Qwen3.5-0.8B-INT-W8A8
INT8 W8A8 quantization of Qwen/Qwen3.5-0.8B,
produced for development purposes with the simplest possible method:
round-to-nearest (RTN) weights — no GPTQ, no calibration data.
Quantization scheme (compressed-tensors W8A8 preset)
| bits | type | strategy | scales | |
|---|---|---|---|---|
| weights | int8 | symmetric | per-channel | static, RTN via memoryless_minmax |
| activations | int8 | symmetric | per-token | dynamic (computed at runtime) |
Ignored (kept in bf16): ['lm_head', 're:model\\.visual\\..*'] — the vision tower stays full-precision;
lm_head is tied to the input embeddings.
Provenance
- Script:
pytorch_runtime/scripts/quantize_w8a8.py(dram-computing repo) - Date: 2026-08-07
- llm-compressor 0.12.0.1, compressed-tensors 0.17.1, transformers 5.10.1, torch 2.12.0+cpu
Usage
from vllm import LLM
llm = LLM("casperhansen/Qwen3.5-0.8B-INT-W8A8") # compressed-tensors backend auto-detected