Gahow Wang 63387f614d Full v3 trace re-profile with layer-wise: matched migrations improve ...

1213/1214 success; matched migrations (4 common) improved -2.6 to -7.2s,
scaling with prefill hidden behind transfer. Trace-level TTFT p90 -6% / p99
-5% (modest: migrations are 2% of reqs and partly queue-bound). Confirms
layer-wise removes the transfer half of migration overhead but not the
control-plane/queue residual. DESIGN.md updated with results.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

2026-05-28 19:16:37 +08:00

..

connector_tax

Full v3 trace re-profile with layer-wise: matched migrations improve

2026-05-28 19:16:37 +08:00

fresh_setup

MB5 driver fixes: bash env-prefix + replayer flag names + python date math

2026-05-27 23:23:23 +08:00

interference

Microbench 1 plots: prefill-decode interference heatmap + lines

2026-05-26 14:21:30 +08:00

lifecycle

PD-sep server-side profiling: vLLM patches + per-request breakdown

2026-05-26 13:59:09 +08:00

patches

PD-sep server-side profiling: vLLM patches + per-request breakdown

2026-05-26 13:59:09 +08:00

__init__.py

Microbench 3 (connector_tax): infrastructure for KV connector substrate tax

2026-05-26 17:27:41 +08:00

ANALYSIS.md

Microbench: prefill-decode interference + PD transfer lifecycle

2026-05-26 00:57:06 +08:00

interference_microbench_design.md

Microbench: prefill-decode interference + PD transfer lifecycle

2026-05-26 00:57:06 +08:00

plot_breakdown_real.py

PD-sep server-side profiling: vLLM patches + per-request breakdown

2026-05-26 13:59:09 +08:00

plot_breakdown.py

PD-sep server-side profiling: vLLM patches + per-request breakdown

2026-05-26 13:59:09 +08:00

plot_interference.py

Microbench 1 plots: prefill-decode interference heatmap + lines

2026-05-26 14:21:30 +08:00

TODO.md

Microbench: prefill-decode interference + PD transfer lifecycle

2026-05-26 00:57:06 +08:00

transfer_lifecycle_design.md

Microbench: prefill-decode interference + PD transfer lifecycle

2026-05-26 00:57:06 +08:00