[FIX] Report the profiler's buffer-load issue only for AMD kernels

[mark14wu]

Summary

The profiler's buffer-load check is meant for AMD GPUs only, but it also fires on NVIDIA. Profiler.__init__ set has_buffer_load = False, and post_warmup_callback only updates it when the warmup result has an amdgcn ASM stage. NVIDIA kernels have no amdgcn stage, so the value stayed False. Then _check_32bit_range (if self.has_buffer_load is False) marked every kernel whose offsets fit in 32 bits as a potential buffer-load issue, and the summary printed the AMD "Potential Buffer Load Issue Detected" warning. That warning was a false positive.

Fix

• has_buffer_load now starts as None ("no AMD kernel seen yet"). Only a captured amdgcn stage sets it to True or False. The existing is False test therefore flags only AMD kernels whose amdgcn stage has no buffer_load.
• When no amdgcn kernel was captured, the Buffer Load section of the summary now prints Buffer Load check skipped: no AMD (amdgcn) kernel was captured. In that case it shows neither the AMD warning nor the "used appropriately" message.
• New unit tests in tests/unit/test_profiler.py pass a fake warmup result to post_warmup_callback, then 32-bit offsets to _check_32bit_range:
  • an NVIDIA-like result (ttir/ptx stages, no amdgcn) is not flagged, and the summary shows the skipped message;
  • an amdgcn stage without buffer_load is flagged;
  • an amdgcn stage with buffer_load is not flagged.

Testing

Run on an RTX 4090 with the GPU visible. I first confirmed that the worktree's tilelens is the one imported.

python -m pytest tests/unit/test_profiler.py tests/end_to_end/test_profiler.py -q -p no:cacheprovider
26 passed in 1.81s

python -m pytest tests/unit/test_profiler.py -p no:cacheprovider -k buffer_load -rA
4 passed, 12 deselected

pre-commit run --files tilelens/clients/profiler/profiler.py tests/unit/test_profiler.py
all hooks Passed/Skipped (ruff, ruff-format, mypy, codespell, ...)

[github-actions]

Performance Benchmark

Benchmark	main (min)	PR (min)	Change	Samples
gemm	0.074s	0.075s	+0.4%	20 / 20
gemm_oob	0.084s	0.083s	-0.4%	20 / 20
indirect_load	0.015s	0.015s	-1.1%	20 / 20
nested_loop	0.156s	0.155s	-0.3%	20 / 20
block_pointer_loop_advance	0.088s	0.088s	+0.3%	20 / 20
liger_jsd	0.103s	0.103s	+0.3%	20 / 20
flaggems_layernorm	0.269s	0.271s	+0.7%	20 / 20
swiglu	0.125s	0.125s	+0.5%	20 / 20
cross_entropy	0.702s	0.703s	+0.2%	20 / 20
fused_linear_jsd	0.156s	0.157s	+0.4%	20 / 20
Total	1.771s	1.775s	+0.2%	N/A

Iterations: 1 warmup + 20 measured
Samples are shown as main / PR; long pytest benchmarks may use fewer samples.