[Config] [tune] DSv4 a8w8 blockscale: add gfx950 configs for three MI355X shapes

[jiacao-amd]

Summary

Adds 106 tuned rows to dsv4_a8w8_blockscale_tuned_gemm.csv for gfx950, covering four DeepSeek-V4 shape families:

N	K	rows	M coverage
6144	7168	15	1 .. 16384
65536	1536	15	1 .. 16384
7168	3072	15	1 .. 16384
7168	7168	61	observed DSv4 values from 4 .. 65536

These shapes previously fell back to an untuned default.

The M=8192, N=65536, K=1536 entry uses the peer-provided CKTile config 11. On a local MI355X A/B run, it reduced latency from 1303.55 us to 947.08 us (27.35% lower, 1.376x speedup), with errRatio=0.0.

Why merge into the existing file

These rows are merged into the existing dsv4_a8w8_blockscale_tuned_gemm.csv, because AITER_CONFIG.get_config_file() globs model_configs/*<tuned_file_name>*.csv and merges every match. A second file would be redundant at load time, and a8w8_..._dsv4 would also be inconsistent with the existing dsv4_a8w8_blockscale_{tuned,untuned}_gemm.csv naming.

Details

• Config-only change; no operator code path changes.
• Final diff against main: 106 insertions, 0 deletions.
• No duplicate (gfx, cu_num, M, N, K) keys or malformed rows.
• All selected kernel names are present in the compiled kernel registry.

Test Plan

Correctness test passed on MI355X (gfx950, 256 CUs): 106/106 OK, 0 mismatches, 0 errors.

python3 csrc/ck_gemm_a8w8_blockscale/gemm_a8w8_blockscale_tune.py \
    -i aiter/configs/a8w8_blockscale_untuned_gemm.csv \
    --run_config .pr4664_added_shapes_plus_peer.csv \
    --warmup 2 --iters 5

[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 4664 --add-label <label>