weili-0234/fs-exp-trace

FlashState experiment traces

Part 2: weili-0234/fs-exp-trace-2 continues this dataset, which reached about 300 GB on 2026-10-30: it holds the rows posted after 2026-10-30 14:45 UTC (from modal/m512k-c40-w60-alphalow) and the raw payload dumps that did not fit here. Its ROUNDS.md lists what it holds.

Raw per-cell outputs of GPU experiments that compare KV/state cache policies for hybrid models in vLLM:

• vLLM default: official vLLM (commit ec8a3725) with automatic prefix caching.
• Marconi port: Marconi's admission and eviction (with its online alpha tuning) ported into vLLM.
• FlashState method: our state admission and eviction method.

Workload: AgentX token-ID live replay of the SemiAnalysis Claude Code traces (semianalysisai/cc-traces-weka-062126, revision 23f152f6f0f9399a85901b89a6458def0ef16729), closed loop with a fixed number of concurrent sessions, output lengths fixed per request (ignore EOS, the trace's max_tokens), thinking kept. Each engine runs with tensor parallel 2 on 2 NVIDIA B200 GPUs. A cell runs 4,500 s; metrics use the window [900, 4,500) s.

Layout

<platform>/<round>/<cell>/

• platform: modal (Modal B200 containers) or ci (CI lease pods with 2 B200).
• round: one load. Examples:
  • mL48-w60-tri: GLM-5.3-Flash NVFP4, 1M context cap, 48 sessions, the three arms started together.
  • m512k-c48-w60-r2: Qwen3.8-Flash-Next (FP8), 512K context cap (YaRN 2.0), 48 sessions.
  • m256k-c24-w60-tri: Qwen3.8-Flash-Next (FP8), 256K context cap, 24 sessions.
• cell: one arm. GLM: V0 vLLM default, O0 Marconi port, p0 FlashState method (the 0 means thinking kept). Qwen: Vagentx, Oagentx, pagentx.
• <platform>/<round>/analysis/tri_<round>.txt: the full analysis output of the row (seven metrics, excess recompute by cause and idle gap, matched and lost requests, the Marconi port's tuned alpha).

Files in a cell

• requests.jsonl.gz: one record per request: session, turn, send time, chunk times, prompt, cached and completion tokens, phase (warmup or profiling). Every reported metric is computed from this file.
• run.json (client run record), argv.json (the serve command), server_info.json (the engine configuration), service.txt, engine.log, check.txt, host.txt (container CPU, GPUs, driver, packages), mem.txt, cell.log, client.out, BUNDLE.json (client bundle record).
• client/: the aiperf log, Prometheus metrics before and after, aiperf exports (per-request profile export, timeslices, server metrics).
• MANIFEST.json and SHA256SUMS: sha256 of every file of the sealed cell. They include client/aiperf_artifacts/profile_export_raw.jsonl.gz (raw request and response payloads, 0.3-2.3 GB per cell). The raw dumps of the main-figure, math agent and Nemotron 3 Super rounds were added here on 2026-10-30; the others are in part 2 (see above) or kept with the original results.
• Access is gated: request access on this page; each request is approved by hand.

Analysis

The analysis scripts, the per-row result tables and the reproduction steps are in the FlashState repository under experiments/method-consolidate-1026/ (campaign-1028/ for the campaign of 2026-10-28).