---
license: other
base_model: Jackrong/Qwopus3.6-27B-v2
quantized_by: XReyRobert
private: true
---

# Qwopus3.6-27B-v2 GPTQ-Pro FOEM 4-bit g128 ns256 v2

Private GPTQ-Pro FOEM 4-bit group-size 128 quantization of Jackrong/Qwopus3.6-27B-v2.

Quantization recipe:
- GPTQ-Pro factory profile
- bits: 4
- group_size: 128
- sym: true
- desc_act: false
- true_sequential: true
- mse: 2.0
- damp_percent: 0.05
- damp_auto_increment: 0.01
- FOEM alpha: 0.25
- FOEM beta: 0.2
- calibration dataset: Salesforce/wikitext wikitext-2-raw-v1
- nsamples: 256
- seqlen: 2048
- dynamic skips: visual, vision, mtp, lm_head, embed_tokens, norm

Validation on RunPod A100 80GB with vLLM 0.18.0:
- max_model_len: 131072
- quantization: gptq_marlin
- kv_cache_dtype: fp8_e5m2
- text-only mode
- loaded successfully at 131072 context
- short completion after warmup: about 64-67 tok/s
- chat no-thinking after warmup: about 64 tok/s
- long context check: 130810 prompt tokens, recovered needle zebra-9901 with enable_thinking=false

Notes:
- This artifact is private and experimental.
- vLLM startup requires ninja in PATH for flashinfer JIT on this environment.
- For normal chat validation, use chat_template_kwargs enable_thinking=false when no reasoning output is desired.