Phi3 support

[martinlyons]

Feature request

Microsoft's new phi3 mode, in particular the 128K context mini model, is not supported by Optimum export.

Error is:
"ValueError: Trying to export a phi3 model, that is a custom or unsupported architecture, but no custom export configuration was passed as custom_export_configs. Please refer to https://huggingface.co/docs/optimum/main/en/exporters/onnx/usage_guides/export_a_model#custom-export-of-transformers-models for an example on how to export custom models. Please open an issue at https://github.com/huggingface/optimum/issues if you would like the model type phi3 to be supported natively in the ONNX export."

Motivation

Phi3-mini is potentially very significant as it has a large context but a small size. This could be used in lots of scenarios if it has good performance.

Your contribution

Unlikely I could do a PR as ONNX work is not my forte.

[Whadup]

optimum/optimum/utils/normalized_config.py

Line 254 in 56aabbe

"phi": NormalizedTextConfig,

Add "phi3": NormalizedTextConfig in this dict and you seem to be all set for phi3-mini

[IlyasMoutawwakil]

patching the TasksManager and NormalizedConfigManager works (until it's added naively):

from transformers import AutoTokenizer

from optimum.exporters import TasksManager
from optimum.exporters.onnx import main_export
from optimum.onnxruntime import ORTModelForCausalLM
from optimum.utils import NormalizedConfigManager

TasksManager._SUPPORTED_MODEL_TYPE["phi3"] = TasksManager._SUPPORTED_MODEL_TYPE["phi"]
NormalizedConfigManager._conf["phi3"] = NormalizedConfigManager._conf["phi"]

# output = "phi3_onnx"
# main_export(
#     model_name_or_path="microsoft/Phi-3-mini-4k-instruct",
#     task="text-generation-with-past",
#     trust_remote_code=True,
#     output=output,
# )

model = ORTModelForCausalLM.from_pretrained("microsoft/Phi-3-mini-4k-instruct", trust_remote_code=True, export=True)
tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
inputs = tokenizer(["Hello, my dog is cute"], return_tensors="pt")

outputs = model.generate(**inputs, max_length=50)
print(tokenizer.batch_decode(outputs, skip_special_tokens=True))