feat(attention): add Gluon forward attention backend

[bangtianliu]

Description

Adds an opt-in Gluon backend for dense Flash Attention forward on gfx950.

Gluon is not part of the default selection or cross-backend autotuning. It is forward-only, and a pinned unsupported call fails before launch instead of falling back. The port does not load an LLVM pass plugin, pin LLVM/Triton, or select behavior from a compiler version.

Fixes # (N/A)

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

• Vendor the gfx950 FP16/BF16 Gluon forward kernel and record its provenance.
• Register BackendType.GLUON in the dense attention dispatcher.
• Add an opaque PyTorch custom-op boundary with fake-tensor and torch.compile(fullgraph=True) support.
• Add fail-fast eligibility and lifecycle checks for unsupported inputs and backward requests.
• Add correctness regressions and a correctness-gated forward benchmark.

Supported envelope

Dense rank-4 FP16/BF16 attention on gfx950; BSHD or BHSD storage; MHA, GQA, and MQA; head dimension up to 256; causal and noncausal. Causal requires Sq <= Skv.

Performance

MI350X (gfx950), PyTorch 2.13.0+rocm7.1, Triton 3.8.0+fb, FP16, BSHD, D128, warm cache. The 30-shape set covers causal/noncausal MHA, GQA, and MQA from S128 through S16384. Each row must pass the FP32-reference check at greater than 40 dB output SNR before it is timed. Backends are pinned; no fallback result is included. Gluon uses a fresh autotune cache: 26 general shape keys use the 36-config autotuner; four short-causal rows use the fixed specialization. Tuning and correctness complete before 20 warmups and 100 measured launches. TFLOPS use the attended FLOP count, including the causal fraction.

Backend	Correct	Arithmetic mean TFLOPS
Gluon	30/30	480.52
AITER	30/30	459.69

The geometric mean of the 30 per-shape Gluon/AITER throughput ratios is 0.830x; the long, high-throughput rows dominate the arithmetic mean. The result is shape-dependent. Selected causal D128 rows:

B	S	Hq/Hkv	Gluon TFLOPS	AITER TFLOPS	Ratio
32	512	64/64	165.26	294.21	0.562x
16	1024	64/64	446.05	411.21	1.085x
8	2048	64/64	485.22	507.75	0.956x
4	4096	64/64	643.15	573.14	1.122x
2	8192	64/64	700.78	610.26	1.148x
1	8192	32/8	725.04	635.29	1.141x
1	2048	32/1	185.17	465.37	0.398x
1	8192	32/1	724.71	658.79	1.100x

Checklist

• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my feature works
• New and existing focused tests pass locally with my changes

[RuibinCheung]

Is forward FA kernel only need for client ?

[bangtianliu]

This PR adds forward first. Gluon backward is the next focus and will follow separately.

[RuibinCheung]

Please remove debug env.

[RuibinCheung]

Please merge this file into test_attention.py.

[RuibinCheung]

Please use this func.

Primus-Turbo/primus_turbo/pytorch/core/utils.py

Lines 24 to 25 in e29a9b3

	def is_gfx950() -> bool:
	return get_device_compute_capability() == (9, 5)

[RuibinCheung]

What's the purpose of this unit test file ?

[bangtianliu]

This was a self-added vendoring-policy check rather than existing CI coverage. I removed this test and the corresponding checker to keep the PR focused on the Gluon backend.