EnsueAI/metal-int4-sdpa

Fused int4 SDPA for Apple Silicon

Core attention kernel from Open-TQ-Metal.

Open-TQ-Metal is a Metal-native implementation of fused compressed-domain attention built by Ensue. The full release enables Llama 3.1 70B at 128K context on a single 64GB Mac, 48x faster attention at 128K context, and includes a C++ inference engine, multiple attention kernels, a 330-experiment cross-architecture analysis, and a paper.

• Paper: https://arxiv.org/pdf/2604.16957
• Write-up: https://ensue.dev/blog/introducing-open-tq-metal/
• Llama 3.1 70B engine: https://github.com/mutable-state-inc/turboquant-llama3.170B                                                                                                             
• Gemma 4 31B engine: https://github.com/mutable-state-inc/gemma4metal

This is the Metal implementation of TurboQuant's fused compressed-domain attention (Zandieh et al., ICLR 2026): A fused int4 Scaled Dot-Product Attention Metal kernel for Apple Silicon (M1/M2/M3/M4).

Computes softmax(Q @ dequant(K_int4)^T * scale) @ dequant(V_int4) in a single kernel dispatch with online softmax — never materializing full K/V or score matrices.

Performance

Benchmarked on M1 Max (32 GPU cores, 64GB unified memory). 56 experiments conducted.

vs MLX int4 dequant+SDPA (apples-to-apples)

Config	N	Fused int4	MLX int4	Speedup
Gemma 4 sliding (sw=1024)	4096	0.34ms	3.81ms	11.1x
Gemma 4 sliding (sw=1024)	8192	0.40ms	5.79ms	14.3x
Gemma 4 full (D=512)	8192	0.98ms	7.87ms	8.1x
Llama 3.1 (D=128)	8192	0.89ms	2.53ms	2.9x

vs PyTorch native SDPA (FP32)

Config	N	Fused int4	Native SDPA	Speedup
Gemma 4 (D=256)	2048	0.45ms	1.35ms	3.0x
Llama 3.1 (D=128)	8192	0.88ms	1.53ms	1.7x

Long-context scaling (Gemma 4 31B, 60 layers pipelined)

Context	SDPA/token	SDPA ceiling	int4 KV
4K	5.1ms	195 tok/s	30 MB
8K	7.2ms	138 tok/s	60 MB
32K	19.6ms	51 tok/s	240 MB
128K	88ms	11 tok/s	960 MB

6.4x KV cache compression — 128K context in 960 MB int4 vs 6.1 GB FP32.

Quality

Metric	Value
Cosine similarity (int4 vs FP32 output)	0.992
Attention KL divergence	0.004
Top-10 token overlap	90%
Kernel numerical precision	< 1e-6 relative error
Deterministic	Yes (bit-identical across runs)

Features

• Online softmax — O(1) memory in sequence length
• qdot pattern — eliminates per-nibble shift operations in K dot product
• exp2 optimization — native Apple GPU instruction for 3-12% speedup
• Adaptive split-K — doubles GPU utilization for models with few attention heads
• Sliding window fast-skip — jumps directly to valid tokens
• GQA support — arbitrary grouped-query attention factors (1-32 tested)
• Batched decode — speculative decoding via flattened Q tensor
• bfloat16 scales/biases — native MLX KV cache format support
• D = 128, 256, 512 — all major model head dimensions

Reliability

• 40 correctness tests (reference + Metal + architecture-specific + batched)
• 43 stress tests (N=1-131072, GQA=1-32, sliding window edge cases)
• 12 Gemma 4 31B end-to-end eval tests (factual, instruction, coherence, perplexity)
• Zero memory leaks (10000 calls tested)
• No thermal throttling (60s sustained load)
• Deterministic (100 calls bit-identical)

int4 Format

• uint32 packs 8 x 4-bit nibbles
• Per-group (64 elements) asymmetric quantization: value = scale * nibble + bias
• 6.4x compression vs FP32

Usage

import kernels

sdpa_int4 = kernels.get_kernel("EnsueAI/metal-int4-sdpa", "sdpa_int4")

# Single-token decode
output = sdpa_int4.sdpa_int4(
    queries,       # (num_heads, D) float32, MPS
    k_quant,       # (num_kv_heads, N, D//8) uint32, MPS
    k_scales,      # (num_kv_heads, N, D//64) float32, MPS
    k_biases,      # (num_kv_heads, N, D//64) float32, MPS
    v_quant,       # (num_kv_heads, N, D//8) uint32, MPS
    v_scales,      # (num_kv_heads, N, D//64) float32, MPS
    v_biases,      # (num_kv_heads, N, D//64) float32, MPS
    gqa_factor=4,  # num_heads // num_kv_heads
    N=2048,        # sequence length
    scale=0.0625,  # 1/sqrt(D)
    sliding_window=0,  # 0 = full attention
)

Tested Models

• Llama 3.1 70B — 80 layers, 64 heads, 8 KV heads, D=128, GQA=8
• Gemma 4 31B — 12/12 end-to-end eval tests pass (9.2 tok/s on M1 Max)
• Llama 3.1 8B — 32 heads, 8 KV heads, D=128, GQA=4
• Qwen2 7B — 28 heads, 4 KV heads, D=128, GQA=7
• Mistral 7B — 32 heads, 8 KV heads, D=128, sliding window=4096
• Mixtral 8x7B — 32 heads, 8 KV heads, D=128, MoE with sliding window
• Arbitrary GQA factors (1-32), non-power-of-2 head counts

Origin

Ported from TurboQuant's sdpa_int4_vector kernel.

Paper: TurboQuant (arxiv.org/abs/2504.19874), ICLR 2026.