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

NVFP4 (W4A4) quantization of huihui-ai/Huihui-LFM2.5-8B-A1B-abliterated — the abliterated (refusal-reduced) build of Liquid AI's 8.3B-total / 1.5B-active mixture-of-experts reasoner (131 K context). Runs on one 16 GB Blackwell GPU with room for a ~1.1 M-token KV cache (fp8), and scales cleanly to 2–4 GPUs.

Quantized by Lna-Lab with NVIDIA TensorRT Model-Optimizer (modelopt). To our knowledge this is the first NVFP4 build of an abliterated lfm2_moe.

⚠️ Uncensored model. The base is abliterated by huihui.ai — its safety filtering is significantly reduced. See Usage warnings below. (Per the base card, the abliteration touches the dense path; the MoE experts were not ablated.)

Why it's nice: 8.3B total / 1.5B active MoE + a hybrid backbone (only 6 of 24 layers are attention; the rest are short-convolution) means the KV cache is tiny. Shrink the weights to 4-bit and the freed VRAM turns straight into concurrency — one card happily serves a stack of parallel sessions, and TP fans that out further.

---

📊 Measured throughput — 1 / 2 / 4× RTX PRO 2000 Blackwell (16 GB, SM120)

docker vllm/vllm-openai:v0.22.0, --quantization modelopt, fp8 KV, max-model-len 32768, 256-token decode via /v1/completions (ignore_eos+min_tokens). Aggregate output tokens/s at concurrency C:

C	TP=1 (1 GPU)	TP=2 (2 GPU)	TP=4 (4 GPU)
1	130	210	305
2	206	323	452
4	363	595	808
8	687	1147	1557
16	1166	1831	2463
32	2111	3273	4545

	TP=1	TP=2	TP=4
weights / GPU	~5.3 GB	~2.7 GB	~1.4 GB
KV pool (fp8)	1.16 M tok	3.17 M tok	7.82 M tok
max concurrency @ 32 K ctx	35×	97×	239×

• Single-stream scales with TP: 130 → 210 → 305 tok/s (1→2→4 GPU).
• Aggregate scales near-linearly with concurrency; TP=4 reaches ~4.5 K tok/s at C=32 and still has KV headroom for hundreds of sessions.
• Numbers are single-run and indicative (±~10–15 % at high concurrency). bf16 KV trades ~half the KV pool for marginally different speed.

Coherence spot-check (greedy / EN-JA-code): MoE explanation, Japanese, and a linked-list reversal all correct; reasoning, multilingual and code generation preserved.

---

🔗 Multi-GPU (TP) on no-NVLink Blackwell — important

These workstation cards have no NVLink (PCIe-only). With plain --tensor-parallel-size 2/4, NCCL deadlocks at communicator init (GPUs spin at 100 % util, no progress). Launch TP with P2P disabled:

docker run -d --gpus '"device=0,1,3,4"' --ipc=host \
  -e NCCL_P2P_DISABLE=1 -e NCCL_CUMEM_ENABLE=0 \
  -v $PWD:/model:ro -p 8000:8000 --shm-size 16g \
  --entrypoint vllm vllm/vllm-openai:v0.22.0 \
  serve /model --served-model-name lfm25-abl \
    --quantization modelopt --tensor-parallel-size 4 \
    --disable-custom-all-reduce \
    --kv-cache-dtype fp8 --max-model-len 32768 \
    --max-num-seqs 256 --gpu-memory-utilization 0.90 \
    --host 0.0.0.0 --port 8000

P2P-off staging is fine for this small MoE (per-token all-reduce volume is tiny). On NVLink hardware you can drop NCCL_P2P_DISABLE / --disable-custom-all-reduce.

---

🧠 It's a reasoning model — chat template

LFM2.5 uses a ChatML-like template and emits an explicit <think> … </think> chain of thought before the final answer. tokenizer.apply_chat_template(...) renders, e.g.:

<|startoftext|><|im_start|>user
What is the capital of France?<|im_end|>
<|im_start|>assistant

In vLLM, pass --reasoning-parser deepseek_r1 — it matches the </think> delimiter so the OpenAI API returns the CoT in reasoning_content and the answer in content. Omit it for raw text (think tags included).

🛠 Tool use (agentic)

Pass tools via apply_chat_template(..., tools=[...]). By default the model emits Pythonic calls between <|tool_call_start|> and <|tool_call_end|>:

<|tool_call_start|>[get_weather(city="Tokyo")]<|tool_call_end|>

For OpenAI-API auto-parsing add --enable-auto-tool-choice --tool-call-parser pythonic (if your vLLM build's Pythonic parser handles the wrapper); otherwise parse the special-token block yourself. The bundled chat_template.jinja renders tools / calls / tool-role results.

---

🚀 Serve with vLLM

Single GPU:

CUDA_VISIBLE_DEVICES=0 vllm serve sakamakismile/Huihui-LFM2.5-8B-A1B-abliterated-NVFP4 \
    --served-model-name lfm25-abl \
    --quantization modelopt \
    --kv-cache-dtype fp8 \
    --max-model-len 128000 \
    --max-num-seqs 16 \
    --gpu-memory-utilization 0.90 \
    --reasoning-parser deepseek_r1 \
    --port 8000

(For TP=2/4 see the no-NVLink launch above.)

flag	what it does for this model
--quantization modelopt	required — reads hf_quant_config.json (NVFP4). Omit it and weights load as garbage.
--kv-cache-dtype fp8	~2× KV capacity (used for the table above).
--max-num-seqs	concurrency; KV is cheap here, so 16–256 is comfortable.
--max-model-len	up to 131072 (native).
--reasoning-parser deepseek_r1	separates <think> CoT from the answer.
--tensor-parallel-size	1 fits one card; 2/4 for more single-stream speed + KV (mind the NCCL note).

Offline

from vllm import LLM, SamplingParams
llm = LLM("sakamakismile/Huihui-LFM2.5-8B-A1B-abliterated-NVFP4", quantization="modelopt",
          max_model_len=32768, gpu_memory_utilization=0.90, max_num_seqs=8)
tok = llm.get_tokenizer()
chat = tok.apply_chat_template(
    [{"role": "user", "content": "日本語で自己紹介して。"}],
    tokenize=False, add_generation_prompt=True)
print(llm.generate([chat], SamplingParams(
    temperature=0.2, top_k=80, repetition_penalty=1.05, max_tokens=512))[0].outputs[0].text)

Sampling (Liquid's recommendation): temperature=0.2, top_k=80, repetition_penalty=1.05. It thinks first, so give it max_tokens ≥ 512.

---

⚠️ Usage notes & caveats

• Needs Blackwell (SM120) + a recent vLLM (≥0.21 with NVFP4/modelopt) and flashinfer — the FP4 GEMM and MoE run on FlashInfer-CUTLASS kernels.
• ModuleNotFoundError: No module named 'trinity_turbo' in the logs is harmless.
• If a MoE backend objects to the FP4 scales, force Marlin: VLLM_USE_FLASHINFER_MOE_FP4=0.
• Use --quantization modelopt only — not fp8/awq/gptq.
• A handful of rarely-routed ("cold") experts are calibrated from limited activation coverage and exported with a weight-derived scale fallback; for the overwhelming majority of tokens, output tracks the source closely.
• Uncensored / abliterated (refusal-reduced) — inherited from the base. Safety filtering is significantly reduced; outputs may be sensitive or inappropriate. Use for research / controlled settings, monitor and review outputs, and comply with local law. See the base model card for the full warnings. Quantization changes precision only — no behavioral change vs the abliterated base.

🔬 What's quantized

NVFP4 = e2m1 weights, 16-wide blocks, FP8-e4m3 block scales + FP32 global scale, static per-tensor input_scale.

• → NVFP4: all 32 MoE experts (per layer) + the 2 dense MLP layers.
• kept BF16: attention (q/k/v/out), short-conv projections, the MoE router (feed_forward.gate), token embeddings, and lm_head.

Full recipe + scripts: Lna-Lab lnarizer/recipes/lfm2_moe/ (the one modelopt calibration patch for lfm2_moe, plus the expert-key remap for vLLM). The exact same recipe produced the non-abliterated LFM2.5-8B-A1B-NVFP4 unchanged.

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 → this NVFP4 build.

Credits

• Base model: Liquid AI — LFM2.5-8B-A1B.
• Abliteration: huihui.ai — Huihui-LFM2.5-8B-A1B-abliterated.
• NVFP4 quantization & packaging: Lna-Lab.
• Tooling: NVIDIA TensorRT Model-Optimizer, vLLM, FlashInfer.

💖 Support the Base Model Author

This NVFP4 build 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 · NVFP4 for Blackwell · LLMs without colored glasses, in 4-bit, for the edge

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