Mistral-Small-24B-Instruct-2501-FP8-dynamic

Model Overview

• Model Architecture: Mistral-Small-24B-Instruct-2501
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Weight quantization: FP8
  • Activation quantization: FP8
• Release Date: 3/1/2025
• Version: 1.0
• Model Developers: Neural Magic

Quantized version of Mistral-Small-24B-Instruct-2501. It achieves an average score of 78.88 on the OpenLLM benchmark (version 1), whereas the unquantized model achieves 79.45.

Model Optimizations

This model was obtained by quantizing the weights and activations to FP8 data type, ready for inference with vLLM. 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 are quantized.

Deployment

Use with vLLM

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

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

max_model_len, tp_size = 4096, 1
model_name = "neuralmagic/Mistral-Small-24B-Instruct-2501-FP8-Dynamic"
tokenizer = AutoTokenizer.from_pretrained(model_name)
llm = LLM(model=model_name, tensor_parallel_size=tp_size, max_model_len=max_model_len, trust_remote_code=True)
sampling_params = SamplingParams(temperature=0.3, max_tokens=256, stop_token_ids=[tokenizer.eos_token_id])

messages_list = [
    [{"role": "user", "content": "Who are you? Please respond in pirate speak!"}],
]

prompt_token_ids = [tokenizer.apply_chat_template(messages, add_generation_prompt=True) for messages in messages_list]

outputs = llm.generate(prompt_token_ids=prompt_token_ids, sampling_params=sampling_params)

generated_text = [output.outputs[0].text for output in outputs]
print(generated_text)

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

Deploy on Red Hat AI Inference Server

$ podman run --rm -it --device nvidia.com/gpu=all -p 8000:8000 \
 --ipc=host \
--env "HUGGING_FACE_HUB_TOKEN=$HF_TOKEN" \
--env "HF_HUB_OFFLINE=0" -v ~/.cache/vllm:/home/vllm/.cache \
--name=vllm \
registry.access.redhat.com/rhaiis/rh-vllm-cuda \
vllm serve \
--tensor-parallel-size 8 \
--max-model-len 32768  \
--enforce-eager --model RedHatAI/Mistral-Small-24B-Instruct-2501-FP8-dynamic

​​See Red Hat AI Inference Server documentation for more details.

Deploy on Red Hat Enterprise Linux AI

# Download model from Red Hat Registry via docker
# Note: This downloads the model to ~/.cache/instructlab/models unless --model-dir is specified.
ilab model download --repository docker://registry.redhat.io/rhelai1/mistral-small-24b-instruct-2501-fp8-dynamic:1.5

# Serve model via ilab
ilab model serve --model-path ~/.cache/instructlab/models/mistral-small-24b-instruct-2501-fp8-dynamic --gpu 1 -- --trust-remote-code
  
# Chat with model
ilab model chat --model ~/.cache/instructlab/models/mistral-small-24b-instruct-2501-fp8-dynamic

See Red Hat Enterprise Linux AI documentation for more details.

Deploy on Red Hat Openshift AI

# Setting up vllm server with ServingRuntime
# Save as: vllm-servingruntime.yaml
apiVersion: serving.kserve.io/v1alpha1
kind: ServingRuntime
metadata:
 name: vllm-cuda-runtime # OPTIONAL CHANGE: set a unique name
 annotations:
   openshift.io/display-name: vLLM NVIDIA GPU ServingRuntime for KServe
   opendatahub.io/recommended-accelerators: '["nvidia.com/gpu"]'
 labels:
   opendatahub.io/dashboard: 'true'
spec:
 annotations:
   prometheus.io/port: '8080'
   prometheus.io/path: '/metrics'
 multiModel: false
 supportedModelFormats:
   - autoSelect: true
     name: vLLM
 containers:
   - name: kserve-container
     image: quay.io/modh/vllm:rhoai-2.20-cuda # CHANGE if needed. If AMD: quay.io/modh/vllm:rhoai-2.20-rocm
     command:
       - python
       - -m
       - vllm.entrypoints.openai.api_server
     args:
       - "--port=8080"
       - "--model=/mnt/models"
       - "--served-model-name={{.Name}}"
     env:
       - name: HF_HOME
         value: /tmp/hf_home
     ports:
       - containerPort: 8080
         protocol: TCP

# Attach model to vllm server. This is an NVIDIA template
# Save as: inferenceservice.yaml
apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  annotations:
    openshift.io/display-name: mistral-small-24b-instruct-2501-fp8-dynamic # OPTIONAL CHANGE
    serving.kserve.io/deploymentMode: RawDeployment
  name: mistral-small-24b-instruct-2501-fp8-dynamic         # specify model name. This value will be used to invoke the model in the payload
  labels:
    opendatahub.io/dashboard: 'true'
spec:
  predictor:
    maxReplicas: 1
    minReplicas: 1
    model:
      args:
        - "--tokenizer-mode=mistral"
        - "--config-format=mistral"
        - "--load-format=mistral"
        - "--tool-call-parser=mistral"
        - "--enable-auto-tool-choice"
        - "--limit-mm-per-prompt=image=10"
        - "--max-model-len=16384"
        - "--uvicorn-log-level=debug"
        - "--trust-remote-code"
      modelFormat:
        name: vLLM
      name: ''
      resources:
        limits:
          cpu: '2'			# this is model specific
          memory: 8Gi		# this is model specific
          nvidia.com/gpu: '1'	# this is accelerator specific
        requests:			# same comment for this block
          cpu: '1'
          memory: 4Gi
          nvidia.com/gpu: '1'
      runtime: vllm-cuda-runtime	# must match the ServingRuntime name above
      storageUri: oci://registry.redhat.io/rhelai1/modelcar-mistral-small-24b-instruct-2501-fp8-dynamic:1.5
    tolerations:
    - effect: NoSchedule
      key: nvidia.com/gpu
      operator: Exists

# make sure first to be in the project where you want to deploy the model
# oc project <project-name>

# apply both resources to run model

# Apply the ServingRuntime
oc apply -f vllm-servingruntime.yaml

# Apply the InferenceService
oc apply -f qwen-inferenceservice.yaml

# Replace <inference-service-name> and <cluster-ingress-domain> below:
# - Run `oc get inferenceservice` to find your URL if unsure.

# Call the server using curl:
curl https://<inference-service-name>-predictor-default.<domain>/v1/chat/completions
        -H "Content-Type: application/json" \
        -d '{
    "model": "mistral-small-24b-instruct-2501-fp8-dynamic",
    "stream": true,
    "stream_options": {
        "include_usage": true
    },
    "max_tokens": 1,
    "messages": [
        {
            "role": "user",
            "content": "How can a bee fly when its wings are so small?"
        }
    ]
}'

See Red Hat Openshift AI documentation for more details.

Creation

This model was created with llm-compressor by running the code snippet below.

import argparse
from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.transformers import oneshot
import os

def main():
    parser = argparse.ArgumentParser(description='Quantize a transformer model to FP8')
    parser.add_argument('--model_id', type=str, required=True,
                        help='The model ID from HuggingFace (e.g., "meta-llama/Meta-Llama-3-8B-Instruct")')
    parser.add_argument('--save_path', type=str, default='.',
                        help='Custom path to save the quantized model. If not provided, will use model_name-FP8-dynamic')
    args = parser.parse_args()

    # Load model
    model = AutoModelForCausalLM.from_pretrained(
        args.model_id, device_map="auto", torch_dtype="auto", trust_remote_code=True,
    )
    tokenizer = AutoTokenizer.from_pretrained(args.model_id)

    # Configure the quantization algorithm and scheme
    recipe = QuantizationModifier(
        targets="Linear", scheme="FP8_DYNAMIC", ignore=["lm_head"]
    )

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

    save_path = os.path.join(args.save_path, args.model_id.split("/")[1] + "-FP8-dynamic")
    os.makedirs(save_path, exist_ok=True)

    # Save to disk in compressed-tensors format
    model.save_pretrained(save_path)
    tokenizer.save_pretrained(save_path)
    print(f"Model and tokenizer saved to: {save_path}")

if __name__ == "__main__":
    main()

Evaluation

The model was evaluated on OpenLLM Leaderboard V1 and V2, using the following commands:

OpenLLM Leaderboard V1:

lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic/Mistral-Small-24B-Instruct-2501-FP8-Dynamic",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=1,gpu_memory_utilization=0.8,enable_chunked_prefill=True,trust_remote_code=True \
  --tasks openllm \
  --write_out \
  --batch_size auto \
  --output_path output_dir \
  --show_config

OpenLLM Leaderboard V2:

lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic/Mistral-Small-24B-Instruct-2501-FP8-Dynamic",dtype=auto,add_bos_token=False,max_model_len=4096,tensor_parallel_size=1,gpu_memory_utilization=0.8,enable_chunked_prefill=True,trust_remote_code=True \
  --apply_chat_template \
  --fewshot_as_multiturn \
  --tasks leaderboard \
  --write_out \
  --batch_size auto \
  --output_path output_dir \
  --show_config

Accuracy

OpenLLM Leaderboard V1 evaluation scores

Metric	mistralai/Mistral-Small-24B-Instruct-2501	nm-testing/Mistral-Small-24B-Instruct-2501-FP8-dynamic
ARC-Challenge (Acc-Norm, 25-shot)	72.18	71.76
GSM8K (Strict-Match, 5-shot)	90.14	89.01
HellaSwag (Acc-Norm, 10-shot)	85.05	84.65
MMLU (Acc, 5-shot)	80.69	80.55
TruthfulQA (MC2, 0-shot)	65.55	64.85
Winogrande (Acc, 5-shot)	83.11	82.48
Average Score	79.45	78.88
Recovery (%)	100.00	99.28

OpenLLM Leaderboard V2 evaluation scores

Metric	mistralai/Mistral-Small-24B-Instruct-2501	nm-testing/Mistral-Small-24B-Instruct-2501-FP8-dynamic
IFEval (Inst-and-Prompt Level Strict Acc, 0-shot)	73.27	73.53
BBH (Acc-Norm, 3-shot)	45.18	44.39
MMLU-Pro (Acc, 5-shot)	38.83	37.28
Average Score	52.42	51.73
Recovery (%)	100.00	98.68
Math-Hard (Exact-Match, 4-shot)	6.35	2.99
GPQA (Acc-Norm, 0-shot)	8.29	6.97
MUSR (Acc-Norm, 0-shot)	7.84	8.04

Results on Math-Hard, GPQA, and MUSR are not considred for accuracy recovery calculation because the unquantized model has close to random prediction accuracy (6.35, 8.29, 7.84) which doesn't provide a reliable baseline for recovery calculation.