granite-34b-code-instruct-8k-FP8

FP8 quantized version of IBM's Granite 34B Code model for efficient inference

This is an FP8 (E4M3) quantized version of ibm-granite/granite-34b-code-instruct-8k using compressed_tensors format. Quantized by TevunahAi on enterprise-grade hardware.

🎯 Recommended Usage: vLLM (Required)

For 34B models, vLLM is essential for practical deployment. FP8 quantization makes this flagship model accessible on high-end consumer GPUs.

Quick Start with vLLM

pip install vllm

Python API:

from vllm import LLM, SamplingParams

# vLLM auto-detects FP8 from model config
llm = LLM(model="TevunahAi/granite-34b-code-instruct-8k-FP8", dtype="auto")

# Generate
prompt = "Write a Python function to calculate fibonacci numbers:"
sampling_params = SamplingParams(temperature=0.7, max_tokens=256)

outputs = llm.generate([prompt], sampling_params)
for output in outputs:
    print(output.outputs[0].text)

OpenAI-Compatible API Server:

vllm serve TevunahAi/granite-34b-code-instruct-8k-FP8 \
    --dtype auto \
    --max-model-len 8192

Then use with OpenAI client:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="token-abc123",  # dummy key
)

response = client.chat.completions.create(
    model="TevunahAi/granite-34b-code-instruct-8k-FP8",
    messages=[
        {"role": "user", "content": "Write a Python function to calculate fibonacci numbers"}
    ],
    temperature=0.7,
    max_tokens=256,
)

print(response.choices[0].message.content)

vLLM Benefits

• ✅ Weights, activations, and KV cache in FP8
• ✅ ~34GB VRAM (50% reduction vs BF16's ~68GB)
• ✅ Single high-end GPU deployment (H100, RTX 6000 Ada, A100 80GB)
• ✅ Native FP8 tensor core acceleration
• ✅ Production-grade performance

⚠️ Transformers: Not Practical

At 34B parameters, transformers will decompress to ~68GB+ VRAM, requiring multi-GPU setups or data center GPUs. This is not recommended for deployment.

Transformers Example (Multi-GPU Required - Click to expand)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Requires multi-GPU or 80GB+ single GPU
model = AutoModelForCausalLM.from_pretrained(
    "TevunahAi/granite-34b-code-instruct-8k-FP8",
    device_map="auto",  # Will distribute across GPUs
    torch_dtype="auto",
    low_cpu_mem_usage=True,
)
tokenizer = AutoTokenizer.from_pretrained("TevunahAi/granite-34b-code-instruct-8k-FP8")

# Generate
prompt = "Write a Python function to calculate fibonacci numbers:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Requirements:

pip install torch>=2.1.0 transformers>=4.40.0 accelerate compressed-tensors

System Requirements:

• ~68GB+ VRAM (decompressed to BF16)
• Multi-GPU setup or A100 80GB / H100 80GB
• Not practical for most deployments

⚠️ Critical: Use vLLM instead. Transformers is only viable for research/testing with multi-GPU setups.

📊 Quantization Details

Property	Value
Base Model	ibm-granite/granite-34b-code-instruct-8k
Quantization Method	FP8 E4M3 weight-only
Framework	llm-compressor + compressed_tensors
Calibration Dataset	open_platypus (512 samples)
Storage Size	~34GB (sharded safetensors)
VRAM (vLLM)	~34GB
VRAM (Transformers)	~68GB+ (decompressed to BF16)
Target Hardware	NVIDIA H100, A100 80GB, RTX 6000 Ada
Quantization Time	31.0 minutes

Quantization Infrastructure

Professional hardware ensures consistent, high-quality quantization:

• CPUs: Dual Intel Xeon Max 9480 (112 cores / 224 threads, 128GB HBM2e)
• GPU: NVIDIA RTX 5000 Ada Generation (32GB VRAM, native FP8 support)
• Memory: 256GB DDR5 + 128GB HBM2e = 384GB total system memory
• Software Stack: Ubuntu 25.10 | Python 3.12 | PyTorch 2.8 | CUDA 13.0 | llm-compressor

Performance Notes

Optimal HBM2e Utilization:

• This 34B model demonstrates ideal sizing for our dual Xeon Max architecture
• Full CPU/HBM2e processing path for maximum efficiency
• Superior per-parameter performance (0.91 min/B vs 1.1 min/B for 20B)
• Counterintuitively faster quantization than smaller models due to pure HBM2e workflow
• Sweet spot for our hardware infrastructure

🔧 Why FP8 for 34B Models?

With vLLM/TensorRT-LLM:

• ✅ Enables single-GPU deployment (~34GB vs ~68GB BF16)
• ✅ 50% memory reduction across weights, activations, and KV cache
• ✅ Faster inference via native FP8 tensor cores
• ✅ Makes flagship model accessible on high-end consumer/prosumer GPUs
• ✅ Minimal quality loss for code generation tasks

Without FP8:

• ❌ BF16 requires ~68GB VRAM (H100 80GB or multi-GPU)
• ❌ Limited deployment options
• ❌ Higher infrastructure costs

FP8 quantization transforms 34B from "data center only" to "high-end workstation deployable".

💾 Model Files

This model is sharded into multiple safetensors files (all required for inference). The compressed format enables efficient storage and faster downloads.

🚀 Granite Code Model Family

IBM's Granite Code models are specifically trained for enterprise code generation. The 34B version represents the flagship tier:

Model	VRAM (vLLM)	Quality	Use Case
8B-FP8	~8GB	Good	Fast iteration, prototyping
20B-FP8	~20GB	Better	Complex tasks, better reasoning
34B-FP8	~34GB	Best	Flagship performance, production

34B Benefits:

• ✅ State-of-the-art code quality for Granite family
• ✅ Superior reasoning and complex problem solving
• ✅ Enterprise-grade completions for mission-critical applications
• ✅ Best context understanding across the model family
• ✅ 8K context window for larger codebases

🔬 Quality Assurance

• Professional calibration: 512 diverse code samples
• Validation: Tested on code generation benchmarks
• Format: Standard compressed_tensors for broad compatibility
• Optimization: Hardware-optimized quantization workflow

📚 Original Model

This quantization is based on ibm-granite/granite-34b-code-instruct-8k by IBM.

For comprehensive information about:

• Model architecture and training methodology
• Supported programming languages
• Evaluation benchmarks and results
• Ethical considerations and responsible AI guidelines

Please refer to the original model card.

🔧 Hardware Requirements

Minimum (vLLM):

• GPU: NVIDIA A100 40GB or RTX 6000 Ada (48GB)
• VRAM: 34GB minimum, 40GB+ recommended
• CUDA: 11.8 or newer

Recommended (vLLM):

• GPU: NVIDIA H100 (80GB) / A100 80GB / RTX 6000 Ada (48GB)
• VRAM: 40GB+
• CUDA: 12.0+

Transformers:

• GPU: Multi-GPU setup (2x A100 40GB) or single A100/H100 80GB
• VRAM: 68GB+ total
• Not recommended - use vLLM instead

📖 Additional Resources

• vLLM Documentation: docs.vllm.ai
• TensorRT-LLM: github.com/NVIDIA/TensorRT-LLM
• TevunahAi Models: huggingface.co/TevunahAi
• llm-compressor: github.com/vllm-project/llm-compressor
• IBM Granite: github.com/ibm-granite

📄 License

This model inherits the Apache 2.0 License from the original Granite model.

🙏 Acknowledgments

• Original Model: IBM Granite team
• Quantization Framework: Neural Magic's llm-compressor
• Quantized by: TevunahAi

📝 Citation

If you use this model, please cite the original Granite work:

@misc{granite2024,
  title={Granite Code Models},
  author={IBM Research},
  year={2024},
  url={https://huggingface.co/ibm-granite/granite-34b-code-instruct-8k}
}

---

Professional AI Model Quantization by TevunahAi

Making flagship models accessible through enterprise-grade quantization

View all models | Contact for custom quantization