You can deploy this model and start earning money today!

NotoriousH2/
kanana-2-30b-a3b-instruct-2601-awq-w4a16

Name: kanana-2-30b-a3b-instruct-2601-awq-w4a16
Availability: InStock

Views

No views yet

Kanana-2-30B-A3B-Instruct AWQ (W4A16)

This is the AWQ quantized version of kakaocorp/kanana-2-30b-a3b-instruct-2601.

Model Description

Base Model: kakaocorp/kanana-2-30b-a3b-instruct-2601
Quantization Method: AWQ (Activation-aware Weight Quantization)
Quantization Scheme: W4A16 (4-bit weights, 16-bit activations)
Calibration Dataset: ChuGyouk/Asan-AMC-Healthinfo
Calibration Samples: 64
Max Sequence Length: 512

Quantization Details

This model was quantized using LLM Compressor with the following configuration:

Ignored Layers (not quantized):
- lm_head: Output layer
- mlp.gate: MoE router gates
- mlp.shared_expert_gate: Shared expert gates

python

from llmcompressor.modifiers.awq import AWQModifier

recipe = [
    AWQModifier(
        ignore=["lm_head", "re:.*mlp.gate$", "re:.*mlp.shared_expert_gate$"],
        scheme="W4A16",
        targets=["Linear"],
    ),
]

Usage

With vLLM (Recommended)

python

from vllm import LLM, SamplingParams

model = LLM(
    model="NotoriousH2/kanana-2-30b-a3b-instruct-2601-awq-w4a16",
    trust_remote_code=True,
)

sampling_params = SamplingParams(temperature=0.7, max_tokens=512)
output = model.generate("고혈압 환자의 식이요법에 대해 설명해주세요.", sampling_params)
print(output[0].outputs[0].text)

With Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "NotoriousH2/kanana-2-30b-a3b-instruct-2601-awq-w4a16",
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "NotoriousH2/kanana-2-30b-a3b-instruct-2601-awq-w4a16",
    trust_remote_code=True,
)

messages = [
    {"role": "user", "content": "고혈압 환자의 식이요법에 대해 설명해주세요."}
]

input_ids = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
    add_generation_prompt=True,
).to(model.device)

output = model.generate(input_ids, max_new_tokens=512)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Acknowledgements

Original model by Kakao Corp
Quantization powered by LLM Compressor