8BitStudio/Roblox-luau-coding_L1
Viewer • Updated • 12.3k • 327 • 3
🔄 Aniimage model coming soon
Jackson Yeskie
8BitStudio
P(doom) <0.01%
AI & ML interests
Text to image, Language models
Recent Activity
updated a dataset about 2 hours ago
8BitStudio/Roblox-luau-coding_L1 liked a Space about 2 hours ago
tardellirs/model-pulse upvoted a changelog 2 days ago
P(doom) on user profilesOrganizations
None yet
updated a
dataset about 2 hours ago
liked a
Space about 2 hours ago
upvoted a changelog 2 days ago
Hugging Face Changelog
P(doom) on user profiles
• 242
WAKE UP
11
#5 opened 3 days ago
by 
GGUFGuy
Thats cool! If you can say, how big is the model, im curious
Hugging Science is DEAD???
➕ 2
38
#2 opened 12 days ago
by 
XBFDI
Message board idea
👍 5
7
#3 opened 7 days ago
by 
el4
updated a
model 6 days ago
Line count
10
#2 opened about 1 month ago
by 
Sterling-Ai
reacted to harshitkgupta's post with 🚀 10 days ago
Post
8737
Fine-tuned Qwen 2.5 (0.5B → 3B) on real coding-agent traces, 10 controlled runs, one 16GB Mac. Compared PyTorch MPS vs. Apple MLX for local LoRA SFT — and the honest answer is "it depends on what you're optimizing for":
• PyTorch MPS: 2.2x–5.7x faster raw throughput, but hits a hard memory wall — can't load a 3B model in FP16 on 16GB.
• Apple MLX: 4-bit QLoRA fits 3B+ models with almost flat memory scaling as context grows (+109 MB going from 1k→4k tokens).
• 4-bit quantization doesn't cost you convergence — eval loss tracks closely across backends.
• The bigger surprise: most of MLX's slowdown isn't the 4-bit dequant tax. Two of the 10 runs went unquantized to isolate it — dequant only explains 1.07x–1.4x of the gap. A ~4.1–4.6x framework-level gap remains either way.
All 10 LoRA adapters + Trackio logs are public so the numbers are checkable, not just claimed.
Full writeup: https://huggingface.co/blog/harshitkgupta/fine-tuning-coding-agents-on-mac-pytorch-mps-mlx
• PyTorch MPS: 2.2x–5.7x faster raw throughput, but hits a hard memory wall — can't load a 3B model in FP16 on 16GB.
• Apple MLX: 4-bit QLoRA fits 3B+ models with almost flat memory scaling as context grows (+109 MB going from 1k→4k tokens).
• 4-bit quantization doesn't cost you convergence — eval loss tracks closely across backends.
• The bigger surprise: most of MLX's slowdown isn't the 4-bit dequant tax. Two of the 10 runs went unquantized to isolate it — dequant only explains 1.07x–1.4x of the gap. A ~4.1–4.6x framework-level gap remains either way.
All 10 LoRA adapters + Trackio logs are public so the numbers are checkable, not just claimed.
Full writeup: https://huggingface.co/blog/harshitkgupta/fine-tuning-coding-agents-on-mac-pytorch-mps-mlx
- 7 replies
replied to Banaxi-Tech's post 12 days ago
does BananaMind have a website? Would love to easily test these on it if so
liked a
model 12 days ago
YOO WHAT?
60
#1 opened 14 days ago
by 
Bc-AI
I think they are fine, and they are cool, but is there a clear use for them? What are you for, @BananaMindBot
liked a
model 15 days ago
Saw it being used as a computer automation tool (i.e. open an app, type, something, etc) I think once more open jev type models come out we will have crazy things
I cant imagine what we will do when the human brain is mapped out in 20 years
With the usage these days you basically have to 😭
liked a
dataset 22 days ago
liked a
model 23 days ago
