edge-llm-benchmarks / README.md
kleinbem's picture
chore(data): update edge GGUF quality & power benchmark dataset
2273a4a verified
|
Raw History Blame Contribute Delete
2.43 kB
---
license: mit
task_categories:
- text-generation
tags:
- benchmark
- quantization
- kl-divergence
- gguf
- llama-cpp
- jetson-orin-nano
- energy-efficiency
- edge-ai
pretty_name: Edge & Local GGUF Quantization, KL-Divergence & Power Telemetry
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train
path: data/benchmarks.parquet
---
# Edge & Local GGUF Quantization, KL-Divergence & Power Telemetry
Hardware throughput (`pp512`, `tg128`), exact **FP16 logit Kullback-Leibler divergence (`mean_kld`, `p99_kld`, `same_top1_pct`)**, unified-memory OOM boundaries, and **INA3221 / RAPL energy per token (`Joules/token`)** collected reproducibly via [`github.com/kleinbem/hf-bench`](https://github.com/kleinbem/hf-bench).
- **Interactive Explorer Space**: [`spaces/kleinbem/edge-llm-pareto`](https://huggingface.co/spaces/kleinbem/edge-llm-pareto)
## Evaluated Base Models
```yaml
base_models:
- Qwen/Qwen2.5-Coder-7B-Instruct
- Qwen/Qwen3-4B
```
## Hardware Targets
1. **`orin-nano`**: NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, Ampere 1024-core Tegra GPU `sm_87`, NixOS + JetPack 6 CUDA 12.6, INA3221 hardware power rail sampling).
2. **`nixos-nvme`**: 13th Gen Intel Core i3-1315U (`x86_64` Raptor Lake AVX2, 62 GiB DDR5, NixOS idle scheduler priority, Intel RAPL package energy sampling).
## Schema Highlights
| Column | Type | Description |
|---|---|---|
| `model_name` | `string` | Base model architecture name |
| `quant` | `string` | `llama.cpp` quantization (`Q8_0` .. `IQ3_M`, plus `Q4_K_M-std` uncalibrated control) |
| `imatrix_used` | `bool` | Whether agentic + code importance matrix calibration was applied |
| `bpw` | `float64` | Effective bits per weight (`weight_bytes * 8 / params`) |
| `status` | `string` | `ok` or `oom` when exceeding the device's unified memory cliff |
| `mean_kld` | `float64` | Mean Kullback-Leibler divergence (nats) vs exact FP16 base logits |
| `same_top1_pct` | `float64` | Percentage of evaluation tokens sharing the exact FP16 argmax prediction |
| `pp512_tok_s` | `float64` | 512-token prompt prefill speed (`tokens/sec`) |
| `tg128_tok_s` | `float64` | 128-token autoregressive decode speed (`tokens/sec`) |
| `avg_watts` | `float64` | Mean board/package power draw during inference (`Watts`) |
| `joules_per_token` | `float64` | Energy consumed per generated token (`avg_watts / tg128_tok_s`) |