nebius/MTP-DeepSeek-V3-0324

Model Description

This model is a fine-tuned version of DeepSeek-V3's native MTP module, optimized for speculative decoding using LK losses — training objectives that directly target acceptance rate rather than using KL divergence as a proxy.

The original DeepSeek-V3 MTP module was trained primarily for first-token prediction and reused autoregressively for later positions, causing degraded acceptance rates at later draft positions. Our fine-tuning addresses this mismatch, substantially improving multi-token speculation performance.

Training Details

• Base model: deepseek-ai/DeepSeek-V3-0324
• Draft architecture: MTP
• Training data: Infinity-Instruct-0625 with DeepSeek-V3 generated responses
• Training objective: Hybrid LK loss with adaptive λ scheduling (η=3)
• Training: 1 epoch from pretrained MTP weights
• Draft length: K = 6 speculative tokens

Performance

Average acceptance length (τ) measured across MT-bench, HumanEval, and GSM8K with K = 7:

Configuration	Temperature = 0	Temperature = 1
Original MTP	3.20	3.09
MTP + KL fine-tuning	4.79	4.43
MTP + LK fine-tuning (ours)	4.83	4.68

Note: Earlier vLLM versions sampled draft tokens greedily regardless of temperature, which underestimated acceptance rates at temperature > 0. Stochastic draft sampling was introduced in v0.18.0, and from v0.21.0 it can be enabled via speculative_config using rejection_sample_method and draft_sample_method. The acceptance metrics reported above were measured under standard rejection sampling and are reproducible with the configuration below.

Usage with vLLM

from vllm import LLM, SamplingParams

llm = LLM(
    model="deepseek-ai/DeepSeek-V3-0324",
    speculative_config={
        "method": "deepseek_mtp",
        "model": "nebius/MTP-DeepSeek-V3-0324",
        "num_speculative_tokens": 6,
        "rejection_sample_method": "standard",
        "draft_sample_method": "gumbel",
    },
    tensor_parallel_size=8,
    max_num_seqs=1
)

sampling_params = SamplingParams(temperature=0.7)
outputs = llm.generate(["Explain speculative decoding in simple terms."], sampling_params)

License

CC BY 4.0

Citation

@misc{samarin2026lklosses,
  title     = {LK Losses: Direct Acceptance Rate Optimization for Speculative Decoding},
  author    = {Alexander Samarin and Sergei Krutikov and Anton Shevtsov and Sergei Skvortsov and Filipp Fisin and Alexander Golubev},
  year      = {2026},
  eprint    = {2602.23881},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url       = {https://arxiv.org/abs/2602.23881}
}