Zero-copy CPU import: mx.array(host_buffer, copy=False) on unified memory#3872
Open
HaoXuAI wants to merge 7 commits into
Open
Zero-copy CPU import: mx.array(host_buffer, copy=False) on unified memory#3872HaoXuAI wants to merge 7 commits into
HaoXuAI wants to merge 7 commits into
Conversation
Expose a copy kwarg on the array constructor and adopt page-aligned CPU host buffers on unified memory instead of copying. In the CPU import branch, when copy=False, wrap the host pointer via the Metal allocator (newBufferWithBytesNoCopy) rather than throwing; fall back to an explicit error when Metal is unavailable, the buffer is not page-aligned, or a dtype conversion is required. A deleter keeps the source buffer alive for the adopted array's lifetime. Enables true zero-copy Arrow/NumPy -> MLX on Apple Silicon.
MLX maps some numpy dtypes to a different-width mlx dtype (e.g. float64 -> float32), so the dst==src check does not catch a reinterpret mismatch. Add an itemsize == size_of(dst_dtype) guard (mirroring the Metal path) before adopting; otherwise copy=False raises. Verified: all zero-copy tests pass on a local Metal build (M4 Max).
HaoXuAI
marked this pull request as ready for review
July 19, 2026 18:37
Empirically newBufferWithBytesNoCopy adopts non-page-aligned host buffers correctly on current macOS, so the 'requires page alignment' comment was overstated and the unaligned->raises test premise does not hold. Keep the null-buffer check as a portable defensive fallback; generalize its message. Verified: remaining 4 tests pass on a local Metal build (M4 Max).
The zero-copy CPU import deleter released the adopted buffer with allocator::free(), which recycles it into the allocator's buffer cache. A make_buffer() buffer wraps caller-owned external memory, so recycling it hands that memory to an unrelated array on a later allocation -> corruption/crash (surfaces after several iterations, e.g. a training loop). Use allocator::release() as documented for make_buffer() buffers. Adds a loop regression test that adopts fresh buffers and computes each iteration.
angeloskath
approved these changes
Jul 21, 2026
angeloskath
left a comment
angeloskath
left a comment
Member
There was a problem hiding this comment.
That's great!
I refactored it a bit, the main point being to remove the copy argument from mx.array following also the discussion in #3531. We can use the no-copy cast via mx.asarray or mx.from_dlpack and mx.array always copies. If there is need we can add the copy in the array constructor later.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
When I tried loading Parquet/Arrow columns into MLX on Apple Silicon:
mx.array(host_buffer)always copies the host bytes into MLX-owned memory. But on unified memory the CPU and GPU share the same DRAM, so that copy is avoidable — the GPU can address the existing bytes. Pure overhead for large resident columns (feature tables, embeddings).What
A keyword-only
copyon thearrayconstructor.copy=Falseadopts a CPU host buffer (via MetalnewBufferWithBytesNoCopy) instead of copying it.(M4 Max. Default is
copy=None— unchanged behavior.)How
In the CPU branch of
nd_array_to_mlx, wrap the host pointer withmake_bufferandset_data, using a deleter that keeps the source alive and releases the wrapper viaallocator::release. It raises (rather than silently copying) when Metal is unavailable, the source element size != dtype size, or the buffer can't be adopted.Validation
All tests pass on a local Metal build (M4 Max): values, zero-copy sharing, dtype-mismatch raise, source-lifetime, and adopt-in-a-loop. Fork Linux CI green (lint / Fedora / ASAN / UBSAN); the macOS/Metal matrix only runs on the upstream repo.
Review questions