gemma-3n-E4B-it-pte
executorch .pte export of google/gemma-3n-E4B-it for on-device mobile inference
available models
| variant | dtype | size | file |
|---|---|---|---|
| bf16 | bfloat16 | 13.1 gb | Gemma3n-E4B-IT-text-only.pte |
| int8 | int8 weights | 9.6 gb | Gemma3n-E4B-text-only-int8.pte |
model details
| property | value |
|---|---|
| source model | google/gemma-3n-E4B-it |
| text parameters | 7.40b |
| transformer layers | 35 |
| format | executorch .pte |
text-only export
this export contains only the text decoder components extracted from the full multimodal gemma-3n model
included:
- language_model (transformer decoder)
- lm_head (output projection)
not included:
- vision_tower (image encoder)
- audio_tower (audio encoder)
use this export for text-only inference tasks. if you need multimodal capabilities use the original huggingface model
quantization
- bf16: full bfloat16 precision weights
- int8: int8 weight-only quantization via torchao - recommended for mobile deployment
note: int4 quantization requires gpu for inference and is not suitable for cpu-only mobile deployment
export configuration
- fixed sequence length: 32 tokens
- torch.export with strict=False
- executorch to_edge conversion
usage
from executorch.runtime import Runtime
runtime = Runtime.get()
program = runtime.load_program("Gemma3n-E4B-text-only-int8.pte")
method = program.load_method("forward")
# input_ids shape: [1, 32] dtype: torch.long
output = method.execute([input_ids])
# output shape: [1, 32, 262400] dtype: torch.bfloat16
required patches
the transformers library requires two patches before export. see maceip/gemma3n-executorch for details
benchmarks
coming soon
links
- source code: https://github.com/maceip/gemma3n-executorch
- original model: https://huggingface.co/google/gemma-3n-E4B-it