LFM2.5-230M-ToMoE

A Mixture-of-Experts (MoE) conversion of LiquidAI/LFM2.5-230M, obtained with the ToMoE method (Dynamic Structural Pruning + Hypernetwork).

The dense 230M base model is converted into a channel-MoE where:

• MLP (14 layers): per-token expert routing over 8 channel-experts (top-1 via Gumbel-softmax), the FFN is cut to the union of the expert channel masks.
• Full attention (6 layers): Q/K kept at full width; the V channels are dynamically masked per token (deterministic masks).
• ShortConv (8 layers): statically pruned (a single channel index shared by the B/C/x projections, the conv and the output projection).

Metrics

Metric	Dense base	ToMoE MoE
Total params	229.7M	200.9M stored
Active params per token	229.7M	~151M
PPL wikitext-2 (2048 tok)	20.4	~815
Embeddings (always active)	67.1M	67.1M

The pruned budget is p = 0.3 with the ToMoE regularization (lam 64→256, Gumbel base annealed 1.0→−0.25, 8 experts). The attention QK is excluded from the budget (kept full) because the small head-dim (64) collapses under aggressive QK pruning.

Usage (transformers, trust_remote_code)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Nichonauta/LFM2.5-230M-ToMoE",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="cuda:0",
    attn_implementation="eager",
)
tokenizer = AutoTokenizer.from_pretrained("Nichonauta/LFM2.5-230M-ToMoE")

prompt = "The capital of France is"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda:0")
out = model.generate(**inputs, max_new_tokens=30)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Note: this model requires trust_remote_code=True — the custom modeling_lfm2_moe_final.py defines the Lfm2MoEForCausalLM class.

Conversion details

• Base: LiquidAI/LFM2.5-230M (LFM Open License v1.0)
• Method: ToMoE — a bidirectional-GRU hypernetwork generates differentiable channel masks per sublayer; the model weights are frozen and only the hypernetwork is trained (online KD against the dense teacher, wikitext-103 stream). The final pruned weights are extracted from the trained masks.
• Training: 1,200 steps, seq 2048, online knowledge distillation (2× forward KL), gradient checkpointing.

Limitations

• The model is significantly degraded versus the dense base (PPL 815 vs 20.4) — the pruning aggressively reduces the conv layers (12–22% width) and the MLP per-token width (62%), and generation tends to repeat.
• This is a research artifact demonstrating the ToMoE pipeline on a small LFM2 architecture.

License

This model is a derivative of LiquidAI/LFM2.5-230M and is released under the LFM Open License v1.0 (see LICENSE).

Files

• model.safetensors — the pruned MoE weights (fp32)
• modeling_lfm2_moe_final.py — the custom model definition (trust_remote_code)
• config.json — model configuration (with auto_map)
• tokenizer.json, tokenizer_config.json, chat_template.jinja, generation_config.json

GGUF quantizations

Quantized GGUF files (Q4_K_M, Q8_0, F16) are available in the companion repository: Nichonauta/LFM2.5-230M-ToMoE-GGUF.