$ the-wire · showcase
DTensor any/all stops all-gathering, Helion fixes silent wrong results
By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology
The day's real work is correctness: DTensor changes the placement contract for any/all, Helion lands 26 compiler fixes for silent wrong results, and PyTorch's sparse CSR path gets a wrong-value and out-of-bounds fix.
[DTensor] Reduce sharded any/all locally with safe partials (#204895) pytorch/pytorch
The reduce now runs locally on shards with safe partials and drops the all-gather before any/all, so the output placement changes from R to P. The PR flags this as potentially BC-breaking and points to the behavior before the single-dim migration.
[compiler] Fix in-place updates, dead-code elimination, barriers and scans (#4453) pytorch/helion
Thirteen fixes to device-code tracing and lowering, each with a test, and most address silent wrong results: an in-place method in a loop or branch was dropped, a host statement with effects was eliminated as dead, and a load after an hl.inline_triton snippet or atomic ran without a barrier.
[compiler] Mask padded hl.arange lanes and partial tiles; fix config slots (#4454) pytorch/helion
Seven of these thirteen fixes are masking bugs where lanes past the end of a padded hl.arange or a partial tile took part in means, reductions, loop carries, loads and stores; the other six correct config slot, block size, thread query and fact assignment.
[sparse] Fix wrong CSR results for strided COO indices (#202122) (#202883) pytorch/pytorch
Converting a coalesced sparse COO tensor whose indices are a strided view (such as a transposed (nnz, 2) coordinate list) to CSR or BSR returned wrong values and made the sparse product wrong too. On CPU the same read could also run past the end of the output, and the linked issue reports crashes from it.
[rl] Run Verifiers groups in rollout workers pytorch/torchtitan
Verifiers groups in torchtitan's RL path now run in rollout worker processes like native groups, with each worker running its own GenerationServer so the controller receives finished groups instead of serving every model call and decoding every trace. The rollouts are reported unchanged: on CPU, every training field of 16 rollouts matched.
[scripts] Ask nvidia-smi again when it fails instead of reporting no GPUs (#4448) pytorch/helion
Elsewhere: gpu_board.py treated any failed nvidia-smi call as a machine without GPUs, and retries now happen because under heavy CPU load calls ran 11 to 20 seconds and one blew the timeout, reporting eight GPUs as none.