File size: 14,212 Bytes

---
tags:
- int8
- vllm
- chat
- neuralmagic
- llmcompressor
language:
- en
- de
- fr
- it
- pt
- hi
- es
- th
pipeline_tag: text-generation
license: llama3.3
base_model: meta-llama/Llama-3.3-70B-Instruct
---

# Llama-3.3-70B-Instruct-quantized.w8a8

## Model Overview
- **Model Architecture:** Llama
  - **Input:** Text
  - **Output:** Text
- **Model Optimizations:**
  - **Activation quantization:** INT8
  - **Weight quantization:** INT8
- **Intended Use Cases:** Intended for commercial and research use multiple languages. Similarly to [Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct), this models is intended for assistant-like chat.
- **Out-of-scope:** Use in any manner that violates applicable laws or regulations (including trade compliance laws).
- **Release Date:** 01/20/2025
- **Version:** 1.0
- **Model Developers:** Neural Magic

Quantized version of [Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct).
It was evaluated on a several tasks to assess the its quality in comparison to the unquatized model, including multiple-choice, math reasoning, and open-ended text generation.
Llama-3.3-70B-Instruct-quantized.w8a8 achieves 99.4% recovery for OpenLLM v1 (using Meta's prompting when available) and 100% for both HumanEval and HumanEval+ pass@1.

### Model Optimizations

This model was obtained by quantizing the weights and activations of [Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) to INT8 data type.
This optimization reduces the number of bits used to represent weights and activations from 16 to 8, reducing GPU memory requirements (by approximately 50%) and increasing matrix-multiply compute throughput (by approximately 2x).
Weight quantization also reduces disk size requirements by approximately 50%.

Only weights and activations of the linear operators within transformers blocks are quantized.
Weights are quantized with a symmetric static per-channel scheme, where a fixed linear scaling factor is applied between INT8 and floating point representations for each output channel dimension.
Activations are quantized with a symmetric dynamic per-token scheme, computing a linear scaling factor at runtime for each token between INT8 and floating point representations.

## Deployment

This model can be deployed efficiently using the [vLLM](https://docs.vllm.ai/en/latest/) backend, as shown in the example below.

```python
from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = "neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8"
number_gpus = 1
max_model_len = 8192

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, max_model_len=max_model_len)

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](https://docs.vllm.ai/en/latest/) for more details.


## Creation

This model was created by using the [llm-compressor](https://github.com/vllm-project/llm-compressor) library as presented in the code snipet below.

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
from datasets import Dataset
from llmcompressor.transformers import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
import random

model_id = "meta-llama/Meta-Llama-3.1-8B-Instruct"

num_samples = 1024
max_seq_len = 8192

tokenizer = AutoTokenizer.from_pretrained(model_id)

max_token_id = len(tokenizer.get_vocab()) - 1
input_ids = [[random.randint(0, max_token_id) for _ in range(max_seq_len)] for _ in range(num_samples)]
attention_mask = num_samples * [max_seq_len * [1]]
ds = Dataset.from_dict({"input_ids": input_ids, "attention_mask": attention_mask})

recipe = GPTQModifier(
  targets="Linear",
  scheme="W8A8",
  ignore=["lm_head"],
  dampening_frac=0.01,
)

model = SparseAutoModelForCausalLM.from_pretrained(
  model_id,
  device_map="auto",
)

oneshot(
  model=model,
  dataset=ds,
  recipe=recipe,
  max_seq_length=max_seq_len,
  num_calibration_samples=num_samples,
)

model.save_pretrained("Llama-3.3-70B-Instruct-quantized.w8a8")
```

## Evaluation

This model was evaluated on the well-known OpenLLM v1, OpenLLM v2, HumanEval, and HumanEval+ benchmarks.
In all cases, model outputs were generated with the [vLLM](https://docs.vllm.ai/en/stable/) engine.

OpenLLM v1 and v2 evaluations were conducted using Neural Magic's fork of [lm-evaluation-harness](https://github.com/neuralmagic/lm-evaluation-harness/tree/llama_3.1_instruct) (branch llama_3.1_instruct).
This version of the lm-evaluation-harness includes versions of MMLU, ARC-Challenge and GSM-8K that match the prompting style of [Meta-Llama-3.1-Instruct-evals](https://huggingface.co/datasets/meta-llama/Meta-Llama-3.1-8B-Instruct-evals) and a few fixes to OpenLLM v2 tasks.

HumanEval and HumanEval+ evaluations were conducted using Neural Magic's fork of the [EvalPlus](https://github.com/neuralmagic/evalplus) repository.

### Accuracy

<table>
  <tr>
   <th>Category
   </th>
   <th>Benchmark
   </th>
   <th>Llama-3.3-70B-Instruct
   </th>
   <th>Llama-3.3-70B-Instruct-quantized.w8a8 (this model)
   </th>
   <th>Recovery
   </th>
  </tr>
  <tr>
   <td rowspan="8" ><strong>OpenLLM v1</strong>
   </td>
   <td>MMLU (5-shot)
   </td>
   <td>81.60
   </td>
   <td>81.19
   </td>
   <td>99.5%
   </td>
  </tr>
  <tr>
   <td>MMLU (CoT, 0-shot)
   </td>
   <td>86.58
   </td>
   <td>85.92
   </td>
   <td>99.2%
   </td>
  </tr>
  <tr>
   <td>ARC Challenge (0-shot)
   </td>
   <td>49.23
   </td>
   <td>48.04
   </td>
   <td>97.6%
   </td>
  </tr>
  <tr>
   <td>GSM-8K (CoT, 8-shot, strict-match)
   </td>
   <td>94.16
   </td>
   <td>94.01
   </td>
   <td>99.8%
   </td>
  </tr>
  <tr>
   <td>Hellaswag (10-shot)
   </td>
   <td>86.49
   </td>
   <td>86.47
   </td>
   <td>100.0%
   </td>
  </tr>
  <tr>
   <td>Winogrande (5-shot)
   </td>
   <td>84.77
   </td>
   <td>83.74
   </td>
   <td>98.8%
   </td>
  </tr>
  <tr>
   <td>TruthfulQA (0-shot, mc2)
   </td>
   <td>62.75
   </td>
   <td>63.09
   </td>
   <td>99.5%
   </td>
  </tr>
  <tr>
   <td><strong>Average</strong>
   </td>
   <td><strong>77.94</strong>
   </td>
   <td><strong>77.49</strong>
   </td>
   <td><strong>99.4%</strong>
   </td>
  </tr>
  <tr>
   <td rowspan="7" ><strong>OpenLLM v2</strong>
   </td>
   <td>MMLU-Pro (5-shot)
   </td>
   <td>51.89
   </td>
   <td>xxxx
   </td>
   <td>xxxx%
   </td>
  </tr>
  <tr>
   <td>IFEval (0-shot)
   </td>
   <td>90.89
   </td>
   <td>xxxx
   </td>
   <td>xxxx%
   </td>
  </tr>
  <tr>
   <td>BBH (3-shot)
   </td>
   <td>63.15
   </td>
   <td>xxxx
   </td>
   <td>xxxx%
   </td>
  </tr>
  <tr>
   <td>Math-lvl-5 (4-shot)
   </td>
   <td>0.17
   </td>
   <td>xxxx
   </td>
   <td>N/A
   </td>
  </tr>
  <tr>
   <td>GPQA (0-shot)
   </td>
   <td>46.10
   </td>
   <td>xxxx
   </td>
   <td>xxxx%
   </td>
  </tr>
  <tr>
   <td>MuSR (0-shot)
   </td>
   <td>44.35
   </td>
   <td>xxxx
   </td>
   <td>xxxx%
   </td>
  </tr>
  <tr>
   <td><strong>Average</strong>
   </td>
   <td><strong>49.42</strong>
   </td>
   <td><strong>xxxx</strong>
   </td>
   <td><strong>xxxx%</strong>
   </td>
  </tr>
  <tr>
   <td rowspan="2" ><strong>Coding</strong>
   </td>
   <td>HumanEval pass@1
   </td>
   <td>83.20
   </td>
   <td>83.30
   </td>
   <td>100.1%
   </td>
  </tr>
  <tr>
   <td>HumanEval+ pass@1
   </td>
   <td>78.40
   </td>
   <td>78.60
   </td>
   <td>100.3%
   </td>
  </tr>
  <tr>
   <td rowspan="9" ><strong>Multilingual</strong>
   </td>
   <td>Portuguese MMLU (5-shot)
   </td>
   <td>79.76
   </td>
   <td>79.47
   </td>
   <td>99.6%
   </td>
  </tr>
  <tr>
   <td>Spanish MMLU (5-shot)
   </td>
   <td>79.33
   </td>
   <td>79.23
   </td>
   <td>99.9%
   </td>
  </tr>
  <tr>
   <td>Italian MMLU (5-shot)
   </td>
   <td>79.15
   </td>
   <td>78.80
   </td>
   <td>99.6%
   </td>
  </tr>
  <tr>
   <td>German MMLU (5-shot)
   </td>
   <td>77.94
   </td>
   <td>77.92
   </td>
   <td>100.0%
   </td>
  </tr>
  <tr>
   <td>French MMLU (5-shot)
   </td>
   <td>75.69
   </td>
   <td>75.79
   </td>
   <td>100.1%
   </td>
  </tr>
  <tr>
   <td>Hindi MMLU (5-shot)
   </td>
   <td>73.81
   </td>
   <td>73.49
   </td>
   <td>99.6%
   </td>
  </tr>
  <tr>
   <td>Thai MMLU (5-shot)
   </td>
   <td>71.97
   </td>
   <td>71.44
   </td>
   <td>99.2%
   </td>
  </tr>
</table>

### Reproduction

The results were obtained using the following commands:

#### MMLU
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,max_model_len=3850,max_gen_toks=10,tensor_parallel_size=1 \
  --tasks mmlu_llama_3.1_instruct \
  --fewshot_as_multiturn \
  --apply_chat_template \
  --num_fewshot 5 \
  --batch_size auto
```

#### MMLU-CoT
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,max_model_len=4064,max_gen_toks=1024,tensor_parallel_size=1 \
  --tasks mmlu_cot_0shot_llama_3.1_instruct \
  --apply_chat_template \
  --num_fewshot 0 \
  --batch_size auto
```

#### ARC-Challenge
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,max_model_len=3940,max_gen_toks=100,tensor_parallel_size=1 \
  --tasks arc_challenge_llama_3.1_instruct \
  --apply_chat_template \
  --num_fewshot 0 \
  --batch_size auto
```

#### GSM-8K
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,max_model_len=4096,max_gen_toks=1024,tensor_parallel_size=1 \
  --tasks gsm8k_cot_llama_3.1_instruct \
  --fewshot_as_multiturn \
  --apply_chat_template \
  --num_fewshot 8 \
  --batch_size auto
```

#### Hellaswag
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=1 \
  --tasks hellaswag \
  --num_fewshot 10 \
  --batch_size auto
```

#### Winogrande
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=1 \
  --tasks winogrande \
  --num_fewshot 5 \
  --batch_size auto
```

#### TruthfulQA
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=1 \
  --tasks truthfulqa \
  --num_fewshot 0 \
  --batch_size auto
```

#### OpenLLM v2
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,max_model_len=4096,tensor_parallel_size=1,enable_chunked_prefill=True \
  --apply_chat_template \
  --fewshot_as_multiturn \
  --tasks leaderboard \
  --batch_size auto
```

#### MMLU Portuguese
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,max_model_len=3850,max_gen_toks=10,tensor_parallel_size=1 \
  --tasks mmlu_pt_llama_3.1_instruct \
  --fewshot_as_multiturn \
  --apply_chat_template \
  --num_fewshot 5 \
  --batch_size auto
```

#### MMLU Spanish
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,max_model_len=3850,max_gen_toks=10,tensor_parallel_size=1 \
  --tasks mmlu_es_llama_3.1_instruct \
  --fewshot_as_multiturn \
  --apply_chat_template \
  --num_fewshot 5 \
  --batch_size auto
```

#### MMLU Italian
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,max_model_len=3850,max_gen_toks=10,tensor_parallel_size=1 \
  --tasks mmlu_it_llama_3.1_instruct \
  --fewshot_as_multiturn \
  --apply_chat_template \
  --num_fewshot 5 \
  --batch_size auto
```

#### MMLU German
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,max_model_len=3850,max_gen_toks=10,tensor_parallel_size=1 \
  --tasks mmlu_de_llama_3.1_instruct \
  --fewshot_as_multiturn \
  --apply_chat_template \
  --num_fewshot 5 \
  --batch_size auto
```

#### MMLU French
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,max_model_len=3850,max_gen_toks=10,tensor_parallel_size=1 \
  --tasks mmlu_fr_llama_3.1_instruct \
  --fewshot_as_multiturn \
  --apply_chat_template \
  --num_fewshot 5 \
  --batch_size auto
```

#### MMLU Hindi
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,max_model_len=3850,max_gen_toks=10,tensor_parallel_size=1 \
  --tasks mmlu_hi_llama_3.1_instruct \
  --fewshot_as_multiturn \
  --apply_chat_template \
  --num_fewshot 5 \
  --batch_size auto
```

#### MMLU Thai
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8",dtype=auto,max_model_len=3850,max_gen_toks=10,tensor_parallel_size=1 \
  --tasks mmlu_th_llama_3.1_instruct \
  --fewshot_as_multiturn \
  --apply_chat_template \
  --num_fewshot 5 \
  --batch_size auto
```

#### HumanEval and HumanEval+
##### Generation
```
python3 codegen/generate.py \
  --model neuralmagic-ent/Llama-3.3-70B-Instruct-quantized.w8a8 \
  --bs 16 \
  --temperature 0.2 \
  --n_samples 50 \
  --root "." \
  --dataset humaneval
```
##### Sanitization
```
python3 evalplus/sanitize.py \
  humaneval/neuralmagic-ent--Llama-3.3-70B-Instruct-quantized.w8a8_vllm_temp_0.2
```
##### Evaluation
```
evalplus.evaluate \
  --dataset humaneval \
  --samples humaneval/neuralmagic-ent--Llama-3.3-70B-Instruct-quantized.w8a8_vllm_temp_0.2-sanitized
```