gemma4-E4B-Q40-custom

<div align="center"> <img src=https://ai.google.dev/gemma/images/gemma4_banner.png> </div> <p align="center"> <a href="https://huggingface.co/collections/google/gemma-4" target="_blank">Hugging Face</a> | <a href="https://github.com/google-gemma" target="_blank">GitHub</a> | <a href="https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemma-4-12B/" target="_blank">Launch Blog</a> | <a href="https://ai.google.dev/gemma/docs/core" target="_blank">Documentation</a> <br> <b>License</b>: <a href="https://ai.google.dev/gemma/docs/gemma_4_license" target="_blank">Apache 2.0</a> | <b>Authors</b>: <a href="https://deepmind.google/models/gemma/" target="_blank">Google DeepMind</a> </p>

Gemma 4-E4B

This is a custom quantized version of the Gemma 4-E4B model, quantized to Q4_0 with custom OVERRIDE file. It is designed to achieve fast inference on Qualcomm Hexagon NPU while maintaining adequate accuracy.

how model is generated

Built with llama.cpp commit 7c158fb.

Three steps, run from the repo root.

Step 1 — download the unquantized HF model

hf download google/gemma-4-E4B-it-qat-q4_0-unquantized --local-dir ./hf-model

Step 2 — convert HF → F16 GGUF

convert_hf_to_gguf.py ./hf-model --outfile model-f16.gguf --outtype f16

Step 3 — follow the OVERRIDE file and quantize to Q4_0

build/bin/llama-quantize --tensor-type-file <override-file> \
    model-f16.gguf model-q4_0-override.gguf q4_0

Performance Measurement Commands

CPU uses --device none -ngl 0; HTP uses --device HTP0 -ngl 99. For each (model, backend, CTX ∈ {512, 1024, 4096}) two llama-bench runs were issued — one for prefill, one for decode:

# environment on device
export LD_LIBRARY_PATH=./lib
export ADSP_LIBRARY_PATH=./lib

# Prefill (Prefill tok/s; TTFT = CTX / Prefill × 1000)
./bin/llama-bench --device <none|HTP0> -m <model.gguf> \
  --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 --ubatch-size 1024 -fa on \
  -ngl <0|99> -p <CTX> -n 0

# Decode at depth = CTX (Decode tok/s)
./bin/llama-bench --device <none|HTP0> -m <model.gguf> \
  --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 --ubatch-size 1024 -fa on \
  -ngl <0|99> -p 0 -n 128 -d <CTX>

Performance Metrics

Performance on IQ9 (QCS9075M)

Accuracy Metrics

The MMLU-Pro is measured:

License

Apache 2.0