[gfx1250][FlyDSL] Refactor GEMMs with layout-based API

[aoli26]

Motivation

Replace legacy gfx1250 GEMM implementations with focused, layout-based kernels.

Technical Details

Add A8W8 PTPC, and A8W8 blockscale kernels and remove legacy implementations.

Test Plan

Ran a8w8 blockscale/PTPC GEMM correctness tests and AOT pre-compilation on gfx1250.

Test Result

Results match the previous kernel path; AOT compilation passes.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 4374 --add-label <label>

[aoli26]

Depends on #4210

[Copilot]

Pull request overview

This PR refactors and extends gfx1250 FP8 GEMM support by introducing new FlyDSL layout-based WMMA kernels (PTPC and blockscale bpreshuffle), updating quantization scale shuffling semantics, and refreshing tuned-config coverage plus AOT compilation support for the new kernel families.

Changes:

• Add gfx1250 FlyDSL WMMA kernels for a8w8 PTPC bpreshuffle GEMM and a8w8 blockscale bpreshuffle GEMM, and update gfx1250 bpreshuffle dispatch to use the new launchers.
• Update quantization kernels/dispatch to distinguish MX-hardware e8m0 swizzle (group_size==32) vs. transpose (other group sizes), and expose an fp8-e8m0 scale output option via scale_type.
• Update tuned CSVs (new gfx1250 entries) and extend AOT precompile tooling plus op_tests to exercise the fp8-e8m0 scale path.

Reviewed changes

Copilot reviewed 15 out of 16 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
op_tests/test_gemm_a8w8_blockscale.py	Adds optional fp8-e8m0 scale generation/testing path and a --flydsl flag for blockscale tests.
csrc/kernels/quant_kernels.cu	Refines e8m0 scale padding/shuffle behavior to be MX-layout-specific (group_size==32) and uses transpose otherwise.
aiter/ops/quant.py	Adds scale_type option to per_group_quant_hip and routes fp8-e8m0 scale requests through per-group quant op.
aiter/ops/gemm_op_a8w8.py	Adds gfx1250 support and dispatch for FlyDSL blockscale bpreshuffle path when scales are fp8-e8m0.
aiter/ops/flydsl/kernels/tdm_oob.py	Removes vendored OOB-capable TDM descriptor builder fallback.
aiter/ops/flydsl/kernels/gemm_mxscale_gfx1250.py	Tightens buffer resource bounds for grouped-contiguous-M metadata loads.
aiter/ops/flydsl/kernels/gemm_a8w8_ptpc_gfx1250.py	New FlyDSL WMMA PTPC (per-token/per-channel fp32 scale) preshuffle GEMM kernel for gfx1250.
aiter/ops/flydsl/kernels/gemm_a8w8_blockscale_gfx1250.py	New FlyDSL WMMA blockscale (e8m0 scale) preshuffle GEMM kernel for gfx1250.
aiter/ops/flydsl/gemm_tune/flydsl_gemm_a8w8_bpreshuffle_wmma_common.py	Reworks gfx1250 PTPC candidate generation and shape-fitting logic.
aiter/ops/flydsl/gemm_tune/flydsl_gemm_a8w8_blockscale_bpreshuffle_wmma_common.py	Adds gfx1250 blockscale WMMA candidate catalog and shape-fitting logic.
aiter/ops/flydsl/bpreshuffle_gemm_gfx1250.py	Updates gfx1250 PTPC runner to call the new layout-based WMMA launcher and adjusts split-k limits.
aiter/ops/flydsl/blockscale_bpreshuffle_gemm_gfx1250.py	New gfx1250 blockscale bpreshuffle runner and kernelName parser/dispatcher.
aiter/configs/model_configs/dsv4_a8w8_blockscale_bpreshuffle_tuned_gemm.csv	Adds many gfx1250 flydsl tuned entries for blockscale bpreshuffle GEMM.
aiter/configs/a8w8_bpreshuffle_tuned_gemm.csv	Updates gfx1250 flydsl tuned entries for PTPC bpreshuffle GEMM.
aiter/aot/flydsl/gemm.py	Extends AOT to detect/compile the new gfx1250 WMMA kernel families and adds optional ARCH-based job filtering.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.