remove per kernel fp8 triton quantized benchmarks

[iamzainhuda]

Summary

This removes the redundant bench_triton_fp8_* quant benchmark scripts under benchmarks/prototype/blockwise_fp8_training and consolidates their correctness coverage into benchmark_quant_kernel_bandwidth.py.

The consolidated benchmark now has an optional --check-correctness mode that compares each Triton quant kernel against the corresponding Torch reference implementation, including layout/stride checks where applicable. The weight-kernel references were adjusted to match Triton’s column-major outputs and scales, which was required for weight_quant_rhs and weight_quant_transposed_rhs.

Testing

• python -m py_compile benchmarks/prototype/blockwise_fp8_training/benchmark_quant_kernel_bandwidth.py
• python -m benchmarks.prototype.blockwise_fp8_training.benchmark_quant_kernel_bandwidth --help
• python -m benchmarks.prototype.blockwise_fp8_training.benchmark_quant_kernel_bandwidth --check-correctness on H100, completed successfully with no correctness failures.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/4166

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 9717b75 with merge base bca91b0 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[danielvegamyhre]

can we also update this script to only measure against peak achievable bandwidth as defined in the gpu_spec in roofline_utils.py we import (not measure against the "advertised max theoretical bandwidth" number)

[iamzainhuda]

right now script reports both, do you want me to remove the max theoretical bandwidth numbers altogether?

[danielvegamyhre]

let's just keep vs_achievable

[iamzainhuda]

done, added print to make it clear to user as well what achievable % is used