Llama-4-Maverick-17B-128E-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-Maverick-17B-128E-Instruct.

Model Optimizations

This model was obtained by quantizing the weights and activations of meta-llama/Llama-4-Maverick-17B-128E-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 nm-testing/Llama-4-Maverick-17B-128E-Instruct-block-FP8 --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 = "nm-testing/Llama-4-Maverick-17B-128E-Instruct-block-FP8"

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-Maverick-17B-128E-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 OpenLLM leaderboard task, using lm-evaluation-harness. vLLM was used for all evaluations.

Evaluation details

Openllm V1

lm_eval \
  --model vllm \
  --model_args pretrained="nm-testing/Llama-4-Maverick-17B-128E-Instruct-block-FP8",dtype=auto,add_bos_token=True,max_model_len=16384,tensor_parallel_size=8,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="nm-testing/Llama-4-Maverick-17B-128E-Instruct-block-FP8",dtype=auto,add_bos_token=False,max_model_len=16384,tensor_parallel_size=8,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 "nm-testing/Llama-4-Maverick-17B-128E-Instruct-block-FP8" \
                  --dataset "humaneval" \
                  --backend vllm \
                  --tp 8 \
                  --greedy
evalplus.evaluate --model "nm-testing/Llama-4-Maverick-17B-128E-Instruct-block-FP8" \
                --dataset "mbpp" \
                --backend vllm \
                --tp 8 \
                --greedy

Multimodal Evaluation

lm_eval \
--model vllm-vlm \
--model_args pretrained="nm-testing/Llama-4-Maverick-17B-128E-Instruct-block-FP8",dtype=auto,add_bos_token=False,max_model_len=1000000,tensor_parallel_size=8,gpu_memory_utilization=0.9,enable_chunked_prefill=True,trust_remote_code=True,max_images=10 \
--tasks mmlu \
--apply_chat_template \
--batch_size auto

Accuracy

Category	Metric	meta-llama/Llama-4-Maverick-17B-128E-Instruct	nm-testing/Llama-4-Maverick-17B-128E-Instruct-block-FP8	Recovery (%)
OpenLLM V1	ARC-Challenge (Acc-Norm, 25-shot)	73.38	73.38	100.00
	GSM8K (Strict-Match, 5-shot)	93.03	92.72	99.67
	HellaSwag (Acc-Norm, 10-shot)	87.39	87.33	99.93
	MMLU (Acc, 5-shot)	86.03	86.15	100.13
	TruthfulQA (MC2, 0-shot)	62.76	62.90	100.23
	Winogrande (Acc, 5-shot)	79.56	79.40	99.80
	Average Score	80.36	80.31	99.94
OpenLLM V2	IFEval (Inst Level Strict Acc, 0-shot)	89.93	90.89	101.07
	BBH (Acc-Norm, 3-shot)	70.53	71.03	100.71
	Math-Hard (Exact-Match, 4-shot)	64.73	65.26	100.82
	GPQA (Acc-Norm, 0-shot)	31.29	30.54	97.59
	MUSR (Acc-Norm, 0-shot)	46.56	46.03	98.86
	MMLU-Pro (Acc, 5-shot)	64.11	63.95	99.75
	Average Score	61.19	61.28	100.15
Coding	HumanEval pass@1	abc	88.40	xyz
	HumanEval+ pass@1	abc	79.30	xyz
	MBPP pass@1	abc	90.20	xyz
	MBPP+ pass@1	abc	75.10	xyz
Multi-modal	MMMU (val)	79.08	78.50	99.26