Skip to content

Llama33 70b awq draft llama32 1b

SPD Overview

This SPD defines a dual-model configuration for speculative decoding, where the Target Model ensures high-quality multilingual responses and the Draft Model accelerates inference.

Target Model – llama-3.3-70b-instruct-awq

This is the AWQ version of the Llama 3.3 70B Instruct model. The Meta Llama 3.3 multilingual large language model (LLM) is an instruction tuned generative model in 70B (text in/text out). The Llama 3.3 instruction tuned text only model is optimized for multilingual dialogue use cases and outperforms many of the available open source and closed chat models on common industry benchmarks.

  • Model Architecture: Llama 3.3 is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions use supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align with human preferences for helpfulness and safety.
  • Model Release Date: December 6, 2024.
  • Model Source: casperhansen/llama-3.3-70b-instruct-awq
  • Model Repository: casper-hansen/AutoAWQ
  • License: llama3.3
  • Supported languages: English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.

QPC Configurations

Precision SoCs / Tensor slicing NSP-Cores (per SoC) Full Batch Size Chunking Prompt Length Context Length (CL) QPC URL QPC Size QPC Download Onnx URL Onnx Download Generation Date
MXFP6 2 8 1 128 4096 https://dc00tk1pxen80.cloudfront.net/SDK1.21.4.0/casperhansen/llama-3.3-70b-instruct-awq/casperhansen_llama-3.3-70b-instruct-awq_qpc_8cores_128pl_4096cl_2devices_11Stokens_target.tar.gz 38GB Download https://dc00tk1pxen80.cloudfront.net/SDK1.21.4.0/casperhansen/llama-3.3-70b-instruct-awq/casperhansen_llama-3.3-70b-instruct-awq_ONNX_target.tar.gz Download 8-May-2026

Draft Model – Llama 3.2 1B Instruct

The Meta Llama 3.2-1B Instruct is a smaller, instruction-tuned model used here as the Draft Model for speculative decoding. It provides fast draft token generation to reduce latency when paired with the Target 70B model.

QPC Configurations

Precision SoCs / Tensor slicing NSP-Cores (per SoC) Full Batch Size Chunking Prompt Length Context Length (CL) QPC URL QPC Size QPC Download Onnx URL Onnx Download Generation Date
MXFP6 2 8 1 128 4096 https://dc00tk1pxen80.cloudfront.net/SDK1.21.4.0/meta-llama/Llama-3.2-1B-Instruct/meta-llama_Llama-3.2-1B-Instruct_qpc_8cores_128pl_4096cl_2devices_11Stokens_draft.tar.gz 2.9GB Download https://dc00tk1pxen80.cloudfront.net/SDK1.21.4.0/meta-llama/Llama-3.2-1B-Instruct/meta-llama_Llama-3.2-1B-Instruct_ONNX_draft.tar.gz Download 8-May-2026

Run This Model

Download QPCs

mkdir -p casperhansen/llama-3.3-70b-instruct-awq
cd casperhansen/llama-3.3-70b-instruct-awq

# Download Target QPC
wget <Target_QPC_Download_URL>
tar xzvf <target_qpc_filename.tar.gz>

mkdir -p meta-llama/Llama-3.2-1B-Instruct
cd meta-llama/Llama-3.2-1B-Instruct

# Download Draft QPC
wget <Draft_QPC_Download_URL>
tar xzvf <draft_qpc_filename.tar.gz>

Run QPC

Replace target_qpc_path, draft_qpc_path with actual extracted QPC directories.

python -m vllm.entrypoints.openai.api_server \
  --host 0.0.0.0 \
  --port <PORT> \
  --device-group <device_group> \
  --tensor-parallel-size 2 \
  --model casperhansen/llama-3.3-70b-instruct-awq \
  --max-model-len <ctx_len> \
  --max-seq-len-to-capture <prefill_seq_len> \
  --max-num-seqs 1 \
  --kv-cache-dtype mxint8 \
  --quantization mxfp6 \
  --block-size 32 \
  --gpu-memory-utilization 0.5 \
  --override-qaic-config "qpc_path=<target_qpc_path> num_cores=8 mxint8_kv_cache=True" \
  --speculative-config '{"num_speculative_tokens":11,"model":"meta-llama/Llama-3.2-1B-Instruct","quantization":"mxfp6","draft_override_qaic_config":{"qpc_path":"<draft_qpc_path>","device_group":<device_group>,"num_cores":8,"kv_cache_dtype":"mxint8","mos":2}}'

Run Inference

Once the server is running, send a request to the OpenAI-compatible endpoint:

curl http://127.0.0.1:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": casperhansen/llama-3.3-70b-instruct-awq,
    "messages": [
      {"role": "user", "content": "Explain speculative decoding in 3 sentences."}
    ],
    "max_tokens": 128,
    "temperature": 0
  }'