[FLYDSL][GEMM] Full re-tuning for mixed stream-k a16w16 gemm & enhance co-issue

[xytpai]

A16W16 Gemm Tuning Distribution

Summary

Rank	Model	flydsl Share	flydsl Rows / Total Rows
1	gptoss	83.15%	74 / 89
2	kimi & kimik2	77.70%	115 / 148
3	glm5	74.71%	65 / 87
4	deepseekv3	56.90%	33 / 58
5	deepseekv4 + opus backend	50.86%	178 / 350
-	Overall	63.52%	465 / 732

gptoss

N	K	flydsl Rows	Total Rows	flydsl Share
128	2880	13	13	100.00%
640	2880	13	13	100.00%
2560	2880	11	13	84.62%
2880	512	6	13	46.15%
2880	2048	12	13	92.31%
2880	4096	8	12	66.67%
5120	2880	11	12	91.67%
Overall	All	74	89	83.15%

kimi&kimik2

N	K	flydsl Rows	Total Rows	flydsl Share
384	7168	61	64	95.31%
1024	7168	30	49	61.22%
2112	7168	10	15	66.67%
3072	1536	1	3	33.33%
4096	512	3	5	60.00%
7168	512	10	12	83.33%
Overall	All	115	148	77.70%

glm5

N	K	flydsl Rows	Total Rows	flydsl Share
32	6144	0	11	0.00%
128	6144	10	11	90.91%
256	6144	9	11	81.82%
2624	6144	8	11	72.73%
4096	2048	10	11	90.91%
6144	3072	10	11	90.91%
6144	4096	9	10	90.00%
6144	6144	8	10	80.00%
38720	6144	1	1	100.00%
Overall	All	65	87	74.71%

deepseekv3

N	K	flydsl Rows	Total Rows	flydsl Share
256	7168	10	13	76.92%
2112	7168	3	6	50.00%
3072	1536	12	13	92.31%
7168	2048	8	13	61.54%
16160	7168	0	13	0.00%
Overall	All	33	58	56.90%

deepseekv4 (+opus backend)

N	K	flydsl Rows	Total Rows	flydsl Share
64	7168	63	68	92.65%
384	7168	15	53	28.30%
512	7168	30	68	44.12%
1024	7168	40	93	43.01%
2048	7168	30	68	44.12%
Overall	All	178	350	50.86%

---

gptoss acc

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     3|exact_match|↑  |0.9014|±  |0.0082|
|     |       |strict-match    |     3|exact_match|↑  |0.3146|±  |0.0128|

CMD for re-tune

# gptoss
python3 gradlib/gradlib/gemm_tuner.py --tuned_file aiter/configs/model_configs/gptoss_bf16_tuned_gemm.csv --input_file aiter/configs/model_configs/gptoss_bf16_untuned_gemm.csv

# glm5
python3 gradlib/gradlib/gemm_tuner.py --tuned_file aiter/configs/model_configs/glm5_bf16_tuned_gemm.csv --input_file aiter/configs/model_configs/glm5_untuned_gemm_bf16.csv

# dsv3
python3 gradlib/gradlib/gemm_tuner.py --tuned_file aiter/configs/model_configs/dsv3_bf16_tuned_gemm.csv --input_file aiter/configs/model_configs/dsv3_bf16_untuned_gemm.csv

# kimi
python3 gradlib/gradlib/gemm_tuner.py --tuned_file aiter/configs/model_configs/kimi_bf16_tuned_gemm.csv --input_file aiter/configs/model_configs/kimi_bf16_untuned_gemm.csv

# kimik2
python3 gradlib/gradlib/gemm_tuner.py --tuned_file aiter/configs/model_configs/kimik2_bf16_tuned_gemm.csv --input_file aiter/configs/model_configs/kimik2_bf16_untuned_gemm.csv

# dsv4
python3 gradlib/gradlib/gemm_tuner.py --tuned_file aiter/configs/model_configs/dsv4_bf16_tuned_gemm.csv --input_file aiter/configs/model_configs/dsv4_bf16_untuned_gemm.csv

---

# llama70b
python3 gradlib/gradlib/gemm_tuner.py --tuned_file aiter/configs/model_configs/llama70B_bf16_tuned_gemm.csv --input_file aiter/configs/model_configs/llama70B_untuned_gemm_bf16.csv

python3 gradlib/gradlib/gemm_tuner.py --tuned_file aiter/configs/model_configs/llama405B_bf16_tuned_gemm.csv --input_file aiter/configs/model_configs/llama405B_bf16_untuned_gemm.csv

# qwen32b
python3 gradlib/gradlib/gemm_tuner.py --tuned_file aiter/configs/model_configs/qwen32B_bf16_tuned_gemm.csv --input_file aiter/configs/model_configs/qwen32B_untuned_gemm_bf16.csv

python3 gradlib/gradlib/gemm_tuner.py --tuned_file /home/yuxu/workspace/prof/dsv4_tuned.csv --input_file /home/yuxu/workspace/prof/dsv4.csv --libtype triton

[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 3469 --add-label <label>