File size: 4,777 Bytes
bfcbe0b 8b052ff bfcbe0b 477f5bc bfcbe0b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 |
---
license: apache-2.0
base_model:
- google/gemma-3-4b-it-qat-int4-unquantized
tags:
- OpenArc
- OpenVINO
- Optimum-Intel
- image-text-to-text
---
## Gemma 3 for OpenArc has landed!
My Project [OpenArc](https://github.com/SearchSavior/OpenArc), an inference engine for OpenVINO, now supports this model and serves inference over OpenAI compatible endpoints for text to text *and* text with vision!
We have a growing Discord community of others interested in using Intel for AI/ML.
[](https://discord.gg/maMY7QjG)
This model was converted to the OpenVINO IR format using the following Optimum-CLI command:
```
optimum-cli export openvino -m ""input-model"" --task image-text-to-text --weight-format int8 ""converted-model""
```
- Find documentation on the Optimum-CLI export process [here](https://huggingface.co/docs/optimum/main/en/intel/openvino/export)
- Use my HF space [Echo9Zulu/Optimum-CLI-Tool_tool](https://huggingface.co/spaces/Echo9Zulu/Optimum-CLI-Tool_tool) to build commands and execute locally
### What does the test code do?
Well, it demonstrates how to inference in python *and* what parts of that code are important for benchmarking performance.
Text generation offers different challenges than text-generation with images; for examples, vision encoders often use different strategies for handling properties an image can have; to get good performance **be mindful of image resolution**.
In practice this can translate to higher memory usage, reduced throughput and greater variety in results. Gemma-3 uses SigLIP 2 which has many SOTA optimizations; even so, effort in the preprocessing stage of a pipeline makes a world of difference.
To run the test code:
- Install device specific drivers
- Build Optimum-Intel for OpenVINO from source
- Find your spiciest images to get that AGI refusal smell
```
pip install optimum[openvino]+https://github.com/huggingface/optimum-intel
```
```
import time
from PIL import Image
from transformers import AutoProcessor
from optimum.intel.openvino import OVModelForVisualCausalLM
model_id = "Echo9Zulu/gemma-3-4b-it-int8_asym-ov" # Can be an HF id or a path
ov_config = {"PERFORMANCE_HINT": "LATENCY"} # Optimizes for first token latency and locks to single CPU socket
print("Loading model... this should get faster after the first generation due to caching behavior.")
print("")
start_load_time = time.time()
model = OVModelForVisualCausalLM.from_pretrained(model_id, export=False, device="CPU", ov_config=ov_config) # For GPU use "GPU.0"
processor = AutoProcessor.from_pretrained(model_id) # Instead of using AutoTokenizers we use AutoProcessor which routes to the appropriate input processor i.e, how does a model expect image tokens.
# Under the hood this takes care of model specific preprocessing and has functionality overlap with AutoTokenizers.
end_load_time = time.time()
image_path = r"" # This script expects .png
image = Image.open(image_path)
image = image.convert("RGB") # Required by gemma3. In practice this would need to be handled at the engine level OR in model-specifc pre-processing.
conversation = [
{
"role": "user",
"content": [
{
"type": "image"
},
{"type": "text", "text": "Describe this image."},
],
}
]
text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
inputs = processor(text=[text_prompt], images=[image], padding=True, return_tensors="pt")
input_token_count = len(inputs.input_ids[0])
print(f"Sum of image and text tokens: {len(inputs.input_ids[0])}")
start_time = time.time()
output_ids = model.generate(**inputs, max_new_tokens=1024)
generated_ids = [output_ids[len(input_ids) :] for input_ids, output_ids in zip(inputs.input_ids, output_ids)]
output_text = processor.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
num_tokens_generated = len(generated_ids[0])
load_time = end_load_time - start_load_time
generation_time = time.time() - start_time
tokens_per_second = num_tokens_generated / generation_time
average_token_latency = generation_time / num_tokens_generated
print("\nPerformance Report:")
print("-"*50)
print(f"Input Tokens : {input_token_count:>9}")
print(f"Generated Tokens : {num_tokens_generated:>9}")
print(f"Model Load Time : {load_time:>9.2f} sec")
print(f"Generation Time : {generation_time:>9.2f} sec")
print(f"Throughput : {tokens_per_second:>9.2f} t/s")
print(f"Avg Latency/Token : {average_token_latency:>9.3f} sec")
print(output_text)
``` |