[gfx1201] Add tuned kernel configs and FP8 attention support for AMD RDNA4 (Radeon AI PRO R9700)

[vllmellm]

Motivation

Following the correctness fixes in #1681 (ISA patches for v_pk_mul_f32, DPP broadcast, buffer_load_lds, and RMSNorm kernel operand syntax), this PR adds the tuning configs and FP8 attention enablement needed to run performant FP8 inference on gfx1201 (AMD Radeon AI PRO R9700) via vLLM.

Evaluated on Qwen3-0.6B-FP8 (dense) and Qwen3-30B-A3B-FP8 (MoE): up to +33% throughput, +25% TPOT improvement over vLLM default.

Technical Details

1. FP8 Triton Attention (aiter/ops/triton/_triton_kernels/flash_attn_triton_amd/utils.py)

Added "gfx1201" to the FP8_ARCHS frozenset. Without this, the flash attention Triton dispatcher silently falls back to BF16/FP16 arithmetic even when FP8 tensors are provided, missing the FP8 fast-path.

2. CK GEMM Tuned Configs (aiter/configs/a8w8_blockscale_tuned_gemm.csv, a8w8_blockscale_untuned_gemm.csv)

Added ~370 tuned kernel configurations for gfx1201 covering common M/N/K dimensions from Qwen3-0.6B-FP8 and Qwen3-30B-A3B-FP8. Each entry maps a tensor shape to an optimal CK (Composable Kernel) instance for block-scaled FP8 GEMM. Without these, the dispatcher has no valid mapping and falls back to suboptimal kernels or errors.

3. Triton GEMM Configs (aiter/ops/triton/configs/gemm/gfx1201-*.json)

Added 37 Triton GEMM config files for gfx1201 across the following variants:

• GEMM-A8W8 / GEMM-A8W8_BLOCKSCALE / GEMM-A8W8_BLOCKSCALE_PRESHUFFLED / GEMM-A8W8_PER_TOKEN_SCALE
• GEMM-A16W16 / GEMM-A16W16-ATOMIC
• BATCHED_GEMM-A8W8

Includes both default configs and shape-specific overrides for common N/K dimensions (1024/2048/3072/4096/6144 × 1024).

Test Plan

• Unit tests: op_tests/test_gemm_a8w8_blockscale.py, op_tests/test_gemm_a8w8.py, op_tests/triton_tests/gemm/basic/test_gemm_a8w8_blockscale.py, op_tests/triton_tests/gemm/basic/test_gemm_a8w8_per_token_scale.py on gfx1201
• End-to-end: vLLM serving benchmark (benchmarks/benchmark_serving.py) on Qwen3-0.6B-FP8 and Qwen3-30B-A3B-FP8

Test Result

Benchmarked on AMD Radeon AI PRO R9700 (gfx1201) via vLLM serving benchmark.

Mean TTFT (s) — lower is better

ISL/OSL	1. Default	2. aiter CK	3. aiter CK + Attn	4. aiter CK + Attn + Norm
1024/1024	0.482	0.307	0.657	0.340
2048/2048	0.551	0.608	0.647	0.680
4096/4096	2.007	2.020	2.410	2.522
8192/1024	12.306	10.451	12.137	12.294
16384/2048	141.984	130.585	134.655	136.155

Mean TPOT (s) — lower is better

ISL/OSL	1. Default	2. aiter CK	3. aiter CK + Attn	4. aiter CK + Attn + Norm
1024/1024	0.0221	0.0168	0.0158	0.0167
2048/2048	0.0319	0.0266	0.0248	0.0257
4096/4096	0.0508	0.0456	0.0436	0.0445
8192/1024	0.0624	0.0590	0.0644	0.0654
16384/2048	0.0660	0.0610	0.0612	0.0623

Total Token Throughput (tok/s) — higher is better

ISL/OSL	1. Default	2. aiter CK	3. aiter CK + Attn	4. aiter CK + Attn + Norm
1024/1024	2830	3749	3900	3766
2048/2048	1989	2376	2543	2455
4096/4096	1214	1365	1424	1395
8192/1024	3434	3753	3454	3397
16384/2048	1793	1952	1924	1896

Qwen3-30B-A3B-FP8 (MoE)

Mean TTFT (s) — lower is better

ISL/OSL	1. Default	2. aiter CK	3. aiter CK + Attn	4. aiter CK + Attn + Norm
1024/1024	0.926	0.886	0.942	0.940
2048/2048	1.565	1.585	1.575	1.566
4096/4096	5.333	5.387	5.300	5.291
8192/1024	14.066	13.701	13.112	13.088
16384/2048	144.387	136.012	131.844	129.665

Mean TPOT (s) — lower is better

ISL/OSL	1. Default	2. aiter CK	3. aiter CK + Attn	4. aiter CK + Attn + Norm
1024/1024	0.0378	0.0334	0.0322	0.0313
2048/2048	0.0377	0.0367	0.0333	0.0335
4096/4096	0.0456	0.0427	0.0413	0.0410
8192/1024	0.0788	0.0765	0.0761	0.0742
16384/2048	0.0715	0.0675	0.0694	0.0701

Total Token Throughput (tok/s) — higher is better

ISL/OSL	1. Default	2. aiter CK	3. aiter CK + Attn	4. aiter CK + Attn + Norm
1024/1024	1652	1868	1930	1986
2048/2048	1662	1706	1879	1869
4096/4096	1360	1451	1498	1511
8192/1024	2856	2990	3067	3133
16384/2048	1715	1845	1858	1856

Notes

• Depends on correctness patches from [Draft] [Preview] Support gfx1201 #1681

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.