[Question] Best Practice With Multiple GPUs on 1 node

Hello, I'm creating a fine-tuned LLM based off of StableLM 3b Tuned model. I used lora.py with 4 Nvidia L4(24 GB) on GCP(I don’t have access to A100 but anything else I do). My observation is that for the amount of compute I'm providing the fine-tuning takes too long and consumes too much VRAM which has led to reducing max-seq-length to 2000 tokens instead of the max of 4096. The fine-tuning dataset is lighteval/legal-summarization on hugging Face and I adapted it to a alpaca format. Here are my questions:

1. Should I be using different GPU's as I looked at the benchmarks on the github page and it showed 3 min for 1000 iterations (1 A100) while for me it was around 40 minutes?
2. My dataset's input goes well beyond 4096 tokens, does the lora.py file do any type of chunking or does it just take the first 4096 tokens?
3. For micro-batch-size of 4 and batches of 128 and max-seq-length of 500 tokens, the gpu utilization on all gpu's was around 18GB (using bf16-true). Shouldn't this be lower for 3B parameter model?
4. What can I do to make this fine-tuning process faster (should i spawn a cluster of nodes or 1 instance with multiple gpus is enough)?

Start by reading lora.py and the repository's documented multi-GPU fine-tuning entry points, then compare the reported 4×L4 setup with the benchmark cited in the issue. Check how lighteval/legal-summarization inputs are handled at max-seq-length and how bf16-true affects memory; done means a maintainer-confirmed explanation and actionable guidance for the four questions.