Ground Segment Budgets¶

GPU VRAM: run nvidia-smi during training
Training time: logged by HuggingFace Trainer at end of each epoch
Quantization time: wall-clock the llama-quantize command
Validation time: wall-clock python validation.py

Resource requirements for training, quantization, and dataset generation.

Training Hardware¶

Resource	Requirement	Notes
GPU	CUDA-capable, 8+ GB VRAM	QLoRA loads base model in 4-bit (NF4)
GPU VRAM usage	TBD	Micro-batch size 1, gradient checkpointing enabled
System RAM	16+ GB recommended	For data loading and preprocessing
Training time per epoch	TBD	3 epochs total, ~480 training samples
Total training time	TBD	Depends on GPU

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`	~700 MB	Deployed to Pi
Vision projector	`orion-mmproj-f16.gguf`	~854 MB	FP16, deployed to Pi

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

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

Resource	Requirement	Notes
GPU VRAM	Same as training (~8+ GB)	Model loaded in 4-bit for inference
Test samples	~60 (20% of 300 targets)	4 conditions x 60 = 240 inferences total
Time per inference	TBD	Single image + prompt per sample
Total validation time	TBD	240 inferences per script

Item	Size	Notes
Target images	~30 MB	300 PNG images at ~100 KB each
train_dataset.jsonl	~1 MB	~480 records (240 targets x 2 with coordinate dropout)
test_dataset.jsonl	~200 KB	~60 records
Total dataset	~31 MB	Images + JSONL
Generation time	~2.5 min	300 images from SimSat at ~2 req/s

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