You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

KVC raw data

Raw experimental data for KVC. KVC studies KV-cache management for LLM inference, focusing on the energy cost of KV-cache operations and their interaction with inference latency.

  • Each folder holds the complete raw run behind one entry of the GitHub repository and uses the same global ID.
  • Processed plotting points, plotting scripts, experiment metadata and conclusions are kept in the GitHub repository (results/<entry>/), and the experiment code in scripts/<entry>/.
  • Each result's metadata.json on GitHub records the exact folder and revision of its raw data here.

Index

ID Date Raw data GitHub result Description
000 2026-09-29 βœ“ βœ“ 45-min KV lifecycle run lifecycle_45min_20260927_231128 (telemetry, traces, server logs, configs, code snapshot, figures); used for the 128K SSD restore vs recompute energy breakdown; Llama-3.1-8B, A100 (MIG 7g), vLLM 0.11.0 + LMCache 0.3.9
003 2026-09-30 βœ“ βœ“ H100 power-measurement validity (1A sensor calibration, 1B phase alignment incl. nsys reports, 1C KV-transfer power); Llama-3.1-8B + Llama-3.3-70B TP4, vLLM 0.11.0 + LMCache 0.3.9
004 2026-09-30 βœ“ βœ“ Energy vs throughput (2A static/dynamic split, 2B KV strategies at fixed load); 8B and 70B TP4 sessions with per-request timestamps and telemetry
005 2026-10-01 βœ“ βœ“ Iso-batch control + P0 checks (write-back, nsys, pinned path); per-step vLLM records, closed loop / Poisson / gated release / injected stall
006 2026-10-01 βœ“ βœ“ Mooncake/Azure trace replays, clock-governor events, lock/unlock micro-benchmark, rule 2 on the SSD tier; input traces included
007 2026-10-01 βœ“ βœ“ Modelling grid sessions (L 2K-124K, prefix, concurrency, output, tier); the 8K layer is in 005
013 2026-10-19 βœ“ βœ“ Iso-power response modes (firmware cap vs clock tables / admission PI / KV decisions), 8B + lmsys controls, dynamic cuts; old + new host
014 2026-10-19 βœ“ βœ“ KV-decision increment B0 vs C0 (Mooncake EXP5, SWE-bench EXP5b / EXP5c incl. disk-tier capacity cells); new host
015 2026-10-19 βœ“ βœ“ Per-server power to rack capacity (60-min node traces, rack oversubscription simulation); new host
016 2026-10-19 βœ“ βœ“ Blocking vs async KV loading under memory pressure (SWE-bench, mem 0.55, lambda*_blk search); new host
017 2026-10-19 βœ“ βœ“ H100 clock halving under locked clocks with gappy load (micro-benchmark grids on both hosts, post-hoc cap cells)
021 2026-11-02 βœ“ βœ“ Restoration-planner selection (DP / enumeration / continuity-aware B&B / hybrid / Cake / all-read): closed-loop simulator runs v8 (72 cells, 2,287 runs, per-iteration logs, solver-instance dumps), offline replays, lambda sweep, A100 profile + cost model, prepared traces; superseded rounds v1-v7 included; Llama-3.1-8B, A100, vLLM 0.11.0
034 2026-11-08 βœ“ βœ“ JouleKV v2 overnight benchmark, site L (4 x A100-SXM4-80GB, Llama-3.1-8B, vLLM 0.11.0 + LMCache 0.3.9): ~230 serving runs on ShareGPT multi-turn, LooGLE, LooGLE-LMC and L-Eval x 4 loads x {plain vLLM, all-read, all-read + generic DVFS, JouleKV} plus A/B batches (per-request records, NVML energy, per-step scheduler logs, engine traces), capacity calibrations, workload traces, per-clock profiles, driver scripts and logs

Entries 003-007, 013-017 and 021 are on GitHub branches (exp-021-kv-restoration-solver-selection, exp-003-004-h100-measurement-and-energy-space, exp-005-007-batch1, exp-013-017-batch3) until they are merged into main. Entries 013-017 share one raw-data tree (batch3-newhost-2026-10-15/, batch3-snapshot-2026-10-15/); their folders index it.

Downloads last month
505