17 Commits

Author SHA1 Message Date
2b7880efa7 fix platform basic test compatibility 2026-07-15 14:25:52 +08:00
e374b14bf9 add moe bucket profiling and record bottlenecks 2026-07-15 11:34:59 +08:00
59e497fdc9 disable profiling trace by default 2026-07-15 04:43:46 +08:00
cb962589e9 revert moe router topk softmax experiment 2026-07-15 04:21:03 +08:00
eea0653ace experiment moe router topk softmax 2026-07-15 04:19:56 +08:00
af375830c7 align platform build config with working submission 2026-07-15 00:49:24 +08:00
2c6cabcc63 optimize moe tiny batch tokenwise path 2026-07-14 21:49:55 +08:00
0597fa6c6d profile decode path and optimize tiny batch moe 2026-07-14 20:44:05 +08:00
1902c81fdd fix issue of loading weight 2026-06-30 09:55:13 +08:00
c84151eef9 fix issues 2026-06-26 12:55:02 +08:00
b5806731e0 some op overhead optimization 2026-06-19 11:19:39 +08:00
50e3a05fb0 fix incorrect MoE step to ensure decoding speed 2026-06-12 11:44:50 +08:00
629f878c28 initial commit for qwen3.6-moe adaptation 2026-06-12 10:10:49 +08:00
2d1ef50992 chunked prefill support and memory opts 2026-06-05 16:03:34 +08:00
8c047a70ea some modifications to ensure 50K context input 2026-06-04 17:56:29 +08:00
3ef8227384 initial version of adding chunked attention, ensuring 20K context 2026-05-29 16:49:33 +08:00
0e89906481 Qwen3.6-27B iluvatar bi-v100 adaptation 2026-05-21 16:37:24 +08:00