| base_model: Qwen/QwQ-32B | |
| gate_mode: random | |
| architecture: qwen | |
| dtype: bfloat16 | |
| experts: | |
| - source_model: Qwen/QwQ-32B | |
| - source_model: Qwen/QwQ-32B | |
| - source_model: Qwen/QwQ-32B | |
| - source_model: Qwen/QwQ-32B | |
| - source_model: Qwen/QwQ-32B | |
| - source_model: Qwen/QwQ-32B | |
| - source_model: Qwen/QwQ-32B | |
| - source_model: Qwen/QwQ-32B | |
| shared_experts: | |
| - source_model: Qwen/QwQ-32B | |
| residual_scale: 0.1 # downweight output from shared expert to prevent overcooking the model | |