File size: 1,919 Bytes
b8825d6 | 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 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 | {
"arch_id": "qwen3-next-mtp",
"artifact_role": "small-q8-speed-test",
"base_trunk": "mlx-community/Qwen3.5-4B-MLX-8bit",
"exactness_baseline": {
"artifact": "/Users/youssof/.mtplx/qwen35_4b_q8trunk_mtp1_gate.json",
"gate": "mtp1-greedy-ar-equivalence",
"max_tokens": 16,
"matches": 1,
"scope": "greedy AR and MTP1 generated identical token sequence on the warm_code_continuation prompt",
"seed": 0,
"status": "passed",
"total": 1
},
"mtp_depth_max": 2,
"mtp_sidecar": "official Qwen/Qwen3.5-4B MTP tensors; quantization recorded in config.json",
"mtplx_version": "0.1.0-preview",
"recommended_draft_lm_head": {
"bits": 4,
"group_size": 64,
"mode": "affine"
},
"recommended_draft_sampler": {
"temperature": 0.7,
"top_k": 20,
"top_p": 0.95
},
"recommended_profile": "performance-cold",
"sampler": {
"temperature": 0.6,
"top_k": 20,
"top_p": 0.95
},
"speed_evidence": {
"acceptance_by_depth": [
0.7647058823529411,
0.47058823529411764
],
"accepted_drafts": 21,
"ar_tok_s": 75.62701983984475,
"artifact": "/Users/youssof/.mtplx/qwen35_4b_depth_grid_q8_body4_draftt0p7.json",
"draft_lm_head": "8-bit tied embedding requantized as 4-bit affine group64 draft head",
"draft_temperature": 0.7,
"enable_thinking": false,
"generated_tokens": 48,
"mtp_depth": 2,
"mtp_tok_s": 105.20912101297236,
"note": "Q8 trunk gives the best multiplier in the local one-prompt matrix, but the Q4 trunk remains faster in absolute tok/s.",
"sampler": {
"temperature": 0.6,
"top_k": 20,
"top_p": 0.95
},
"speedup_vs_ar": 1.391157832686963
},
"verified_on": {
"hardware": "Apple Silicon local MTPLX workstation",
"machine_arch": "arm64",
"model": "Qwen3.5-4B-MTPLX-Optimized-Speed-Q8Trunk",
"timestamp": "2026-05-04T06:17:00+0100"
}
}
|