Sebesky's picture
Update README.md
86ffe4b verified
|
Raw
History Blame Contribute Delete
967 Bytes
metadata
license: mit
base_model:
  - Inferact/MiniMax-M3-EAGLE3
tags:
  - eagle3
  - speculative-decoding
  - compressed-tensors
  - int4

MiniMax-M3 EAGLE3 drafter — RTN INT4

EAGLE3 speculative-decoding draft head for Sebesky/MiniMax-M3-W4A16-GPTQ: a single Llama-style decoder layer (hidden 6144, 64-head full MHA, head_dim 128) with its own embed_tokens and lm_head (vocab 200,064) — not shared with the target. Everything is RTN-quantized to int4 (compressed-tensors pack-quantized, group 128): Linear, embedding, and lm_head. ~1.6 GB.

Usage (vLLM)

--speculative-config '{"method":"eagle3","model":"Sebesky/MiniMax-M3-EAGLE3-RTN-INT4","num_speculative_tokens":3}'

Validated on 2x DGX Spark (GB10) at TP2 with draft_tensor_parallel_size: 2: acceptance length 3.36 (78.6% draft acceptance) at 120k-token context, decode 39.7 / 30.4 / 26.4 tok/s @512/65k/120k with 4-bit KV quantization.