RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-block

Llama-4-Scout-17B-16E-Instruct-BLOCK-FP8

Model Overview

• Model Architecture: Llama4ForConditionalGeneration
  • Input: Text, Image
  • Output: Text
• Model Optimizations:
  • Weight quantization: FP8
  • Activation quantization: FP8
• Release Date:
• Version: 1.0
• Model Developers:: Red Hat

Quantized version of meta-llama/Llama-4-Scout-17B-16E-Instruct.

Model Optimizations

This model was obtained by quantizing the weights and activations of meta-llama/Llama-4-Scout-17B-16E-Instruct to FP8 data type. This optimization reduces the number of bits per parameter from 16 to 8, reducing the disk size and GPU memory requirements by approximately 50%. Only the weights and activations of the linear operators within transformers blocks of the language model are quantized.

Deployment

Use with vLLM

1. Initialize vLLM server:

vllm serve RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-block --tensor_parallel_size 8

2. Send requests to the server:

from openai import OpenAI

# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://<your-server-host>:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

model = "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-block"

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"},
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

outputs = client.chat.completions.create(
    model=model,
    messages=messages,
)

generated_text = outputs.choices[0].message.content
print(generated_text)

Creation

This model was quantized using the llm-compressor library as shown below.

Creation details

from transformers import AutoProcessor, LlamaForCausalLM, AutoModelForImageTextToText

from llmcompressor import oneshot
from llmcompressor.modeling import replace_modules_for_calibration
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.utils import dispatch_for_generation

MODEL_ID = "meta-llama/Llama-4-Scout-17B-16E-Instruct"

# Load model.
model = AutoModelForImageTextToText.from_pretrained(MODEL_ID, dtype="auto")
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = replace_modules_for_calibration(model)

# Configure the quantization algorithm and scheme.
# In this case, we:
#   * quantize the weights to fp8 with per-block quantization
#   * quantize the activations to fp8 with dynamic token activations
ecipe = QuantizationModifier(
    targets="Linear",
    scheme="FP8_BLOCK",
    ignore=[
        "re:.*lm_head",
        "re:.*self_attn",
        "re:.*router",
        "re:.*vision_model.*",
        "re:.*multi_modal_projector.*",
        "Llama4TextAttention",
    ],
)

# Apply quantization.
oneshot(model=model, recipe=recipe
dispatch_for_generation(model)


# Save to disk in compressed-tensors format.
SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-block"
model.save_pretrained(SAVE_DIR)
processor.save_pretrained(SAVE_DIR)

Evaluation

The model was evaluated on the OpenLLMv1 leaderboard task, using lm-evaluation-harness, on reasoning tasks using lighteval. vLLM was used for all evaluations.

Evaluation details

Openllm V1

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-block",dtype=auto,add_bos_token=True,max_model_len=16384,tensor_parallel_size=4,gpu_memory_utilization=0.9,enable_chunked_prefill=True,trust_remote_code=True \
  --tasks openllm \
  --write_out \
  --batch_size auto \
  --show_config

Openllm V2

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-block",dtype=auto,add_bos_token=False,max_model_len=16384,tensor_parallel_size=4,gpu_memory_utilization=0.7,disable_log_stats=True,enable_chunked_prefill=True,trust_remote_code=True \
  --tasks leaderboard \
  --apply_chat_template \
  --fewshot_as_multiturn \
  --write_out \
  --batch_size auto \
  --show_config

Coding Benchmarks

evalplus.evaluate --model "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-block" \
                  --dataset "humaneval" \
                  --backend vllm \
                  --tp 4 \
                  --greedy
evalplus.evaluate --model "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-block" \
                --dataset "mbpp" \
                --backend vllm \
                --tp 4 \
                --greedy

Accuracy

Category	Metric	meta-llama/Llama-4-Scout-17B-16E-Instruct	RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-block	Recovery (%)
OpenLLM V1	ARC-Challenge (Acc-Norm, 25-shot)	69.62	68.60	98.53
	GSM8K (Strict-Match, 5-shot)	90.52	90.90	100.42
	HellaSwag (Acc-Norm, 10-shot)	85.27	85.24	99.96
	MMLU (Acc, 5-shot)	80.49	80.48	99.99
	TruthfulQA (MC2, 0-shot)	61.28	61.30	100.03
	Winogrande (Acc, 5-shot)	77.82	77.35	99.39
	Average Score	77.50	77.31	99.75
OpenLLM V2	IFEval (Inst Level Strict Acc, 0-shot)	89.09	89.93	100.94
	BBH (Acc-Norm, 3-shot)	65.02	65.11	100.13
	Math-Hard (Exact-Match, 4-shot)	57.93	57.85	99.87
	GPQA (Acc-Norm, 0-shot)	30.45	30.70	100.83
	MUSR (Acc-Norm, 0-shot)	42.99	43.39	100.92
	MMLU-Pro (Acc, 5-shot)	55.74	55.58	99.70
	Average Score	56.87	57.09	100.39
Coding	HumanEval pass@1	abc	82.30	xyz
	HumanEval+ pass@1	abc	75.60	xyz
	MBPP pass@1	abc	80.70	xyz
	MBPP+ pass@1	abc	64.80	xyz