Gahow Wang gahow
  • Joined on 2026-04-03
gahow pushed to main at gahow/xserv 2026-05-22 06:29:29 +00:00
d8493bd70f phase 12: implement real continuous batching scheduler
7d05ececa0 docs: split Phase 12 and Phase 13 into separate design documents
da043554ba phase 12+13: HTTP API server with OpenAI-compatible endpoint (Milestone ③)
2be27d6d94 perf: GPU transpose/reshape/repeat_kv kernels (eliminate CPU round-trips)
2d48f25e66 phase 11: GPU-resident KV cache
Compare 12 commits »
gahow pushed tag phase0-1 to gahow/xserv 2026-05-22 06:29:29 +00:00
gahow pushed tag phase10 to gahow/xserv 2026-05-22 06:29:29 +00:00
gahow pushed tag phase11 to gahow/xserv 2026-05-22 06:29:29 +00:00
gahow pushed tag phase13 to gahow/xserv 2026-05-22 06:29:29 +00:00
gahow pushed tag phase2 to gahow/xserv 2026-05-22 06:29:29 +00:00
gahow pushed tag phase3 to gahow/xserv 2026-05-22 06:29:29 +00:00
gahow pushed tag phase4 to gahow/xserv 2026-05-22 06:29:29 +00:00
gahow pushed to main at gahow/agentic-kvc 2026-05-22 05:50:27 +00:00
1d2eeb4925 Elastic P2P offload: TTFT p50 -49% vs baseline (0.551 vs 1.080)
gahow pushed to main at gahow/agentic-kvc 2026-05-22 05:25:36 +00:00
e9e313f9c5 P2P cache analysis: external KV correctly registered in prefix cache
gahow pushed to main at gahow/agentic-kvc 2026-05-22 04:28:26 +00:00
1b9268ba4c P2P prefill offload: TTFT p50 -13% but p90 +59% (median-vs-tail tradeoff)
gahow pushed to main at gahow/agentic-kvc 2026-05-22 02:59:01 +00:00
7f93d36970 System profile: 4 mechanisms why PD-Sep loses to session-sticky combined
gahow pushed to main at gahow/agentic-kvc 2026-05-22 02:35:19 +00:00
42bcd31976 TP=2 DP=4 + hybrid routing: best TTFT at cost of TPOT
gahow pushed to main at gahow/agentic-kvc 2026-05-22 02:15:09 +00:00
a65ec42467 Update report: adaptive v2 confirms no KV transfer helps single-machine
gahow pushed to main at gahow/agentic-kvc 2026-05-22 02:14:12 +00:00
2fee355626 Adaptive v2 (selective Mooncake offload): worse than baseline
gahow pushed to main at gahow/agentic-kvc 2026-05-21 19:14:06 +00:00
4bf0b999ff Final GPU comparison: hybrid routing matches baseline latency with better APC
gahow pushed to main at gahow/agentic-kvc 2026-05-21 18:54:49 +00:00
795edc6c66 Overnight work report: routing optimization achieves +4.7pp APC
gahow pushed to main at gahow/agentic-kvc 2026-05-21 18:53:46 +00:00
012d73f596 Hybrid routing: session-sticky + load-aware override achieves best results
gahow pushed to main at gahow/agentic-kvc 2026-05-21 18:13:18 +00:00
efe984477a Balanced routing result: APC +4pp but latency +23% (cache-load tradeoff)
gahow pushed to main at gahow/agentic-kvc 2026-05-21 17:50:28 +00:00
32f09d32cd Balanced session-sticky routing + agentic workload pattern analysis