Krishna KalyanandMatthias Jouanneaux 105c479648 [docs] Document CUDA-to-CPU transfers on NVLink-C2C systems (#199737) ...

On NVLink-C2C systems (Grace Hopper, Grace Blackwell), CPU↔GPU copies already use C2C automatically. The bottleneck is the CPU destination: `tensor.cpu()` allocates fresh pageable memory, so the first copy is limited by page allocation and faults rather than link bandwidth. Activation offloading that calls `.cpu()` every step can pay that cost every time.

This PR documents the fix (allocate pinned CPU buffers once and reuse them), the synchronization rules, and the caveat that pinned memory can still fault under OS memory compaction. It also links the new section from `Tensor.cpu()`.

[Slack Thread](https://pytorch.slack.com/archives/C3PDTEV8E/p1790188671754669?thread_ts=1788293519.083089&cid=C3PDTEV8E)

Pull Request resolved: https://github.com/pytorch/pytorch/pull/199737
Approved by: https://github.com/MatthiasKohl, https://github.com/albanD

Co-authored-by: Matthias Jouanneaux <matthias.kohl@hotmail.de>

2026-10-27 12:53:49 +00:00

..

amp_examples.md

[Docathon] Convert notes/amp_examples.rst from rST to MyST Markdown (#182575)

2026-05-07 17:34:44 +00:00

autograd.md

docs: render inline math on the autograd mechanics page (#192562)

2026-08-11 14:22:35 +00:00

broadcasting.md

[Docathon] Convert broadcasting.rst to MyST Markdown [#182482] (#182807)

2026-05-08 19:05:30 +00:00

cpu_threading_torchscript_inference.md

Convert stub docs pages to MyST Markdown (#183498)

2026-05-13 03:41:29 +00:00

cuda.md

[docs] Document CUDA-to-CPU transfers on NVLink-C2C systems (#199737)

2026-10-27 12:53:49 +00:00

custom_operators.md

Convert stub docs pages to MyST Markdown (#183498)

2026-05-13 03:41:29 +00:00

ddp.md

Convert DDP note to MyST Markdown (#183936)

2026-05-16 19:09:42 +00:00

extending.func.md

[Docathon 2026] Convert notes/extending.func.rst to MyST Markdown (#182608)

2026-05-06 18:44:49 +00:00

extending.md

Expose InputBuffer as ctx.input_grad_buffers in autograd.Function backward (#196491)

2026-09-16 21:47:23 +00:00

faq.md

Fix typo in docs/source/notes/faq.md (#193361)

2026-08-14 13:25:17 +00:00

get_start_xpu.md

[docs][xpu] clarify Intel GPU prerequisites wording in get_start_xpu (#198820)

2026-09-29 08:53:59 +00:00

gradcheck.md

[Docathon 2026] Convert notes/gradcheck.rst to MyST Markdown (#182557)

2026-05-06 04:30:44 +00:00

hip.rst

[ROCm] Always eagerly allocate hipBLAS(Lt) workspaces (#196920)

2026-10-01 14:33:19 +00:00

large_scale_deployments.md

docs: convert notes/large_scale_deployments to MyST (#182553)

2026-05-06 02:40:25 +00:00

libtorch_stable_abi.md

[Stable-ABI] Add torch::stable::add. (#196762)

2026-10-19 17:38:12 +00:00

local_tensor_tutorial.md

[docs] Convert notes/local_tensor_tutorial.rst from rST to MyST Markdown (#182787)

2026-05-07 17:51:45 +00:00

mkldnn.md

docs: fix PyTorch brand name consistency (#189248)

2026-07-08 18:46:57 +00:00

modules.md

Convert notes/modules.rst to .md (#183351)

2026-05-12 16:26:44 +00:00

mps.md

[MPS] Require macOS 15.0+ and report missing GPU separately (#200062)

2026-10-08 04:19:08 +00:00

multiprocessing.md

Fix wording and typos (#191862)

2026-08-03 15:24:36 +00:00

numerical_accuracy.md

Add BF16x9 precision mode for CUDA FP32 matmul (#195301)

2026-09-02 05:59:39 +00:00

out.md

[Docathon] Convert notes/out.rst from rST to MyST Markdown (#182775)

2026-05-07 19:32:19 +00:00

randomness.md

[Docathon] Convert notes/randomness.rst from rST to MyST Markdown (#183793)

2026-05-14 22:56:48 +00:00

serialization.md

[Docathon] Convert notes/serialization.rst from rST to MyST Markdown (#182905)

2026-05-08 16:15:24 +00:00

tensor_iterator.md

[TensorIterator] Expose the build pipeline to Python (kwargs API) (#184603)

2026-05-22 11:30:48 +00:00

windows.md

Convert docs/source/notes/windows.rst to MyST Markdown (#182538)

2026-05-06 18:14:06 +00:00