[mark14wu]

Summary

The profiler crashes on every tensor-descriptor load or store:

src = tl.make_tensor_descriptor(in_ptr, shape=[N], strides=[1], block_shape=[BLOCK])
x = src.load([pid * BLOCK])
# AttributeError: 'memoryview' object has no attribute 'size'
#   tilelens/clients/profiler/profiler.py, in _get_mask_stats: total_count = mask.data.size

Triton's interpreter implements create_descriptor_load/create_descriptor_store by calling the masked load/store with the bounds mask from TensorDescHandle.materialize_pointers, which is a plain NumPy array, not a TensorHandle. The profiler's load/store callbacks read mask.data; for a NumPy array that is a memoryview, which has no .size. On Triton 3.7, tl.make_block_ptr loads and stores reach the same path through create_tensor_pointer_load/create_tensor_pointer_store, so block-pointer kernels crashed there too; on Triton 3.8 they pass a TensorHandle and are unaffected.

This PR adds _mask_data, which returns a NumPy-array mask as is and a TensorHandle's .data otherwise, and uses it in _get_mask_stats and _report_load_store_bytes. (np.count_nonzero happened to accept the memoryview, so only .size failed.)

Tests

• New in tests/end_to_end/test_profiler.py: test_mask_percentage_of_tensor_descriptor_accesses copies a 100-element tensor through descriptors in 4 blocks of 32 and checks the mask counts (128 total, 28 masked, for loads and stores) and the byte counts. It fails on main with the error above.
• pytest tests/: 366 passed, 6 skipped on main; 367 passed, 6 skipped with this PR (Python 3.12, Triton 3.8.0).
• ruff check and ruff format --check (v0.1.9) pass on the changed files.

🤖 Generated with Claude Code

[github-actions]

Performance Benchmark

Benchmark	main (min)	PR (min)	Change	Samples
gemm	0.102s	0.101s	-0.4%	20 / 20
gemm_oob	0.113s	0.113s	-0.3%	20 / 20
indirect_load	0.021s	0.021s	+1.1%	20 / 20
nested_loop	0.224s	0.224s	-0.0%	20 / 20
block_pointer_loop_advance	0.121s	0.121s	+0.3%	20 / 20
liger_jsd	0.133s	0.133s	+0.2%	20 / 20
flaggems_layernorm	0.380s	0.380s	-0.1%	20 / 20
swiglu	0.164s	0.163s	-0.8%	20 / 20
cross_entropy	0.944s	0.944s	-0.0%	20 / 20
fused_linear_jsd	0.203s	0.203s	+0.2%	20 / 20
Total	2.405s	2.404s	-0.1%	N/A

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