Yang, Bo

Avoid monkey patching vLLM

1

Currently, vLLM's `vllm.worker.worker.Worker` is replaced with `openrlhf.trainer.ray.vllm_worker_wrap.WorkerWrap` on fly as a monkey patch. The monkey patch is avoidable by making `init_process_group` and `update_weight` global functions and invoke them via `__ray_call__`....

Search drivers on NixOS and WSL

3

NixOS usually links drivers to `/run/opengl-driver/lib`, and WSL mounts drivers under `/usr/lib/wsl/lib`. Let's also search these paths.

These files are required to run CUDA on WSL 2, mounted automatically at /usr/lib/wsl/lib

[REQUEST] Supporting custom generation loop (outlines, LMQL, guidance) in DeepSpeedHybridEngine

2

**Is your feature request related to a problem? Please describe.** Currently, `DeepSpeedHybridEngine.generate` internally calls `self._generate` to run inference loop, which is intialized as the module's `generate` method (usually it's `transformers.PreTrainedModel.generate`):...

AutoML does not pass proper `objective` to `estimator_class` when `metric` is non-default.

3

Currently, neither `search_space` nor `get_params` in `LGBMEstimator` passes `objective` to the `params`. https://github.com/microsoft/FLAML/blob/a68d073ccf708005c244b3a91f51a5c30fbe6604/flaml/automl/model.py#L1266-L1309 As a result, the objective of the underlying model is always the default value `reg:squarederror`, even when...

# What this PR does / why we need it: Implement offline_write_batch for postgresql offline store. # Which issue(s) this PR fixes: Fixes https://github.com/feast-dev/feast/issues/3759 # Misc This PR is based...

Disable `populateBin` when `envfs` is enabled

feat: add `devenv.flakeArgs` to avoid `--no-pure-eval`

5

This PR avoids `--no-pure-eval` in `devenv-devShell.nix`, in favor of `"--override-input" "devenv-root" "path:.devenv/state/pwd"`.