File size: 1,361 Bytes
39e7656 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 | {
"_comment": "Optimal runtime configuration for TokForge speculative decoding",
"recommended": {
"target_backend": "opencl",
"draft_backend": "cpu",
"draft_predict_length": 3,
"draft_thread_num": 2,
"draft_power": "high",
"draft_sampler_type": "greedy"
},
"compatible_targets": {
"qwen3_8b": {
"enabled": true,
"uplift": "+24-40% (3x avg, 500-tok prose)"
},
"qwen3_14b": {
"enabled": true,
"uplift": "+40-70% (3x avg, 500-tok prose)"
},
"qwen3_4b": {
"enabled": false,
"reason": "Degenerates — KL trained from 8B teacher"
},
"qwen3_5": {
"enabled": false,
"reason": "Not compatible — different architecture (LinearAttention)"
}
},
"notes": {
"thread_num": "2 threads keeps WALT governor on performance cores. 4+ threads spread load and governor drops to min frequency.",
"power_high": "Forces Android performance hint. Critical for draft model speed.",
"draft_backend_cpu": "CPU draft avoids GPU memory contention with OpenCL target. OpenCL draft OOMs on 16GB devices.",
"predict_length_3": "d=3 is optimal across all tested devices. d=2 too conservative, d=4 wastes draft overhead."
}
}
|