Skip to content

Mobile Evaluation

This document collects all on-device training/peak-RAM measurements. It lists the original results reported in the paper (Google Pixel 6) and the newly collected measurements (Samsung Galaxy S21 FE 5G and additional Pixel 6 TinyLlama RAM measurements). All measurements are reported as mean ± standard deviation.


Measurement methodology

  • Discard the first warm-up step. Continue training for nine forward–backward passes with optimizer updates. Measure the time to complete a single training step (forward+backward+update). Repeat the entire process five times and report the mean and standard deviation of the per-step time.
  • Reported values for training step time are in seconds (mean ± std).
  • Peak RAM usage is reported in GB.
  • Batch size (bs): All experiments used bs = 4, except the Qwen measurement on Samsung Galaxy S21 FE 5G which used bs = 1 due to OOM at larger batch sizes (this affects the per-step timing and should be noted when comparing numbers).

Devices / hardware

  • Google Pixel 6 — 8 GB RAM (~ 6.7 GB available)
  • Samsung Galaxy S21 FE 5G — 6 GB RAM (~ 2.9 GB available)

Units and naming

  • Timing values: seconds per training step (mean ± std)
  • Peak RAM: GB
  • Methods: LoRA, MARS OPT0, MARS Q-OPT1 (MARS OPT1)
  • Ranks reported as r (2, 8, 32)

Training step time

Google Pixel 6

Models: SmolLM2 (360M) and Qwen-2 (0.5B).

SmolLM2 (360M) — step time (s)

Rank LoRA (s) MARS Q-OPT0 (s) MARS Q-OPT1 (s)
r=2 18.94 ± 0.12 18.49 ± 0.22 18.07 ± 0.12
r=8 19.12 ± 0.22 19.39 ± 0.17 18.33 ± 0.07
r=32 21.94 ± 0.11 21.29 ± 0.04 19.76 ± 0.14
Average 20.01 19.72 18.75

Qwen-2 (0.5B) — step time (s)

Rank LoRA (s) MARS Q-OPT0 (s) MARS Q-OPT1 (s)
r=2 21.87 ± 0.10 21.55 ± 0.12 20.43 ± 0.20
r=8 22.11 ± 0.19 21.64 ± 0.13 20.96 ± 0.08
r=32 23.74 ± 0.21 22.10 ± 0.04 21.55 ± 0.23
Average 22.57 21.76 20.98

Samsung Galaxy S21 FE 5G

Notes: SmolLM2 measurements used bs = 4. Qwen-2 measurements on this device used bs = 1 due to OOM at larger batch sizes; timing for Qwen is therefore not directly comparable to bs=4 measurements.

SmolLM2 (bs = 4) — step time (s)

Rank LoRA (s) MARS OPT0 (s) MARS OPT1 (s)
r=2 16.20 ± 0.10 15.05 ± 0.08 14.44 ± 0.18
r=8 18.72 ± 0.11 15.82 ± 0.22 15.30 ± 0.33
r=32 23.10 ± 0.43 17.81 ± 0.51 16.10 ± 0.39
Average 19.34 16.22 15.28

Qwen-2 (bs = 1 on Samsung) — step time (s)

Rank LoRA (s) MARS OPT0 (s) MARS OPT1 (s)
r=2 6.91 ± 0.07 6.75 ± 0.02 6.39 ± 0.19
r=8 7.38 ± 0.09 7.17 ± 0.11 6.96 ± 0.21
r=32 10.10 ± 0.31 9.29 ± 0.38 8.98 ± 0.43
Average 8.13 7.74 7.44

Important: Qwen timings above were measured with bs = 1 on Samsung due to OOM at larger batch sizes; these numbers will appear faster compared to bs=4 runs and must be treated accordingly when comparing across devices/experiments.


Peak RAM usage measurements

All reported peak memory is the observed maximum resident memory during a training step (GB). Ranks reported where available.

Device (RAM) Model Rank LoRA Peak RAM MARS OPT0 Peak RAM MARS OPT1 Peak RAM
Google Pixel 6 (8 GB) TinyLlama r=32 2.8 GB 2.8 GB 2.6 GB
Samsung Galaxy S21 FE 5G (6 GB) SmolLM2 r=32 2.4 GB 2.4 GB 2.2 GB

Important notes / summary of measurement caveats

  • Batch size: Unless otherwise noted, measurements used bs = 4. The Qwen run on Samsung used bs = 1 (OOM otherwise) — this influences step-time comparisons.
  • Repetitions: Each reported mean/std is computed from 5 independent repeats of the 9-step timing procedure (first warm-up step discarded), measuring one-step duration each repeat.
  • Devices: Original paper values were collected on Google Pixel 6 (8GB). New measurements were collected on Samsung Galaxy S21 FE 5G (6GB) except the TinyLlama Peak-RAM entry which was measured on Google Pixel 6.
  • Units: Step time in seconds (mean ± std). Peak RAM in GB.