feat(gemm_a8w8_blockscale_bpreshuffle): add GLM-5.2 tuned configs (gfx950)

[Raiden-Makoto]

Motivation

The a8w8_blockscale_bpreshuffle GEMM has per-model tuned configs under
aiter/configs/model_configs/ for dsv3, mm2.5, and qwen — but none for GLM-5.2.
GLM-5.2's two 128-aligned MLA dense projections on gfx950 — q_b_proj
(N=4096, K=2048) and o_proj (N=6144, K=4096) — therefore miss the tuned-config
lookup and fall back to the default solution (aiter logs
not found tuned config in ... will use default config).

Note aiter already ships a8w8_bpreshuffle_tuned_gemm_glm5.2.csv, but that is the
per-tensor bpreshuffle path (schema has q_dtype_w). These projections use the
block-scale (128x128) path (a8w8_blockscale_bpreshuffle), which has no GLM-5.2
config. This PR adds one. Config coverage only — no code changes.

Technical Details

• Adds one file:
  aiter/configs/model_configs/a8w8_blockscale_bpreshuffle_tuned_gemm_glm5.2.csv
  • 120 tuned rows for gfx950 / cu_num=256 covering both shapes
    (q_b_proj 4096,2048 and o_proj 6144,4096).
  • M coverage: prefill token counts (1024–65536) and decode / cuda-graph batch
    sizes (1–512).
  • Schema matches the base a8w8_blockscale_bpreshuffle_tuned_gemm.csv:
    gfx,cu_num,M,N,K,libtype,kernelId,splitK,us,kernelName,tflops,bw,errRatio
    (libtype ∈ {asm, ck, cktile}).
• No code changes. AiterConfigs.get_config_file(...) globs
  model_configs/*a8w8_blockscale_bpreshuffle_tuned_gemm*.csv and
  update_config_files(...) merges + dedups (lowest us per shape) with the base
  at runtime, so the new file is picked up automatically.
• Rows produced by the standard tuner
  (csrc/ck_gemm_a8w8_blockscale/gemm_a8w8_blockscale_tune.py --preshuffle), no
  manual edits.

Test Plan

• Run the tuner for both shapes across the M range on gfx950 (MI355X, cu256).
• Load GLM-5.2 and confirm the two shapes no longer emit the default-config
  fallback warning (i.e. get_config_file merges the new model_config).
• op_tests/test_gemm_a8w8_blockscale.py on the added shapes.

Test Result

• Tuner completed with 0 failures across all M for both shapes (gfx950, cu256).
• With the file merged, the not found tuned config … will use default warnings
  for 4096,2048 and 6144,4096 no longer appear at load; both resolve to their
  tuned entries.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
• Formatted with pre-commit (black/ruff).
• CI green.

[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 4243 --add-label <label>

[Raiden-Makoto]

@yifehuan following up on your review — the duplicated a8w8_blockscale cases you flagged (the shapes overlapping qwen3.5_397b) were removed in f3130ae, so the merge is dedup-clean now and CI is fully green: 29 checks passing, 13 skipped, 0 failures, and the PR is mergeable.

This is the same tuned-config pattern you already reviewed and merged in #4236 (and #4264), just the remaining GLM-5.2 o_proj rows. Could you re-review and approve/merge when you get a chance? Thanks!

[yifehuan]

LGTM, no more duplicate shape now