weili-0234/fs-exp-trace-4

FlashState experiment traces, part 4: Kimi-K3.1-Pro on GB300

Client-side request and response dumps, with each run's logs and results, of the FlashState experiments on Kimi-K3.1-Pro on NVIDIA GB300 from 2026-11-05 to 2026-11-08. Parts 1 to 3 (weili-0234/fs-exp-trace, weili-0234/fs-exp-trace-2 and weili-0234/fs-exp-trace-3) hold the B200 runs and describe the trace replay in more detail.

Study

FlashState keeps the states of a hybrid model (attention KV and recurrent checkpoints) for agent sessions that wait for tool calls. These cells measure it on Kimi-K3.1-Pro with the deployment's decode path, against official vLLM: first without a host store, then with a Mooncake host store on both sides at several store sizes, and on 2026-11-08 against SGLang with HiCache as well. Each cell's arm.env holds the exact settings of its arm; the arm names start with official (official vLLM), fs (FlashState) or sgl (SGLang). The FlashState code of the 2026-11-08 sweep is released as the flashstate-k31pro-gb300 branches of Weili-0234/vllm and Weili-0234/Mooncake, and as the image weili0234/flashstate:kimi-k31-cu134-266e88c-fs3f8611c-mc4592063-gb300.

Setup (all cells)

• Model: moonshotai/Kimi-K3.1-Pro. Speculative decoding: DSpark (5 draft tokens) on vLLM, DFLASH on SGLang.
• One engine: two GB300 trays (8 GPUs) of one NVL72 domain, tensor parallel 8 and decode context parallel 8.
• Workload: AgentX coding-agent sessions (semianalysisai/cc-traces-weka-062126), replayed live with token IDs (the model's own outputs go into the next turn); output lengths are fixed by the trace (ignore_eos).
• A level runs a fixed number of concurrent sessions for 900 s (600 s in smoke cells) after the engine's caches and the host store are emptied. Levels named warm-* are warm-ups and are left out of every result.

Layout

k31/<cell>/                  arm.env, cell.log, levels.jsonl (one line per level), engine and store logs
k31/<cell>/<level>/          requests.jsonl.gz (one line per request), run.json, server_info.json
k31/<cell>/<level>/client/aiperf_artifacts/profile_export_raw.jsonl.gz
CELLS.md                     every cell with its arm, start time, levels, status and raw dumps
SHA256SUMS                   sha256 of every file under k31/

Each raw dump has one JSON line per request: metadata (session, conversation and turn, request timestamps in ns, warm-up or profiling phase), payload (the prompt as token IDs and the sampling parameters), responses (the streamed packets with their timestamps, including the returned token IDs), request_headers and status.

Placeholders

Cluster-specific strings in the logs and result files were replaced before upload: the work directory by $W, other cluster paths by <lustre>, user names by <user>, the account by <account>, login hosts by <host>, private network addresses by <ip> and email addresses by <email>. The raw dumps are unchanged.

Checks

Every raw dump passed a scan for cluster identifiers and credentials before upload, and every file's sha256 (in SHA256SUMS) was checked against the hub after upload.

Access is gated: request access on this page; each request is approved by hand.