FINDINGS.md

metadata

license: mit
task_categories:
  - text-generation
tags:
  - benchmark
  - throughput
  - quality
  - vllm
  - gentoo
  - nanbeige
  - awq
language:
  - en
pretty_name: Nanbeige4.2-3B AWQ gentoo gauntlet

Nanbeige4.2-3B AWQ — gentoo quality / speed / context / concurrency gauntlet

Separate dataset from LostGentoo/gentoo-small-model-throughput-vllm. Same harness (conc + ctx throughput + lighteval chat_core), single model: AWQ W4A16 of Nanbeige4.2-3B on the Nanbeige vLLM fork.

Setup

Setting	Value
Host	gentoo (3× RTX 5060 Ti 16 GB)
GPU	CUDA device 2
Engine	Nanbeige/vllm @nanbeige42
Model	LostGentoo/Nanbeige4.2-3B-AWQ-W4A16-ASYM
max_model_len	4096
gpu_memory_utilization	0.70
Throughput gen length	128 tokens
Warmup / repeats	2 / 2
Concurrency levels	1, 2, 4, 6, 8, 10, 12, 16, 20
Context tiers	128, 512, 1024, 2048, 4096
Quality suite	lighteval chat_core = ifeval, gpqa:diamond, mmlu_pro
Quality concurrent_requests	16
Quality note	chat_template thinking default OFF; no --reasoning-parser (answers in content)

Key findings (throughput)

1. Peak aggregate decode @ conc=20: ~1442 sum-of-stream tok/s (wall agg ~1347 tok/s). Scales smoothly with 0 errors.
2. Single-stream decode: 87 tok/s @ conc=1 — slower than MiniCPM5-1B (200) / Qwen3.5-0.8B (238) on the same box; closer to Qwen3.5-2B (106).
3. Context @ conc=10: decode falls from ~78 tok/s/stream @ 128-tier to ~41 @ 2048-tier. 4096 tier n/a (prompt+128 exceeds max_model_len), same pattern as Qwen in the small-model dataset.
4. Serve caveat: requires Nanbeige vLLM fork + trust_remote_code; stock vLLM cannot load NanbeigeForCausalLM.

Concurrency sweep

conc	TTFT (ms)	mean decode (tok/s)	sum decode (tok/s)	wall agg (tok/s)	err
1	17.8	87.1	87.1	86.1	0
2	25.0	86.6	173.2	169.3	0
4	34.4	85.2	340.8	330.6	0
6	46.4	84.9	509.1	490.9	0
8	63.1	84.4	675.0	645.5	0
10	67.4	80.4	804.0	767.5	0
12	80.0	80.1	961.2	910.0	0
16	87.8	78.5	1255.5	1179.9	0
20	109.0	72.1	1441.9	1346.6	0

Context sweep

tier	conc	TTFT (ms)	mean decode (tok/s)	sum decode (tok/s)	wall agg (tok/s)	err
128	1	18.4	86.9	86.9	85.9	0
128	10	49.4	78.2	781.8	702.3	0
512	1	19.8	85.0	85.0	80.0	0
512	10	70.3	66.3	663.4	476.9	0
1024	1	22.0	82.5	82.5	73.9	0
1024	10	76.1	54.7	547.2	330.6	0
2048	1	25.5	79.2	79.2	64.7	0
2048	10	108.7	40.8	408.5	200.6	0
4096	1	n/a	n/a	n/a	n/a	1
4096	10	n/a	n/a	n/a	n/a	10

Quality (chat_core)

• source: quality/LostGentoo_Nanbeige4.2-3B-AWQ-W4A16-ASYM/results/openai/LostGentoo/Nanbeige4.2-3B-AWQ-W4A16-ASYM/results_2026-07-22T12-13-40.991373.json
• wall seconds: 16886.941249988973
• max_samples: None

task	metric	value
`ifeval	0`	prompt_level_strict_acc
`ifeval	0`	inst_level_strict_acc
`ifeval	0`	prompt_level_loose_acc
`ifeval	0`	inst_level_loose_acc
`gpqa:diamond	0`	gpqa_pass@k:k=1
`mmlu_pro	0`	extractive_match

Quality section updated 2026-07-22T12:14:58

Files

Files

File	Description
*_scale_conc.json	Concurrency sweep
*_scale_ctx.json	Context-length sweep
quality/	lighteval chat_core outputs
SUMMARY.md	Run log / pass-fail
README.md	This card

Generated 2026-07-22T07:43:40