$ the-wire · showcase
TorchRL 0.15 ships dataset exports and TD-MPC2; ROCm drops a 2021 FFT workaround
By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology
TorchRL 0.15 adds ReplayBuffer.export with Hugging Face and LeRobot adapters alongside TD-MPC2 and Flow Q-Learning objectives, while PyTorch removes the ROCm clone_input override that has forced a full contiguous copy on every out-of-place FFT since 2021.
TorchRL 0.15: Dataset Exports, TD-MPC2, Robot Deployment, and Flow Q-Learning pytorch/rl
TorchRL v0.15.0 adds ReplayBuffer.export with Hugging Face and LeRobot adapters, export_onnx_policy, MenagerieEnv, LowLevelController, TransformerModule, RewardModelLoss and StopOnSignal, plus TD-MPC2, Flow Q-Learning and Transitive RL objectives. Offline RL datasets and policies now have a direct path out of the buffer without a custom serializer.
[ROCm] Remove the ROCm clone_input override in CuFFTPlanCache (#199834) pytorch/pytorch
PyTorch forced clone_input = true on ROCm since 2021 to work around hipFFT clobbering the input buffer of out-of-place transforms; that override is gone. The copy accounted for a full contiguous clone of the signal on every forward FFT, and because clone_input also selected the plan layout it pinned ROCm to cufft_simple_embed instead of as_cufft_embed, materializing strided inputs even where cu...
Allow multinomial with >2**24 entries (#204153) pytorch/pytorch
multinomial_out used to reject more than 2**24 categories on every device and code path, justified by a stale comment claiming the index tensor was float. The no-replacement and n_sample == 1 paths compute argmax/topk of p / Exp(1) and return correct int64 indices past 2**24 on CPU and CUDA, so the cap is lifted.
Qualcomm AI Engine Direct - Exact integer floor_divide and remainder pytorch/executorch
Integer // and % on Qualcomm HTP went through float div plus floor, which is inexact: 23999 // 80 returns 300 in fp16, which the floor_divide test carried as a rounding TODO. Integer remainder now lowers to QNN ElementWiseMod and integer floor division to int32 (x - x mod y) / y for both floor_divide and div(rounding_mode="floor"); DecomposeFloorDivide is removed, float paths are unchanged, and...
[rl] Periodic validation beside training (#5228) pytorch/torchtitan
Validation in the TorchTitan RL path becomes its own actor so generation runs beside training instead of blocking it. The PR's stated scope covers only this: gather signal without a separate checkpoint-loading job, since validation on start and end currently holds up training until it finishes.
Provide the ESP-NN delegate and esp_nn:: op library in the ESP-IDF component pytorch/executorch
The rest is plumbing: the ESP-IDF component now vendors the ESP-NN delegate runtime and esp_nn:: op library as a fourth kernel library for gen_kernels.py, can build a kernel bundle at configure time when EXECUTORCH_PYTHON imports executorch, and test-infra stopped cutting nightly branches for the eight meta-pytorch domains that now build from main on their own 11:30 UTC schedule.