampler to support multi-epoch training, so training runs using identical parameters
  will not give the same results anymore because the shuffle logic has changed (specifically the seed value is modified slightly per epoch).
  If you run CI/regression type tests, then be warned that the test may break due to this shuffle change.

Bug Fixes

• Fixed a potential issue when the base model's model.data.data_prefix config is a list and is about to be overridden with
  a dictionary from the training configuration.
• exp_manager.max_time_per_run is now respected, the trainers will save and run validation before exiting if we've reached the time limit.
• Fixed crash in PPO when using a separate reward model server (i.e., with combine_rm_and_critic_server=False).
• Fixed crash when LR scheduler is not specified

Container

docker pull nvcr.io/nvidia/nemo:24.01.framework

To get access:

1. Sign up to get free and immediate access to NVIDIA NeMo Framework container. If you don’t have an NVIDIA NGC account, you will be prompted to sign up for an account before proceeding.
2. If you don’t have an NVIDIA NGC API key, sign into NVIDIA NGC, selecting organization/team: ea-bignlp/ga-participants and click Generate API key. Save this key for the next step. Else, skip this step.
3. On your machine, docker login to nvcr.io using

docker login nvcr.io
Username: $oauthtoken
Password: <Your Saved NGC API Key>

PyPi

https://pypi.org/project/nemo-aligner/0.2.0/

NVIDIA NeMo-Aligner v0.1.0

Highlights

First open source release of NeMo-Aligner. Featuring:

• Support for the full Reinforcement Learning from Human Feedback(RLHF) pipeline including SFT, Reward Model Training and Reinforcement Learning
• Support for the SteerLM technique
• Support for Direct Preference Optimization
• Support for all Megatron Core GPT models such as LLAMA2 70B

Container

docker pull nvcr.io/ea-bignlp/ga-participants/nemofw-training:23.11

To get access:

1. Sign up to get free and immediate access to NVIDIA NeMo Framework container. If you don’t have an NVIDIA NGC account, you will be prompted to sign up for an account before proceeding.
2. If you don’t have an NVIDIA NGC API key, sign into NVIDIA NGC, selecting organization/team: ea-bignlp/ga-participants and click Generate API key. Save this key for the next step. Else, skip this step.
3. On your machine, docker login to nvcr.io using

docker login nvcr.io
Username: $oauthtoken
Password: <Your Saved NGC API Key>

PyPi

https://pypi.org/project/nemo-aligner/0.1.0/