Spaces:
Running
Is "jedisct1/Qwen3-Coder-30B-A3B-Instruct-q4-mlx" a valid/good conversion of Unsloth's dyn 4bit quant of Qwen3 30B Coder ?
Downloaded the MLX flavour of Unsloth's dynamic 4bit quant of Qwen3 30B Coder model (this one: "jedisct1/Qwen3-Coder-30B-A3B-Instruct-q4-mlx" to be precise) from HF, but in LM-Studio the "Load Model" button remains disabled. I'm running LM Studio 0.3.30 (Build 2), on a Mac Mini M4 with 24GB Unified RAM.
When I select the "Load Anyway" option, the red "Load Anyway" button appears, which when clicked, does appear to load the model. However, send sending a text prompt via the LM-Studio chat interface to it, leads to the following error: The model has crashed without additional information. (Exit code: 6)
Note that I have not changed the context-window length (still at 4K), but I've changed Temperature to 0.7, changed top-K, min-P, repetition-penalty, as suggested by Unsloth to maximize performance.
Is there any information on whether the said model is actually valid one, as it is not from mlx_community but from jedisctl ? Or any other way to check ? I don't have access to other GPU based systems or cloud GPU for now.