$ the-wire · showcase
Triton conv routing lands in ExecuTorch, CUDA 13.5 enters the nightly matrix
By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology
ExecuTorch's CUDA path gained a Triton convolution backend end to end, PyTorch's nightly validation now covers cu135 wheels on Linux, and two torchrl bug fixes correct masks and sampler device selection in shipped wheels.
[executorch][cuda] Route 2-D depthwise and few-channel grouped convolutions to a Triton direct-stencil kernel pytorch/executorch
2-D depthwise and few-channel grouped convolutions on CUDA now route to a Triton direct-stencil kernel, the last piece of a four-PR stack that also adds triton::convolution and the routing framework, a planned-replacement pass, and a deprecated triton_kernel_mode=OFF path. The ExecuTorch CUDA export path is moving off the previous convolution lowering, so kernel-mode settings that relied on OFF...
Add CUDA 13.5 to the nightly build and validation matrix (Linux) pytorch/test-infra
CUDA 13.5 joins the nightly build and validation matrix for Linux x86_64 and aarch64. pytorch/pytorch already publishes cu135 nightlies up to 2.16.0.dev20261116, and until now nothing in the validation matrix exercised them; the s3_management targets already existed.
[BugFix] Keep CPU priority trees for CUDA storage when the CUDA trees are not built pytorch/rl
PrioritizedSampler used to select a CUDA segment tree for any CUDA storage and then raise when the _torchrl extension had no CUDA trees built, which is the case for the PyPI wheels on x86_64 and aarch64. _tree_device_from_storage now returns the CUDA device only when CudaSumSegmentTreeFp32 is available and keeps the CPU trees otherwise, matching the device docstring.
[BugFix] Preserve action masks when expanding categorical specs pytorch/rl
Categorical.expand discarded its mask, letting forbidden random actions through, while MultiCategorical.expand added a spurious mask dimension and raised on valid input. Both now pass the existing mask to constructors that broadcast it, and the outcome-mask preservation is documented.
Prevent external constant name collisions across methods (#24679) pytorch/executorch
Methods in one ExecuTorch program can independently create identically named CONSTANT_TENSOR inputs, and externalizing them into a shared weights file could let one method's name-to-buffer entry overwrite another's data; matching dtype and shape let the program load and compute silently wrong outputs. The emitter now compares same-named external constants by raw bytes, dtype, shape, and dimensi...