Instructions to use konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4")
model = AutoModelForCausalLM.from_pretrained("konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4

SGLang

How to use konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4 with Docker Model Runner:
```
docker model run hf.co/konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4
```

LFM2.5-8B-A1B-REAP-50-AWQ-INT4

Model Description

This is the REAP-pruned + AWQ INT4-quantized variant of LiquidAI/LFM2.5-8B-A1B, representing the full compression pipeline:

REAP pruning (50% experts removed) → konic-labs/LFM2.5-8B-A1B-REAP-50
AWQ INT4 quantization (W4A16_ASYM, group_size=128) → this model

Base: 8.47B params, 16.94 GB (bf16)
This model: 2.79 GB on disk (83.5% total compression)
Quant format: compressed-tensors pack-quantized (INT4 weights, bf16 activations, asymmetric, group_size=128)
Runtime VRAM: ~8.6 GB (weights decompress to bf16 in memory)
1,122 weight tensors quantized; routers, layer norms, embeddings, and Conv1d internals left unquantized

⚠️ Loading Requires a Small Patch

The compressed-tensors pack-quantized format has two bugs in current transformers that prevent loading asymmetric-quantized MoE expert weights. A self-contained patch is included in patches/apply_int4_patch.py.

Upstream status: These bugs are fixed upstream in PR huggingface/transformers#47430. The fix is general for any MoE model with fused expert nn.Parameter weights (LFM2.5, Qwen2-MoE family) quantized with an asymmetric compressed-tensors scheme. Once merged, this patch script becomes unnecessary — re-run it only until your installed transformers release includes the fix.

# Apply the patch (creates .orig backups, idempotent, --check to verify)
python patches/apply_int4_patch.py

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4",
    trust_remote_code=True,
    dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(
    "konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4", trust_remote_code=True
)

Re-run the patch after every pip install -U transformers. See patches/README.md for details and how to decompress to a vLLM-compatible bf16 checkpoint if you need vLLM inference (the INT4 pack-quantized format is not yet loadable by vLLM directly).

AWQ Quantization Details

Method: AWQ (Activation-aware Weight Quantization) via llm-compressor
Scheme: W4A16_ASYM (4-bit weights, 16-bit activations, asymmetric, group_size=128)
Calibration: 256 samples from evol-codealpaca-v1/train.jsonl
Custom LFM2.5 AWQ mappings: operator_norm → attention q/k/v projections, ffn_norm → MoE expert w1/w3 projections (LFM2.5 uses non-standard layer-norm names). v_proj→out_proj and conv.in_proj→conv.out_proj were skipped due to GQA dimension mismatch and Conv1d intermediary respectively.
Average AWQ smoothing error: ~5-7 × 10⁻⁴ (well-converged)

Benchmark Results

Benchmark	Base (published)	REAP-50 (pruned)	This model (REAP+AWQ-INT4)	Retention
MATH500	88.76%	77.0%	72.0%	81.1%
BFCLv3 (single-turn)	64.79%	59.07%	57.36%	88.5%

MATH500 per-difficulty (this model)

Level	Accuracy
L1	90.7%
L2	84.4%
L3	81.9%
L4	72.7%
L5	49.3%

BFCLv3 per-category (this model)

Category	Accuracy
simple_python	69.25%
simple_java	44.00%
simple_javascript	38.00%
multiple	79.00%
parallel	69.50%
parallel_multiple	66.00%
irrelevance	83.75%
live_simple	51.94%
live_multiple	48.24%
live_parallel	12.50%
live_parallel_multiple	45.83%
live_irrelevance	75.57%
live_relevance	50.00%

Key finding: Quantization adds only ~5pp MATH500 and ~1.7pp BFCLv3 degradation beyond pruning alone, while delivering a 67% disk-size reduction on top of REAP. L1-L3 math is completely unaffected; advanced/competition math (L4-L5) bears the cost.

Throughput (vLLM, NVIDIA L4)

Concurrency	Tokens/sec
64	2,227
128	3,334
256	4,809 (sweetspot)
384	4,287
512	4,273

(Measured on the AWQ-scaled bf16 decompressed variant, which is functionally identical to this INT4 model at runtime.)

Intended Use

Research and deployment of compressed MoE models. This is the full-pipeline output of REAP pruning + AWQ INT4 quantization — the smallest deployable variant.

Citation

@software{reap-cuda,
  title  = {REAP: Ranking and Eliminating Experts via Activation Patterns},
  author = {Konic Labs},
  url    = {https://github.com/egesabanci/reap-cuda}
}

License

Apache-2.0 (inherited from base model).

Downloads last month: 61

Safetensors

Model size

5B params

Tensor type

I64

F32

I32

BF16

Model tree for konic-labs/LFM2.5-8B-A1B-REAP-50-AWQ-INT4

Base model

LiquidAI/LFM2.5-8B-A1B-Base

Finetuned

LiquidAI/LFM2.5-8B-A1B

Quantized

(59)

this model