Update README.md
Browse files
README.md
CHANGED
|
@@ -13,9 +13,15 @@ tags:
|
|
| 13 |
- mlx
|
| 14 |
---
|
| 15 |
|
| 16 |
-
# bobig/DeepScaleR-1.5B-6.5bit
|
| 17 |
|
| 18 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
converted to MLX format from [agentica-org/DeepScaleR-1.5B-Preview](https://huggingface.co/agentica-org/DeepScaleR-1.5B-Preview)
|
| 20 |
using mlx-lm version **0.21.4**.
|
| 21 |
|
|
@@ -28,7 +34,7 @@ pip install mlx-lm
|
|
| 28 |
```python
|
| 29 |
from mlx_lm import load, generate
|
| 30 |
|
| 31 |
-
model, tokenizer = load("bobig/DeepScaleR-1.5B-6.5bit
|
| 32 |
|
| 33 |
prompt = "hello"
|
| 34 |
|
|
|
|
| 13 |
- mlx
|
| 14 |
---
|
| 15 |
|
| 16 |
+
# bobig/DeepScaleR-1.5B-6.5bit
|
| 17 |
|
| 18 |
+
This works as a draft model for speculative decoding in [LMstudio 3.10 beta](https://lmstudio.ai/docs/advanced/speculative-decoding)
|
| 19 |
+
|
| 20 |
+
Try it with: [mlx-community/FuseO1-DeepSeekR1-Qwen2.5-Coder-32B-4.5bit](https://huggingface.co/mlx-community/FuseO1-DeepSeekR1-Qwen2.5-Coder-32B-4.5bit)
|
| 21 |
+
|
| 22 |
+
you should see 30% faster TPS for math/code prompts even with "thinking"
|
| 23 |
+
|
| 24 |
+
The Model [bobig/DeepScaleR-1.5B-6.5bit](https://huggingface.co/bobig/DeepScaleR-1.5B-6.5bit) was
|
| 25 |
converted to MLX format from [agentica-org/DeepScaleR-1.5B-Preview](https://huggingface.co/agentica-org/DeepScaleR-1.5B-Preview)
|
| 26 |
using mlx-lm version **0.21.4**.
|
| 27 |
|
|
|
|
| 34 |
```python
|
| 35 |
from mlx_lm import load, generate
|
| 36 |
|
| 37 |
+
model, tokenizer = load("bobig/DeepScaleR-1.5B-6.5bit")
|
| 38 |
|
| 39 |
prompt = "hello"
|
| 40 |
|