granite-4.0-h-small-ZeroClawToolUse-MagicQuant-GGUF

Fine-tuned GGUF quantization of granite-4.0-h-small, trained on ./data/zeroclaw_training_data.jsonl with QLoRA and quantized using MagicQuant hybrid evolutionary search.

Base Model

This is a fine-tuned and quantized derivative of granite-4.0-h-small. All credit for the base model architecture and weights goes to the original authors. The base model's license applies to this derivative.

Quantization Method

Quantized using MagicQuant hybrid evolutionary per-tensor quantization, based on the methodology by magiccodingman:

• Tensors are classified into sensitivity groups (Embeddings, Head, Query, Key, Output, FFN Up/Down, MoE Experts, Router)
• An evolutionary search finds the optimal quantization type per group, balancing size vs. perplexity
• Q4/Q5/Q6 tier targets are produced with different size-quality tradeoffs
• Small-row tensors and sensitivity-critical layers (embeddings, output head, router) are kept at F32/F16/BF16
• This is NOT a uniform quantization -- each tensor group gets its own optimal type

Training Details

Parameter	Value
Method	QLoRA with completion-only loss masking
LoRA rank (r)	32
LoRA alpha	64
LoRA dropout	0.05
Epochs	3
Learning rate	0.0002
LR scheduler	cosine
Batch size	2 (effective 8 with gradient accumulation)
Optimizer	adamw_8bit
Training sequence length	4096
Precision	BF16
Dataset	./data/zeroclaw_training_data.jsonl
Hardware	AMD Ryzen AI Max+ 395 (Strix Halo), 128 GB unified memory (GTT), ROCm
Training pipeline	Custom fast QLoRA with shard-by-shard BnB 4-bit quantization

Completion-only loss: Only assistant response turns contribute to the training loss. System and user turns are masked, so the model learns to generate responses rather than memorizing prompts.

GGUF Files

File	Size	Quant
granite-4.0-h-small-Q4-MXFP4_MOE-EH-BF16-OQ-Q80-K-Q6K-U-MXFP4MOE-D-IQ4NL.gguf	18.6 GB	Q4 hybrid
granite-4.0-h-small-Q5-MXFP4_MOE-EHKO-BF16-Q-Q80-U-Q6K-D-MXFP4MOE.gguf	18.9 GB	Q5 hybrid
granite-4.0-h-small-Q6-MXFP4_MOE-EHKOU-BF16-Q-Q80-D-MXFP4MOE.gguf	19.5 GB	Q6 hybrid

Usage

LM Studio

1. Download the GGUF file of your preferred quantization tier
2. Place it in your LM Studio models directory
3. Load the model in LM Studio -- it will auto-detect the chat template
4. The model supports the base model's full context length

llama.cpp

# Interactive chat
llama-cli -m granite-4.0-h-small-ZeroClawToolUse-MagicQuant-GGUF-Q5.gguf -c 8192 --chat-template chatml -cnv

# Single prompt
llama-cli -m granite-4.0-h-small-ZeroClawToolUse-MagicQuant-GGUF-Q5.gguf -c 8192 -p "Your prompt here"

# Server mode
llama-server -m granite-4.0-h-small-ZeroClawToolUse-MagicQuant-GGUF-Q5.gguf -c 8192 --port 8080

Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(model_path="./granite-4.0-h-small-ZeroClawToolUse-MagicQuant-GGUF-Q5.gguf", n_ctx=8192)
output = llm.create_chat_completion(
    messages=[
        {"role": "user", "content": "Hello, how are you?"}
    ]
)
print(output["choices"][0]["message"]["content"])

Caveats

• This is a personal fine-tune, not an official release from the base model authors
• The base model's license (apache-2.0) applies to all derivative files
• Quality depends on the training data and may not generalize to all tasks
• Quantization reduces precision -- verify outputs for your specific use case
• The hybrid quantization assigns different precision to different tensor groups, which means quality characteristics may differ from uniform quantizations

Limitations

• Training data used sequences up to 4096 tokens; the model retains the base model's full context window
• Performance on tasks not represented in the training data may be degraded
• Quantized models may exhibit subtle differences from the full-precision fine-tune
• This model inherits any limitations and biases present in the base model

---

Generated with the MagicQuant Pipeline