Qwen3.5-27B_PTQ

Qwen3.5-27B_PTQ (W4A16 and W8A16, Post-Training Quantization)

Overview

This repository provides a W4A16 and W8A16 PTQ (Post-Training Quantized) version of Qwen3.5-27B.

Unlike AWQ/GPTQ workflows, this model was quantized using a true PTQ pipeline with no calibration dataset.
The quantization is applied in a one-shot pass, making it extremely fast and simple while still maintaining strong fidelity.

---

Key Highlights

• Quantization Type: PTQ (Post-Training Quantization)
• Scheme: W4A16
  • Weights: INT4 (per-channel symmetric)
  • Activations: FP16/BF16 (unchanged)
• Scheme: W8A16
  • Weights: INT8 (per-channel symmetric)
  • Activations: FP16/BF16 (unchanged)
• Calibration Dataset: ❌ None (not required)
• Method: llmcompressor.oneshot pipeline
• Target Layers: Linear layers only
• Ignored Layers:
  • lm_head
  • visual modules
  • linear_attn
  • mtp

---

Quantization Details

This quant was created using a QuantizationModifier recipe:

• Targets: Linear layers
• Scheme: W4A16, W8A16
• Approach: One-shot PTQ (no iterative calibration)
• Preserves: Model structure, tokenizer, and chat template

---

PTQ Quality Metrics

W4A16

• Mean KLD: 0.054260
• Total Positions: 204,700

W8A16

• Mean KLD: 0.001895
• Total Positions: 204,700

---

Example Usage (vLLM)

pip install -U vllm

vllm serve TheHouseOfTheDude/Qwen3.5-27B-Writer_PTQ \
    --quantization compressed-tensors \
    --tensor-parallel-size 2 \
    --dtype bfloat16

---

Notes

• No calibration dataset required
• Extremely fast quantization pipeline
• Designed for vLLM runtime

---

Credits

• Base Model: Qwen3.5-27B
• Quantization: TheHouseOfTheDude