An EAGLE-3 / MTP draft head for speculative decoding of sakamakismile/Huihui-LFM2.5-8B-A1B-abliterated-NVFP4 in vLLM. It is a tiny (≈0.4 B) single-layer drafter that proposes tokens the NVFP4 target then verifies, so single-stream decode runs faster with identical-quality (greedy-lossless) output.

This is a companion model — it does nothing on its own; load it as the speculative_config draft for the NVFP4 target above.

Measured speedup (greedy, compiled vLLM, 256-tok decode)

config	base t/s	with this MTP draft	speedup	draft acceptance
TP=1 (1× RTX PRO 2000)	129.8	154.2	1.19×	59.8 %
TP=4 (4× RTX PRO 2000)	301.9	339.8	1.13×	59.8 %

Mean accepted length ≈ 1.60 tok/step at num_speculative_tokens=1. Acceptance is a property of the draft (TP-independent). Speedup is larger at lower TP (the target is slower there, so the draft saves relatively more).

Usage (vLLM)

python

from vllm import LLM, SamplingParams

llm = LLM(
    model="sakamakismile/Huihui-LFM2.5-8B-A1B-abliterated-NVFP4",
    quantization="modelopt",
    speculative_config={
        "method": "eagle3",
        "model": "sakamakismile/Huihui-LFM2.5-8B-A1B-abliterated-MTP-NVFP4",
        "num_speculative_tokens": 1,
    },
    max_model_len=8192,
)
# This is a reasoning/chat model — ALWAYS chat-template prompts.
sp = SamplingParams(temperature=0.0, max_tokens=256)
print(llm.chat([{"role": "user", "content": "Explain speculative decoding."}], sp)[0].outputs[0].text)

Server + launch flags (`vllm serve`)

TP=1 (single GPU):

bash

vllm serve sakamakismile/Huihui-LFM2.5-8B-A1B-abliterated-NVFP4 \
  --quantization modelopt \
  --speculative-config '{"method":"eagle3","model":"sakamakismile/Huihui-LFM2.5-8B-A1B-abliterated-MTP-NVFP4","num_speculative_tokens":1}' \
  --max-model-len 8192 \
  --reasoning-parser deepseek_r1          # optional: the model emits <think>…</think>

TP=4 (4 GPUs, no-NVLink / PCIe box):

bash

NCCL_P2P_DISABLE=1 NCCL_CUMEM_ENABLE=0 \
vllm serve sakamakismile/Huihui-LFM2.5-8B-A1B-abliterated-NVFP4 \
  --quantization modelopt \
  --tensor-parallel-size 4 \
  --disable-custom-all-reduce \
  --speculative-config '{"method":"eagle3","model":"sakamakismile/Huihui-LFM2.5-8B-A1B-abliterated-MTP-NVFP4","num_speculative_tokens":1}' \
  --max-model-len 8192

flag	why
`--quantization modelopt`	the target is NVFP4 (ModelOpt FP4)
`--speculative-config '{"method":"eagle3","model":"…MTP-NVFP4","num_speculative_tokens":1}'`	turn on EAGLE-3 spec-decode with this draft (k=1)
`--tensor-parallel-size N` + `--disable-custom-all-reduce`	multi-GPU
`NCCL_P2P_DISABLE=1 NCCL_CUMEM_ENABLE=0` (env)	required on no-NVLink/PCIe boxes or NCCL deadlocks at init
`--max-model-len 8192`	context window (raise as needed)
do not* pass `--enforce-eager`*	the compiled path is where spec-decode is actually faster; eager misleads

Offline LLM(...) equivalents: quantization="modelopt", speculative_config={…}, tensor_parallel_size=N, disable_custom_all_reduce=True, enforce_eager=False, max_model_len=8192.

⚠️ Requires a vLLM with lfm2_moe speculative-decode support. Stock vLLM does not yet wire EAGLE-3 / short_conv spec-decode correctly for this hybrid (short-conv + attention + MoE) arch. The fix is upstream PR vllm-project/vllm#44296 (+ a model-side SupportsEagle3 patch). Until merged, apply those patches (4 files). On unpatched vLLM the draft loads but acceptance is ~0 %. For multi-GPU on no-NVLink boxes: NCCL_P2P_DISABLE=1 NCCL_CUMEM_ENABLE=0 + disable_custom_all_reduce.

How it was trained

Method: EAGLE-3 — one Llama-style decoder layer (qkv input = 2·H: token-embed ⊕ fused aux hidden), fc: 3·H→H fusing aux from target layers (2, 12, 21), own lm_head, d2t vocab map.
Reduced draft vocab = 32 000 (top-frequency, covers 99.8 % of tokens). This is what makes spec-decode actually faster — a full 128 k draft head is as costly as the target's and caps the speedup.
Training data: 105 371 sequences of NVFP4-matched hidden states — captured by serving the quantized target itself (not a BF16 proxy), which is what lets serving acceptance track training. Single-step (k=1), 3 epochs, reduced-vocab CE + SmoothL1 hidden-regression.
In-sample top-1 ≈ 0.84.

Config (vLLM `Eagle3LlamaForCausalLM`)

hidden 2048 · 1 layer · 32 heads / 8 KV · head_dim 64 · inter 8192 · draft_vocab 32000 · target_hidden 2048 · rope_theta 5e6 · eagle_config.use_aux_hidden_state: true.

Limitations

k=1 only for now. num_speculative_tokens ≥ 2 currently diverges on the short_conv kernel (a separate multi-query conv bug); higher k is the main lever left for bigger speedups (the ~400 t/s TP=4 target needs it).
Inherits the base model's abliterated/uncensored behavior and its license.

License

Inherits the base license (LFM Open License v1.0, license_name: lfm1.0) — see the bundled LICENSE. Lineage: LiquidAI/LFM2.5-8B-A1B → huihui-ai/Huihui-LFM2.5-8B-A1B-abliterated → …-NVFP4 (the target) → this MTP / EAGLE-3 draft. The draft only proposes tokens the target verifies — greedy output is unchanged vs the abliterated base.

Credits

Base model: Liquid AI — LFM2.5-8B-A1B.
Abliteration: huihui.ai — Huihui-LFM2.5-8B-A1B-abliterated.
NVFP4 target + EAGLE-3 / MTP draft training & packaging: Lna-Lab.
Tooling: vLLM, EAGLE-3, NVIDIA TensorRT Model-Optimizer, FlashInfer.

💖 Support the Base Model Author

This draft stands on huihui.ai's abliteration work. If it's useful to you, please consider supporting them — a cup of coffee goes a long way:

Ko-fi: https://ko-fi.com/huihuiai
Bitcoin: bc1qqnkhuchxw0zqjh2ku3lu4hq45hc6gy84uk70ge

🔬 Lna-Lab · EAGLE-3 / MTP speculative decoding for Blackwell · faster tokens, same output, in 4-bit

Trained & benchmarked on 7× RTX PRO 2000 Blackwell · 2026

sakamakismile/Huihui-LFM2.5-8B-A1B-abliterated-MTP-NVFP4

Huihui-LFM2.5-8B-A1B-abliterated-MTP-NVFP4