Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

[REQUEST] XPU: register native pinned host memory with the device runtime (follow-up to #8283)

Open
  • Is something wrong?
#8,315 1 comment 2 reactions 1 assignee View on GitHub

Maintainers usually reply within 3 days

@delock is already working on this.

Since Aug 26, 2026.

  • #8322 by @delock — closed without merging

Assessment

This issue has not been assessed yet.

Description

Is your feature request related to a problem? Please describe.

#8283 taught the native pinned-memory allocator (DS_PIN_MEMORY_BACKEND=native, posix_memalign + mlock) to register host memory with the device runtime so that copy_(..., non_blocking=True) can go through the DMA engine. On CUDA this closed a large bandwidth gap (H200: ~10-17 GB/s with mlock only vs. ~53-55 GB/s after cudaHostRegister).

Registration goes through the new register_host_memory / unregister_host_memory accelerator hooks, but only CUDA_Accelerator overrides them. XPU_Accelerator inherits the base-class no-op, so on Intel GPUs native pinned buffers stay mlock-only and H2D/D2H transfers run at pageable bandwidth even with DS_PIN_MEMORY_REGISTER_DEVICE=1. The degradation is silent: registration just returns False and nothing logs.

Describe the solution you'd like

  • Implement register_host_memory / unregister_host_memory in accelerator/xpu_accelerator.py, e.g. via Level Zero zeMemRegisterHostMemory / zeMemUnregisterMemory on the allocation range, honoring the same DS_PIN_MEMORY_REGISTER_DEVICE opt-out and the graceful "log once and continue with mlock" fallback on failure.
  • Extend tests/unit/v1/pin_memory/test_pin_memory.py with an XPU equivalent of the CUDA cudart-mock test.
  • Optionally extend benchmarks/pin_memory/h2d_d2h_bench.py so the before/after H2D/D2H bandwidth can be quantified on XPU.

Describe alternatives you've considered

  • Keep mlock-only on XPU (status quo): functionally correct, but leaves H2D/D2H at pageable speeds for native-pin users on Intel GPUs.

Additional context

  • Hooks added in #8283: no-op defaults in accelerator/abstract_accelerator.py, CUDA implementation in accelerator/cuda_accelerator.py (torch.cuda.cudart().cudaHostRegister / cudaHostUnregister).
  • Call sites: NativePinnedMemory.pin / _release in deepspeed/utils/pin_memory.py.
Dominant language
Python
Stars
43.3k
Forks
5.1k
Avg merge
4d 20h
Merged PRs (30d)
69

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from deepspeedai/DeepSpeed

All issues in deepspeedai/DeepSpeed

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.