Eagle model for CohereLabs/North-mini-code-1.0.
Key highlights include:
- three layers with dense transformer blocks
- all draft layers use sliding window attention
- compatible with fp8/w4a16 target model checkpoints
vllm
To start vllm server with this draft model, run the following
vllm serve CohereLabs/North-Mini-Code-1.0 \
-tp 2 \
--max-model-len 320000 \
--tool-call-parser cohere_command4 \
--reasoning-parser cohere_command4 \
--enable-auto-tool-choice \
--speculative_config "{\"method\": \"eagle\", \"model\": \"CohereLabs/North-Mini-Code-1.0-eagle\", \"num_speculative_tokens\": 3, \"draft_tensor_parallel_size\": 2}"