Hugging Face's logo

RedHatAI
/
Llama-4-Scout-17B-16E-Instruct-NVFP4

Text Generation
Safetensors
llama4
fp4
vllm
conversational
8-bit precision
compressed-tensors
Model card Files
xet
Community

Llama-4-Scout-17B-16E-Instruct-NVFP4

Model Overview

  • Model Architecture: Meta-Llama-3.1
    • Input: Text / Image
    • Output: Text
  • Model Optimizations:
    • Weight quantization: FP4
    • Activation quantization: FP4
  • Intended Use Cases: Intended for commercial and research use in multiple languages.
  • Out-of-scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in languages other than English.
  • Release Date: 7/15/25
  • Version: 1.0
  • License(s): llama3.1
  • Model Developers: RedHatAI

This model is a quantized version of Llama-4-Scout-17B-16E-Instruct. It was evaluated on a several tasks to assess the its quality in comparison to the unquatized model.

Model Optimizations

This model was obtained by quantizing the weights and activations of Llama-4-Scout-17B-16E-Instruct to FP4 data type, ready for inference with vLLM>=0.9.1 This optimization reduces the number of bits per parameter from 16 to 4, reducing the disk size and GPU memory requirements by approximately 25%.

Only the weights of the linear operators within transformers blocks are quantized using LLM Compressor.

Deployment

Use with vLLM

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

Model Usage Code 
from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = "RedHatAI/Llama-4-Scout-17B-16E-Instruct-NVFP4"
number_gpus = 2

sampling_params = SamplingParams(temperature=0.6, top_p=0.9, max_tokens=256)

tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]

prompts = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

llm = LLM(model=model_id, tensor_parallel_size=number_gpus)

outputs = llm.generate(prompts, sampling_params)

generated_text = outputs[0].outputs[0].text
print(generated_text)

vLLM aslo supports OpenAI-compatible serving. See the documentation for more details.

Creation

This model was created by applying LLM Compressor with calibration samples from neuralmagic/calibration dataset, as presented in the code snipet below.

Model Creation Code 
from transformers import Llama4ForConditionalGeneration, Llama4Processor
from transformers.quantizers.quantizers_utils import get_module_from_name
import torch
from datasets import load_dataset

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.utils.dev import skip_weights_initialize
from transformers.models.llama4.modeling_llama4 import Llama4TextMLP
from llmcompressor.modifiers.quantization import QuantizationModifier
import gc
from llmcompressor.modifiers.smoothquant import SmoothQuantModifier

def convert_model_for_quantization(model):
    to_delete = []
    for name, module in model.named_modules():
        module_class_name = module.__class__.__name__
        if module_class_name == "Llama4TextMoe":
            parent_module, module_name = get_module_from_name(model, name)
            parent_module._modules[module_name] = SequentialLlama4TextMoe(
                model.config.get_text_config(),
                module,
            )
            to_delete.append(module)
            print(f"Patched {name} with SequentialLlama4TextMoe", flush=True)

    for module in to_delete:
        del module
        gc.collect()
        torch.cuda.empty_cache()


class SequentialLlama4TextMoe(torch.nn.Module):
    def __init__(self, config, original_moe):
        super().__init__()
        self.top_k = config.num_experts_per_tok
        self.hidden_dim = config.hidden_size
        self.num_experts = config.num_local_experts
        self.experts = SequentialLlama4TextExperts(config, original_moe.experts)
        self.router = original_moe.router
        self.shared_expert = original_moe.shared_expert

    def forward(self, hidden_states):
        hidden_states = hidden_states.reshape(-1, self.hidden_dim)
        router_logits = self.router(hidden_states)

        router_top_value, router_indices = torch.topk(router_logits, self.top_k, dim=1)

        router_scores = (
            torch.full_like(router_logits, float("-inf")).scatter_(1, router_indices, router_top_value).transpose(0, 1)
        )
        router_scores = torch.sigmoid(router_scores.float()).to(hidden_states.dtype)

        out = self.shared_expert(hidden_states)
        for i in range(self.num_experts):
            out += self.experts[i](hidden_states) * router_scores[i].reshape(-1, 1)

        return out, router_scores


class SequentialLlama4TextExperts(torch.nn.ModuleList):
    def __init__(self, config, original_experts):
        self.num_experts = original_experts.gate_up_proj.shape[0]
        with skip_weights_initialize():
            super().__init__([Llama4TextMLP(config) for _ in range(self.num_experts)])

        intermediate_size = original_experts.down_proj.shape[1]

        for i in range(self.num_experts):
            gate_up = original_experts.gate_up_proj[i]
            down = original_experts.down_proj[i]

            gate_proj = gate_up[:, :intermediate_size]
            up_proj = gate_up[:, intermediate_size:]

            self[i].gate_proj.weight.data = gate_proj.t().clone().contiguous()
            self[i].up_proj.weight.data = up_proj.t().clone().contiguous()
            self[i].down_proj.weight.data = down.t().clone().contiguous()

        original_experts.gate_up_proj = None
        original_experts.down_proj = None
        gc.collect()
        torch.cuda.empty_cache()


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

model = Llama4ForConditionalGeneration.from_pretrained(
    model_id, torch_dtype=torch.bfloat16  # load on cpu
)
processor = Llama4Processor.from_pretrained(model_id)

convert_model_for_quantization(model)

# Oneshot arguments
DATASET_ID = "neuralmagic/calibration"
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 8192

ds = load_dataset(DATASET_ID, name="LLM", split=f"train[:{NUM_CALIBRATION_SAMPLES}]")

def preprocess_function(example):
    messgages = []
    for message in example["messages"]:
        messgages.append(
            {
                "role": message["role"], 
                "content": [{"type": "text", "text": message["content"]}]
            }
        )
    
    return processor.apply_chat_template(
        messgages, 
        return_tensors="pt", 
        padding=False, 
        truncation=True, 
        max_length=MAX_SEQUENCE_LENGTH,
        tokenize=True,
        add_special_tokens=False,
        return_dict=True,
        add_generation_prompt=False,
    ).to("cuda:0")

ds = ds.map(
    preprocess_function,
    batched=False,
    remove_columns=ds.column_names
)

# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {
        key: torch.tensor(value) if key != "pixel_values" else torch.tensor(value, dtype=torch.bfloat16).squeeze(0)
        for key, value in batch[0].items()
    }

# Recipe
recipe = QuantizationModifier(targets="Linear", scheme="NVFP4", 
            ignore=[
                're:.*lm_head',
                're:.*self_attn',
                're:.*router',
                're:.*vision_model',
                're:.*multi_modal_projector',
                're:.*multi_modal_projector',
                "Llama4TextAttention",
            ],
            sequential_targets=["Llama4TextMLP"],
        )

SAVE_DIR = f"{model_id.split('/')[1]}-{recipe.scheme}"

# Perform oneshot
oneshot(
    model=model,
    tokenizer=model_id,
    dataset=ds,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
    data_collator=data_collator,
    output_dir=SAVE_DIR
)

# Save to disk compressed.
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)

Evaluation

This model was evaluated on the well-known OpenLLM v1, OpenLLM v2, HumanEval, and HumanEval_64 benchmarks. All evaluations were conducted using lm-evaluation-harness.

Category Metric Llama-4-Scout-17B-16E-Instruct (A100) Llama-4-Scout-17B-16E-Instruct-NVFP4 (B200) Recovery (%)
OpenLLM V1 ARC Challenge (LLaMA) 93.39 92.10 98.62%
GSM8K (LLaMA) 92.87 94.31 101.55%
MMLU (LLaMA) 81.01 79.37 97.98%
MMLU-CoT (LLaMA) 85.99 84.58 98.36%
Hellaswag 79.13 78.47 99.17%
TruthfulQA-mc2 62.53 60.83 97.28%
Winogrande 73.56 73.01 99.25%
Average 81.21 80.38 98.89%
OpenLLM V2 MMLU-Pro 55.64 53.84 96.76%
IFEval 89.09 89.93 100.94%
BBH 65.14 64.00 98.25%
Math-Hard 52.64 56.12 106.61%
GPQA 32.21 31.88 98.98%
MuSR 42.20 42.99 101.87%
Average 56.15 56.46 100.55%
Coding HumanEval Instruct pass@1 81.71 76.22 93.29%
HumanEval 64 Instruct pass@2 83.49 81.10 97.14%
HumanEval 64 Instruct pass@8 87.71 88.66 101.08%
HumanEval 64 Instruct pass@16 88.71 90.11 101.58%
HumanEval 64 Instruct pass@32 89.38 90.91 101.71%
HumanEval 64 Instruct pass@64 89.63 91.46 102.04%

Reproduction

The results were obtained using the following commands:

Model Evaluation Commands

MMLU_LLAMA 

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-4-Scout-17B-16E-Instruct-NVFP4",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True \
  --tasks mmlu_llama \
  --apply_chat_template \
  --fewshot_as_multiturn \
  --batch_size auto

MMLU_COT_LLAMA 

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-4-Scout-17B-16E-Instruct-NVFP4",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True \
  --tasks mmlu_cot_llama \
  --apply_chat_template \
  --fewshot_as_multiturn \
  --batch_size auto

ARC-Challenge 

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-4-Scout-17B-16E-Instruct-NVFP4",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True \
  --tasks arc_challenge_llama \
  --apply_chat_template \
  --batch_size auto

GSM-8K 

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-4-Scout-17B-16E-Instruct-NVFP4",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True \
  --tasks gsm8k_llama \
  --apply_chat_template \
  --fewshot_as_multiturn \
  --batch_size auto

Hellaswag 

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-4-Scout-17B-16E-Instruct-NVFP4",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True \
  --tasks hellaswag \
  --apply_chat_template \
  --fewshot_as_multiturn \
  --batch_size auto

Winogrande 

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-4-Scout-17B-16E-Instruct-NVFP4",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True \
  --tasks winogrande \
  --apply_chat_template \
  --fewshot_as_multiturn \
  --batch_size auto

TruthfulQA 

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-4-Scout-17B-16E-Instruct-NVFP4",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True \
  --tasks truthfulqa \
  --apply_chat_template \
  --fewshot_as_multiturn \
  --batch_size auto

OpenLLM v2 

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-4-Scout-17B-16E-Instruct-NVFP4",dtype=auto,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True\
  --apply_chat_template \
  --fewshot_as_multiturn \
  --tasks leaderboard \
  --batch_size auto

HumanEval and HumanEval_64 

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-4-Scout-17B-16E-Instruct-NVFP4",dtype=auto,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True\
  --apply_chat_template \
  --fewshot_as_multiturn \
  --tasks humaneval_instruct \
  --batch_size auto


lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-4-Scout-17B-16E-Instruct-NVFP4",dtype=auto,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True\
  --apply_chat_template \
  --fewshot_as_multiturn \
  --tasks humaneval_64_instruct \
  --batch_size auto
Downloads last month
151
Safetensors
Model size
63.7B params
Tensor type
F32
路
BF16
路
F8_E4M3
路
U8
路
Inference Providers NEW
Text Generation
This model isn't deployed by any Inference Provider. 馃檵 Ask for provider support

Model tree for RedHatAI/Llama-4-Scout-17B-16E-Instruct-NVFP4

Base model

meta-llama/Llama-4-Scout-17B-16E
Finetuned
meta-llama/Llama-4-Scout-17B-16E-Instruct
Quantized
(27)
this model