gemma-4-26B-A4B-it_q8

gemma-4-26B-A4B-it_q8

This repository contains a vLLM-compatible, expert-only W8A16 quantization of google/gemma-4-26B-A4B-it.

Quantization

• Quantized modules: routed MoE expert gate_proj, up_proj, and down_proj
• Weight format: signed INT8
• Activations: BF16 (W8A16)
• GPTQ group size: 32
• Symmetric, static weight quantization
• Attention, router, shared MLP, embeddings, LM head, and vision tower remain BF16
• Calibration: 32 coding-oriented samples, sequence length 2,048
• Checkpoint format: compressed-tensors / pack-quantized
• Weight shards: 7
• Weight size: approximately 28.13 GiB

The compressed-tensors ignore metadata includes the missing v_proj aliases for Gemma 4 global-attention layers so that vLLM can consistently fuse Q/K/V.

Tested configuration

• vLLM: 0.21.0
• GPU: NVIDIA A100 PCIe 40GB
• Context length: 8,192
• GPU memory while loaded: approximately 37.1GB
• Backend: CompressedTensorsWNA16MarlinMoEMethod
• OpenAI-compatible /v1/models and /v1/chat/completions: tested
• Observed generation throughput: approximately 32–42.5 tokens/s for a single short request

Start with NVIDIA Docker

docker run -d \
  --name gemma4-vllm \
  --gpus '"device=0"' \
  --ipc=host \
  --ulimit memlock=-1 \
  -p 8000:8000 \
  -v "$PWD:/model:ro" \
  vllm/vllm-openai:v0.21.0 \
  --model /model \
  --served-model-name gemma-4-26B-A4B-it_q8 \
  --host 0.0.0.0 \
  --port 8000 \
  --dtype bfloat16 \
  --max-model-len 8192 \
  --max-num-batched-tokens 4096 \
  --gpu-memory-utilization 0.90

Tested rootless Podman command

docker run -d \
  --name gemma4-vllm \
  --userns=keep-id \
  --user "$(id -u):$(id -g)" \
  --device nvidia.com/gpu=1 \
  --ipc=host \
  --ulimit memlock=-1 \
  -p 18080:8000 \
  -v "$PWD:/model:ro,Z" \
  --entrypoint /usr/bin/python3 \
  localhost/vllm/vllm-openai:v0.21.0-cu129-ubuntu2404-media \
  -m vllm.entrypoints.openai.api_server \
  --model /model \
  --served-model-name gemma-4-26B-A4B-it_q8 \
  --host 0.0.0.0 \
  --port 8000 \
  --dtype bfloat16 \
  --max-model-len 8192 \
  --max-num-batched-tokens 4096 \
  --gpu-memory-utilization 0.90

Gemma 4 requires --max-num-batched-tokens to be greater than its multimodal item token budget. The tested value is 4,096.

OpenAI-compatible request

curl http://127.0.0.1:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma-4-26B-A4B-it_q8",
    "messages": [
      {"role": "user", "content": "Write a Python two_sum function in O(n)."}
    ],
    "max_tokens": 256,
    "temperature": 0.2
  }'

Notes

• A 40GB GPU is recommended for the tested 8K configuration.
• This is an expert-only Q8 checkpoint, not a full-model Q8 conversion.
• Text generation and code generation were tested. Multimodal inference was not functionally evaluated.
• The model produced coherent code during testing, but Traditional Chinese instruction following was weaker than expected and should be evaluated for your application.
• Usage remains subject to the Gemma license and the base model terms.