← TinyGPT · docs · devlog · roadmap · speedup
source: docs/pace-handoff-2026-06-10.md · view on GitHub ↗

Pace handoff — tinygpt deliverables status (2026-06-10)

tinygpt (the factory) refocuses on local-LLM research. This doc freezes the state of everything tinygpt owes or has delivered to Pace (the consumer), so either repo can pick up any thread without re-deriving context.

Delivered — done, shipped, in use

DeliverableWhereState
v9 planner LoRA (shipping specialist)~/.cache/tinygpt/runs/pace-planner-v9/33.3% v2 / 70% compose; not Pace’s runtime default (see below)
ANE M8 28-block chaintinygpt ANE scripts17 tok/s, ’ Paris’ top-1 correct
v10 action registry + schemagrammars/v10-actions/ → copied into pace/leanring-buddy/Resources/Pace bundles its own copy; registry validates at app startup
Serve: grammar constraint, EOS stops, prompt cache, TTFW 119ms warm, partial-JSON streamingTinyGPTServePace consumes via localhost HTTP
WhisperKit qualificationmemory + benchmarkslarge-v3-turbo: 1.5GB, 9× realtime — Pace’s provider scaffold awaits streaming wiring
Qwen3-Embedding-0.6B qualificationmemoryqualified for RAG; Pace’s v1 retrieval shipped lexical instead — embedding upgrade is a future option
Stage A dictation post-processorpace/leanring-buddy/PaceDictationPostProcessor.swift + 26 testscommitted 2026-06-10
v11 eval suites (60 unhappy fixtures + 96 BFCL-12)pace/evals/fm-fixtures-{oos,ambig,destructive} + ~/.cache/tinygpt/datasets/bfcl/BFCL_v3_pace12.jsonheld-out; baselines locked in docs/v11-baselines-2026-06-09.md

Ready to fire — one command, needs GPU free

v11 planner train-and-gate: bash scripts/v11_pipeline.sh [--amplify]

Owed — open tinygpt work Pace is waiting on

#WorkWhy Pace needs itSize
#266VLM in-process runtime (Qwen3-VL or UI-Venus M4 port)Pace’s ScreenAnalysisProvider: inProcess/coreML/mlx options are stubs falling back to LM Studio HTTPdays — the next big factory project
#308VLM A/B (UI-Venus vs Qwen3-VL-2B)picks the #266 port target1-2 days; needs screenshots + GPU
#292-adjWhisperKit streaming runtime wiringPace’s TranscriptionProvider: whisperKit scaffold falls back to Apple Speechmostly Pace-side; tinygpt advises
#306macOS 26 int8 ANE handoff into M8~1.8× ANE speedup primitiveperf research track
#303/#305Quantization pass + Swift QuantizedLinearsmaller shipping modelsperf research track

Decisions frozen (do not re-litigate without new data)

Coordination notes