Devstral-Small-2507-quantized.w4a16

Model Overview

• Model Architecture: MistralForCausalLM
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Activation quantization: INT4
  • Weight quantization: None
• Release Date: 08/29/2025
• Version: 1.0
• Model Developers: Red Hat (Neural Magic)

Model Optimizations

This model was obtained by quantizing weights of Devstral-Small-2507 to INT4 data type. This optimization reduces the number of bits used to represent weights from 16 to 4, reducing GPU memory requirements (by approximately 75%). Weight quantization also reduces disk size requirements by approximately 75%.

Deployment

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

vllm serve RedHatAI/Devstral-Small-2507-quantized.w4a16 --tensor-parallel-size 1 --tokenizer_mode mistral

Creation

This model was created with [llm-compressor](https://github.com/vllm-project/llm-compressor) by running the code snippet below.

python quantize.py --model_path mistralai/Devstral-Small-2507 --calib_size 1024 --dampening_frac 0.1 --observer mse --sym false --actorder weight

import argparse
import os
from datasets import load_dataset
from transformers import AutoModelForCausalLM
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.transformers import oneshot
from compressed_tensors.quantization import QuantizationScheme, QuantizationArgs, QuantizationType, QuantizationStrategy
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
from mistral_common.protocol.instruct.request import ChatCompletionRequest
from mistral_common.protocol.instruct.messages import (
  SystemMessage, UserMessage
)

def load_system_prompt(repo_id: str, filename: str) -> str:
  file_path = os.path.join(repo_id, filename)
  with open(file_path, "r") as file:
      system_prompt = file.read()
  return system_prompt

def parse_actorder(value):
  if value.lower() == "false":
      return False
  elif value.lower() == "weight":
      return "weight"
  elif value.lower() == "group":
      return "group"
  else:
      raise argparse.ArgumentTypeError("Invalid value for --actorder.")

def parse_sym(value):
  if value.lower() == "false":
      return False
  elif value.lower() == "true":
      return True
  else:
      raise argparse.ArgumentTypeError(f"Invalid value for --sym. Use false or true, but got {value}")


parser = argparse.ArgumentParser()
parser.add_argument('--model_path', type=str)
parser.add_argument('--calib_size', type=int, default=256)
parser.add_argument('--dampening_frac', type=float, default=0.1)
parser.add_argument('--observer', type=str, default="minmax")
parser.add_argument('--sym', type=parse_sym, default=True)
parser.add_argument(
  '--actorder',
  type=parse_actorder,
  default=False,
  help="Specify actorder as 'group' (string) or False (boolean)."
)
args = parser.parse_args()


model = AutoModelForCausalLM.from_pretrained(
  args.model_path,
  device_map="auto",
  torch_dtype="auto",
  use_cache=False,
  trust_remote_code=True,
)

ds = load_dataset("garage-bAInd/Open-Platypus", split="train")
ds = ds.shuffle(seed=42).select(range(args.calib_size))

SYSTEM_PROMPT = load_system_prompt(args.model_path, "SYSTEM_PROMPT.txt")
tokenizer = MistralTokenizer.from_hf_hub("mistralai/Devstral-Small-2507")

def tokenize(sample):
  tmp = tokenizer.encode_chat_completion(
      ChatCompletionRequest(
          messages=[
              SystemMessage(content=SYSTEM_PROMPT),
              UserMessage(content=sample['instruction']),
          ],
      )
  )
  return {'input_ids': tmp.tokens}

ds = ds.map(tokenize, remove_columns=ds.column_names)

quant_scheme = QuantizationScheme(
  targets=["Linear"],
  weights=QuantizationArgs(
      num_bits=4,
      type=QuantizationType.INT,
      symmetric=args.sym,
      group_size=128,
      strategy=QuantizationStrategy.GROUP,
      observer=args.observer,
      actorder=args.actorder
  ),
  input_activations=None,
  output_activations=None,
)

recipe = [
  GPTQModifier(
      targets=["Linear"],
      ignore=["lm_head"],
      dampening_frac=args.dampening_frac,
      config_groups={"group_0": quant_scheme},
  )
]

oneshot(
  model=model,
  dataset=ds,
  recipe=recipe,
  num_calibration_samples=args.calib_size,
  max_seq_length=8192,
)

save_path = args.model_path + "-quantized.w4a16"
model.save_pretrained(save_path)

Evaluation

The model was evaluated on popular coding tasks (HumanEval, HumanEval+, MBPP, MBPP+) via EvalPlus and vllm backend (v0.10.1.1). For evaluations, we run greedy sampling and report pass@1. The command to reproduce evals:

evalplus.evaluate --model "RedHatAI/Devstral-Small-2507-quantized.w4a16" \
                  --dataset [humaneval|mbpp] \
                  --base-url http://localhost:8000/v1 \
                  --backend openai --greedy

Accuracy

	Recovery (%)	mistralai/Devstral-Small-2507	RedHatAI/Devstral-Small-2507-quantized.w4a16 (this model)
HumanEval	98.65	89.0	87.8
HumanEval+	100.0	81.1	81.1
MBPP	98.97	77.5	76.7
MBPP+	102.12	66.1	67.5
Average Score	99.81	78.43	78.28