Gemma3 models trained with BF16 can produce residual stream activations that exceed FP16's representable range (±65,504). FP16 inference is a hard requirement for ANE. This causes:

Overflow to inf in FP16 computation
NaN propagation through subsequent layers
Complete model failure on ANE (which uses FP16)

Model Quality Benchmarks

FP16 Scaling for ANE Compatibility

Gemma3 4B QAT models produce activations that exceed FP16 range (±65,504) during inference. We apply weight scaling (α=0.1875) to prevent overflow:

Embedding weights scaled by α=0.1875 (3/16)
LM head logits divided by α to restore original scale
Zero runtime overhead - transformation applied at conversion time
100% token match with BF16 reference

Quantization Results

Configuration	KL Divergence	Correlation	Match Rate	Notes
FP16 baseline (no LUT)	0.0006	0.995	99.86%	Best quality
FFN LUT4,4 + LM LUT6,4	0.196	0.959	90%	This model
FFN LUT4,8 only	0.284	0.971	87%	Larger size
FFN LUT4,8 + LM LUT6,4	0.279	0.970	86%	-

Metric Guidelines

Metric	Healthy	Concerning
KL Divergence	< 0.3	> 0.5
Correlation	> 0.95	< 0.90
Match Rate	> 85%	< 75%

Reference

HF Model: google/gemma-3-4b-it-qat-int4-unquantized
Scaling: α=0.1875 (FP16 overflow prevention)
Context: 4096 tokens
Sliding Window: 1024

LUT Quantization

ANEMLL applies LUT (Lookup Table) palettization to compress weights. This model uses per-channel grouping (per_channel=4) where every 4 output channels share one lookup table for finer granularity and better quality.

Model Architecture

4-Function Chunk Design

Each FFN chunk contains 4 CoreML functions to support Gemma 3's sliding window attention:

Function	Purpose
`infer`	Single-token inference (position < sliding_window)
`prefill`	Batch prefill processing (position < sliding_window)
`infer_rotate`	Single-token inference after KV rotation (position ≥ sliding_window)
`prefill_rotate`	Batch prefill after KV rotation (position ≥ sliding_window)

Why 4 functions? When position reaches the sliding window boundary (1024), KV cache must be rotated. The *_rotate functions handle inference after rotation, enabling efficient sliding window attention without recomputing the entire cache.

Gemma 3 Hybrid Local-Global Attention

Gemma 3 uses a hybrid attention architecture that combines local and global attention:

Local Attention (Sliding Window)

Applied to most layers
Window size: 1024 tokens
Efficient O(n × w) complexity where w = window size
Bounded memory usage for long sequences

Global Attention

Applied every 4th layer (layers 4, 8, 12, ...)
Full context access up to 4096 tokens
Preserves long-range dependencies

Advantages:

Memory Efficiency: ~75% smaller KV cache compared to full attention
Speed: Faster inference for long sequences
Quality: Global attention layers maintain document-level coherence
Scalability: Efficiently handles longer contexts than pure global attention

iOS/macOS Distribution

This folder contains pre-compiled CoreML models (.mlmodelc directories) ready for iOS and macOS deployment. No zip extraction required.

Requirements

macOS 15 (Sequoia) or later with Apple Silicon
iOS 17+ for mobile deployment
8GB+ RAM recommended
Python 3.9+ for testing
CoreML Tools 8.x+ and HuggingFace Transformers

Installation

bash

# Install Git LFS (required for large model files)
brew install git-lfs
git lfs install

# Install Python dependencies
pip install coremltools transformers

# Clone the repository
git clone https://huggingface.co/anemll/anemll-google-gemma-3-4b-it-qat-int4-unquantized-ctx4096_0.3.5

# Navigate to iOS/macOS distribution folder
cd anemll-google-gemma-3-4b-it-qat-int4-unquantized-ctx4096_0.3.5/ios

# Verify models are ready (should see .mlmodelc directories)
ls -la *.mlmodelc

You should see four .mlmodelc directories:

gemma3_embeddings.mlmodelc
gemma3_FFN_PF_lut4_chunk_01of02.mlmodelc
gemma3_FFN_PF_lut4_chunk_02of02.mlmodelc
gemma3_lm_head_lut6.mlmodelc

Test Inference

bash

# Basic chat interface
python chat.py --meta ./meta.yaml

# Or with conversation history
python chat_full.py --meta ./meta.yaml

Controls:

Ctrl-D to exit
Ctrl-C to interrupt generation

Note: First load takes time as macOS places the model on ANE. Subsequent loads are instant.

Model Specifications

Parameter	Value
Base Model	`google/gemma-3-4b-it-qat-int4-unquantized`
Context Length	4096 tokens
Sliding Window	1024 tokens
Batch Size	64
Number of Chunks	2
FFN Quantization	LUT4 (4-bit, per_channel=4)
LM Head Quantization	LUT6 (6-bit, per_channel=4)
Embeddings	FP16 (unquantized)
FP16 Scaling	α=0.1875

Model Files

File	Size	Description
`gemma3_embeddings.mlmodelc`	1.3 GB	Token embeddings (FP16)
`gemma3_FFN_PF_lut4_chunk_01of02.mlmodelc`	788 MB	FFN layers 1-13 + prefill
`gemma3_FFN_PF_lut4_chunk_02of02.mlmodelc`	788 MB	FFN layers 14-26
`gemma3_lm_head_lut6.mlmodelc`	488 MB	Language model head

Total Size: ~3.4 GB

iOS/macOS App Integration

Copy .mlmodelc directories to your Xcode project
Include config.json for offline tokenizer
Reference: ANEMLL ChatBot TestFlight

iOS: Requires unzipped .mlmodelc (this distribution) + config.json macOS: Supports both zipped and unzipped formats

License

ANEMLL Pipeline: MIT License
Gemma Model: Google Gemma Terms of Use

anemll/anemll-google-gemma-3-4b-it-qat-int4-unquantized-ctx4096_0.3.5

ANEMLL