A 1-layer EAGLE3 speculative-decoding draft head for MiniMax-M2.7, with the lm_head vocab-pruned to the top 32,000 tokens (~99.1% coverage). The trimmed [32000, 3072] lm_head is 6.25× smaller than the full-vocab head, giving a faster per-step draft forward — the deployment variant of the full-vocab draft.

Performance (mean speculative accept length)

Config num_steps=3, topk=1, draft_tokens=4, bf16 draft:

benchmark	mean accept length
HumanEval	2.59

Pruning trades a small amount of accept length for a 6.25× smaller lm_head and faster draft forward.

Architecture

LlamaForCausalLMEagle3, num_hidden_layers=1, hidden_size=3072, vocab_size=200064, draft_vocab_size=32000, bf16. Carries t2d/d2t token maps in the safetensors.

Usage (SGLang)

bash

python -m sglang.launch_server \
  --model-path MiniMaxAI/MiniMax-M2.7 --tp 4 --trust-remote-code \
  --reasoning-parser minimax --tool-call-parser minimax-m2 \
  --attention-backend triton --speculative-draft-attention-backend triton \
  --speculative-algorithm EAGLE3 \
  --speculative-draft-model-path asherszhang/MiniMax-M2.7-EAGLE3-draft-vocab32k \
  --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
  --speculative-draft-model-quantization unquant

Serve the draft as bfloat16; pass unquant so it isn't force-quantized to the target's FP8.

License

Draft head weights released under apache-2.0. Derivative of MiniMax-M2.7 (embeds its embed_tokens) — see the base model for its terms.