Skip to content
Problems.

PyTorch GPU access and drivers are painful

MLOps3 posts from 3 people3d active+00 posts in the last 7 days, 0 the 7 days before (fading)

Posts per day

Posts per day3 posts, Aug 28 to Sep 25

The posts behind it

3, newest first
PostDate
[PT2] Device-TMA reduction stores lose tma_min_block_sizes...the bug ## Describe the bug `store_reduction()` bypasses `codegen_block_ptr()` and fails to record an in-kernel device-TMA descriptor. The resulting kernel can select XBLOCK=1 for an fp32 descriptor store, which Triton rejects because 1 * 4 < 16 bytes. ## Minimal reproducer...pytorch/pytorchhdhaiSep 102 weeks ago
[FSDP2][NCCL] Large slowdown for non-16-byte-aligned per-rank collective payloads### 🐛 Describe the bug ## Summary Using PyTorch's NCCL tensor collectives on a flat buffer with the same per-rank layout as FSDP2's default non-symmetric-memory communication path, a per-rank payload whose byte count is not a multiple of 16 causes a substantial latency...pytorch/pytorch0x777a6cAug 313 weeks ago
[MPS] bmm/baddbmm/addbmm silently wrong on M1/M2 when K >= 32767 (2-D paths already routed by #183535)### 🐛 Describe the bug > This report was prepared with the help of an AI assistant; everything in it was verified by me on an M2 Max running macOS 26.6.2. On Apple7/8 GPUs (M1/M2), MPSGraph matmuls with a reduction dimension K >= 32767 and both output dims >= 16 can sum memory...pytorch/pytorchscott-kirilaAug 284 weeks ago

Companies and products named

Company or productPosts naming it
PyTorch3
NVIDIA3
GitHub2
Docker1
About this problem

Evidence

3 posts from 3 people in 1 place, about 2 a week over 14 days. Mostly on pytorch/pytorch. Tools named alongside: NVIDIA, GitHub, Docker.

Frustration Frustration 0 of 3· Seen on GitHub issues

How it was grouped

Posts that state a pain and match the "gpu-access" rule about PyTorch. First post Aug 28, 2026, latest Sep 10, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method

History

  • 2026-09-25 Added: PyTorch GPU access and drivers are painful (3 posts)

Rising problems by email

Mondays: the problems in data, tech and AI that grew fastest that week.

Double opt-in. Unsubscribe any time.