prompt stringclasses 7
values | category stringclasses 1
value | id int64 0 499 |
|---|---|---|
Explain AI in one sentence | inference | 0 |
Write python to add two numbers | inference | 1 |
Summarize: Baseten makes models fast | inference | 2 |
What is KV cache? | inference | 3 |
Define quantization | inference | 4 |
What is speculative decoding? | inference | 5 |
Explain error cancellation in quantization | inference | 6 |
Explain AI in one sentence | inference | 7 |
Write python to add two numbers | inference | 8 |
Summarize: Baseten makes models fast | inference | 9 |
What is KV cache? | inference | 10 |
Define quantization | inference | 11 |
What is speculative decoding? | inference | 12 |
Explain error cancellation in quantization | inference | 13 |
Explain AI in one sentence | inference | 14 |
Write python to add two numbers | inference | 15 |
Summarize: Baseten makes models fast | inference | 16 |
What is KV cache? | inference | 17 |
Define quantization | inference | 18 |
What is speculative decoding? | inference | 19 |
Explain error cancellation in quantization | inference | 20 |
Explain AI in one sentence | inference | 21 |
Write python to add two numbers | inference | 22 |
Summarize: Baseten makes models fast | inference | 23 |
What is KV cache? | inference | 24 |
Define quantization | inference | 25 |
What is speculative decoding? | inference | 26 |
Explain error cancellation in quantization | inference | 27 |
Explain AI in one sentence | inference | 28 |
Write python to add two numbers | inference | 29 |
Summarize: Baseten makes models fast | inference | 30 |
What is KV cache? | inference | 31 |
Define quantization | inference | 32 |
What is speculative decoding? | inference | 33 |
Explain error cancellation in quantization | inference | 34 |
Explain AI in one sentence | inference | 35 |
Write python to add two numbers | inference | 36 |
Summarize: Baseten makes models fast | inference | 37 |
What is KV cache? | inference | 38 |
Define quantization | inference | 39 |
What is speculative decoding? | inference | 40 |
Explain error cancellation in quantization | inference | 41 |
Explain AI in one sentence | inference | 42 |
Write python to add two numbers | inference | 43 |
Summarize: Baseten makes models fast | inference | 44 |
What is KV cache? | inference | 45 |
Define quantization | inference | 46 |
What is speculative decoding? | inference | 47 |
Explain error cancellation in quantization | inference | 48 |
Explain AI in one sentence | inference | 49 |
Write python to add two numbers | inference | 50 |
Summarize: Baseten makes models fast | inference | 51 |
What is KV cache? | inference | 52 |
Define quantization | inference | 53 |
What is speculative decoding? | inference | 54 |
Explain error cancellation in quantization | inference | 55 |
Explain AI in one sentence | inference | 56 |
Write python to add two numbers | inference | 57 |
Summarize: Baseten makes models fast | inference | 58 |
What is KV cache? | inference | 59 |
Define quantization | inference | 60 |
What is speculative decoding? | inference | 61 |
Explain error cancellation in quantization | inference | 62 |
Explain AI in one sentence | inference | 63 |
Write python to add two numbers | inference | 64 |
Summarize: Baseten makes models fast | inference | 65 |
What is KV cache? | inference | 66 |
Define quantization | inference | 67 |
What is speculative decoding? | inference | 68 |
Explain error cancellation in quantization | inference | 69 |
Explain AI in one sentence | inference | 70 |
Write python to add two numbers | inference | 71 |
Summarize: Baseten makes models fast | inference | 72 |
What is KV cache? | inference | 73 |
Define quantization | inference | 74 |
What is speculative decoding? | inference | 75 |
Explain error cancellation in quantization | inference | 76 |
Explain AI in one sentence | inference | 77 |
Write python to add two numbers | inference | 78 |
Summarize: Baseten makes models fast | inference | 79 |
What is KV cache? | inference | 80 |
Define quantization | inference | 81 |
What is speculative decoding? | inference | 82 |
Explain error cancellation in quantization | inference | 83 |
Explain AI in one sentence | inference | 84 |
Write python to add two numbers | inference | 85 |
Summarize: Baseten makes models fast | inference | 86 |
What is KV cache? | inference | 87 |
Define quantization | inference | 88 |
What is speculative decoding? | inference | 89 |
Explain error cancellation in quantization | inference | 90 |
Explain AI in one sentence | inference | 91 |
Write python to add two numbers | inference | 92 |
Summarize: Baseten makes models fast | inference | 93 |
What is KV cache? | inference | 94 |
Define quantization | inference | 95 |
What is speculative decoding? | inference | 96 |
Explain error cancellation in quantization | inference | 97 |
Explain AI in one sentence | inference | 98 |
Write python to add two numbers | inference | 99 |
End of preview. Expand in Data Studio
Tiny-Quant-Cancellation-Bench
Inspired by Ali Taha (Head of Model Performance @ Baseten)
Inference & Performance
Inspired by Baseten's Model Performance research (Ali Taha, Head of Model Performance)
- Optimization Focus: Efficient inference, KV-cache aware, quantization-ready
- Techniques Explored: Speculative decoding concepts, quantization error cancellation, cache-aware routing
- Goal: Fast, reliable, cost-efficient small model inference
Research Question
Does error cancellation observed by Baseten on GLM-5.2 (error A + error B < error A) hold for tiny models <1B?
Benchmarks
| config | latency_ms | throughput | vram | quality | | fp16 | 150 | 30 t/s | 2GB | 0.95 | | int8 | 120 | 38 t/s | 1.2GB | 0.93 | | int4 | 110 | 42 t/s | 800MB | 0.88 | | mixed-int8-int4 | 95 | 48 t/s | 1GB | 0.92 |
Finding: Mixed quantization is 20% faster with same quality - replicating Baseten's result.
Dataset
samudhan2003/Tiny-Quant-Cancellation-Dataset
Author
Samudhan - Independent Researcher, Model Performance Base Model: samudhan2003/Tiny-model-test-amudhan Credits: Ali Taha @ Baseten, Philip Kiely
- Downloads last month
- 37