FineEnvs/HF_ML_Tasksmith
HF ML Tasksmith: Harbor dataset on Hugging Face with 50 tasks. Fifty PR-derived Harbor tasks from Accelerate, Diffusers, PEFT, Transformers and TRL, including CPU and GPU tasks. Contains 50 Harbor tasks generated with the owned tasksmith recipe in Repo2RLEnv. Browse the complete task bundles in…
Tasks
- Add support for periodic reference model synchronization (TR-DPO style) to KTOConfig and KTOTrainer , matching the existing DPO…
- When using LoRA with target parameters , it must now be possible to add multiple adapters to the same model, subject to one constraint…
- Refactor environment and per-rollout tool setup in GRPOTrainer , DPPOTrainer , GRPOWithReplayBufferTrainer , and AsyncRolloutLoop so…
- In DistillationTrainer.log() , add support for logging completion tables to the trackio backend alongside the existing wandb support.
- Update trl/chat template utils.py to support the new response-parsing API in transformers 5.13+, while keeping backward compatibility with…
- Add a memory-efficient "chunked nll" loss type to SFTConfig and SFTTrainer .
- Add ZImageTransformer2DModel and ZImagePipeline to the diffusers library.
- Fix three bugs in OnlineDPOTrainer. generate vllm server so that it produces the same block-layout batch format as the colocate and…
- Add the Helium model (a lightweight causal language model by Kyutai) to the transformers library.
- When set peft model state dict is called on a plain model (one returned by inject adapter in model , without a PeftModel wrapper), it…
- Add a data seed option that lets users control the random seed used by SeedableRandomSampler independently of the global PyTorch seed.
- Improve Accelerate's FSDP2 preparation for language models whose embeddings and final normalization currently share an unnecessarily large…
- The concatenate function in accelerate.utils.operations recursively concatenates tensors held in nested lists, tuples, or dicts. Fix it so…
- Add a Mistral3 multimodal vision-language model to the library. The implementation should include:
- Correct reward-margin reporting in the experimental KTO trainer ( trl/experimental/kto/kto trainer.py ). A logged margin should summarize…
- Add a new SanaSprintImg2ImgPipeline class to the diffusers library as an image-to-image variant of the existing SanaSprintPipeline .
- Add a memory-efficient chunked LM-head log-probability utility to trl/trainer/utils.py and export it from trl/trainer .
- Add a dynamo plugin parameter to Accelerator. init that accepts a TorchDynamoPlugin instance (or None ).
- Add an experimental Geometric-Mean Policy Optimization trainer, available as GMPOConfig and GMPOTrainer from trl.experimental.gmpo . It…
- Add a new MetaCLIP 2 model family to the transformers library and update CLIPProcessor to accept any tokenizer.
- Fix a regression in FullyShardedDataParallelPlugin.set auto wrap policy : transformer-based auto-wrapping can attempt to resolve…
- Add a public utility function get grad scaler(distributed type=None, kwargs) in accelerate/utils/modeling.py and export it from…
- Add a Dinov2WithRegisters model family to the transformers library. This is a variant of DINOv2 that inserts additional "register" tokens…
- Fix notebook launcher so that on PyTorch builds that include torch.numa , the parent process does not initialize the CUDA driver before…
- Add two new keyword parameters to set module tensor to device in src/accelerate/utils/modeling.py :
- When using LoRA's target parameters feature via get peft model , attempting to target an nn.Parameter registered directly on the top-level…
- Add LoRA loading support for Ideogram4Pipeline via a new Ideogram4LoraLoaderMixin and a state-dict conversion utility for non-diffusers…
- The SDXL-derived pipelines have an upcast vae() compatibility helper whose selective casting can leave parts of a VAE in the original…
- Add a DoraCaching helper to peft.helpers that caches two expensive intermediate results computed by DoRA modules during inference: the…
- Add the PVeRA (Probabilistic Vector-based Random Matrix Adaptation) adapter to the PEFT library, exposed as PveraConfig and PveraModel in…
- Add FluxKontextPipeline to the diffusers library as an image-conditioned Flux denoising pipeline for in-context image editing (style…
- Fix dataset preprocessing cache reuse in DPOTrainer and SFTTrainer . Reprocessing the same raw dataset with the same tokenization settings…
- Add LoRA support for the tensor-parallel models supplied by the installed Transformers integration.
- Remove the following deprecated public APIs from the accelerate library. After the change, calling any of these must fail hard…
- Add a get device mesh(device type=None) method to ParallelismConfig that lazily builds and caches the device mesh on the instance.
- Fix an intermittent hang during GRPO generation with a PEFT/LoRA model and colocated vLLM tensor parallelism on multiple GPUs without…
- Add a get cosine scaled reward factory function to trl/rewards and export it from trl.rewards .
- Add Generalized Low Rank Adaptation (GLoRA) to PEFT for ordinary torch.nn.Linear layers. Expose GloraConfig, GloraModel and PeftType.GLORA…
- Add two new public utilities to accelerate.utils — compile regions and has compiled regions — and extend TorchDynamoPlugin with a use…
- Extend ModularPipeline.save pretrained() to save component model weights in addition to the pipeline configuration. Previously, only the…
- Add a BlockRefinementScheduler , a LLaDA2Pipeline , and a compute confidence aware loss utility to the diffusers library for discrete…
- Add a new AyaVision vision-language model family to the transformers library. The implementation must expose the following public APIs…
- When inject adapter in model(config, model, state dict=sd) is called and either (a) the state dict keys carry a orig mod. prefix (e.g…
- Update the batch-size reduction factor in find executable batch size (in src/accelerate/utils/memory.py ).
- Add two public functions to PEFT that convert a non-LoRA PEFT adapter into an equivalent LoRA adapter using truncated SVD on each layer's…
- Implement Krea2Transformer2DModel and Krea2Pipeline for text-to-image generation with the Krea 2 architecture.
- Extend BatchSamplerShard (in src/accelerate/data loader.py ) to support variable-length (dynamic) batch sizes — samplers where batch size…
- Add LoRA loading/saving support to HiDreamImagePipeline and add a force inference output option to HiDreamImageTransformer2DModel .
- Support loading LoRA adapters trained with Transformers v4 onto models whose architecture changed in Transformers v5, including Mixtral…
- Fix an intermittent hang in TRL's colocated vLLM generation when a PEFT model uses tensor parallelism across multiple GPUs. The…