Repository navigation
[kimi k3][pd disagg] support pp prefill + dcp decode with dspark - #40045
Conversation
YAMY1234
left a comment
There was a problem hiding this comment.
Went through the three parts (K3/kimi_linear PP-aware aux capture, non-last-rank DSparkWorkerV2, draft-head slicing in send_kvcache_dcp). The logic looks consistent for prefill PP + decode DCP, and the deferred boundary-layer capture in K3 checks out. A few comments inline.
Could you also fill in Motivation / Modifications and add the eval numbers (prefill-PP + decode-DCP vs. PP1)? Since this touches the KV transfer path it'd be good to have them on record.
|
/rerun-test test/registered/unit/managers/test_disagg_idle_step_counters.py |
|
Results for 🚀 |
|
/rerun-test test/registered/unit/managers/test_disagg_idle_step_counters.py |
|
Results for 🚀 |
|
/rerun-test test/registered/unit/managers/test_disagg_idle_step_counters.py |
|
Results for 🚀 |
Motivation
Modifications
Accuracy Tests
aime26 verified
Speed Tests and Profiling
Checklist
Review and Merge Process
/tag-and-rerun-ci,/tag-run-ci-label,/rerun-failed-ciCI States
Latest PR Test (Base): 🚫 Run #35496307454
Latest PR Test (Extra): ❌ Run #35496307325
Latest PR Test (AMD ROCm 10): 🚫 Run #35496307450