classify
large · Q4_K_M · 1 threads · batch 64
- Meets quality floor 0.950 with 1.000.
- Selected from 1 Pareto-efficient candidates using the balanced preset.
- Median latency: 0.0% lower than baseline.
- Peak memory: 0.0% lower than baseline.
Measured workflow compilation.
VERIFIED ARM64 RUN
DEPLOYABLE
Every workflow stage has a candidate that passed its quality gate.
Run run-1488b44fa409 · commit b414b9e8d52ea72dc909b92d87ba8a0efbd9c2d4
Arm system evidence
Architectureaarch64
CPUaarch64
Cores4
Memory15.57 GB
NEONdetected
SVEdetected
MATMUL_INT8detected
Performixunavailable
Before vs compiled routing
| Stage | Baseline | A64Forge | Baseline median | Compiled median | Change | Quality |
|---|---|---|---|---|---|---|
| classify | large | large / Q4_K_M | 6,404.47 ms | 6,404.47 ms | +0.0% | 1.00 |
| extract | large | medium / Q4_K_M | 15,294.76 ms | 2,124.38 ms | +86.1% | 0.83 |
| tool_select | large | large / Q4_K_M | 6,373.08 ms | 6,338.98 ms | +0.5% | 1.00 |
| reason | large | small / Q4_0 | 17,979.78 ms | 3,244.15 ms | +82.0% | 0.83 |
| summarize | large | small / Q4_0 | 15,190.19 ms | 1,626.49 ms | +89.3% | 1.00 |
Optimization decisions
extract
medium · Q4_K_M · 4 threads · batch 128
- Meets quality floor 0.800 with 0.833.
- Selected from 2 Pareto-efficient candidates using the balanced preset.
- Median latency: 86.1% lower than baseline.
- Peak memory: 52.8% lower than baseline.
tool_select
large · Q4_K_M · 1 threads · batch 128
- Meets quality floor 0.950 with 1.000.
- Selected from 2 Pareto-efficient candidates using the balanced preset.
- Median latency: 0.5% lower than baseline.
- Peak memory: 0.0% higher than baseline.
reason
small · Q4_0 · 1 threads · batch 64
- Meets quality floor 0.800 with 0.833.
- Selected from 3 Pareto-efficient candidates using the balanced preset.
- Median latency: 82.0% lower than baseline.
- Peak memory: 80.5% lower than baseline.
summarize
small · Q4_0 · 1 threads · batch 128
- Meets quality floor 0.950 with 1.000.
- Selected from 2 Pareto-efficient candidates using the balanced preset.
- Median latency: 89.3% lower than baseline.
- Peak memory: 80.8% lower than baseline.
Method
All candidates used the stage dataset hash stored in results.json. Each record stores warm-up count, measured count, model hash, host, architecture, configuration, median and p95 latency, process resource observations, and quality. Missing fields remain “Not measured yet.” Development fixtures are never verified.