For your situation, I would suggest looking at things from a budget and value perspective since you're rocking a single NVIDIA GeForce RTX 3090 24GB Graphics Card. I've been fine-tuning models for a few years now, and honestly, the "best" framework realy depends on how much you wanna fight with your hardware limits. Here is how the top contenders stack up for DeepSeek: 1. Unsloth - Pros: This is the low-cost champion for consumer gear. It's literally 2x faster and uses way less VRAM than the standard scripts. It's definitely the most stable way to fine-tune a DeepSeek-R1-Distill-Qwen-7B or DeepSeek-R1-Distill-Llama-8B model without hitting OOM every five minutes.
- Cons: It's a bit more specialized, so it might lack some of the super niche experimental features you'd find in more complex setups. 2. Axolotl
- Pros: Massive flexibility. If you have a complex dataset strategy or want to experiment with every optimizer under the sun, this is the one. It's what the pros use for diverse training runs.
- Cons: It's definitely harder to keep VRAM usage low. I've found it's better for multi-GPU setups tbh. I mean, it's a great tool, but might be overkill for a single card. Basically, if you want to save time and keep your sanity on a single card, Unsloth is the clear winner. I've learned the hard way that fighting memory errors for hours is just not worth it... anyway, hope that helps!! gl!

Honestly, I spent hours fighting OOMs until I tried Unsloth. It realy ran DeepSeek-R1-Distill-7B on my NVIDIA GeForce RTX 3090 24GB perfectly for $0. Basically a lifesaver, lol!

For your situation, I would suggest Unsloth. Been around for years but I'm still learning... it's reallyyy good for NVIDIA GeForce RTX 3090 24GB VRAM. Just be careful!

Just catching up on this thread and everyone is spot on about Unsloth being the speed king. But honestly, I've had some real headaches with stability when doing long-haul training. If you're like me and you've ever woken up to a crashed script after 8 hours of progress, you might want to consider LLaMA-Factory Framework instead. It's a bit of a different vibe but super reliable for long sessions. I remember when I first tried tweaking DeepSeek models on my NVIDIA GeForce RTX 3090 24GB, I kept hitting these weird memory spikes that felt random. I switched over to LLaMA-Factory and the built-in monitoring really saved my skin. It uses bitsandbytes 4-bit quantization in a way that feels a lot more robust for consumer gear imo. Just be careful tho, because even with 24GB of VRAM, things can get dicey if your context window is too big. I always suggest starting with a smaller sequence length first just to make sure your card can handle the heat... literally. My 3090 used to sound like a jet engine until I dialed in the settings properly in the config. It's not as fast as the alternatives but I havent had a single failed run since I made the switch.