Slipstream-Max/CPsyCounX-InternLM2-Chat-7B-GGUF-fp16

Model Details

Model Description

• Developed by: AITA
• Model type: Full-Precision Text Generation LLM (FP16 GGUF format)
• Original Model: https://huggingface.co/CAS-SIAT-XinHai/CPsyCounX
• Precision: FP16 (non-quantized full-precision version)

Repository

• GGUF Converter: llama.cpp
• Huggingface Hub: https://huggingface.co/Slipstream-Max/CPsyCounX-InternLM2-Chat-7B-GGUF-fp16

Usage

Method 1: llama.cpp Backend Server + Chatbox

Step 1: Start .llama.cpp Server

./llama-server \
  -m /path/to/model.gguf \
  -c 2048 \          # Context length
  --host 0.0.0.0 \   # Allow remote connections
  --port 8080 \      # Server port
  --n-gpu-layers 35  # GPU acceleration (if available)

Step 2: Connect via Chatbox

1. Download Chatbox
2. Configure API endpoint:

   API URL: http://localhost:8080
   Model: (leave empty)
   API Type: llama.cpp
   

3. Set generation parameters:

   {
     "temperature": 0.7,
     "max_tokens": 512,
     "top_p": 0.9
   }
   

Method 2: LM Studio

1. Download LM Studio
2. Load GGUF file:
   • Launch LM Studio
   • Search Slipstream-Max/Emollm-InternLM2.5-7B-chat-GGUF-fp16
3. Configure settings:

   Context Length: 2048
   GPU Offload: Recommended (enable if available)
   Batch Size: 512
   

4. Start chatting through the built-in UI

Precision Details

Filename	Precision	Size	Characteristics
CPsyCounX.gguf	FP16	[15.5GB]	Full original model precision

Hardware Requirements

Minimum:

• 24GB RAM (for 7B model)
• CPU with AVX/AVX2 instruction set support

Recommended:

• 32GB RAM
• CUDA-capable GPU (for acceleration)
• Fast SSD storage (due to large model size)

Key Notes

1. Requires latest llama.cpp (v3+ recommended)
2. Use --n-gpu-layers 35 for GPU acceleration (requires CUDA-enabled build)
3. Initial loading takes longer (2-5 minutes)
4. Requires more memory/storage than quantized versions
5. Use --mlock to prevent swapping

Advantages

• Preserves original model precision
• Ideal for precision-sensitive applications
• No quantization loss
• Suitable for continued fine-tuning