sensenova/SenseNova-6.8-Code2Video - GSQ-RCO-IQ3_S GGUF
Non-uniform mixed-precision GGUF quantization of sensenova/SenseNova-6.8-Code2Video following the GSQ + RCO methodology pioneered by ISTA-DASLab.
What is GSQ-RCO Quantization?
Instead of applying a single uniform quantization type to all weights, GSQ-RCO utilizes Riemannian Constrained Optimization (RCO) to assign a tailored quantization type to each tensor:
- Attention Projections (
attn_qkv,attn_gate): Higher precision (IQ4_XS,Q5_K,Q6_K) to safeguard reasoning and retrieval capabilities. - Feed-Forward Layers (
ffn_down,ffn_up): Aggressively quantized (IQ2_S,IQ2_XXS) where parameter redundancy is highest. - SSM & Normalization Layers: Preserved at full
BF16andF32precision for numerical stability.
Specifications
- Target Quantization Profile:
GSQ-RCO-IQ3_S - Tensor Allocation Map: Derived from
ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF - Importance Matrix:
imatrix-qwen3.8-27b.gguf
Usage
# Terminal Chat
./llama-cli -m SenseNova-6.8-Code2Video-GSQ-RCO-IQ3_S.gguf -p "Generate a video script:" -ngl 99
# OpenAI-Compatible Server
./llama-server -m SenseNova-6.8-Code2Video-GSQ-RCO-IQ3_S.gguf --port 8080 -ngl 99 -c 32768