metal : add GATED_LINEAR_ATTN op

[TheTom]

Add Metal backend support for GGML_OP_GATED_LINEAR_ATTN (GLA). Supports head_size 64 and 128, f32 only.

Overview

Metal is currently the only major GPU backend without GLA support. Without this, RWKV6Qwen2 models fall back to CPU for this op on every layer, causing GPU/CPU sync overhead that hurts performance significantly on lower-end Apple Silicon.

The kernel follows the existing RWKV WKV6 Metal structure: one thread per head element, threadgroup shared memory for k/q/gate, and a float4-vectorized inner loop.

supports_op restricts execution to f32 with head sizes 64 or 128.

Correctness

test-backend-ops (7/7 passed on both devices):

Before (master a6cc43c):

GATED_LINEAR_ATTN: not supported [MTL0] (0/0 tests)

After:

GATED_LINEAR_ATTN(head_size=64,  n_seq_tokens=1,   n_seqs=1): OK
GATED_LINEAR_ATTN(head_size=64,  n_seq_tokens=32,  n_seqs=1): OK
GATED_LINEAR_ATTN(head_size=64,  n_seq_tokens=32,  n_seqs=4): OK
GATED_LINEAR_ATTN(head_size=64,  n_seq_tokens=128, n_seqs=4): OK
GATED_LINEAR_ATTN(head_size=128, n_seq_tokens=1,   n_seqs=1): OK
GATED_LINEAR_ATTN(head_size=128, n_seq_tokens=32,  n_seqs=1): OK
GATED_LINEAR_ATTN(head_size=128, n_seq_tokens=32,  n_seqs=4): OK
7/7 tests passed

Performance

Benchmarked with QRWKV6-7B-Instruct Q4_K_M, llama-bench, 3 runs each.

M2 Mac Mini 32GB (Apple8)

Test	Baseline (CPU fallback)	Metal GLA	Speedup
pp1	4.73 t/s	25.19 t/s	+432%
pp512	194.88 t/s	214.69 t/s	+10.2%
pp2048	200.33 t/s	214.95 t/s	+7.3%
pp8192	201.46 t/s	215.38 t/s	+6.9%
tg128	20.35 t/s	24.93 t/s	+22.5%

M5 Max 128GB (Apple10)

Test	Baseline (CPU fallback)	Metal GLA	Delta
pp1	72.49 t/s	71.14 t/s	within noise
pp512	1260.14 t/s	1262.71 t/s	within noise
pp2048	1261.07 t/s	1243.36 t/s	within noise
pp8192	1195.14 t/s	1190.42 t/s	within noise
tg128	70.64 t/s	73.19 t/s	+3.6%

The M5 Max has enough GPU bandwidth that GLA is not the bottleneck at this model size. The M2 Mini shows the real impact of CPU fallback: 5.3x on single-token prefill and 22.5% on decode.

Perplexity

wikitext-2, 5 chunks, ctx=512:

Device	Baseline	Metal GLA
M5 Max	7.7383	7.7383
M2 Mini	7.7386	7.7382

Zero quality impact.

Notes

docs/ops.md and docs/ops/Metal.csv regenerated using test-backend-ops support --output csv and ./scripts/create_ops_docs.py.

Mentions #14909

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. AI used for research and guidance, all code manually reviewed and understood.

[TheTom]

Regarding the automated flags:

• Multiple open PRs: I currently have another open PR (metal: add opt-in V skip for negligible attention weights #21119). I have converted it to draft it to comply with the new contributor limit of one open PR.

• Large PR: The functional change is ~170 lines across the Metal backend plus 3 test lines. The remaining diff is from regenerating docs/ops/Metal.csv and docs/ops.md via the required workflow, which picked up unrelated updates from recent changes in main.

[CISC]

You seem to have included your other PR in this one.

[TheTom]

Thank you @CISC . When fixing the conflict with ./scripts/create_ops_docs.py I accidentally included other fixes for another branch. Will address.

[TheTom]

Requesting review from @ggml-org/ggml-metal

[aminya]

@CISC Would you mind taking another look? The issue is addressed.
Note that, this PR is a precursor to adding TurboQuant support
TheTom#113

[CISC]

@CISC Would you mind taking another look? The issue is addressed.

I cannot review Metal PRs (I don't even have the necessary hardware), wait for @ggml-org/ggml-metal

[TheTom]

@ggerganov any chance I can get a review?