mps-sdpa 0.1.0

Fast drop-in scaled_dot_product_attention for Apple silicon (MPS) — wraps Apple's native MPSGraph SDPA with a zero-copy C++ bridge. 5-7x inference, 2-2.5x training.

mps-sdpa

Fast drop-in scaled_dot_product_attention for Apple silicon (MPS). Wraps Apple's native MPSGraph.scaledDotProductAttention op with a zero-copy C++ / Objective-C++ bridge. 5–7× inference speedup, 2–2.5× training speedup, 16–170× less driver memory per call — with identical math and checkpoint compatibility.

from mps_sdpa import sdpa_opt

# Drop-in replacement — same signature as torch.nn.functional.scaled_dot_product_attention
out = sdpa_opt(query, key, value,
               attn_mask=None, dropout_p=0.0, is_causal=False, scale=None,
               backend="auto")

---

Why

PyTorch's MPS backend dispatches torch.nn.functional.scaled_dot_product_attention to sdpa_general_mps, which builds a naive matmul → softmax → matmul graph. It does not call Apple's dedicated MPSGraph.scaledDotProductAttention op, which is present on macOS 15+ and is significantly faster.

mps-sdpa wraps that op directly, with an autograd.Function for training, shape-threshold auto-calibration, graceful fallbacks, and thread-safe graph caching. The C++ extension uses ATen's getMTLBufferStorage to hand torch tensors to MPSGraph without CPU memcpy.

Measured performance

All numbers on M4 / macOS 26.4.1 / torch 2.13 nightly, bfloat16.

Inference (forward only, B=1, H=8, D=64)

L	stock	mps-sdpa	speedup
1024	5.78 ms	0.90 ms	6.42×
2048	19.0 ms	3.82 ms	4.97×
4096	76.1 ms	11.79 ms	6.45×
8192	317 ms	44.3 ms	7.17×

Weighted geomean across a realistic audio-model shape suite: 4.88×.

Training (forward + backward, same shapes)

L	stock	mps-sdpa	speedup
1024	9.93 ms	5.06 ms	1.96×
2048	38.6 ms	17.1 ms	2.25×
4096	154 ms	64.8 ms	2.38×
8192	608 ms	247 ms	2.46×

Training with dropout (dropout_p=0.1)

L	stock	mps-sdpa	speedup
1024	14.19 ms	7.63 ms	1.86×
2048	55.83 ms	28.49 ms	1.96×
4096	228 ms	101 ms	2.26×

Driver memory per call

Apple's fused op doesn't materialize the [Lq, Lkv] attention matrix. The zero-copy C++ bridge removes the CPU-side intermediate buffer too.

L	stock	mps-sdpa	reduction
2048	1024 MB	<1 MB	≫128×
4096	1024 MB	<1 MB	≫64×
8192	1024 MB	32 MB	32×

Install

pip install mps-sdpa

Requires macOS 15+ on Apple silicon (M1–M4) with PyTorch ≥ 2.11. ninja is pulled in automatically for the zero-copy backend's JIT compile (~6s on first import, then cached to ~/.cache/torch_extensions).

Development install from source:

git clone https://github.com/crlandsc/mps-sdpa.git
cd mps-sdpa
pip install -e ".[dev]"
pytest tests/

Usage

The API exactly mirrors torch.nn.functional.scaled_dot_product_attention:

import torch
from mps_sdpa import sdpa_opt

q = torch.randn(1, 8, 2048, 64, dtype=torch.bfloat16, device="mps")
k = torch.randn(1, 8, 2048, 64, dtype=torch.bfloat16, device="mps")
v = torch.randn(1, 8, 2048, 64, dtype=torch.bfloat16, device="mps")

out = sdpa_opt(q, k, v)                                  # basic
out = sdpa_opt(q, k, v, is_causal=True)                  # causal mask
out = sdpa_opt(q, k, v, attn_mask=bool_mask)             # bool mask
out = sdpa_opt(q, k, v, dropout_p=0.1)                   # training dropout
out = sdpa_opt(q, k, v, backend="mpsgraph_zc")           # force specific backend

As a drop-in swap inside an existing model:

import torch.nn as nn
import torch.nn.functional as F
from mps_sdpa import sdpa_opt

class MyAttention(nn.Module):
    def __init__(self, d_model, n_heads):
        super().__init__()
        self.d_model = d_model
        self.n_heads = n_heads
        self.qkv = nn.Linear(d_model, 3 * d_model, bias=False)
        self.out = nn.Linear(d_model, d_model, bias=False)
        self.dropout_p = 0.1

    def forward(self, x):
        B, L, _ = x.shape
        qkv = self.qkv(x).reshape(B, L, 3, self.n_heads, -1).permute(2, 0, 3, 1, 4)
        q, k, v = qkv.unbind(0)
        p = self.dropout_p if self.training else 0.0  # standard pattern
        y = sdpa_opt(q, k, v, dropout_p=p)            # <— was F.scaled_dot_product_attention
        y = y.transpose(1, 2).reshape(B, L, self.d_model)
        return self.out(y)

Gate the behavior at model-construction time (useful for A/B comparison):

class Attn(nn.Module):
    def __init__(self, ..., use_mps_sdpa_opt: bool = False):
        self.use_opt = use_mps_sdpa_opt

    def forward(self, q, k, v):
        if self.use_opt:
            return sdpa_opt(q, k, v, ...)
        return F.scaled_dot_product_attention(q, k, v, ...)

Checkpoints are fully interchangeable — use_opt changes behavior only, not parameters.

CLI

mps-sdpa self-test                      # quick validation (<1s)
mps-sdpa list-backends                  # show available backends
mps-sdpa correctness --backend mpsgraph_zc --device mps --suite realistic
mps-sdpa benchmark --backend mpsgraph_zc --baseline stock --device mps --suite realistic

Backends

backend="auto" picks the best available. Available backends (in preference order):

Name	Impl	Best for
mpsgraph_zc	Obj-C++ torch extension, getMTLBufferStorage zero-copy	Default — forward, training, masks, dropout
mpsgraph	pyobjc + CPU-copy bridge	Fallback when the ext can't build
stock	torch.nn.functional.scaled_dot_product_attention	Always available final fallback
metal_proto	naive single-thread Metal kernel via torch.mps.compile_shader	Reference / experimentation, not auto-selected

Fallbacks are transparent — a call that can't go through mpsgraph_zc (e.g. short seq, CPU device, unsupported dtype) routes down the list automatically.

Correctness contract

Same tolerance bar as CUDA flash-attention vs math:

• fp32: atol=5e-6, rtol=5e-5
• fp16/bf16: atol=5e-3, rtol=5e-2
• gradients: 2× the forward tolerance

Not bitwise identical. Checkpoints trained with one backend load cleanly into the other (verified).

Scope / supported configs

Category	Tested	Planned	Not planned
Apple silicon	M4	M3	M1, M2 (should work via auto-calibration)
macOS	26.4.1	15.x	14.x (documented unsupported — backend registers unavailable)
torch	2.11 stable + 2.13 nightly	—	—
dtypes	bf16, fp16, fp32	—	fp64 (MPS doesn't support it)

Architecture

See ARCHITECTURE.md for the internal structure: backend registry, graph cache, threshold auto-calibration, C++ extension build system, autograd wiring, dropout path.

What doesn't work (yet)

• GQA (Hq ≠ Hkv): routes to stock with repeat_interleave. mpsgraph op is MHA-only. One-time warning.
• Second-order gradients (create_graph=True on our output's grad): raises a clear error. Backward uses MPSGraph which is opaque to torch autograd; true higher-order would require a differentiable backward impl.
• torch.compile (full graph capture): untested. The JIT-compiled C++ extension is compatible with eager mode; inductor integration is future work.
• macOS 14: the MPSGraph.scaledDotProductAttention method doesn't exist on Sequoia's predecessor. Backend registers as unavailable with a clear reason; calls fall back to stock.

Correctness — what's tested

244 tests across 20 files. Highlights:

• Shape matrix: D ∈ {32, 64, 96, 128, 192, 256}; H ∈ {1..32}; B ∈ {1..32}; Lq, Lkv ∈ {powers of 2, 777, 1345, 3141}.
• Masks: causal, bool, additive float, all-True, mostly-False, per-head [B,H,Lq,Lkv], per-batch, broadcast variants, dtype coercion.
• Autograd: partial requires_grad subsets, AMP autocast (bf16/fp16), torch.utils.checkpoint re-entry, retain_graph, grad accumulation, second-order detection, once_differentiable guard.
• Training: generic transformer smoke (forward + backward), 1000-step long-horizon convergence, mid-run backend toggle, fp16 500-step AMP.
• Numerical extremes: Q/K std ∈ {10, 1e-4}, grad_out ∈ {1e3, 1e4} — all NaN/Inf-free.
• Edge cases: degenerate shapes, non-contiguous inputs, strided views, cache thrash (9 shapes), 200-call leak probe, 16k-seq OOM recovery.

See tests/ for the full test suite.

Citation / credit

• Apple's fused SDPA op: MPSGraph.scaledDotProductAttention... (macOS 15+).
• Related prior art: philipturner/metal-flash-attention (Swift, not vetted here); PyPI mps-flash-attn (8 stars, unverified). Neither is used or wrapped.

License

MIT.

Contributing

See CONTRIBUTING.md. Bug reports with reproducible shapes are the most helpful; the full-suite pytest tests/ should pass on any supported config.