Instructions to use SubconsciousDev/glm-5.2-fp8-dflash-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use SubconsciousDev/glm-5.2-fp8-dflash-v1 with Transformers:
# Load model directly from transformers import AutoTokenizer, AutoModel tokenizer = AutoTokenizer.from_pretrained("SubconsciousDev/glm-5.2-fp8-dflash-v1") model = AutoModel.from_pretrained("SubconsciousDev/glm-5.2-fp8-dflash-v1", device_map="auto") - Notebooks
- Google Colab
- Kaggle
| library_name: transformers | |
| tags: | |
| - dflash | |
| - speculative-decoding | |
| - sglang | |
| - glm-5.2 | |
| # GLM-5.2 FP8 DFLASH v1 | |
| ## Overview | |
| This is a DFLASH speculative draft model for GLM-5.2 FP8 serving. The checkpoint | |
| uses DFLASH block size 16 and is intended to be loaded as the draft model in | |
| SGLang speculative decoding. | |
| Training data: `SubconsciousDev/Subconscious-Dflash-Training-Dataset-short-glm52-v1`. | |
| ## SGLang Usage | |
| Add these arguments to the SGLang launch command: | |
| ```bash | |
| --speculative-algorithm DFLASH \ | |
| --speculative-draft-model-path SubconsciousDev/glm-5.2-fp8-dflash-v1 \ | |
| --speculative-num-draft-tokens 16 \ | |
| ``` | |
| ## Evaluation | |
| Concurrency is 1. | |
| | Benchmark | EAGLE (5,1,6) Speedup | EAGLE (5,1,6) Accept Length | DFLASH Speedup | DFLASH Accept Length | | |
| |---|---:|---:|---:|---:| | |
| | GSM8K | 2.2x | 3.99 | 2.3x | 4.87 | | |
| | MATH500 | 3.4x | 4.92 | 4.3x | 7.44 | | |
| | MBPP | 3.5x | 5.21 | 4.9x | 9.03 | | |
| | LiveCodeBench | 2.9x | 4.17 | 3.1x | 5.35 | | |
| | MT-Bench | 2.5x | 3.71 | 2.4x | 3.96 | | |