puyuan-yang commited on
Commit
0a52d38
·
verified ·
1 Parent(s): b869a6a

README: source, validated platform, baseline table, AVO version/config and reference results; add baseline_mi308x.csv and AVO task template

Browse files
MANIFEST.sha256 CHANGED
@@ -1,26 +1,28 @@
 
 
 
1
  332a949a82a0ce983290f9db1b121e1ebc76fc75229cd185b4ae875ba3a67d79 ./adapters/glm-5.3-flash__fused_moe_kernel.py
2
  6f6eaa3835f38da5c37ff54114dbaa4adab54ad01f393cf6cbb38a8ca6907e8d ./adapters/kimi-k3__fwd_grouped_kernel_stage1.py
3
  dd5c35f14114fc6fa35dd899c31c38d97d7716f132218a8d9a23f5de81b07071 ./adapters/minimax-m3__gqa_share_sparse_fwd_kernel.py
4
  97904137ad7fae2c9cb8ece0f77ff50b3314712c7eabc036fefa52418ec70e2c ./docs/2026-10-26_headkernel-mi308-筛选报告.md
5
  01fbf829552aea14b0c3b68e65aae0af43507d6fce29a4b057e2fd4f70e6e155 ./docs/2026-10-26_headkernel-mi308-筛选方案.md
 
6
  b87b9ace0639c8cbaf2ec00ae607db311074a41219d3673ce05683944ff09828 ./docs/screening_WORKFLOW.md
7
  f169fdd6d32cdd900872c48878d681dd37b44a7e129a3a3b2f0d862386086924 ./harness/child.py
8
  9e0f66c7ab2bc993c997b074bf9f32aa26d8271c16863bdb9a6b47999ecb58b0 ./harness/run.py
9
- 43070e2d4e532684de521b885f385d0841030efa2b1a20bafb76133a5e1379c1 ./LICENSE
 
10
  172dc74a35e1752df75ecfb2b2cf9326d2852bb1379868ebeec9571654489679 ./model_cfg/GLM-5.2-FP8-config/chat_template.jinja
11
  d1539d36be7546a1d827fe9cf74c55874695652efb6a5aaa3e60cde1c76ba819 ./model_cfg/GLM-5.2-FP8-config/config.json
12
  ac76b43d8683d3b930126870fc8be73d8679308fe752fa1f381096d8354f6a55 ./model_cfg/GLM-5.2-FP8-config/generation_config.json
13
- f4a18c6ae40b0a8e7d2b7667f52f6e1994e54a46430d2e172b73cb8c9b5eb0d7 ./model_cfg/GLM-5.2-FP8-config/LICENSE
14
- de23c1b7cab43a99f0fedf4edf10de0d165882a2ecb2e2bad6c5796fcabf2e46 ./model_cfg/GLM-5.2-FP8-config/README.md
15
- 98b1271574f41abf89427ae2dda030d94dc9478f0edc5a8bd240db213c6fd5fc ./model_cfg/GLM-5.2-FP8-config/tokenizer_config.json
16
  19e773648cb4e65de8660ea6365e10acca112d42a854923df93db4a6f333a82d ./model_cfg/GLM-5.2-FP8-config/tokenizer.json
17
- f2bc19f9842ebe2fed45ed8ed401275a46d01cc5af829c1684616ae2dcc0b6d2 ./NOTICE
18
- 630ea78fe4b8ebfb9b4695d0e1b81ab692403c4696e2b4246432730b13fd7cd6 ./README.md
19
  d4adecea1b2384e7626a8deb4bdb61c9585669cf71506904429371c669088c31 ./refs/glm-5.3-flash__fused_moe_kernel.py
20
  bdb7692f6befc224627c9ca16863b5ec3ee00b9346932f61a256a81d7728a774 ./refs/kimi-k3__fwd_grouped_kernel_stage1.py
21
  90fd66b30da37f06dab71497728f0ffdc63465984270f9fe7cfb060d757b77d4 ./refs/minimax-m3__gqa_share_sparse_fwd_kernel.py
22
  de0505d47e3131bb43e5c8b12e1969fdc14a7e70c2a2153f785bb1ae63d8ee64 ./refs/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode.py
23
  fdc4a8355670a69fb3af57d30c276109ceebc6a89eaf2f0e557b055d0aa971ba ./refs/qwen3.8-2.4t__gemma_fused_add_rmsnorm.py
 
24
  0b643b1f087e7bf7ab6cedf6c70cd4df7fdd6912dd42868bc73303fe40a87a7a ./results/screening_20261026/final_run.log
25
  b542b90f8a45a77fca7293ce57e35e2f3921f22bb6facd16fea8bff3d2f2c494 ./results/screening_20261026/logs/deepseek-v4-pro__dsa_sparse_mla_attn.log
26
  7a6b82415d904802684ba8565fa2f40b9de0b6e971fb1fecd92e8621ee45c547 ./results/screening_20261026/logs/deepseek-v4-pro__moe_stage1_grouped_gemm_silu_flydsl.log
@@ -38,12 +40,12 @@ f4110a19550c35a8dca96a430fc80b9306754206799779cdfce5758bdd7d3906 ./results/scre
38
  25bd0384a1452053b2fd8abba68c73024fcee5c6b5b4c657d3f2a74712eac124 ./results/screening_20261026/logs/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode.log
39
  3a96d01b568564a055cd3bdc70e9f887073ae1b6ac61d887c2fdf5adff99653e ./results/screening_20261026/logs/qwen3.8-2.4t__gemma_fused_add_rmsnorm.log
40
  10321407f8fc7172c9d5e5f7ef4d5ad18c42171929c4845a42624ce92bbe5195 ./results/screening_20261026/logs/qwen3.8-2.4t__paged_attention_decode.log
 
41
  2936be32a4f279046bb6d807bd9727621680b080dd82c2c2666fcd9038fa0f21 ./results/screening_20261026/results/deepseek-v4-pro__dsa_sparse_mla_attn.json
42
  2b6660d6c29656d14dae6d16e6072707cd8c6fa9ffb7e4731e8c4a2d1d3de815 ./results/screening_20261026/results/deepseek-v4-pro__moe_stage1_grouped_gemm_silu_flydsl.json
43
  1c0beae123528156c76ca01d29f76b47bcb559c0270b94d5f63624353ec2c7ab ./results/screening_20261026/results/deepseek-v4-pro__moe_stage2_down_proj_reduce_opus_a8w4.json
44
  bc6f78c0ab5e9308ee8cdda09f06d1d4cbe6de5d827a445fe8d61ca164a72cb5 ./results/screening_20261026/results/glm-5.3-flash__elementwise_copy_cluster.json
45
  d010b336662cf38023da409a36ccc09885228b918f27177950a9d04078848de0 ./results/screening_20261026/results/glm-5.3-flash__fused_moe_kernel.json
46
- 49ceca9213adbeea3d524d6f137fc0aea80dd380455acd448ef9dbeda627cf09 ./results/screening_20261026/results.json
47
  eae3913f406adb11a87f7920da66bc9d0eb8676abf44f921ed18cfe98362a0b1 ./results/screening_20261026/results/kimi-k3__fwd_grouped_kernel_stage1.json
48
  c1a5be2fbed0a62ed986e44ec4113a5fd7c11d2c4e4ddf80c057ad1c45fa052b ./results/screening_20261026/results/kimi-k3__moe_gemm1_stage1.json
49
  c53cd3a3d28268201cf4c9b855ee2f0a6aa649981ff4209bc41d84d1d58c599a ./results/screening_20261026/results/kimi-k3__moe_gemm2_stage2.json
@@ -180,12 +182,13 @@ e0ade21d25f49a77fc9342f5497d52e3675afbf78e9c6a7cd8e2111de5be225d ./results/vali
180
  1eac19a6b7ced90d0b910fa62edca1d6fe494a8836d987210596291e2e0efc28 ./results/validation_20261103_glm_fnuz/perf_3/glm-5.3-flash__fused_moe_kernel.performance.log
181
  3dadde573642525a306b4aac74c9dc53b070aaac226d60ef93b94e8364389897 ./scripts/compare_runs.py
182
  5f2fe07f20e3c63d6981edc1e601d294cc136916c803569da947597ed6243f58 ./scripts/run_in_container.sh
 
183
  88abdf0030b98a3c57547469cb6253cf371a15afbf89632023d807ac9cb89624 ./tasks/glm-5.3-flash__fused_moe_kernel/config.upstream.yaml
184
  715a08adfa21477ef04c731ca4505ea6240c3caf3457cbece6ee698377035458 ./tasks/glm-5.3-flash__fused_moe_kernel/config.yaml
185
- 829185490c5a4adf60fbfc47a7d2101639fc8f28b53bb4eef923fbd77a9fca3d ./tasks/glm-5.3-flash__fused_moe_kernel/README.md
186
  aafe92b3f7c809f96a87930f5dd8e60e1c2d161fdf57debb5a2a1bcd2cc94419 ./tasks/glm-5.3-flash__fused_moe_kernel/scripts/_bench.py
187
  30b267c28419c03a55571e87e4a125161d86926ebf28577287c14a72b25b1630 ./tasks/glm-5.3-flash__fused_moe_kernel/scripts/task_runner.py
188
  4ad1d5a4faad153021c7c765939ef26117e7ba755b555e16e107c11b1b5da2da ./tasks/glm-5.3-flash__fused_moe_kernel/source/fused_moe.py
 
189
  264088409bcb95752aef4e96d7c1ed5222dcba333560b4274971b930b7af5ac6 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/baseline_overlay/_overlay_manifest.json
190
  d8b5c55e926eea158154d30ee1a796144058e50a102c4c762cbd7c0f7203fd5b ./tasks/glm-5.3-flash__fused_moe_kernel/ut/baseline_overlay/sitecustomize.py
191
  9626f13eab79d55865162164117d55c8e7f3d6a06a708d3cd2728de6a71735ed ./tasks/glm-5.3-flash__fused_moe_kernel/ut/baseline_overlay/sitecustomize.py.eager_bak
@@ -199,18 +202,18 @@ a35dde641f8072d042f71dfb52bd16adc2c462f17ee994cb10bbb4e42c73a465 ./tasks/glm-5.
199
  7cbf3d70eb2c2921c9f128c2c072706c3f009de818f9e5c53cbaf8bcfeec71fa ./tasks/glm-5.3-flash__fused_moe_kernel/ut/overlay_setup.py
200
  0343b5f4b5f48aabb93f392f25d278cfe72109bb7655309257b0b6f63519d544 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/profile_roofline.json
201
  aa0238def5af8f368db2e5f79c96d3ca8af14138ede9126377afcdf63930a09a ./tasks/glm-5.3-flash__fused_moe_kernel/ut/profile_roofline.md
202
- 13b22178fc8ecc3341a559d3e016fd4bba038804df4f9a770ae16c2a5e3a17c9 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/README.md
203
  e5e1dcafdfb58e8350dcc00cbf87db71f47d859606c6ec20269898708def07f7 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/regime.json
204
  3dc80ed88b7d491dbc455ae054152fdb567575b5332612f11d46aeddabe535cb ./tasks/glm-5.3-flash__fused_moe_kernel/ut/selection_validation.json
205
  d3895d82bed71271b53d439bb75788f657882bc152331a8caef4b3152c269246 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/sglang_bootstrap.py
206
  6b5e5baa762a0b14359611502dc32ffda7c56d3f3e897fa94c3d001a8e228355 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/unittest.py
207
  9f5beb4e30d9fcda8b174c9663f326700349af2540277fe17d6471a0fab039d3 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/workload.json
 
208
  3a273b102f60fdb32cff7b7e4fb0a8785c421a0389c658fa8bb50fbba50276d8 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/config.upstream.yaml
209
  6c224899fb567dd4488f275bcbe2b1385ff214a85785d7ab274fcd378d056c2c ./tasks/kimi-k3__fwd_grouped_kernel_stage1/config.yaml
210
- e403276b3b337460a6ae1c5a5717e4995fec32745625721eea141e367ea57872 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/README.md
211
  aafe92b3f7c809f96a87930f5dd8e60e1c2d161fdf57debb5a2a1bcd2cc94419 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/scripts/_bench.py
212
  44b9311111588a4c66c5f31df03951c578b05c62af39f96a67717a833ddf6865 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/scripts/task_runner.py
213
  15b0456e621c184d7f11c704001b63822a5c9008f07ee3d374653d9f981d37fe ./tasks/kimi-k3__fwd_grouped_kernel_stage1/source/geak_mla_stage1.py
 
214
  760da1b3f9e3720b4f5343f846c15a57c8ce7ac5631034d69c3dadd5f8bd5b02 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/baseline_overlay/_overlay_manifest.json
215
  c407056e04297d658fa6126d0657624fa685055fba984669a9f14a85a97bb7e5 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/baseline_overlay/_patched/sglang.kernels.ops.attention.decode_attention.py
216
  5680c72bf60dedd7b0d5843b5c3a002ef3a43247f8ac4ca51a4e598dd05a9b84 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/baseline_overlay/_patched/sglang.srt.models.kimi_k3.py
@@ -223,22 +226,22 @@ fe8e18f4cba1313cc5033ae47950fb6154f6680e050d77436f7c5bb4095fd7f2 ./tasks/kimi-k
223
  8477f7d73b440fbb654a9c822389a7d78e3a24e6ace15a062128c903fec02e74 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/opbench_result.json
224
  eee0c5761359b3a3884489ce405bf665b98617d83d32d00a252e20db631f42cf ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/profile_roofline.json
225
  4f115aa44f2a7f7cf4278606a78b4c419773b88aacbd0fec66e0f399dd69f8d3 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/profile_roofline.md
226
- 79cea973f45e4d8aa179402992cf760e4f5c213960285132d2c30ab0148415bd ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/README.md
227
  558991da4f908f27ee37d4afa9a0bd384a34a67ccfc1aff259eac93bf01e2dfb ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/regime.json
228
  83d90056afdd7abdfe006c802e0434a9f0749da096b1e6e75ca2ed30269ed940 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/run_unittest.sh
229
  ac123439ca83a4cb8a4dd4453c781e1cc942da119fc0e0c0d80db565feac5d7d ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/selection_validation.json
230
  9bc5d7f57f3c03f094db2f78d6d738966e16258d23f7b6754fa36205b7f14648 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/unittest.py
231
  99cc27b298cf41abef541491020e6859507d753f895a613f9f1e30fb734d42a2 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/workload.json
 
232
  49252ff803966cc96f7ab434af2a9715f15b99608df318fe1d362a1bae511f5a ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/config.upstream.yaml
233
  3475b5f847418c87f6780a1b5e2803fcba26b4887ef2f2c83fe32e39e5178f7f ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/config.yaml
234
- 244c876ae934db54bf35ba94e09835873925ac426714f6f800145b57a1268d75 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/README.md
235
  aafe92b3f7c809f96a87930f5dd8e60e1c2d161fdf57debb5a2a1bcd2cc94419 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/scripts/_bench.py
236
  30b267c28419c03a55571e87e4a125161d86926ebf28577287c14a72b25b1630 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/scripts/task_runner.py
237
  1bc59d10771c5f4ecf98b0addd50e7de88cc7aeeddb9121532ccd259458c3688 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/source/topk_sparse.py
 
 
238
  264088409bcb95752aef4e96d7c1ed5222dcba333560b4274971b930b7af5ac6 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/baseline_overlay/_overlay_manifest.json
239
  9626f13eab79d55865162164117d55c8e7f3d6a06a708d3cd2728de6a71735ed ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/baseline_overlay/sitecustomize.py
240
  1bc59d10771c5f4ecf98b0addd50e7de88cc7aeeddb9121532ccd259458c3688 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/baseline_ref/topk_sparse.py.orig
241
- 31f5a5f770d84d7411129d2963700768d78bce1a27b709fbed3d586fd50dc892 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/_capture_meta.json
242
  fc1c03c341aa771d1ca1b793f6c64e0112975f336e546360c17caa035d58bcbd ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/cases.py
243
  d196db4cfee8441467a8ef05fb4fead5b79183f02562b8716043bca24ecc9245 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/harness_lib.py
244
  1bc59d10771c5f4ecf98b0addd50e7de88cc7aeeddb9121532ccd259458c3688 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/kernel_src/topk_sparse.py
@@ -247,14 +250,13 @@ a35dde641f8072d042f71dfb52bd16adc2c462f17ee994cb10bbb4e42c73a465 ./tasks/minima
247
  5934423a52035a9f6b68e238dc236e1b5434935dceba2c4281cc6f9833c98525 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/overlay_setup.py
248
  66e43c4908898317cfac79de31419d02d7befc71dc8c289cfd20ece1a43fc24c ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/profile_roofline.json
249
  a23f4bc63c89c8439f6f6000ef3eb2651248fe9c6c1c67ebf50e93be7e7ef109 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/profile_roofline.md
250
- 9ab824bc9303f32ae1766e4c8c5873a97bc67d0c019f5112b41b1b14a50edfb0 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/README.md
251
  7c3216058c69ec1f23d93c8faadeca26e4271fcc67fecc3b4c4e9a80fb48510f ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/regime.json
252
  6b4bb6a8e1ac2fddab318fc4e8c2f99d40a0e03c07d4e871f8a43597fceabf60 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/selection_validation.json
253
  c36cef314dcaaf8dd2ec386182849a31b524dd040deec6a311ff5258d5deb661 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/unittest.py
254
  51d1f540391d9ae95560c51e731da1317e80089f588629bea0ca2d0fc2ab3b20 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/workload.json
 
255
  a15c003fa4143f99cf7c61064f79cd30ffe121477df0e809bfa8d51e819385f6 ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/config.upstream.yaml
256
  504fc0bf19b0f4bad3ca17e58fc6e8a47daec86a3f3f4878352434442f5f95f5 ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/config.yaml
257
- 13fcb149eb67f73137de35bf45091f6578b12888b48c83a93774c2297cbe57ca ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/README.md
258
  aafe92b3f7c809f96a87930f5dd8e60e1c2d161fdf57debb5a2a1bcd2cc94419 ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/scripts/_bench.py
259
  44b9311111588a4c66c5f31df03951c578b05c62af39f96a67717a833ddf6865 ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/scripts/task_runner.py
260
  cf5e980d86174a631bcd0872f8ebf7a6ab7c21c3435c097f66f8ba0e686debc0 ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/source/fused_recurrent.py
@@ -276,9 +278,9 @@ d5a94a4febf017507411b8031c888c019ecb7b4491b7ef737079bd77e8d2bf7e ./tasks/qwen3.
276
  9248a9ae73697c1d5c49c7027271fdd2b36844bfccc072a6b899cd7722cb49ab ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/ut/selection_validation.json
277
  2cd8de513e55e555ca69ea1a2cb3e9ddeb390c7d2c4a20fc566cd29b7c68b47a ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/ut/selection_validation.probe.json
278
  eec5c197c6afa019612c02d4b3150e06a9864964463ddc158dd099d7c9718e37 ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/ut/unittest.py
 
279
  f5a19c4ba6f91374c628248bc022b80d4a07e047c0a6dd2173d60e0076f8c894 ./tasks/qwen3.8-2.4t__gemma_fused_add_rmsnorm/config.upstream.yaml
280
  02cf3eca23ab3482602c5170444eaf410bce860eec52861f8e45d9fbda316e0d ./tasks/qwen3.8-2.4t__gemma_fused_add_rmsnorm/config.yaml
281
- 5a0ee68d605579bebe6b10872aac894b1da41f4da265c6c3afb4149d9b3fc244 ./tasks/qwen3.8-2.4t__gemma_fused_add_rmsnorm/README.md
282
  aafe92b3f7c809f96a87930f5dd8e60e1c2d161fdf57debb5a2a1bcd2cc94419 ./tasks/qwen3.8-2.4t__gemma_fused_add_rmsnorm/scripts/_bench.py
283
  44b9311111588a4c66c5f31df03951c578b05c62af39f96a67717a833ddf6865 ./tasks/qwen3.8-2.4t__gemma_fused_add_rmsnorm/scripts/task_runner.py
284
  e7e81662d1989eacd46b757b0240111ef9caae31a77fb3035ca2b9196307198d ./tasks/qwen3.8-2.4t__gemma_fused_add_rmsnorm/source/minimax_m3_rmsnorm.py
 
1
+ 43070e2d4e532684de521b885f385d0841030efa2b1a20bafb76133a5e1379c1 ./LICENSE
2
+ f2bc19f9842ebe2fed45ed8ed401275a46d01cc5af829c1684616ae2dcc0b6d2 ./NOTICE
3
+ 1958a65b3044c6223cfc5261f19b1dfadc5c72edf20ed85e81f7731d8e0fd592 ./README.md
4
  332a949a82a0ce983290f9db1b121e1ebc76fc75229cd185b4ae875ba3a67d79 ./adapters/glm-5.3-flash__fused_moe_kernel.py
5
  6f6eaa3835f38da5c37ff54114dbaa4adab54ad01f393cf6cbb38a8ca6907e8d ./adapters/kimi-k3__fwd_grouped_kernel_stage1.py
6
  dd5c35f14114fc6fa35dd899c31c38d97d7716f132218a8d9a23f5de81b07071 ./adapters/minimax-m3__gqa_share_sparse_fwd_kernel.py
7
  97904137ad7fae2c9cb8ece0f77ff50b3314712c7eabc036fefa52418ec70e2c ./docs/2026-10-26_headkernel-mi308-筛选报告.md
8
  01fbf829552aea14b0c3b68e65aae0af43507d6fce29a4b057e2fd4f70e6e155 ./docs/2026-10-26_headkernel-mi308-筛选方案.md
9
+ 9e3b8109bff08ddc68bbde6ca6528ce164d1a9ef1ea16468b4ff0ff986101afa ./docs/avo_task_template.md
10
  b87b9ace0639c8cbaf2ec00ae607db311074a41219d3673ce05683944ff09828 ./docs/screening_WORKFLOW.md
11
  f169fdd6d32cdd900872c48878d681dd37b44a7e129a3a3b2f0d862386086924 ./harness/child.py
12
  9e0f66c7ab2bc993c997b074bf9f32aa26d8271c16863bdb9a6b47999ecb58b0 ./harness/run.py
13
+ f4a18c6ae40b0a8e7d2b7667f52f6e1994e54a46430d2e172b73cb8c9b5eb0d7 ./model_cfg/GLM-5.2-FP8-config/LICENSE
14
+ de23c1b7cab43a99f0fedf4edf10de0d165882a2ecb2e2bad6c5796fcabf2e46 ./model_cfg/GLM-5.2-FP8-config/README.md
15
  172dc74a35e1752df75ecfb2b2cf9326d2852bb1379868ebeec9571654489679 ./model_cfg/GLM-5.2-FP8-config/chat_template.jinja
16
  d1539d36be7546a1d827fe9cf74c55874695652efb6a5aaa3e60cde1c76ba819 ./model_cfg/GLM-5.2-FP8-config/config.json
17
  ac76b43d8683d3b930126870fc8be73d8679308fe752fa1f381096d8354f6a55 ./model_cfg/GLM-5.2-FP8-config/generation_config.json
 
 
 
18
  19e773648cb4e65de8660ea6365e10acca112d42a854923df93db4a6f333a82d ./model_cfg/GLM-5.2-FP8-config/tokenizer.json
19
+ 98b1271574f41abf89427ae2dda030d94dc9478f0edc5a8bd240db213c6fd5fc ./model_cfg/GLM-5.2-FP8-config/tokenizer_config.json
 
20
  d4adecea1b2384e7626a8deb4bdb61c9585669cf71506904429371c669088c31 ./refs/glm-5.3-flash__fused_moe_kernel.py
21
  bdb7692f6befc224627c9ca16863b5ec3ee00b9346932f61a256a81d7728a774 ./refs/kimi-k3__fwd_grouped_kernel_stage1.py
22
  90fd66b30da37f06dab71497728f0ffdc63465984270f9fe7cfb060d757b77d4 ./refs/minimax-m3__gqa_share_sparse_fwd_kernel.py
23
  de0505d47e3131bb43e5c8b12e1969fdc14a7e70c2a2153f785bb1ae63d8ee64 ./refs/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode.py
24
  fdc4a8355670a69fb3af57d30c276109ceebc6a89eaf2f0e557b055d0aa971ba ./refs/qwen3.8-2.4t__gemma_fused_add_rmsnorm.py
25
+ 34351e1457012b6d26dc58e4dddd8a6ebef9737d21d33ca361f790aa7a7ce979 ./results/baseline_mi308x.csv
26
  0b643b1f087e7bf7ab6cedf6c70cd4df7fdd6912dd42868bc73303fe40a87a7a ./results/screening_20261026/final_run.log
27
  b542b90f8a45a77fca7293ce57e35e2f3921f22bb6facd16fea8bff3d2f2c494 ./results/screening_20261026/logs/deepseek-v4-pro__dsa_sparse_mla_attn.log
28
  7a6b82415d904802684ba8565fa2f40b9de0b6e971fb1fecd92e8621ee45c547 ./results/screening_20261026/logs/deepseek-v4-pro__moe_stage1_grouped_gemm_silu_flydsl.log
 
40
  25bd0384a1452053b2fd8abba68c73024fcee5c6b5b4c657d3f2a74712eac124 ./results/screening_20261026/logs/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode.log
41
  3a96d01b568564a055cd3bdc70e9f887073ae1b6ac61d887c2fdf5adff99653e ./results/screening_20261026/logs/qwen3.8-2.4t__gemma_fused_add_rmsnorm.log
42
  10321407f8fc7172c9d5e5f7ef4d5ad18c42171929c4845a42624ce92bbe5195 ./results/screening_20261026/logs/qwen3.8-2.4t__paged_attention_decode.log
43
+ 49ceca9213adbeea3d524d6f137fc0aea80dd380455acd448ef9dbeda627cf09 ./results/screening_20261026/results.json
44
  2936be32a4f279046bb6d807bd9727621680b080dd82c2c2666fcd9038fa0f21 ./results/screening_20261026/results/deepseek-v4-pro__dsa_sparse_mla_attn.json
45
  2b6660d6c29656d14dae6d16e6072707cd8c6fa9ffb7e4731e8c4a2d1d3de815 ./results/screening_20261026/results/deepseek-v4-pro__moe_stage1_grouped_gemm_silu_flydsl.json
46
  1c0beae123528156c76ca01d29f76b47bcb559c0270b94d5f63624353ec2c7ab ./results/screening_20261026/results/deepseek-v4-pro__moe_stage2_down_proj_reduce_opus_a8w4.json
47
  bc6f78c0ab5e9308ee8cdda09f06d1d4cbe6de5d827a445fe8d61ca164a72cb5 ./results/screening_20261026/results/glm-5.3-flash__elementwise_copy_cluster.json
48
  d010b336662cf38023da409a36ccc09885228b918f27177950a9d04078848de0 ./results/screening_20261026/results/glm-5.3-flash__fused_moe_kernel.json
 
49
  eae3913f406adb11a87f7920da66bc9d0eb8676abf44f921ed18cfe98362a0b1 ./results/screening_20261026/results/kimi-k3__fwd_grouped_kernel_stage1.json
50
  c1a5be2fbed0a62ed986e44ec4113a5fd7c11d2c4e4ddf80c057ad1c45fa052b ./results/screening_20261026/results/kimi-k3__moe_gemm1_stage1.json
51
  c53cd3a3d28268201cf4c9b855ee2f0a6aa649981ff4209bc41d84d1d58c599a ./results/screening_20261026/results/kimi-k3__moe_gemm2_stage2.json
 
182
  1eac19a6b7ced90d0b910fa62edca1d6fe494a8836d987210596291e2e0efc28 ./results/validation_20261103_glm_fnuz/perf_3/glm-5.3-flash__fused_moe_kernel.performance.log
183
  3dadde573642525a306b4aac74c9dc53b070aaac226d60ef93b94e8364389897 ./scripts/compare_runs.py
184
  5f2fe07f20e3c63d6981edc1e601d294cc136916c803569da947597ed6243f58 ./scripts/run_in_container.sh
185
+ 829185490c5a4adf60fbfc47a7d2101639fc8f28b53bb4eef923fbd77a9fca3d ./tasks/glm-5.3-flash__fused_moe_kernel/README.md
186
  88abdf0030b98a3c57547469cb6253cf371a15afbf89632023d807ac9cb89624 ./tasks/glm-5.3-flash__fused_moe_kernel/config.upstream.yaml
187
  715a08adfa21477ef04c731ca4505ea6240c3caf3457cbece6ee698377035458 ./tasks/glm-5.3-flash__fused_moe_kernel/config.yaml
 
188
  aafe92b3f7c809f96a87930f5dd8e60e1c2d161fdf57debb5a2a1bcd2cc94419 ./tasks/glm-5.3-flash__fused_moe_kernel/scripts/_bench.py
189
  30b267c28419c03a55571e87e4a125161d86926ebf28577287c14a72b25b1630 ./tasks/glm-5.3-flash__fused_moe_kernel/scripts/task_runner.py
190
  4ad1d5a4faad153021c7c765939ef26117e7ba755b555e16e107c11b1b5da2da ./tasks/glm-5.3-flash__fused_moe_kernel/source/fused_moe.py
191
+ 13b22178fc8ecc3341a559d3e016fd4bba038804df4f9a770ae16c2a5e3a17c9 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/README.md
192
  264088409bcb95752aef4e96d7c1ed5222dcba333560b4274971b930b7af5ac6 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/baseline_overlay/_overlay_manifest.json
193
  d8b5c55e926eea158154d30ee1a796144058e50a102c4c762cbd7c0f7203fd5b ./tasks/glm-5.3-flash__fused_moe_kernel/ut/baseline_overlay/sitecustomize.py
194
  9626f13eab79d55865162164117d55c8e7f3d6a06a708d3cd2728de6a71735ed ./tasks/glm-5.3-flash__fused_moe_kernel/ut/baseline_overlay/sitecustomize.py.eager_bak
 
202
  7cbf3d70eb2c2921c9f128c2c072706c3f009de818f9e5c53cbaf8bcfeec71fa ./tasks/glm-5.3-flash__fused_moe_kernel/ut/overlay_setup.py
203
  0343b5f4b5f48aabb93f392f25d278cfe72109bb7655309257b0b6f63519d544 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/profile_roofline.json
204
  aa0238def5af8f368db2e5f79c96d3ca8af14138ede9126377afcdf63930a09a ./tasks/glm-5.3-flash__fused_moe_kernel/ut/profile_roofline.md
 
205
  e5e1dcafdfb58e8350dcc00cbf87db71f47d859606c6ec20269898708def07f7 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/regime.json
206
  3dc80ed88b7d491dbc455ae054152fdb567575b5332612f11d46aeddabe535cb ./tasks/glm-5.3-flash__fused_moe_kernel/ut/selection_validation.json
207
  d3895d82bed71271b53d439bb75788f657882bc152331a8caef4b3152c269246 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/sglang_bootstrap.py
208
  6b5e5baa762a0b14359611502dc32ffda7c56d3f3e897fa94c3d001a8e228355 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/unittest.py
209
  9f5beb4e30d9fcda8b174c9663f326700349af2540277fe17d6471a0fab039d3 ./tasks/glm-5.3-flash__fused_moe_kernel/ut/workload.json
210
+ e403276b3b337460a6ae1c5a5717e4995fec32745625721eea141e367ea57872 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/README.md
211
  3a273b102f60fdb32cff7b7e4fb0a8785c421a0389c658fa8bb50fbba50276d8 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/config.upstream.yaml
212
  6c224899fb567dd4488f275bcbe2b1385ff214a85785d7ab274fcd378d056c2c ./tasks/kimi-k3__fwd_grouped_kernel_stage1/config.yaml
 
213
  aafe92b3f7c809f96a87930f5dd8e60e1c2d161fdf57debb5a2a1bcd2cc94419 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/scripts/_bench.py
214
  44b9311111588a4c66c5f31df03951c578b05c62af39f96a67717a833ddf6865 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/scripts/task_runner.py
215
  15b0456e621c184d7f11c704001b63822a5c9008f07ee3d374653d9f981d37fe ./tasks/kimi-k3__fwd_grouped_kernel_stage1/source/geak_mla_stage1.py
216
+ 79cea973f45e4d8aa179402992cf760e4f5c213960285132d2c30ab0148415bd ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/README.md
217
  760da1b3f9e3720b4f5343f846c15a57c8ce7ac5631034d69c3dadd5f8bd5b02 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/baseline_overlay/_overlay_manifest.json
218
  c407056e04297d658fa6126d0657624fa685055fba984669a9f14a85a97bb7e5 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/baseline_overlay/_patched/sglang.kernels.ops.attention.decode_attention.py
219
  5680c72bf60dedd7b0d5843b5c3a002ef3a43247f8ac4ca51a4e598dd05a9b84 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/baseline_overlay/_patched/sglang.srt.models.kimi_k3.py
 
226
  8477f7d73b440fbb654a9c822389a7d78e3a24e6ace15a062128c903fec02e74 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/opbench_result.json
227
  eee0c5761359b3a3884489ce405bf665b98617d83d32d00a252e20db631f42cf ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/profile_roofline.json
228
  4f115aa44f2a7f7cf4278606a78b4c419773b88aacbd0fec66e0f399dd69f8d3 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/profile_roofline.md
 
229
  558991da4f908f27ee37d4afa9a0bd384a34a67ccfc1aff259eac93bf01e2dfb ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/regime.json
230
  83d90056afdd7abdfe006c802e0434a9f0749da096b1e6e75ca2ed30269ed940 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/run_unittest.sh
231
  ac123439ca83a4cb8a4dd4453c781e1cc942da119fc0e0c0d80db565feac5d7d ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/selection_validation.json
232
  9bc5d7f57f3c03f094db2f78d6d738966e16258d23f7b6754fa36205b7f14648 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/unittest.py
233
  99cc27b298cf41abef541491020e6859507d753f895a613f9f1e30fb734d42a2 ./tasks/kimi-k3__fwd_grouped_kernel_stage1/ut/workload.json
234
+ 244c876ae934db54bf35ba94e09835873925ac426714f6f800145b57a1268d75 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/README.md
235
  49252ff803966cc96f7ab434af2a9715f15b99608df318fe1d362a1bae511f5a ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/config.upstream.yaml
236
  3475b5f847418c87f6780a1b5e2803fcba26b4887ef2f2c83fe32e39e5178f7f ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/config.yaml
 
237
  aafe92b3f7c809f96a87930f5dd8e60e1c2d161fdf57debb5a2a1bcd2cc94419 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/scripts/_bench.py
238
  30b267c28419c03a55571e87e4a125161d86926ebf28577287c14a72b25b1630 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/scripts/task_runner.py
239
  1bc59d10771c5f4ecf98b0addd50e7de88cc7aeeddb9121532ccd259458c3688 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/source/topk_sparse.py
240
+ 9ab824bc9303f32ae1766e4c8c5873a97bc67d0c019f5112b41b1b14a50edfb0 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/README.md
241
+ 31f5a5f770d84d7411129d2963700768d78bce1a27b709fbed3d586fd50dc892 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/_capture_meta.json
242
  264088409bcb95752aef4e96d7c1ed5222dcba333560b4274971b930b7af5ac6 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/baseline_overlay/_overlay_manifest.json
243
  9626f13eab79d55865162164117d55c8e7f3d6a06a708d3cd2728de6a71735ed ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/baseline_overlay/sitecustomize.py
244
  1bc59d10771c5f4ecf98b0addd50e7de88cc7aeeddb9121532ccd259458c3688 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/baseline_ref/topk_sparse.py.orig
 
245
  fc1c03c341aa771d1ca1b793f6c64e0112975f336e546360c17caa035d58bcbd ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/cases.py
246
  d196db4cfee8441467a8ef05fb4fead5b79183f02562b8716043bca24ecc9245 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/harness_lib.py
247
  1bc59d10771c5f4ecf98b0addd50e7de88cc7aeeddb9121532ccd259458c3688 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/kernel_src/topk_sparse.py
 
250
  5934423a52035a9f6b68e238dc236e1b5434935dceba2c4281cc6f9833c98525 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/overlay_setup.py
251
  66e43c4908898317cfac79de31419d02d7befc71dc8c289cfd20ece1a43fc24c ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/profile_roofline.json
252
  a23f4bc63c89c8439f6f6000ef3eb2651248fe9c6c1c67ebf50e93be7e7ef109 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/profile_roofline.md
 
253
  7c3216058c69ec1f23d93c8faadeca26e4271fcc67fecc3b4c4e9a80fb48510f ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/regime.json
254
  6b4bb6a8e1ac2fddab318fc4e8c2f99d40a0e03c07d4e871f8a43597fceabf60 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/selection_validation.json
255
  c36cef314dcaaf8dd2ec386182849a31b524dd040deec6a311ff5258d5deb661 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/unittest.py
256
  51d1f540391d9ae95560c51e731da1317e80089f588629bea0ca2d0fc2ab3b20 ./tasks/minimax-m3__gqa_share_sparse_fwd_kernel/ut/workload.json
257
+ 13fcb149eb67f73137de35bf45091f6578b12888b48c83a93774c2297cbe57ca ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/README.md
258
  a15c003fa4143f99cf7c61064f79cd30ffe121477df0e809bfa8d51e819385f6 ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/config.upstream.yaml
259
  504fc0bf19b0f4bad3ca17e58fc6e8a47daec86a3f3f4878352434442f5f95f5 ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/config.yaml
 
260
  aafe92b3f7c809f96a87930f5dd8e60e1c2d161fdf57debb5a2a1bcd2cc94419 ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/scripts/_bench.py
261
  44b9311111588a4c66c5f31df03951c578b05c62af39f96a67717a833ddf6865 ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/scripts/task_runner.py
262
  cf5e980d86174a631bcd0872f8ebf7a6ab7c21c3435c097f66f8ba0e686debc0 ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/source/fused_recurrent.py
 
278
  9248a9ae73697c1d5c49c7027271fdd2b36844bfccc072a6b899cd7722cb49ab ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/ut/selection_validation.json
279
  2cd8de513e55e555ca69ea1a2cb3e9ddeb390c7d2c4a20fc566cd29b7c68b47a ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/ut/selection_validation.probe.json
280
  eec5c197c6afa019612c02d4b3150e06a9864964463ddc158dd099d7c9718e37 ./tasks/qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode/ut/unittest.py
281
+ 5a0ee68d605579bebe6b10872aac894b1da41f4da265c6c3afb4149d9b3fc244 ./tasks/qwen3.8-2.4t__gemma_fused_add_rmsnorm/README.md
282
  f5a19c4ba6f91374c628248bc022b80d4a07e047c0a6dd2173d60e0076f8c894 ./tasks/qwen3.8-2.4t__gemma_fused_add_rmsnorm/config.upstream.yaml
283
  02cf3eca23ab3482602c5170444eaf410bce860eec52861f8e45d9fbda316e0d ./tasks/qwen3.8-2.4t__gemma_fused_add_rmsnorm/config.yaml
 
284
  aafe92b3f7c809f96a87930f5dd8e60e1c2d161fdf57debb5a2a1bcd2cc94419 ./tasks/qwen3.8-2.4t__gemma_fused_add_rmsnorm/scripts/_bench.py
285
  44b9311111588a4c66c5f31df03951c578b05c62af39f96a67717a833ddf6865 ./tasks/qwen3.8-2.4t__gemma_fused_add_rmsnorm/scripts/task_runner.py
286
  e7e81662d1989eacd46b757b0240111ef9caae31a77fb3035ca2b9196307198d ./tasks/qwen3.8-2.4t__gemma_fused_add_rmsnorm/source/minimax_m3_rmsnorm.py
README.md CHANGED
@@ -12,6 +12,19 @@ Intended as the starting point for source-level kernel optimization (AVO).
12
  | `qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode` | Triton gated delta-rule state update | 2 |
13
  | `qwen3.8-2.4t__gemma_fused_add_rmsnorm` | Triton fused add + Gemma RMSNorm | 2 |
14
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15
  ## Requirements
16
 
17
  - MI308X (gfx942) GPU, Docker with access to `/dev/kfd` and `/dev/dri`.
@@ -22,6 +35,15 @@ Intended as the starting point for source-level kernel optimization (AVO).
22
  docker pull lmsysorg/sglang@sha256:2a4b6896f2d69e5562545b5abaaf00e8d8dbb14da736fa2f5f5bd47ce00bfb0f
23
  ```
24
 
 
 
 
 
 
 
 
 
 
25
  ## Run
26
 
27
  ```bash
@@ -78,6 +100,81 @@ and block graph capture. If a task requests graph timing and capture fails, the
78
  `HK_WARNING graph capture failed ...` and the per-shape JSON records `timer=cuda_event` instead of
79
  `cuda_event_graph`.
80
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
81
  ## Differences from the upstream tasks
82
 
83
  - No `reference_io.pt`: the upstream captured tensors are in non-public OCI storage. Synthetic inputs
@@ -106,6 +203,6 @@ refs/ PyTorch references
106
  harness/ run.py (entry) and child.py (one task per fresh process)
107
  scripts/ run_in_container.sh, compare_runs.py (timing spread across runs)
108
  model_cfg/ GLM configuration files used by the GLM task
109
- results/ screening_20261026 (16 tasks, rocm/atom image) and validation runs
110
- docs/ screening plan/report and packaging plan (Chinese)
111
  ```
 
12
  | `qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode` | Triton gated delta-rule state update | 2 |
13
  | `qwen3.8-2.4t__gemma_fused_add_rmsnorm` | Triton fused add + Gemma RMSNorm | 2 |
14
 
15
+ ## Source
16
+
17
+ - Tasks: [AMD-AGI/AgentKernelArena](https://github.com/AMD-AGI/AgentKernelArena), branch
18
+ `head_kernels_top5_20260920`, commit `8fa0d73c81e55100d8e32d1963d6b247764fe56a` (Apache-2.0).
19
+ The upstream tasks target MI355X (gfx950), an internal MI355 image, and non-public captured inputs
20
+ (`reference_io.pt`).
21
+ - Selection: 16 upstream tasks were screened on MI308X; 10 run and pass, 6 need gfx950-only instructions
22
+ (FP4, scaled MFMA) or do not compile for gfx942. Five of the 10 are packaged here
23
+ (`docs/2026-10-26_headkernel-mi308-筛选报告.md`).
24
+ - Adaptation for MI308X: seeded synthetic inputs, independent PyTorch references, a negative control, a
25
+ public runtime image, and the changes listed in "Differences from the upstream tasks". The kernel sources
26
+ (`tasks/<task>/source/`) and the upstream tests (`tasks/<task>/ut/`) are unchanged.
27
+
28
  ## Requirements
29
 
30
  - MI308X (gfx942) GPU, Docker with access to `/dev/kfd` and `/dev/dri`.
 
35
  docker pull lmsysorg/sglang@sha256:2a4b6896f2d69e5562545b5abaaf00e8d8dbb14da736fa2f5f5bd47ce00bfb0f
36
  ```
37
 
38
+ ## Validated platform
39
+
40
+ | Item | Value |
41
+ |---|---|
42
+ | GPU | AMD Instinct MI308X (gfx942), 80 CUs, max 1850 MHz, 192 GB HBM3 |
43
+ | Host | Linux 5.15, amdgpu driver 6.16.13, ROCm 7.2.0 |
44
+ | Image | `lmsysorg/sglang@sha256:2a4b6896f2d69e5562545b5abaaf00e8d8dbb14da736fa2f5f5bd47ce00bfb0f` (`versions.lock`) |
45
+ | GPU selection | `ROCR_VISIBLE_DEVICES` (`HIP_VISIBLE_DEVICES` does not filter HSA agents) |
46
+
47
  ## Run
48
 
49
  ```bash
 
100
  `HK_WARNING graph capture failed ...` and the per-shape JSON records `timer=cuda_event` instead of
101
  `cuda_event_graph`.
102
 
103
+ ## Baseline (original kernels)
104
+
105
+ Medians of three performance runs on one idle MI308X (`results/baseline_mi308x.csv`, one row per shape,
106
+ with the run-to-run spread and the source run). GLM uses the FNUZ-weight run of 2026-11-03; the other
107
+ tasks use the fixed-timing run of 2026-10-29.
108
+
109
+ | Task | Shapes | Timer | Fastest shape (ms) | Slowest shape (ms) | Geomean (ms) |
110
+ |---|---:|---|---:|---:|---:|
111
+ | `glm-5.3-flash__fused_moe_kernel` | 4 | eager | 0.537 | 6.310 | 1.085 |
112
+ | `minimax-m3__gqa_share_sparse_fwd_kernel` | 38 | graph | 0.025 | 26.25 | 15.80 |
113
+ | `kimi-k3__fwd_grouped_kernel_stage1` | 4 | graph | 0.037 | 2.523 | 0.351 |
114
+ | `qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode` | 2 | graph | 0.258 | 0.310 | 0.283 |
115
+ | `qwen3.8-2.4t__gemma_fused_add_rmsnorm` | 2 | graph | 0.017 | 0.262 | 0.067 |
116
+
117
+ How to compare:
118
+
119
+ - Absolute latencies depend on the node and on load from neighbouring GPUs (differences of tens of
120
+ percent have been seen). Use the table to check that a setup is sane (within about 10%), not as a
121
+ target.
122
+ - To measure a speedup, time the original and the candidate on the same GPU in the same session,
123
+ interleaved (A B B A), and report per-shape ratios and their geomean.
124
+ - Shapes of a few tens of microseconds vary by up to about 8% between runs (Kimi `oracle_bs1_ctx395`);
125
+ GLM decode under eager timing had one 39% outlier run (`rand_m19_decode`), so take the median of at
126
+ least three runs.
127
+
128
+ ## AVO reference results
129
+
130
+ The five tasks were optimized at source level with AVO. Optimized kernels are not distributed: rerun AVO
131
+ with the version and configuration below, and verify the result as described here.
132
+
133
+ AVO version:
134
+
135
+ | Item | Value |
136
+ |---|---|
137
+ | Repository | [AMD-BRAIN-Internal/PerfSkills-avo](https://github.com/AMD-BRAIN-Internal/PerfSkills-avo) (AMD internal) |
138
+ | Branch | `GEAK-v4-integrated` |
139
+ | Commit | `01b60f937a24b3235f2ea897fda9bdc868f0d9bf` |
140
+ | Entry | `python3 avo-workflow/run_geak_avo.py handoff.json result.json` (validate first with `--dry-run`) |
141
+ | Agent | Claude Code 2.1.308 (Claude Opus 5.5), Node 22 |
142
+
143
+ Configuration (one AVO run per task, two tasks in parallel on one 8-GPU node):
144
+
145
+ | Setting | Value |
146
+ |---|---|
147
+ | `operator` | `team` (default `start_policy=diverse`) |
148
+ | GPUs per task | 4 (`gpu_ids` `0,1,2,3` or `4,5,6,7`) |
149
+ | `budget_s` | 43200 (12 h, counted from workflow start; 10% is reserved for finalization) |
150
+ | `kernel_path` | the repository root (AVO copies it; the harness needs `harness/`, `refs/`, `adapters/`) |
151
+ | `hw` | `gfx942` |
152
+ | `apply_to_original` | `false` |
153
+ | `task` | see `docs/avo_task_template.md` |
154
+
155
+ Results (speedup = original latency / optimized latency; per-shape median of five interleaved rounds on
156
+ one GPU, then the geomean over all shapes; every shape passed the correctness check and the negative
157
+ control):
158
+
159
+ | Task | Speedup | Notes |
160
+ |---|---:|---|
161
+ | `minimax-m3__gqa_share_sparse_fwd_kernel` | 4.39x | AVO stopped at 7.7 h with a workflow error after committing `avo-v5`; `avo-v0..avo-v5` was verified. Per shape 0.95x-5.33x |
162
+ | `glm-5.3-flash__fused_moe_kernel` | 3.19x | Full 12 h. Measured on the FNUZ-weight baseline; the 92x measured on the earlier e4m3fn weights was a weight-format artifact. Per shape 1.38x-9.31x |
163
+ | `kimi-k3__fwd_grouped_kernel_stage1` | 2.64x | Second MI308X node. Starts from a prior optimized candidate |
164
+ | `qwen3.8-2.4t__gemma_fused_add_rmsnorm` | 1.26x | Second MI308X node |
165
+ | `qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode` | 1.02x | Second MI308X node |
166
+
167
+ Notes for rerunning AVO on MI308X with this commit:
168
+
169
+ - Tell AVO to use only `harness/run.py`. With a `workload` key in `meta.json`, its benchmark role prefers
170
+ the task's `ut/unittest.py`, which needs the non-public `reference_io.pt`.
171
+ - Tell AVO the GPU has 80 CUs. Its knowledge file reads the CU count with `rocminfo | grep -m1`, which
172
+ returns the CPU core count on this node, and its roofline table for `gfx942` is MI300X.
173
+ - Tasks with many shapes (MiniMax, 38) can stop with `TypeError: (perCase || []).map is not a function`
174
+ after a commit. Committed versions stay in `avo_repo` (tags `avo-vN`) and can be verified.
175
+ - Review every patch: no output caching, no work skipped on repeated inputs or the fixed seed, no
176
+ shape-specialised branches, no lowered precision.
177
+
178
  ## Differences from the upstream tasks
179
 
180
  - No `reference_io.pt`: the upstream captured tensors are in non-public OCI storage. Synthetic inputs
 
203
  harness/ run.py (entry) and child.py (one task per fresh process)
204
  scripts/ run_in_container.sh, compare_runs.py (timing spread across runs)
205
  model_cfg/ GLM configuration files used by the GLM task
206
+ results/ screening_20261026 (16 tasks, rocm/atom image), validation runs, baseline_mi308x.csv
207
+ docs/ screening plan/report (Chinese), avo_task_template.md (AVO task prompt)
208
  ```
docs/avo_task_template.md ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AVO handoff used for the reference results
2
+
3
+ `handoff.json` given to `avo-workflow/run_geak_avo.py` (AMD-BRAIN-Internal/PerfSkills-avo,
4
+ `GEAK-v4-integrated` @ `01b60f937a24`). Replace `<task>`, `<editable files>`, `<gpus>` and the paths.
5
+ `<editable files>` is the task's `config.yaml` `source_file_path`, prefixed with `tasks/<task>/`.
6
+
7
+ ```json
8
+ {
9
+ "kernel_path": "/bench",
10
+ "exp_root": "/runs/<batch>/<task>",
11
+ "task": "<task text below>",
12
+ "root_dir": "/",
13
+ "claude_bin": "claude",
14
+ "claude_home": "/root",
15
+ "gpu_ids": "<gpus, e.g. 0,1,2,3>",
16
+ "budget_s": 43200,
17
+ "operator": "team",
18
+ "hw": "gfx942",
19
+ "apply_to_original": "false"
20
+ }
21
+ ```
22
+
23
+ Task text (one paragraph):
24
+
25
+ ```text
26
+ Target kernel: tasks/<task>. Edit ONLY: <editable files>. Use the existing harness, run from the workspace
27
+ root: correctness = `python3 harness/run.py --task <task> --mode correctness`; performance =
28
+ `python3 harness/run.py --task <task> --mode performance`, which prints one line
29
+ `GEAK_RESULT_LATENCY_MS=<ms> (<shape id>)` per shape. Select the GPU with ROCR_VISIBLE_DEVICES. The timing
30
+ mode (CUDA-graph replay or eager) is fixed by the task's deployment regime; do not change it. Never modify
31
+ harness/, refs/, adapters/, model_cfg/, tasks/*/ut/, tasks/*/config*.yaml, or any other task.
32
+ harness/run.py is the ONLY correctness and timing harness of this benchmark: it checks every shape against
33
+ an independent PyTorch reference with a negative control. Do NOT run or gate on tasks/*/ut/unittest.py,
34
+ ut/leg_runner.py or scripts/task_runner.py: they need non-public captured tensors (reference_io.pt) that are
35
+ not in this package, so they fail or measure something else. Report the unweighted per-shape geomean.
36
+ Profiling: rocprofv3 counter collection (--pmc) has hung for 10+ minutes on this node; wrap every profiler
37
+ call in `timeout 180` and fall back to --kernel-trace if it times out. Forbidden: caching or memoizing
38
+ outputs across calls; skipping work based on input pointers, repeated inputs or the fixed random seed;
39
+ branches specialised to the benchmark shapes only; lowering numeric precision within the tolerance to gain
40
+ speed (e.g. bf16 accumulation instead of fp32, polynomial approximations of exp/tanh/rsqrt). Keep the
41
+ source's numerics. Hardware: AMD Instinct MI308X (gfx942), 80 CUs (NOT the MI300X's 304), 192 GB HBM3. Size
42
+ grids / persistent-program counts from torch.cuda.get_device_properties().multi_processor_count (do NOT use
43
+ `rocminfo | grep -m1 'Compute Unit'`: its first match is the CPU, 112 here), and treat roofline percentages
44
+ printed by AVO as MI300X-based (MI308X: 80 CUs at max 1850 MHz vs 304 at 2100, so the compute ceiling is
45
+ overstated about 4.3x).
46
+ ```
47
+
48
+ For `qwen3.8-2.4t__gemma_fused_add_rmsnorm`, append: `RMSNorm: the norm must stay computed from the
49
+ unrounded FP32 residual sum.`
50
+
51
+ The container mounts this repository at `/bench`, the output directory at `/runs`, and the AVO checkout
52
+ read-only; `root_dir` `/` lets the agents reach all three.
results/baseline_mi308x.csv ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ task,shape,median_ms,spread_3runs,timer,source
2
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m8192_s1,11.017556,0.0015,cuda_event_graph,results/validation_20261029_fixed_timing
3
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s3,21.963873,0.0007,cuda_event_graph,results/validation_20261029_fixed_timing
4
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s5,22.047792,0.0023,cuda_event_graph,results/validation_20261029_fixed_timing
5
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s7,22.117912,0.0011,cuda_event_graph,results/validation_20261029_fixed_timing
6
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m8205_s8,11.280678,0.0022,cuda_event_graph,results/validation_20261029_fixed_timing
7
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s9,22.231274,0.0009,cuda_event_graph,results/validation_20261029_fixed_timing
8
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s11,22.308195,0.0023,cuda_event_graph,results/validation_20261029_fixed_timing
9
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s13,22.391275,0.0004,cuda_event_graph,results/validation_20261029_fixed_timing
10
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s15,22.470837,0.0014,cuda_event_graph,results/validation_20261029_fixed_timing
11
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s17,22.556000,0.0005,cuda_event_graph,results/validation_20261029_fixed_timing
12
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s19,22.608358,0.0013,cuda_event_graph,results/validation_20261029_fixed_timing
13
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s21,22.719198,0.0005,cuda_event_graph,results/validation_20261029_fixed_timing
14
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s23,22.793921,0.0006,cuda_event_graph,results/validation_20261029_fixed_timing
15
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s25,22.876240,0.0010,cuda_event_graph,results/validation_20261029_fixed_timing
16
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s27,22.972521,0.0002,cuda_event_graph,results/validation_20261029_fixed_timing
17
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s29,23.048800,0.0008,cuda_event_graph,results/validation_20261029_fixed_timing
18
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s31,23.117880,0.0021,cuda_event_graph,results/validation_20261029_fixed_timing
19
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s33,23.195835,0.0012,cuda_event_graph,results/validation_20261029_fixed_timing
20
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s35,23.290161,0.0012,cuda_event_graph,results/validation_20261029_fixed_timing
21
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s37,23.367561,0.0005,cuda_event_graph,results/validation_20261029_fixed_timing
22
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s39,23.445204,0.0008,cuda_event_graph,results/validation_20261029_fixed_timing
23
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s41,23.534245,0.0011,cuda_event_graph,results/validation_20261029_fixed_timing
24
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s43,23.619324,0.0008,cuda_event_graph,results/validation_20261029_fixed_timing
25
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s45,23.705284,0.0010,cuda_event_graph,results/validation_20261029_fixed_timing
26
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s47,23.774244,0.0006,cuda_event_graph,results/validation_20261029_fixed_timing
27
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s49,23.876528,0.0006,cuda_event_graph,results/validation_20261029_fixed_timing
28
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s51,23.942368,0.0013,cuda_event_graph,results/validation_20261029_fixed_timing
29
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s53,24.026953,0.0010,cuda_event_graph,results/validation_20261029_fixed_timing
30
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s55,24.112648,0.0003,cuda_event_graph,results/validation_20261029_fixed_timing
31
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s57,24.196936,0.0009,cuda_event_graph,results/validation_20261029_fixed_timing
32
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s59,24.285688,0.0011,cuda_event_graph,results/validation_20261029_fixed_timing
33
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s61,24.368650,0.0002,cuda_event_graph,results/validation_20261029_fixed_timing
34
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m16384_s63,24.440170,0.0001,cuda_event_graph,results/validation_20261029_fixed_timing
35
+ minimax-m3__gqa_share_sparse_fwd_kernel,prefill_m9185_s64,14.988545,0.0003,cuda_event_graph,results/validation_20261029_fixed_timing
36
+ minimax-m3__gqa_share_sparse_fwd_kernel,replay_even_m16384_s63,26.254461,0.0010,cuda_event_graph,results/validation_20261029_fixed_timing
37
+ minimax-m3__gqa_share_sparse_fwd_kernel,replay_skew_m16384_s63,25.001654,0.0007,cuda_event_graph,results/validation_20261029_fixed_timing
38
+ minimax-m3__gqa_share_sparse_fwd_kernel,small_prefill_m1_s1,0.024960,0.0097,cuda_event_graph,results/validation_20261029_fixed_timing
39
+ minimax-m3__gqa_share_sparse_fwd_kernel,small_prefill_m186_s1,0.054040,0.0045,cuda_event_graph,results/validation_20261029_fixed_timing
40
+ kimi-k3__fwd_grouped_kernel_stage1,oracle_bs1_ctx395,0.036600,0.0754,cuda_event_graph,results/validation_20261029_fixed_timing
41
+ kimi-k3__fwd_grouped_kernel_stage1,oracle_bs64_ctx8193,2.387297,0.0001,cuda_event_graph,results/validation_20261029_fixed_timing
42
+ kimi-k3__fwd_grouped_kernel_stage1,decode_bs1_ctx8704,0.068680,0.0058,cuda_event_graph,results/validation_20261029_fixed_timing
43
+ kimi-k3__fwd_grouped_kernel_stage1,decode_bs64_ctx8704,2.523417,0.0007,cuda_event_graph,results/validation_20261029_fixed_timing
44
+ qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode,decode_b64_live,0.309769,0.0121,cuda_event_graph,results/validation_20261029_fixed_timing
45
+ qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode,decode_b1_live,0.258327,0.0116,cuda_event_graph,results/validation_20261029_fixed_timing
46
+ qwen3.8-2.4t__gemma_fused_add_rmsnorm,prefill_m8192_live_shape,0.262127,0.0163,cuda_event_graph,results/validation_20261029_fixed_timing
47
+ qwen3.8-2.4t__gemma_fused_add_rmsnorm,decode_m64_live_shape,0.017040,0.0187,cuda_event_graph,results/validation_20261029_fixed_timing
48
+ glm-5.3-flash__fused_moe_kernel,rand_m19_decode,0.543404,0.3940,cuda_event,results/validation_20261103_glm_fnuz
49
+ glm-5.3-flash__fused_moe_kernel,rand_m1_decode,0.536725,0.0285,cuda_event,results/validation_20261103_glm_fnuz
50
+ glm-5.3-flash__fused_moe_kernel,rand_m8192_prefill,6.310457,0.0009,cuda_event,results/validation_20261103_glm_fnuz
51
+ glm-5.3-flash__fused_moe_kernel,rand_m64_decode,0.754087,0.0053,cuda_event,results/validation_20261103_glm_fnuz