Ground Segment Budgets

Resource requirements for training, quantization, and dataset generation.

Training Hardware

fine_tune.py is CUDA only. The training stack uses bitsandbytes (4-bit NF4 quantization + paged_adamw_8bit optimizer), which does not support Apple Silicon, AMD GPUs, or CPU-only execution.

Resource	Requirement	Notes
GPU	NVIDIA CUDA-capable, 8+ GB VRAM	QLoRA loads base model in 4-bit (NF4)
CUDA toolkit	12.x	Matched to the GPU driver
System RAM	8+ GB recommended	For data loading and preprocessing
Training time per epoch	830s	3 epochs total, ~480 training samples
Total training time	~2492s	Depends on GPU

Reference hardware (verified working):

Component	Spec
GPU	NVIDIA GeForce RTX 4070 Ti, 12 GB VRAM
Driver	535.261.03
CUDA	12.2
OS	Linux

Model Artifact Sizes

Stage	File	Size	Notes
Base model	LiquidAI/LFM2.5-VL-1.6B	~3.2 GB	Downloaded from Hugging Face
LoRA adapter weights	orion_lora_weights/	~50 MB	r=16, 4 target modules
Merged FP16 model	orion_merged/	~3.2 GB	Full standalone checkpoint
FP16 GGUF	orion-f16.gguf	~3.2 GB	Intermediate conversion
Q4_K_M GGUF	orion-q4_k_m.gguf	~730 MB	Deployed to Pi
Vision projector	orion-mmproj-f16.gguf	~814 MB	FP16, deployed to Pi

Quantization Compute

Step	RAM required	Time	Notes
HF to GGUF conversion	~8 GB	17.5s	Full FP16 model loaded into RAM
mmproj extraction	~4 GB	14s	Vision encoder only
Q4_K_M quantization	~4 GB	18.6s	Reads FP16 GGUF, writes Q4
Total disk (all stages)	~11 GB		Base + merged + F16 GGUF + Q4 GGUF + mmproj

Weight Fusion Compute

Resource	Requirement	Notes
RAM	~8 GB	Full FP16 model loaded on CPU (no GPU required)
Time	98.21s	merge_and_unload() + SafeTensors save
Disk	~3.2 GB output	Merged model saved to orion_merged/

Validation / Ablation Studies

ablation.py (base model) and evaluate.py (fine-tuned) are device-agnostic; hence, they run on CUDA, MPS (Apple Silicon), or CPU at FP16. CPU-only inference is functional but will be 50-100x slower than GPU.

Resource	Requirement	Notes
GPU VRAM	~4 GB (FP16, no quantization in eval)	Or run on CPU/MPS without VRAM budget
Test samples	60 (deterministic IID carve from 360 pool)	4 conditions x 60 = 240 inferences total
Val samples	60 (deterministic IID carve from 360 pool)	Same shape as test, used during training

Dataset

Item	Size	Notes
Target images	~36 MB	360 PNG images at ~100 KB each
train_dataset.jsonl	~1 MB	~480 records (240 targets x 2 with coordinate dropout)
val_dataset.jsonl	~200 KB	60 records
test_dataset.jsonl	~200 KB	60 records
Total dataset	~37 MB	Images + JSONL
Generation time	~3 min	360 images from SimSat at ~2 req/s

Data Transfer (Remote Server)

Operation	Data transferred	Method
Dataset upload	~31 MB (compressed)	upload_to_server.sh via rsync
LoRA weights download	~50 MB	download_weights.sh via rsync