[TEST] Cover Gluon Blackwell tensor memory ops

[Jokeren]

Summary

Adds CPU regression coverage for Blackwell tensor-memory roundtrip operations, stacked on the Hopper WGMMA PR.

The test covers:

• blackwell.allocate_tensor_memory
• TensorMemoryLayout
• tensor-memory descriptor store
• tensor-memory descriptor load
• explicit tensor-memory register layout conversion through get_reg_layout(num_warps=...)

The simulated output is compared against the original torch CPU input tile.

This also makes Gluon builtin patching avoid injecting _semantic into helper functions that are marked builtin but do not accept that keyword, which is needed by Blackwell tensor-memory layout helpers.

Validation

• PYTHONPATH=/mnt/keren/triton-viz pytest tests/end_to_end/test_gluon_blackwell_tensor_memory_ops.py -q
• PYTHONPATH=/mnt/keren/triton-viz pytest tests/end_to_end/test_gluon_blackwell_tensor_memory_ops.py tests/end_to_end/test_gluon_wgmma_ops.py tests/end_to_end/test_gluon_tma_im2col_ops.py tests/end_to_end/test_gluon_blackwell_tma_ops.py tests/end_to_end/test_gluon_tma_ops.py tests/end_to_end/test_gluon_async_copy_ops.py tests/end_to_end/test_gluon_core_ops.py tests/end_to_end/test_gluon.py tests/end_to_end/test_core.py tests/unit/test_adapters.py -q
• pre-commit run --files triton_viz/core/simulation/gluon.py tests/end_to_end/test_gluon_blackwell_tensor_memory_ops.py

[github-actions]

Performance Benchmark

Benchmark	main (min)	PR (min)	Change	Samples
gemm	0.105s	0.105s	-0.1%	20 / 20
gemm_oob	0.118s	0.117s	-0.4%	20 / 20
indirect_load	0.022s	0.022s	-0.7%	20 / 20
nested_loop	0.235s	0.234s	-0.6%	20 / 20
block_pointer_loop_advance	0.206s	0.205s	-0.6%	20 / 20
liger_jsd	0.137s	0.138s	+0.2%	20 / 20
flaggems_layernorm	0.398s	0.395s	-0.8%	20 / 20
swiglu	0.172s	0.172s	-0.2%	20 / 20
cross_entropy	0.980s	0.985s	+0.5%	20 / 20
fused_linear_jsd	0.209s	0.210s	+0.5%	20 / 20
Total	2.583s	2.582s	-0.0%	N/A

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

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4d83822c4a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve FP4 padding when rebuilding tensor-memory layouts

When a Gluon kernel uses a Blackwell tensor-memory layout with fp4_padded=True, _to_ir passes that flag into this builder, but this constructor call now drops it and always returns a layout constructed without the FP4 padding state. That makes the interpreter treat FP4-padded tensor-memory descriptors as ordinary layouts, so any later layout-dependent register conversion or descriptor behavior can diverge from real Gluon semantics. Please propagate fp4_padded here, using a version-aware constructor path if needed for older Triton builds.

Useful? React with 👍 / 👎.