[ROCm] Add AMD Instinct MI300X/MI325X/MI350X/MI355X GPU support

[andyluo7]

Summary

Add GPU detection and performance specs for AMD Instinct GPUs (MI300X, MI325X, MI350X, MI355X) to enable AutoKernel on ROCm.

Problem

On ROCm, torch.cuda.get_device_properties() returns an empty device name and clock_rate=0, causing:

1. GPU not matched in _KNOWN_GPUS table → falls back to estimation
2. clock_rate / 1e6 in fallback path → incorrect zero/near-zero TFLOPS estimates
3. Roofline analysis and performance metrics are meaningless

Solution

GPU Database (bench.py + profile.py)

• Add MI300X (1307.4 TFLOPS, 5300 GB/s, 256 MB L2)
• Add MI325X (1307.4 TFLOPS, 6000 GB/s, 256 MB L2)
• Add MI350X (2300.0 TFLOPS, 8000 GB/s, 256 MB L2)
• Add MI355X (2300.0 TFLOPS, 8000 GB/s, 256 MB L2)

ROCm-aware GPU Detection (bench.py)

• New _KNOWN_AMD_GPUS dict keyed by gcnArchName prefix (e.g. gfx942 → MI300X)
• When device name is empty, fall back to props.gcnArchName for identification
• Guard clock_rate access behind hasattr + > 0 check

Same fixes applied to profile.py fallback detector

Testing

Tested end-to-end on AMD Instinct MI300X (gfx942, ROCm 6.3, PyTorch 2.9):

=== GPU INFO ===
gpu_name: AMD Instinct MI300X
gpu_sm_count: 304
gpu_memory_gb: 192.0
gpu_peak_tflops_fp16: 1307.4
gpu_peak_bandwidth_gb_s: 5300.0
gpu_l2_cache_mb: 256.0

=== CORRECTNESS ===
smoke_test: PASS
FP16 shape_sweep: all PASS
BF16 shape_sweep: all PASS
numerical_stability: PASS
determinism: PASS
edge_cases: PASS

=== PERFORMANCE (xlarge) ===
PyTorch baseline: 607.9 TFLOPS (46.5% peak)
Benchmark harness: runs end-to-end ✅

Zero NVIDIA impact

• Existing _KNOWN_GPUS entries unchanged
• clock_rate path only skipped when clock_rate is 0 or missing (never happens on NVIDIA)
• No new dependencies

[andyluo7]

✅ Verified on AMD Instinct MI350X (gfx950 / CDNA4)

Tested end-to-end on 8x MI350X (ROCm 7.2, PyTorch 2.10.0+rocm7.0, Triton 3.6.0):

GPU Detection

gpu_name: AMD Instinct MI350X VF
gpu_sm_count: 256
gpu_memory_gb: 287.6
gpu_peak_tflops_fp16: 2300.0
gpu_peak_tflops_bf16: 2300.0
gpu_peak_bandwidth_gb_s: 8000.0
gpu_l2_cache_mb: 256.0
gpu_compute_capability: 9.5

GPU correctly identified via gcnArchName: gfx950 → matched "MI350X" in _KNOWN_GPUS.

Correctness

Test	Result
Smoke test	✅ PASS
FP16 shape sweep (10/10)	✅ PASS
BF16 shape sweep (10/10)	✅ PASS
FP32 shape sweep (7/10)	⚠️ 3 fail (xlarge/deep_k/llm_mlp — tight atol, expected)
Numerical stability	✅ PASS
Determinism	✅ PASS
Edge cases	✅ PASS

Performance (starter kernel, not optimized)

size            kernel_us   pytorch_us    speedup     tflops    %peak
------------------------------------------------------------------
tiny                 6.72         9.91     1.473x      0.624     0.0%
small               13.22         9.76     0.738x     20.298     0.9%
large               48.13        32.15     0.668x    356.930    15.5%
xlarge             323.68       129.86     0.401x    424.614    18.5%
llm_mlp            908.00       452.45     0.498x    406.791    17.7%

PyTorch baseline xlarge BF16: 956.5 TFLOPS (41.6% of 2300 TFLOPS peak).

Environment

• ROCm 7.2.0, PyTorch 2.10.0+rocm7.0, Triton 3.6.0
• Native gfx950 support (no HSA_OVERRIDE_GFX_VERSION needed)
• torch.cuda.get_device_properties() returns gcnArchName: gfx950:sramecc+:xnack-

Previously also verified on MI300X (gfx942, ROCm 6.3, PyTorch 2.9) — see PR description.