SnowDrogito-RpR-32B_IQ4-XS

Updates and Description of Files

• Recent files uploaded use ArliAI RpR V3 instead of V1 as indicated in the name.
• All quantizations in this repo use IQ4_XS as a base with Q8 embedding and output tensors.
• (Recommended) SnowDrogito-RpR3-32B_IQ4-XS+Enhanced_Tensors.gguf - largest, highest quality, Q4KM size, quant using recalibrated imatrix on Bartowki's dataset+RP+Tao at 8k context, uses selective quantization with llama-quantize --tensor-type flags to bump up select FFN/self attention tensors between Q6 and Q8 as described here.
• SnowDrogito-RpRv3-32B_IQ4-XS-Q8InOut-Q56Attn.gguf - Q6 and Q5 Attention tensors. This and all quants uploaded prior used imatrix from Snowdrop.

MORE SPEED!

Improve inference speed offloading tensors instead of layers as referenced HERE. --overridetensors ".[13579].ffn_up|.[1-3][13579].ffn_up=CPU Restricts offloading of every third FFN up tensor, saving enough space on GPU to offload all layers on 24gb, taking me from 3.9tps to 10.6 tps. Example:

python koboldcpp.py --gpulayers 65 --quantkv 1 --overridetensors "\.[13579]\.ffn_up|\.[1-3][13579]\.ffn_up=CPU" --threads 10 --usecublas --contextsize 40960 --flashattention --model ~/Downloads/SnowDrogito-RpR3-32B_IQ4-XS+Enhanced_Tensors.gguf

...obviously editing threads, filepaths, etc...

Overview

SnowDrogito-RpR-32B_IQ4-XS is my shot at an optimized imatrix quantization for my QwQ RP Reasoning merge, goal is to add smarts to the popular Snowdrop roleplay model, with a little ArliAI RpR and Deepcogito for the smarts. Built using the TIES merge method, it attempts to combine strengths from multiple fine-tuned QwQ-32B models, quantized to IQ4_XS with Q8_0 embeddings and output layers for enhanced quality, to plus it up just a bit. Uploading because the PPL was lower, have been getting more varied/longer/more creative responses with this, but maybe it lacks contextual awareness compared to snowdrop? Not sure.

Setup for Reasoning and ChatML

• ChatML Formatting: Use ChatML with <|im_start|>role\ncontent<|im_end|>\n (e.g., <|im_start|>user\nHello!<|im_end|>\n).
• Reasoning Settings: Set "include names" to "never." Start reply with <think>\n to enable reasoning.
• Sampler Settings: From Snowdrop: Try temperature 0.9, min_p 0.05, top_a 0.3, TFS 0.75, repetition_penalty 1.03, DRY if available.
• My Settings: Response (tokens): 2048 Context (tokens): 40960 Temperature: 3.25 Top P: 0.98 Min P: 0.04 Top nsigna: 2.5 Repetition Penalty: 1.03 (XTC) Threshold: 0.3 (XTC) Probability: 0.3 Dry Multiplier: 0.8 Dry Base: 1.75 Dry Allowed Length: 4 Dry Penalty Range: 1024

Getting great reasoning results with ST's Start Reply With:

<think>
Chain-of-thought: Alright, what just happened is

For more details, see the setup guides and master import for ST for Snowdrop and other info on ArliAI RpR.

Performance

• Perplexity under identical conditions (IQ4_XS, 40,960 context, Q8_0 KV cache, on a 150K-token chat dataset) SnowDrogito-RpR-32B vs QwQ-32B-Snowdrop-v0:

  4.5597 ± 0.02554  
  4.6779 ± 0.02671

• Fits 40960 context 24GB VRAM using Q8 KV Cache with full GPU offload.

Model Details

• Base Model: Qwen/Qwen2.5-32B
• Architecture: Qwen 2.5 (32B parameters)
• Context Length: 40,960 tokens
• Quantization: IQ4_XS with Q8_0 embeddings and output layers for better quality.
• Used .imatrix file from Snowdrop.

Merge Configuration

This model was created using mergekit with the following TIES merge configuration:

models:
  - model: trashpanda-org/QwQ-32B-Snowdrop-v0
    parameters:
      weight: 0.75
      density: 0.5
  - model: deepcogito/cogito-v1-preview-qwen-32B
    parameters:
      weight: 0.15
      density: 0.5
  - model: ArliAI/QwQ-32B-ArliAI-RpR-v1
    parameters:
      weight: 0.1
      density: 0.5
merge_method: ties
base_model: Qwen/Qwen2.5-32B
parameters:
  weight: 0.9
  density: 0.9
  normalize: true
  int8_mask: true
tokenizer_source: Qwen/Qwen2.5-32B-Instruct
dtype: bfloat16

Quantization Details

• Primary Quantization: IQ4_XS (4-bit integer with extra-small blocks) using an importance matrix (trashpanda-org_QwQ-32B-Snowdrop-v0.imatrix) for high quality at reduced size.
• Embeddings & Output Layers: Quantized to Q8_0 (8-bit) to preserve precision in token embeddings and final output weights, differing from the standard IQ4_XS body. This boosts quality with a modest size increase.

Acknowledgments

• mergekit for merging.
• llama.cpp for quantization.
• Original model creators: Qwen, trashpanda-org, deepcogito, ArliAI.