Commit Graph

353 Commits

Author SHA1 Message Date
10afe4e2e8 Freeze long-context Frontier profile 2026-07-17 10:30:37 +08:00
b6ef6eeae7 Add long-context attention profile closure 2026-07-17 10:26:21 +08:00
202ae718b3 Freeze batch-aware Frontier ablation 2026-07-17 10:22:56 +08:00
a2cf361ffe Generalize Qwen30 fixed-shape real runner 2026-07-17 10:17:21 +08:00
6e8704d525 Add fixed and exact-trace Frontier surfaces 2026-07-17 10:13:25 +08:00
0c747448b6 Normalize frozen profile CSV line endings 2026-07-17 09:58:26 +08:00
1fa203384f Add batch-aware profile and exact trace preparation 2026-07-17 09:55:09 +08:00
95f4af3d99 Add Frontier fidelity envelope campaign 2026-07-17 09:44:49 +08:00
3a59d5df96 Evaluate Qwen30 prefill simulator fidelity 2026-07-17 03:11:45 +08:00
97c2f34700 Add Qwen30 prefill fidelity experiment 2026-07-17 00:32:01 +08:00
76107d3e87 Evaluate vLLM 0.20 profiles against Frontier 2026-07-16 23:59:10 +08:00
008324e70c Summarize vLLM scheduler step profiles 2026-07-16 23:20:04 +08:00
630de9f573 Freeze vLLM 0.20 profiles and capture trace routing 2026-07-16 23:14:27 +08:00
07b1eb4b75 Initialize TP context for router profiling 2026-07-16 22:54:17 +08:00
414838a799 Profile replicated router without TP group 2026-07-16 22:49:31 +08:00
fcca259475 Profile Qwen MoE router on vLLM 0.20 2026-07-16 22:43:55 +08:00
b8f5f1dcf3 Adapt Frontier RMSNorm profiling to vLLM 0.20 2026-07-16 22:34:23 +08:00
885db8527a Initialize vLLM config for Frontier profiling 2026-07-16 22:29:10 +08:00
9b54075e75 Adapt Frontier RoPE profiling to vLLM 0.20 2026-07-16 22:23:56 +08:00
0e9b64893f Add full Frontier linear profile sweep 2026-07-16 22:18:30 +08:00
48d2c2fb80 Add Frontier linear profiler smoke 2026-07-16 22:11:04 +08:00
32ab584789 Initialize collective profiler with model config 2026-07-16 21:59:58 +08:00
3121e35b0e Profile collective dispatch across tensor sizes 2026-07-16 21:57:44 +08:00
4bf9bdf28f Add full Qwen30 MoE profiling matrix 2026-07-16 21:52:43 +08:00
1fb777fc11 Set vLLM config for collective profiling 2026-07-16 21:48:52 +08:00
4fed4329cf Profile MoE TP-local shards without collectives 2026-07-16 21:41:38 +08:00
9bbfb87a85 Keep KV smoke artifacts separate 2026-07-16 21:40:07 +08:00
9bc38d8851 Profile FA3 KV-cache updates separately 2026-07-16 21:37:01 +08:00
a9aed87518 Add FlashInfer TRTLLM all-reduce profile smoke 2026-07-16 21:34:20 +08:00
5a958077a7 Add exact vLLM 0.20 MoE profile smoke 2026-07-16 21:27:35 +08:00
adaf68badc Add full per-TP FlashAttention profile sweep 2026-07-16 21:24:11 +08:00
a70a1fac71 Retry profile smoke with cu129 vLLM wheel 2026-07-16 21:19:22 +08:00
b3e437acb0 Serialize vLLM attention profile results 2026-07-16 21:17:27 +08:00
2f04971c90 Freeze Qwen30 trace profile support 2026-07-16 21:09:51 +08:00
fb12e3b502 Use local cache for vLLM profile smoke 2026-07-16 21:07:24 +08:00
e9a4be6153 Add vLLM 0.20 per-TP FlashAttention profile smoke 2026-07-16 21:00:27 +08:00
13f0713bf2 Document Frontier fixed-cohort fidelity results 2026-07-16 06:21:40 +08:00
344af3a428 Prepare targeted disputed-boundary trial 2026-07-16 04:16:49 +08:00
def9e24c4d Allow targeted SLO boundary repeats 2026-07-16 04:02:29 +08:00
3db632552b Compare Frontier fidelity across real trials 2026-07-16 03:41:41 +08:00
501ceb8171 Skip null SLOs in boundary repeats 2026-07-16 03:09:19 +08:00
163fb3d9e8 Repeat boundaries for every registered SLO 2026-07-16 02:35:41 +08:00
993cce6143 Refine Frontier rank boundaries before repeats 2026-07-16 01:59:37 +08:00
f7052edd14 Report request-level SLO fidelity for Frontier 2026-07-16 01:42:13 +08:00
678388084c Add fleet job for Frontier boundary repeats 2026-07-16 01:28:23 +08:00
84584c719c Add reverse boundary repeats for Frontier comparison 2026-07-16 00:52:13 +08:00
e47bbf3e76 Compare Frontier rank under explicit objectives 2026-07-15 23:32:07 +08:00
ef5c17e6ec Use writable fleet root on dash0 2026-07-15 23:20:01 +08:00
91869c531c Add blind Qwen235B fleet job 2026-07-15 23:17:43 +08:00
33c91109d7 Require batched Frontier EP equivalence 2026-07-15 22:45:40 +08:00