spark-ds41f

GitHub Repo

Release-over-Release Performance Observatory

Driven directly from manifests/suites/*/ across all recorded servebench release, single-stream, and sampled runs.

Source-Text Prefill Throughput (32K → 1M) tok/s (higher is better)
Decode Throughput across Releases output tok/s (higher is better)
Engine Decode Step-Time Grid (Base vs Candidate) trimmed mean ms / step (lower is better)
DSpark Speculative Survival & Prefill Yield draft positions k=0..4 & 131K stall probes
Workload / Metric Base Candidate Delta Notes

Cluster Shapes & KV Pool Sizing

Resolved from config/cluster.json via scripts/recipe.py and scripts/kv_capacity.py (227,840-byte pool blocks).

Shape Ranks & Transport Kernel / Pages Context Limit Max Seqs Batch Tokens KV Pool / Rank Min Free RAM (Head / Workers) 5 GiB Guard Headroom Pool Blocks Full Contexts Held vLLM 0491 Startup Line Qualified Suite

Pinned Upstreams, Patch Stacks & Active Runtime Choices

Synchronized from upstreams.lock.json, manifests/baselines/2026-11-07-spark-ds41f-1107.json, and manifests/sources/.

Component Version / Ref Pinned Revision Patch Series Verified Tree Role
Runtime Choice Active Specification in spark-ds41f-1107

Baseline Manifest History (35 Releases)

Complete immutable promotion history from manifests/baselines/*.json.