A4w4_asm_pro_max_v2

[zufayu]

add new tile shape 224x256 192x256

[Copilot]

Pull Request Overview

This PR adds support for new tile shapes (224x256 and 192x256) to the A4W4 GEMM assembly kernels and improves the tuning infrastructure. The changes enhance the performance optimization capabilities by expanding the available tile configurations and making the tuning process more dynamic.

• Added two new tile shapes (192x256 and 224x256) to the assembly kernel configuration
• Enhanced condition checks for split-K operations to ensure proper validation
• Improved tuning infrastructure to dynamically use all available kernels instead of hardcoded tiles

Reviewed Changes

Copilot reviewed 6 out of 8 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
hsa/gfx950/f4gemm/f4gemm_bf16_per1x32Fp4.csv	Adds kernel configurations for new 192x256 and 224x256 tile shapes
op_tests/test_gemm_a4w4.py	Improves split-K validation and adds comprehensive test cases for tuning
csrc/ck_gemm_a4w4_blockscale/gemm_a4w4_blockscale_tune.py	Enhances tuning to dynamically use all available kernels and fixes split-K validation
aiter/ops/gemm_op_a4w4.py	Updates type annotations for better consistency
aiter/configs/a4w4_blockscale_untuned_gemm_test.csv	Adds new test configuration file with comprehensive test cases
aiter/configs/a4w4_blockscale_tuned_gemm_test.csv	Adds new empty tuned configuration file

Comments suppressed due to low confidence (1)