Laguna XS 2.1 FP8 DFlash results on DGX Spark

#1
by seddonm1 - opened

The important finding is that FP8 must be used consistently at all three layers for DFlash to perform well:

  1. Target model: poolside/Laguna-XS-2.1-FP8
  2. DFlash draft model: poolside/Laguna-XS-2.1-DFlash-FP8
  3. KV cache: --kv-cache-dtype fp8

Using a non-FP8 KV cache caused substantially worse DFlash performance.

Configuration:

  • vLLM with FlashInfer attention
  • Triton MoE backend
  • Concurrency: 1
  • Coding prompt, thinking disabled
  • 128 generated tokens
  • Three runs per context size
  • Decode throughput, excluding TTFT
Draft tokens 4K context 31K context 92K context Average
4 69.60 72.21 57.79 66.53
5 76.54 75.65 71.26 74.48
6 81.37 83.44 76.11 80.31
7 82.05 90.85 72.24 81.71
8 86.75 96.20 85.30 89.42
9 84.33 78.72 77.38 80.14
10 90.73 88.36 74.25 84.45
11 77.79 70.47 66.70 71.65
12 79.21 79.20 67.03 75.15
13 80.45 69.26 70.42 73.38
14 83.80 81.05 61.00 75.28
15 88.81 83.59 73.26 81.89

For this workload, num_speculative_tokens: 8 was the clear winner, averaging 89.42 tok/s.

Sign up or log in to comment