python -m sglang.launch_server \
  --model Qwen/Qwen3.5-9B \
  --speculative-algorithm EAGLE3 \
  --speculative-draft-model-path BLR2/Qwen3.5-9B-Eagle3-ShareGPT \
  --trust-remote-code \
  --mamba-scheduler-strategy extra_buffer

Training Details

Architecture: LlamaForCausalLMEagle3 (single decoder layer)
Target model: Qwen/Qwen3.5-9B
Training data: ShareGPT conversations
Training steps: 71,000
Draft vocabulary: 32,000 most frequent tokens
Precision: BF16