[trainer/algorithm] Implement DAPO and Polaris style dynamic sampling + add DAPO docs + example #130
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @erictang000, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!
I've implemented dynamic sampling capabilities for the PPO training process. This enhancement allows for more intelligent selection or modification of training samples based on their quality, aiming to improve training efficiency and model performance. Two primary strategies, 'filter' (similar to DAPO) and 'replace' (similar to POLARIS), are supported, enabling the system to either discard low-quality samples or substitute them with better ones. This change introduces new configuration options, integrates the sampling logic into the main training loop, and provides robust utility functions for managing the sampling process.
Highlights
- Configuration for Dynamic Sampling: The core configuration for dynamic sampling has been introduced, allowing users to specify the sampling
type(e.g., "filter" or "replace") andmax_sample_batchesdirectly within theppo_base_config.yaml. - Integration into Trainer Class: The
Trainerclass now incorporates adynamic_sampling_stateand integrates the dynamic sampling process directly into its main training loop. This includes a mechanism to continue sampling if needed or to exit the loop if a maximum sampling limit is reached. - Implementation of Sampling Strategies: New utility functions have been added to
trainer_utils.pyto encapsulate the logic for different dynamic sampling strategies. This includeshandle_replace_sampling(POLARIS-style, replacing bad samples) andhandle_filter_sampling(DAPO-style, filtering and accumulating good samples). - Comprehensive Unit Testing: Extensive unit tests have been added for the new dynamic sampling utility functions, covering various scenarios for both replacement and filtering strategies, ensuring the robustness and correctness of the new sampling mechanisms.
Using Gemini Code Assist
The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.
Invoking Gemini
You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.
| Feature | Command | Description |
|---|---|---|
| Code Review | /gemini review |
Performs a code review for the current pull request in its current state. |
| Pull Request Summary | /gemini summary |
Provides a summary of the current pull request in its current state. |
| Comment | @gemini-code-assist | Responds in comments when explicitly tagged, both in issue comments and review comments. |
| Help | /gemini help |
Displays a list of available commands. |
Customization
To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.
Limitations & Feedback
Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments or fill out our survey to provide feedback.
You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.
Footnotes
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Code Review
This pull request introduces dynamic sampling capabilities, with both 'filter' (DAPO-style) and 'replace' (Polaris-style) strategies. The implementation is split between the Trainer class for control flow and trainer_utils.py for the core logic. The changes are well-tested, covering various scenarios for the new sampling methods. My review focuses on improving the robustness and correctness of the implementation. I've identified a few critical issues related to handling token-level rewards and potential errors with empty data, along with a minor issue regarding type hint consistency. Addressing these points will make the new dynamic sampling feature more reliable.
erictang000
changed the title
|
thanks @SumanthRH, addressed comments, should be ready for re-review |
SumanthRH
reviewed
Aug 7, 2025
Comment on lines
30
to
44
| """ | ||
| # this assumes response-level rewards | ||
| scores = token_level_rewards.sum(dim=-1) | ||
|
|
||
| # Overlong punishment params - hardcoded for this script for now | ||
| # TODO (erictang000): make these configurable (in general for all custom registries) | ||
| max_resp_length = 1024 # this is generator.sampling_params.max_generate_length in the `run_dapo_gsm8k.sh` script | ||
| overlong_buffer_len = 512 # overlong buffer is last 512 tokens of the response as an example | ||
| overlong_penalty_factor = ( | ||
| 1.0 # reward penalty increases linearly from 0 to 1.0 as the response length enters the overlong buffer | ||
| ) | ||
|
|
||
| # add soft overlong punishment | ||
| lengths = response_mask.sum(dim=-1) | ||
| buffer_start_idx = max_resp_length - overlong_buffer_len |
There was a problem hiding this comment.
Wait why would be do this here? This is pretty hacky and not what we'd want to show for an example of reward customization?
The ideal place is probably in the Generator, but we don't have any postprocessing hooks there.
For now, we can quickly make this be custom postprocessing in the trainer by subclassing postprocess_generator_output and adding the reward penalty here?
SkyRL/skyrl-train/skyrl_train/trainer.py
Line 672 in bd9d6a9
class DAPOTrainer(RayPPOTrainer):
@torch.no_grad()
def postprocess_generator_output(self, generator_output: GeneratorOutput, uids: List[str]) -> GeneratorOutput:
# modify rewards here
....
# use base class impl for metrics and per-token reward conversion
return super().postprocess_generator_output(....)There was a problem hiding this comment.
yes agree this is cleaner, updated to do this!
SumanthRH
reviewed
Aug 7, 2025
erictang000
commented
Aug 7, 2025
| # main training loop | ||
| pbar = tqdm(total=self.total_training_steps, initial=self.global_step, desc="Training Step Progress") | ||
| pbar = tqdm(total=self.total_training_steps, initial=self.global_step, desc="Training Batches Processed") |
There was a problem hiding this comment.
updated this text because if dynamic sampling is enabled, we don't necessarily take a training step - we just process a batch (and we should update the pbar otherwise it will stop at something like 37/140 which is bad UI).
|
@SumanthRH updated and tested with a cleaner example of creating a custom trainer, ready for another review |
SumanthRH
reviewed
Aug 7, 2025
| @@ -76,7 +77,7 @@ def create_ray_wrapped_inference_engines( | |||
| from skyrl_train.inference_engines.vllm.vllm_engine import VLLMRayActor, AsyncVLLMRayActor | |||
| from skyrl_train.utils import ray_noset_visible_devices, get_all_env_variables, get_ray_pg_ready_with_timeout | |||
|
|
|||
| assert vllm.__version__ >= "0.8.3", "SkyTrainer only supports vLLM >= 0.8.3" | |||
| assert version.parse(vllm.__version__) >= version.parse("0.8.3"), "SkyRL-Train only supports vLLM >= 0.8.3" | |||
There was a problem hiding this comment.
QQ: why this change in this PR?
There was a problem hiding this comment.
i was testing vllm 0.10.0 for fun... (upgrade should be easy) and the assertion failed here lol
|
/gemini review |
There was a problem hiding this comment.
Code Review
This pull request introduces two dynamic sampling strategies, 'filter' (DAPO-style) and 'replace' (Polaris-style), to improve training efficiency. The changes include the core logic for these strategies, configuration options, extensive documentation with examples, and necessary adjustments to the training loop and utility classes like Timer and a new ConditionalWeightsManager. The implementation is well-structured and supported by a comprehensive test suite. My review focuses on a couple of minor issues in the documentation and error message formatting. Overall, this is a high-quality contribution.
vinid
added a commit
to vinid/SkyRL
that referenced
this pull request
Aug 11, 2025
* [Trainer] Support per-token rewards in trainer (NovaSky-AI#109) * Add check for whether p2p access is supported - allows code to run on L4/L40S after NovaSky-AI#73 upgrade to cuda 12.8 (NovaSky-AI#108) # Overview After NovaSky-AI#73, the main code path no longer runs on GPUs without P2P support (potentially due to cuda 12.8 upgrade?) - an error would be thrown like ```bash torch.distributed.DistBackendError: NCCL error in: /pytorch/torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:3353, unhandled cuda error (run with NCCL_DEBUG=INFO for details), NCCL version 2.26.2 ncclUnhandledCudaError: Call to CUDA function failed. Last error: Cuda failure 217 'peer access is not supported between these two devices' ``` This PR adds a check for whether peer access is supported (using torch/cuda) between all GPUs on a node to the ray initialization, and sets relevant NCCL env vars to allow the code to run on these machine types. ```python if not peer_access_supported(): logger.info("Peer access is not supported, disabling P2P and SHM") env_vars["NCCL_P2P_DISABLE"] = "1" env_vars["NCCL_SHM_DISABLE"] = "1" ``` Example running on L40S: <img width="1854" height="227" alt="image" src="https://github.com/user-attachments/assets/1cca46b5-6e16-4ae7-9a33-df52d138bdeb" /> * [dependencies] Upgrade ray to 2.48.0 (NovaSky-AI#106) # What does this PR do Upgrades ray to 2.48.0, which allows us to remove the pip install vllm in the Dockerfile as a fallback for when uv + vllm does not resolve dependencies with the vllm + ray backend correctly. We leave the previous Dockerfile in `docker/Dockerfile.ray244` for backwards compatibility --------- Co-authored-by: Sumanth R Hegde <[email protected]> * fix issue with NovaSky-AI#108 that broke gpu ci (NovaSky-AI#112) missed an argument in `gpu_ci/conftest.py` for `peer_access_supported()` - fix for gpu ci to run Passing now with update: <img width="1811" height="861" alt="image" src="https://github.com/user-attachments/assets/70011c54-1e33-44b5-83a0-616029f891d2" /> And main runs (and disables p2p access) correctly: <img width="2067" height="203" alt="image" src="https://github.com/user-attachments/assets/399aff67-cc51-4588-a632-47698073593c" /> * Add warning for certain uv versions due to `uv run --with` regression (NovaSky-AI#113) # What does this PR do? Adds a warning for uv versions 0.8.0, 0.8.1 and 0.8.2 due to a bug in the uv run --with flag for "Running in ray cluster" section. These are relatively new versions and thus it's better to have this detail in the documentation for users. <img width="692" height="458" alt="Screenshot 2025-07-25 at 6 09 15 PM" src="https://github.com/user-attachments/assets/f1997eac-2867-4552-8ef7-eea8741e32b6" /> <img width="779" height="568" alt="Screenshot 2025-07-25 at 6 09 19 PM" src="https://github.com/user-attachments/assets/5080d328-c934-4864-91a8-932902dea934" /> --------- Signed-off-by: SumanthRH <[email protected]> * [GPU CI] Only trigger workflow for relevant changes in `skyrl-train` (NovaSky-AI#114) * [bug] Loading saved HF weights errors (NovaSky-AI#118) Addresses NovaSky-AI#97 * [DAPO] Add support for overlong filtering (NovaSky-AI#111) ## What does this PR do? Adds `apply_overlong_filtering` to the generator config, and provides a generator utility method `apply_overlong_filtering()` for post-processing the loss mask. I originally implemented this using the `stop_reasons` to determine whether the sequence was truncated, but instead switched to looking for `eos_token` in the response IDs for a more general approach. ## Tests Added CPU tests for the utility method and for SkyRL Gym Generator's use of the utility method. * [skyrl-gym] GSM8k - LLM Judge example (NovaSky-AI#74) * Fix MLFlow logging (NovaSky-AI#121) This is a small change to make the MLFlow integration work. Currently this fails with a Pandas error when trying to flatten an Omega dict; we need to convert to a regular Python dictionary. Can confirm this works on our MLFlow setup: <img width="1406" height="683" alt="image" src="https://github.com/user-attachments/assets/fcee526a-815e-4f08-bf25-d2709779ced7" /> * [Trainer] Support registering custom advantage estimators (NovaSky-AI#115) ## What does this PR do? Adds an `AdvantageEstimatorRegistry` to support custom advantage estimation methods without modifying the skyrl-train package. Added `examples/algorithm/custom_advantage_estimator` folder to give quick example of how to register a custom adv est function. ## Tests Adding cpu test to ensure registration works. * [checkpointing] Add HF model config and tokenizer config to checkpoint folder (NovaSky-AI#124) # Overview Adds the HF model config and tokenizer config to `ckpt_path/huggingface` for deepspeed and fsdp. So now the checkpoint directory will be: ``` {ckpt_path}/ ├── latest_ckpt_global_step.txt # Holds the global step of the latest checkpoint ├── global_step_10/ # Checkpoint at training step 10 │ ├── policy/ # Policy model checkpoint directory │ │ ├── fsdp_config.json # stores fsdp version and world size │ │ ├── huggingface/ │ │ ├── config.json # model config │ │ ├── tokenizer_config.json # tokenizer config │ │ ├── generation_config.json # generation config │ │ ├── ... # other tokenizer config files │ │ ├── model_state.pt # Model parameters │ │ ├── optimizer_state.pt # Optimizer state │ │ └── lr_scheduler_state.pt # Learning rate scheduler state ``` For deepspeed it will be similar but without `fsdp_config.json` ``` {ckpt_path}/ ├── latest_ckpt_global_step.txt # Holds the global step of the latest checkpoint ├── global_step_10/ # Checkpoint at training step 10 │ ├── policy/ # Policy model checkpoint directory │ │ ├── huggingface/ │ │ ├── config.json # model config │ │ ├── tokenizer_config.json # tokenizer config │ │ ├── generation_config.json # generation config │ │ ├── ... # other tokenizer config files │ │ ├── ... # deepspeed checkpointing files ``` * Fix discord link (NovaSky-AI#125) * Fix broken link (NovaSky-AI#128) * [Trainer/Algorithm] Support registering custom policy loss functions + refactor adv estimator registry to allow registration outside ray workers (NovaSky-AI#126) # Overview - Adds support for registering custom policy loss functions, similar to NovaSky-AI#115, - Refactors the policy loss to be a function in `ppo_utils.py` instead of a (`nn.Module` in `worker.py`) - Introduces a breaking change in renaming `trainer.algorithm.ppo_loss_type` to `trainer.algorithm.policy_loss_type` - Addresses Issue NovaSky-AI#116 by creating a new `BaseFunctionRegistry` class that uses a [named actor](https://docs.ray.io/en/latest/ray-core/actors/named-actors.html) to support the following pattern: ```python # Example of custom policy loss: "simple_baseline" def compute_simple_baseline_policy_loss( log_probs: torch.Tensor, ... ): return torch.randn(1, device=log_probs.device), 0.0 # Register the custom policy loss - outside of the ray worker PolicyLossRegistry.register("simple_baseline", compute_simple_baseline_policy_loss) @ray.remote(num_cpus=1) def skyrl_entrypoint(cfg: DictConfig): exp = BasePPOExp(cfg) exp.run() @hydra.main(config_path=config_dir, config_name="ppo_base_config", version_base=None) def main(cfg: DictConfig) -> None: # validate the arguments validate_cfg(cfg) initialize_ray(cfg) ray.get(skyrl_entrypoint.remote(cfg)) ``` this change was necessary for `PolicyLossRegistry` to be accessible, since the worker `actor_loss_fn` attribute is set in `init_model` within the `worker` actor, which is a ray actor created from within the skyrl_entrypoint ray task (and registering within the entrypoint wouldn't propagate down another layer). - updates AdvantageEstimatorRegistry to extend the same `BaseFunctionRegistry` class Example runs: Custom advantage (mean of reward) <img width="956" height="326" alt="image" src="https://github.com/user-attachments/assets/1b7222bc-fbb9-49b1-876d-265b71201087" /> Custom policy loss (reinforce - just (-logprobs * advantages).mean()) <img width="939" height="330" alt="image" src="https://github.com/user-attachments/assets/cbed7ef5-b3e7-4e32-beba-b52b80879f47" /> * [SkyAgent] Upload initial refactored code (NovaSky-AI#131) # What does this PR do? Uploading our initial refactored code for SkyAgent --------- Signed-off-by: SumanthRH <[email protected]> Co-authored-by: Shiyi Cao <[email protected]> Co-authored-by: Dacheng Li <[email protected]> * [trainer] add more robust generation output validation (NovaSky-AI#132) # Overview Adds a `validate_generation_output` function in `trainer_utils.py` with more robust validation of generation output format. Specifically, given ``` class GeneratorOutput(TypedDict): prompt_token_ids: List[List[int]] response_ids: List[List[int]] rewards: Union[List[float], List[List[float]]] loss_masks: List[List[int]] stop_reasons: Optional[List[str]] rollout_metrics: Optional[Dict[str, Any]] ``` We expect - all list attributes should have the same length and be the same length as the input batch of prompts at dim=0 - non zero length lists - response_ids, loss masks, and rewards (if token level rewards) should be the same length - the sum of loss masks should be non-zero (logging a warning if it is not) verified gsm8k run still works: <img width="563" height="330" alt="image" src="https://github.com/user-attachments/assets/eeefebcb-d5fc-486d-b906-f4344b1e2779" /> --------- Co-authored-by: Sumanth R Hegde <[email protected]> * [Trainer] GSPO support (NovaSky-AI#120) This PR adds support for [Group Sequence Policy Optimization (GSPO)](https://arxiv.org/abs/2507.18071), the hotness du jour from Alibaba Qwen. The implementation in this PR is loosely based on [this one](huggingface/trl#3775) from TRL. It adds an `importance_sampling_level` config option which can be `token` (PPO/GRPO) or `sequence` (GSPO). I ran a short/small GSM8k run with Qwen2.5-0.5B and the loss curves look okay: <img width="314" height="240" alt="image" src="https://github.com/user-attachments/assets/f52d7c64-416c-4419-aa96-4a03c9048007" /> However, I had to hack a few things to get this to run on Datadog's cloud infra (including changing some dependency versions) so I'd encourage one of the maintainers to reproduce these results locally before merging. * [SkyAgent] Add initial docs (NovaSky-AI#134) # What does this PR do? Adds initial documentation for SkyAgent. We are still actively cleaning this package up, but I thought initial documentation will be helpful for anyone who stumbles across this. The documentation folder is still in `skyrl-train`, and much of the docs also refer to "SkyRL" when they are really referring to "SkyRL-train", so to avoid any confusion, I have just added this as a simple page on the sidebar. We need to make the docs be mono-repo wide and structure it better but I'm leaving it for a future PR. --------- Signed-off-by: SumanthRH <[email protected]> * [trainer/algorithm] Implement DAPO and Polaris style dynamic sampling + add DAPO docs + example (NovaSky-AI#130) # Overview This PR introduces filter (DAPO) and replace (Polaris/WebSailor) style dynamic sampling strategies. The dynamic sampling strategy can be configured as below: ```yaml # dynamic sampling parameters dynamic_sampling: type: null # filter (DAPO), replace (POLARIS/WebSailor), or null max_sample_batches: 30 # sample at most this many batches before stopping, -1 to sample forever min_replace_ratio: 0.3 # minimum proportion of good samples with which to replace bad samples (for replace strategy only) ``` This PR also adds a docs page describing how to enable all DAPO features, and adds an example GSM8K script where all these features are used. ## Minor Changes Some minor changes to make this dynamic sampling implementation clean: - the utils `Timer` class now updates the dict instead of overwriting in order to correctly track generation time w/ dynamic sampling, which means we need to make sure to reset `all_timings` in any trainer - The use of `self.weights_manager` is a little tricky for the dynamic sampling - introduced the the `ConditionalWeightsManager` to make the added code in the training loop as clean as possible ## Example runs <img width="413" height="264" alt="image" src="https://github.com/user-attachments/assets/072f716a-3632-42bb-a5f7-5f9d6064bd93" /> Generation time for dapo style filtering increases as the training run goes on, while it is stable for polaris and the baseline. <img width="419" height="265" alt="image" src="https://github.com/user-attachments/assets/887df550-e4b9-4623-b578-b4809a9f403f" /> We can see that the training pass @ n metric is 1 for both polaris and dapo style filtering as expected. <img width="421" height="259" alt="image" src="https://github.com/user-attachments/assets/bb63af77-1fbb-4d89-9216-b028f1551ea7" /> For GSM8k + Qwen 1.5B, the sampling strategy (as well as the full DAPO run) results in minimal gains - need larger models/harder dataset to test more fully DAPO sampling Example Run: ```bash (skyrl_entrypoint pid=222117) 2025-08-04 23:13:13.439 | INFO | skyrl_train.trainer:train:245 - Started: 'step' (skyrl_entrypoint pid=222117) 2025-08-04 23:13:13.737 | INFO | skyrl_train.weights_manager:__enter__:76 - Started: 'sync_weights_to_inference_engines' (skyrl_entrypoint pid=222117) 2025-08-04 23:13:16.401 | INFO | skyrl_train.weights_manager:__enter__:76 - Finished: 'sync_weights_to_inference_engines', time cost: 2.66s (skyrl_entrypoint pid=222117) 2025-08-04 23:13:16.401 | INFO | skyrl_train.weights_manager:__enter__:80 - Started: 'offload_policy_model_to_cpu' (skyrl_entrypoint pid=222117) 2025-08-04 23:13:16.842 | INFO | skyrl_train.weights_manager:__enter__:80 - Finished: 'offload_policy_model_to_cpu', time cost: 0.44s (skyrl_entrypoint pid=222117) 2025-08-04 23:13:16.888 | INFO | skyrl_train.trainer:train:261 - Started: 'generate' (AsyncVLLMInferenceEngine pid=223856) INFO 08-04 23:13:13 [executor_base.py:227] It took 0.243244 seconds to wake up tags ['weights']. [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=223854) INFO 08-04 23:13:16 [executor_base.py:227] It took 0.040547 seconds to wake up tags ['kv_cache']. (AsyncVLLMInferenceEngine pid=223856) INFO 08-04 23:13:16 [block_pool.py:316] Successfully reset prefix cache [repeated 7x across cluster] (AsyncVLLMInferenceEngine pid=223855) INFO 08-04 23:13:16 [executor_base.py:227] It took 0.041721 seconds to wake up tags ['kv_cache']. (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.378 | INFO | skyrl_train.trainer:train:261 - Finished: 'generate', time cost: 17.49s (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.395 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:433 - ============= Dynamic sampling filter ============= (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.395 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:434 - Dynamic sampling: 460 < 1024 prompts (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.395 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:435 - Resample batch 1, continue sampling... (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.395 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:436 - ================================================== (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.395 | INFO | skyrl_train.trainer:train:245 - Finished: 'step', time cost: 20.96s (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.407 | INFO | skyrl_train.trainer:train:245 - Started: 'step' (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.445 | INFO | skyrl_train.trainer:train:261 - Started: 'generate' (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.014 | INFO | skyrl_train.trainer:train:261 - Finished: 'generate', time cost: 17.57s (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.029 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:433 - ============= Dynamic sampling filter ============= (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.029 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:434 - Dynamic sampling: 941 < 1024 prompts (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.029 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:435 - Resample batch 2, continue sampling... (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.029 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:436 - ================================================== (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.030 | INFO | skyrl_train.trainer:train:245 - Finished: 'step', time cost: 17.62s (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.033 | INFO | skyrl_train.trainer:train:245 - Started: 'step' (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.074 | INFO | skyrl_train.trainer:train:261 - Started: 'generate' (skyrl_entrypoint pid=222117) 2025-08-04 23:14:08.380 | INFO | skyrl_train.trainer:train:261 - Finished: 'generate', time cost: 16.31s (skyrl_entrypoint pid=222117) 2025-08-04 23:14:08.396 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:439 - ============= Dynamic sampling filter ============= (skyrl_entrypoint pid=222117) 2025-08-04 23:14:08.396 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:440 - Dynamic sampling: collected 1467 >= 1024 prompts (skyrl_entrypoint pid=222117) 2025-08-04 23:14:08.397 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:443 - ================================================== (AsyncVLLMInferenceEngine pid=223856) INFO 08-04 23:13:12 [gpu_worker.py:98] Sleep mode freed 61.88 GiB memory, 4.98 GiB memory is still in use. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=223856) INFO 08-04 23:13:12 [executor_base.py:211] It took 1.264572 seconds to fall asleep. [repeated 3x across cluster] ``` Polaris Style example run: ```bash (skyrl_entrypoint pid=306764) 2025-08-05 00:30:01.648 | INFO | skyrl_train.trainer:train:261 - Started: 'generate' (AsyncVLLMInferenceEngine pid=308521) INFO 08-05 00:29:58 [executor_base.py:227] It took 0.240372 seconds to wake up tags ['weights']. [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=308520) INFO 08-05 00:30:01 [executor_base.py:227] It took 0.040980 seconds to wake up tags ['kv_cache']. (AsyncVLLMInferenceEngine pid=308521) INFO 08-05 00:30:00 [block_pool.py:316] Successfully reset prefix cache [repeated 7x across cluster] (AsyncVLLMInferenceEngine pid=308518) INFO 08-05 00:30:01 [executor_base.py:227] It took 0.041175 seconds to wake up tags ['kv_cache']. (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.663 | INFO | skyrl_train.trainer:train:261 - Finished: 'generate', time cost: 15.01s (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.679 | INFO | skyrl_train.utils.trainer_utils:handle_replace_sampling:316 - Replace sampling: 629 good UIDs out of 1024 total prompts (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.680 | INFO | skyrl_train.utils.trainer_utils:handle_replace_sampling:320 - ============= Dynamic sampling replace =========== (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.680 | INFO | skyrl_train.utils.trainer_utils:handle_replace_sampling:321 - Number of good prompts: 629 (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.680 | INFO | skyrl_train.utils.trainer_utils:handle_replace_sampling:322 - Number of bad prompts: 395 (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.694 | INFO | skyrl_train.utils.trainer_utils:handle_replace_sampling:352 - After replacement - Replaced 395 bad prompts (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.694 | INFO | skyrl_train.utils.trainer_utils:handle_replace_sampling:353 - ================================================== (AsyncVLLMInferenceEngine pid=308520) INFO 08-05 00:29:57 [gpu_worker.py:98] Sleep mode freed 62.14 GiB memory, 6.28 GiB memory is still in use. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=308520) INFO 08-05 00:29:57 [executor_base.py:211] It took 1.331663 seconds to fall asleep. ``` ## Full DAPO example run From example script <img width="417" height="262" alt="image" src="https://github.com/user-attachments/assets/2592a06f-8b8a-4cf1-a29e-321bff819eb0" /> <img width="909" height="325" alt="image" src="https://github.com/user-attachments/assets/50922afd-1424-4183-9329-4f1f340287eb" /> --------- Co-authored-by: Sumanth R Hegde <[email protected]> * [algorithm] Support Dr. GRPO + refactor where policy/critic loss functions are set (NovaSky-AI#133) # Overview ## Dr GRPO Adds `loss_reduction`: `seq_mean_token_sum_norm ` option, and `grpo_norm_by_std` option to support Dr. GRPO So to run Dr. GRPO, set: ```yaml trainer: algorithm: grpo_norm_by_std: false loss_reduction: "seq_mean_token_sum_norm" ... ``` Example run: <img width="906" height="317" alt="image" src="https://github.com/user-attachments/assets/ce9db2ef-253e-45c8-adba-1ef8a270bbd9" /> Reward looks similar <img width="419" height="263" alt="image" src="https://github.com/user-attachments/assets/a4bc4d8c-f3c1-4bad-a497-0297dc30bc27" /> Magnitude of policy loss is lower as expected (since we are normalizing by a larger constant rather than taking the mean) ## Refactor where Critic/Policy Loss are set Changes ppo critic `ValueLoss` to just a function instead of a `nn.Module` for consistency with `policy_loss`, and adds new algorithm field to cfg that require evaluating field values in `utils::validate_cfg` (this runs before entrypoint code, allowing users to modify the cfg further by subclassing `BasePPOExp`) PPO example still running after this refactor: <img width="421" height="262" alt="image" src="https://github.com/user-attachments/assets/88985da3-1403-49c6-8cb5-f1434151fd9e" /> * [fix] move algorithm folder -> algorithms (NovaSky-AI#136) left the algorithm folder in NovaSky-AI#133, move it over * [Logging] Forward mlflow env vars to ray runtime env (NovaSky-AI#135) This PR forward the `MLFLOW_TRACKING_URI` and `MLFLOW_TRACKING_TOKEN` environment variable to the ray runtime env during its initialization. This will enable users to simply provide the above env vars at the driver and be able to use MLFlow for experiment tracking. * data folder * some stuff * updates --------- Signed-off-by: SumanthRH <[email protected]> Co-authored-by: Sumanth R Hegde <[email protected]> Co-authored-by: Eric Tang <[email protected]> Co-authored-by: Tyler Griggs <[email protected]> Co-authored-by: Shu Liu <[email protected]> Co-authored-by: Ben Cohen <[email protected]> Co-authored-by: Shiyi Cao <[email protected]> Co-authored-by: Dacheng Li <[email protected]> Co-authored-by: Etienne Brodu <[email protected]>
fannie1208
pushed a commit
to vinid/SkyRL
that referenced
this pull request
Aug 19, 2025
* [Trainer] Support per-token rewards in trainer (NovaSky-AI#109) * Add check for whether p2p access is supported - allows code to run on L4/L40S after NovaSky-AI#73 upgrade to cuda 12.8 (NovaSky-AI#108) # Overview After NovaSky-AI#73, the main code path no longer runs on GPUs without P2P support (potentially due to cuda 12.8 upgrade?) - an error would be thrown like ```bash torch.distributed.DistBackendError: NCCL error in: /pytorch/torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:3353, unhandled cuda error (run with NCCL_DEBUG=INFO for details), NCCL version 2.26.2 ncclUnhandledCudaError: Call to CUDA function failed. Last error: Cuda failure 217 'peer access is not supported between these two devices' ``` This PR adds a check for whether peer access is supported (using torch/cuda) between all GPUs on a node to the ray initialization, and sets relevant NCCL env vars to allow the code to run on these machine types. ```python if not peer_access_supported(): logger.info("Peer access is not supported, disabling P2P and SHM") env_vars["NCCL_P2P_DISABLE"] = "1" env_vars["NCCL_SHM_DISABLE"] = "1" ``` Example running on L40S: <img width="1854" height="227" alt="image" src="https://github.com/user-attachments/assets/1cca46b5-6e16-4ae7-9a33-df52d138bdeb" /> * [dependencies] Upgrade ray to 2.48.0 (NovaSky-AI#106) # What does this PR do Upgrades ray to 2.48.0, which allows us to remove the pip install vllm in the Dockerfile as a fallback for when uv + vllm does not resolve dependencies with the vllm + ray backend correctly. We leave the previous Dockerfile in `docker/Dockerfile.ray244` for backwards compatibility --------- Co-authored-by: Sumanth R Hegde <[email protected]> * fix issue with NovaSky-AI#108 that broke gpu ci (NovaSky-AI#112) missed an argument in `gpu_ci/conftest.py` for `peer_access_supported()` - fix for gpu ci to run Passing now with update: <img width="1811" height="861" alt="image" src="https://github.com/user-attachments/assets/70011c54-1e33-44b5-83a0-616029f891d2" /> And main runs (and disables p2p access) correctly: <img width="2067" height="203" alt="image" src="https://github.com/user-attachments/assets/399aff67-cc51-4588-a632-47698073593c" /> * Add warning for certain uv versions due to `uv run --with` regression (NovaSky-AI#113) # What does this PR do? Adds a warning for uv versions 0.8.0, 0.8.1 and 0.8.2 due to a bug in the uv run --with flag for "Running in ray cluster" section. These are relatively new versions and thus it's better to have this detail in the documentation for users. <img width="692" height="458" alt="Screenshot 2025-07-25 at 6 09 15 PM" src="https://github.com/user-attachments/assets/f1997eac-2867-4552-8ef7-eea8741e32b6" /> <img width="779" height="568" alt="Screenshot 2025-07-25 at 6 09 19 PM" src="https://github.com/user-attachments/assets/5080d328-c934-4864-91a8-932902dea934" /> --------- Signed-off-by: SumanthRH <[email protected]> * [GPU CI] Only trigger workflow for relevant changes in `skyrl-train` (NovaSky-AI#114) * [bug] Loading saved HF weights errors (NovaSky-AI#118) Addresses NovaSky-AI#97 * [DAPO] Add support for overlong filtering (NovaSky-AI#111) ## What does this PR do? Adds `apply_overlong_filtering` to the generator config, and provides a generator utility method `apply_overlong_filtering()` for post-processing the loss mask. I originally implemented this using the `stop_reasons` to determine whether the sequence was truncated, but instead switched to looking for `eos_token` in the response IDs for a more general approach. ## Tests Added CPU tests for the utility method and for SkyRL Gym Generator's use of the utility method. * [skyrl-gym] GSM8k - LLM Judge example (NovaSky-AI#74) * Fix MLFlow logging (NovaSky-AI#121) This is a small change to make the MLFlow integration work. Currently this fails with a Pandas error when trying to flatten an Omega dict; we need to convert to a regular Python dictionary. Can confirm this works on our MLFlow setup: <img width="1406" height="683" alt="image" src="https://github.com/user-attachments/assets/fcee526a-815e-4f08-bf25-d2709779ced7" /> * [Trainer] Support registering custom advantage estimators (NovaSky-AI#115) ## What does this PR do? Adds an `AdvantageEstimatorRegistry` to support custom advantage estimation methods without modifying the skyrl-train package. Added `examples/algorithm/custom_advantage_estimator` folder to give quick example of how to register a custom adv est function. ## Tests Adding cpu test to ensure registration works. * [checkpointing] Add HF model config and tokenizer config to checkpoint folder (NovaSky-AI#124) # Overview Adds the HF model config and tokenizer config to `ckpt_path/huggingface` for deepspeed and fsdp. So now the checkpoint directory will be: ``` {ckpt_path}/ ├── latest_ckpt_global_step.txt # Holds the global step of the latest checkpoint ├── global_step_10/ # Checkpoint at training step 10 │ ├── policy/ # Policy model checkpoint directory │ │ ├── fsdp_config.json # stores fsdp version and world size │ │ ├── huggingface/ │ │ ├── config.json # model config │ │ ├── tokenizer_config.json # tokenizer config │ │ ├── generation_config.json # generation config │ │ ├── ... # other tokenizer config files │ │ ├── model_state.pt # Model parameters │ │ ├── optimizer_state.pt # Optimizer state │ │ └── lr_scheduler_state.pt # Learning rate scheduler state ``` For deepspeed it will be similar but without `fsdp_config.json` ``` {ckpt_path}/ ├── latest_ckpt_global_step.txt # Holds the global step of the latest checkpoint ├── global_step_10/ # Checkpoint at training step 10 │ ├── policy/ # Policy model checkpoint directory │ │ ├── huggingface/ │ │ ├── config.json # model config │ │ ├── tokenizer_config.json # tokenizer config │ │ ├── generation_config.json # generation config │ │ ├── ... # other tokenizer config files │ │ ├── ... # deepspeed checkpointing files ``` * Fix discord link (NovaSky-AI#125) * Fix broken link (NovaSky-AI#128) * [Trainer/Algorithm] Support registering custom policy loss functions + refactor adv estimator registry to allow registration outside ray workers (NovaSky-AI#126) # Overview - Adds support for registering custom policy loss functions, similar to NovaSky-AI#115, - Refactors the policy loss to be a function in `ppo_utils.py` instead of a (`nn.Module` in `worker.py`) - Introduces a breaking change in renaming `trainer.algorithm.ppo_loss_type` to `trainer.algorithm.policy_loss_type` - Addresses Issue NovaSky-AI#116 by creating a new `BaseFunctionRegistry` class that uses a [named actor](https://docs.ray.io/en/latest/ray-core/actors/named-actors.html) to support the following pattern: ```python # Example of custom policy loss: "simple_baseline" def compute_simple_baseline_policy_loss( log_probs: torch.Tensor, ... ): return torch.randn(1, device=log_probs.device), 0.0 # Register the custom policy loss - outside of the ray worker PolicyLossRegistry.register("simple_baseline", compute_simple_baseline_policy_loss) @ray.remote(num_cpus=1) def skyrl_entrypoint(cfg: DictConfig): exp = BasePPOExp(cfg) exp.run() @hydra.main(config_path=config_dir, config_name="ppo_base_config", version_base=None) def main(cfg: DictConfig) -> None: # validate the arguments validate_cfg(cfg) initialize_ray(cfg) ray.get(skyrl_entrypoint.remote(cfg)) ``` this change was necessary for `PolicyLossRegistry` to be accessible, since the worker `actor_loss_fn` attribute is set in `init_model` within the `worker` actor, which is a ray actor created from within the skyrl_entrypoint ray task (and registering within the entrypoint wouldn't propagate down another layer). - updates AdvantageEstimatorRegistry to extend the same `BaseFunctionRegistry` class Example runs: Custom advantage (mean of reward) <img width="956" height="326" alt="image" src="https://github.com/user-attachments/assets/1b7222bc-fbb9-49b1-876d-265b71201087" /> Custom policy loss (reinforce - just (-logprobs * advantages).mean()) <img width="939" height="330" alt="image" src="https://github.com/user-attachments/assets/cbed7ef5-b3e7-4e32-beba-b52b80879f47" /> * [SkyAgent] Upload initial refactored code (NovaSky-AI#131) # What does this PR do? Uploading our initial refactored code for SkyAgent --------- Signed-off-by: SumanthRH <[email protected]> Co-authored-by: Shiyi Cao <[email protected]> Co-authored-by: Dacheng Li <[email protected]> * [trainer] add more robust generation output validation (NovaSky-AI#132) # Overview Adds a `validate_generation_output` function in `trainer_utils.py` with more robust validation of generation output format. Specifically, given ``` class GeneratorOutput(TypedDict): prompt_token_ids: List[List[int]] response_ids: List[List[int]] rewards: Union[List[float], List[List[float]]] loss_masks: List[List[int]] stop_reasons: Optional[List[str]] rollout_metrics: Optional[Dict[str, Any]] ``` We expect - all list attributes should have the same length and be the same length as the input batch of prompts at dim=0 - non zero length lists - response_ids, loss masks, and rewards (if token level rewards) should be the same length - the sum of loss masks should be non-zero (logging a warning if it is not) verified gsm8k run still works: <img width="563" height="330" alt="image" src="https://github.com/user-attachments/assets/eeefebcb-d5fc-486d-b906-f4344b1e2779" /> --------- Co-authored-by: Sumanth R Hegde <[email protected]> * [Trainer] GSPO support (NovaSky-AI#120) This PR adds support for [Group Sequence Policy Optimization (GSPO)](https://arxiv.org/abs/2507.18071), the hotness du jour from Alibaba Qwen. The implementation in this PR is loosely based on [this one](huggingface/trl#3775) from TRL. It adds an `importance_sampling_level` config option which can be `token` (PPO/GRPO) or `sequence` (GSPO). I ran a short/small GSM8k run with Qwen2.5-0.5B and the loss curves look okay: <img width="314" height="240" alt="image" src="https://github.com/user-attachments/assets/f52d7c64-416c-4419-aa96-4a03c9048007" /> However, I had to hack a few things to get this to run on Datadog's cloud infra (including changing some dependency versions) so I'd encourage one of the maintainers to reproduce these results locally before merging. * [SkyAgent] Add initial docs (NovaSky-AI#134) # What does this PR do? Adds initial documentation for SkyAgent. We are still actively cleaning this package up, but I thought initial documentation will be helpful for anyone who stumbles across this. The documentation folder is still in `skyrl-train`, and much of the docs also refer to "SkyRL" when they are really referring to "SkyRL-train", so to avoid any confusion, I have just added this as a simple page on the sidebar. We need to make the docs be mono-repo wide and structure it better but I'm leaving it for a future PR. --------- Signed-off-by: SumanthRH <[email protected]> * [trainer/algorithm] Implement DAPO and Polaris style dynamic sampling + add DAPO docs + example (NovaSky-AI#130) # Overview This PR introduces filter (DAPO) and replace (Polaris/WebSailor) style dynamic sampling strategies. The dynamic sampling strategy can be configured as below: ```yaml # dynamic sampling parameters dynamic_sampling: type: null # filter (DAPO), replace (POLARIS/WebSailor), or null max_sample_batches: 30 # sample at most this many batches before stopping, -1 to sample forever min_replace_ratio: 0.3 # minimum proportion of good samples with which to replace bad samples (for replace strategy only) ``` This PR also adds a docs page describing how to enable all DAPO features, and adds an example GSM8K script where all these features are used. ## Minor Changes Some minor changes to make this dynamic sampling implementation clean: - the utils `Timer` class now updates the dict instead of overwriting in order to correctly track generation time w/ dynamic sampling, which means we need to make sure to reset `all_timings` in any trainer - The use of `self.weights_manager` is a little tricky for the dynamic sampling - introduced the the `ConditionalWeightsManager` to make the added code in the training loop as clean as possible ## Example runs <img width="413" height="264" alt="image" src="https://github.com/user-attachments/assets/072f716a-3632-42bb-a5f7-5f9d6064bd93" /> Generation time for dapo style filtering increases as the training run goes on, while it is stable for polaris and the baseline. <img width="419" height="265" alt="image" src="https://github.com/user-attachments/assets/887df550-e4b9-4623-b578-b4809a9f403f" /> We can see that the training pass @ n metric is 1 for both polaris and dapo style filtering as expected. <img width="421" height="259" alt="image" src="https://github.com/user-attachments/assets/bb63af77-1fbb-4d89-9216-b028f1551ea7" /> For GSM8k + Qwen 1.5B, the sampling strategy (as well as the full DAPO run) results in minimal gains - need larger models/harder dataset to test more fully DAPO sampling Example Run: ```bash (skyrl_entrypoint pid=222117) 2025-08-04 23:13:13.439 | INFO | skyrl_train.trainer:train:245 - Started: 'step' (skyrl_entrypoint pid=222117) 2025-08-04 23:13:13.737 | INFO | skyrl_train.weights_manager:__enter__:76 - Started: 'sync_weights_to_inference_engines' (skyrl_entrypoint pid=222117) 2025-08-04 23:13:16.401 | INFO | skyrl_train.weights_manager:__enter__:76 - Finished: 'sync_weights_to_inference_engines', time cost: 2.66s (skyrl_entrypoint pid=222117) 2025-08-04 23:13:16.401 | INFO | skyrl_train.weights_manager:__enter__:80 - Started: 'offload_policy_model_to_cpu' (skyrl_entrypoint pid=222117) 2025-08-04 23:13:16.842 | INFO | skyrl_train.weights_manager:__enter__:80 - Finished: 'offload_policy_model_to_cpu', time cost: 0.44s (skyrl_entrypoint pid=222117) 2025-08-04 23:13:16.888 | INFO | skyrl_train.trainer:train:261 - Started: 'generate' (AsyncVLLMInferenceEngine pid=223856) INFO 08-04 23:13:13 [executor_base.py:227] It took 0.243244 seconds to wake up tags ['weights']. [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=223854) INFO 08-04 23:13:16 [executor_base.py:227] It took 0.040547 seconds to wake up tags ['kv_cache']. (AsyncVLLMInferenceEngine pid=223856) INFO 08-04 23:13:16 [block_pool.py:316] Successfully reset prefix cache [repeated 7x across cluster] (AsyncVLLMInferenceEngine pid=223855) INFO 08-04 23:13:16 [executor_base.py:227] It took 0.041721 seconds to wake up tags ['kv_cache']. (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.378 | INFO | skyrl_train.trainer:train:261 - Finished: 'generate', time cost: 17.49s (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.395 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:433 - ============= Dynamic sampling filter ============= (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.395 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:434 - Dynamic sampling: 460 < 1024 prompts (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.395 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:435 - Resample batch 1, continue sampling... (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.395 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:436 - ================================================== (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.395 | INFO | skyrl_train.trainer:train:245 - Finished: 'step', time cost: 20.96s (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.407 | INFO | skyrl_train.trainer:train:245 - Started: 'step' (skyrl_entrypoint pid=222117) 2025-08-04 23:13:34.445 | INFO | skyrl_train.trainer:train:261 - Started: 'generate' (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.014 | INFO | skyrl_train.trainer:train:261 - Finished: 'generate', time cost: 17.57s (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.029 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:433 - ============= Dynamic sampling filter ============= (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.029 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:434 - Dynamic sampling: 941 < 1024 prompts (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.029 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:435 - Resample batch 2, continue sampling... (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.029 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:436 - ================================================== (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.030 | INFO | skyrl_train.trainer:train:245 - Finished: 'step', time cost: 17.62s (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.033 | INFO | skyrl_train.trainer:train:245 - Started: 'step' (skyrl_entrypoint pid=222117) 2025-08-04 23:13:52.074 | INFO | skyrl_train.trainer:train:261 - Started: 'generate' (skyrl_entrypoint pid=222117) 2025-08-04 23:14:08.380 | INFO | skyrl_train.trainer:train:261 - Finished: 'generate', time cost: 16.31s (skyrl_entrypoint pid=222117) 2025-08-04 23:14:08.396 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:439 - ============= Dynamic sampling filter ============= (skyrl_entrypoint pid=222117) 2025-08-04 23:14:08.396 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:440 - Dynamic sampling: collected 1467 >= 1024 prompts (skyrl_entrypoint pid=222117) 2025-08-04 23:14:08.397 | INFO | skyrl_train.utils.trainer_utils:handle_filter_sampling:443 - ================================================== (AsyncVLLMInferenceEngine pid=223856) INFO 08-04 23:13:12 [gpu_worker.py:98] Sleep mode freed 61.88 GiB memory, 4.98 GiB memory is still in use. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=223856) INFO 08-04 23:13:12 [executor_base.py:211] It took 1.264572 seconds to fall asleep. [repeated 3x across cluster] ``` Polaris Style example run: ```bash (skyrl_entrypoint pid=306764) 2025-08-05 00:30:01.648 | INFO | skyrl_train.trainer:train:261 - Started: 'generate' (AsyncVLLMInferenceEngine pid=308521) INFO 08-05 00:29:58 [executor_base.py:227] It took 0.240372 seconds to wake up tags ['weights']. [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=308520) INFO 08-05 00:30:01 [executor_base.py:227] It took 0.040980 seconds to wake up tags ['kv_cache']. (AsyncVLLMInferenceEngine pid=308521) INFO 08-05 00:30:00 [block_pool.py:316] Successfully reset prefix cache [repeated 7x across cluster] (AsyncVLLMInferenceEngine pid=308518) INFO 08-05 00:30:01 [executor_base.py:227] It took 0.041175 seconds to wake up tags ['kv_cache']. (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.663 | INFO | skyrl_train.trainer:train:261 - Finished: 'generate', time cost: 15.01s (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.679 | INFO | skyrl_train.utils.trainer_utils:handle_replace_sampling:316 - Replace sampling: 629 good UIDs out of 1024 total prompts (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.680 | INFO | skyrl_train.utils.trainer_utils:handle_replace_sampling:320 - ============= Dynamic sampling replace =========== (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.680 | INFO | skyrl_train.utils.trainer_utils:handle_replace_sampling:321 - Number of good prompts: 629 (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.680 | INFO | skyrl_train.utils.trainer_utils:handle_replace_sampling:322 - Number of bad prompts: 395 (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.694 | INFO | skyrl_train.utils.trainer_utils:handle_replace_sampling:352 - After replacement - Replaced 395 bad prompts (skyrl_entrypoint pid=306764) 2025-08-05 00:30:16.694 | INFO | skyrl_train.utils.trainer_utils:handle_replace_sampling:353 - ================================================== (AsyncVLLMInferenceEngine pid=308520) INFO 08-05 00:29:57 [gpu_worker.py:98] Sleep mode freed 62.14 GiB memory, 6.28 GiB memory is still in use. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=308520) INFO 08-05 00:29:57 [executor_base.py:211] It took 1.331663 seconds to fall asleep. ``` ## Full DAPO example run From example script <img width="417" height="262" alt="image" src="https://github.com/user-attachments/assets/2592a06f-8b8a-4cf1-a29e-321bff819eb0" /> <img width="909" height="325" alt="image" src="https://github.com/user-attachments/assets/50922afd-1424-4183-9329-4f1f340287eb" /> --------- Co-authored-by: Sumanth R Hegde <[email protected]> * [algorithm] Support Dr. GRPO + refactor where policy/critic loss functions are set (NovaSky-AI#133) # Overview ## Dr GRPO Adds `loss_reduction`: `seq_mean_token_sum_norm ` option, and `grpo_norm_by_std` option to support Dr. GRPO So to run Dr. GRPO, set: ```yaml trainer: algorithm: grpo_norm_by_std: false loss_reduction: "seq_mean_token_sum_norm" ... ``` Example run: <img width="906" height="317" alt="image" src="https://github.com/user-attachments/assets/ce9db2ef-253e-45c8-adba-1ef8a270bbd9" /> Reward looks similar <img width="419" height="263" alt="image" src="https://github.com/user-attachments/assets/a4bc4d8c-f3c1-4bad-a497-0297dc30bc27" /> Magnitude of policy loss is lower as expected (since we are normalizing by a larger constant rather than taking the mean) ## Refactor where Critic/Policy Loss are set Changes ppo critic `ValueLoss` to just a function instead of a `nn.Module` for consistency with `policy_loss`, and adds new algorithm field to cfg that require evaluating field values in `utils::validate_cfg` (this runs before entrypoint code, allowing users to modify the cfg further by subclassing `BasePPOExp`) PPO example still running after this refactor: <img width="421" height="262" alt="image" src="https://github.com/user-attachments/assets/88985da3-1403-49c6-8cb5-f1434151fd9e" /> * [fix] move algorithm folder -> algorithms (NovaSky-AI#136) left the algorithm folder in NovaSky-AI#133, move it over * [Logging] Forward mlflow env vars to ray runtime env (NovaSky-AI#135) This PR forward the `MLFLOW_TRACKING_URI` and `MLFLOW_TRACKING_TOKEN` environment variable to the ray runtime env during its initialization. This will enable users to simply provide the above env vars at the driver and be able to use MLFlow for experiment tracking. * data folder * some stuff * updates --------- Signed-off-by: SumanthRH <[email protected]> Co-authored-by: Sumanth R Hegde <[email protected]> Co-authored-by: Eric Tang <[email protected]> Co-authored-by: Tyler Griggs <[email protected]> Co-authored-by: Shu Liu <[email protected]> Co-authored-by: Ben Cohen <[email protected]> Co-authored-by: Shiyi Cao <[email protected]> Co-authored-by: Dacheng Li <[email protected]> Co-authored-by: Etienne Brodu <[email protected]>
2 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Overview
This PR introduces filter (DAPO) and replace (Polaris/WebSailor) style dynamic sampling strategies. The dynamic sampling strategy can be configured as below:
This PR also adds a docs page describing how to enable all DAPO features, and adds an example GSM8K script where all these features are used.
Minor Changes
Some minor changes to make this dynamic sampling implementation clean:
Timerclass now updates the dict instead of overwriting in order to correctly track generation time w/ dynamic sampling, which means we need to make sure to resetall_timingsin any trainerself.weights_manageris a little tricky for the dynamic sampling - introduced the theConditionalWeightsManagerto make the added code in the training loop as clean as possibleExample runs
Generation time for dapo style filtering increases as the training run goes on, while it is stable for polaris and the baseline.
We can see that the training pass @ n metric is 1 for both polaris and dapo style filtering as expected.
For GSM8k + Qwen 1.5B, the sampling strategy (as well as the full DAPO run) results in minimal gains - need larger models/harder dataset to test more fully
DAPO sampling Example Run:
Polaris Style example run:
Full DAPO example run
From example script
