Avoid appending to external data when running onnx_save

[dbogunowicz]

When running onnx_save(model) and the model is >2Gb, its initializers, and parameters are saved to one external data file model.data. However, when running onnx_save(...) multiple times, by default we are not strictly overwriting the old model.data file (the expected behavior), but we are overwriting the previously seen parameters and appending the unseen ones to the file.

This is why, when exporting a QAT model, the model.data keep growing very large.
It eventually contains:

• contents of its fp32 counterpart external tensors (because we save an fp32 using save_onnx as an intermediate step: https://github.com/neuralmagic/sparseml/blob/main/src/sparseml/pytorch/utils/exporter.py#L574
• AND THEN also call save_onnx appends quant/sparse external tensors to model.data: https://github.com/neuralmagic/sparseml/blob/main/src/sparseml/pytorch/utils/exporter.py#L587

The problem of exploding model.data does not concern us in the context of FP32 model, because onnx_save is being called only once.

Testing:

import onnx
import os
from sparsezoo.utils.onnx import save_onnx
model = onnx.load("/network/damian/sparsegpt_webinar_fp32/sparsegpt_1.3b/model.onnx")
# model.ByteSize() -> 5675170625 which is less then 5.6G
save_onnx(model, model_path = "tmp.onnx")
# The ONNX model is too large to be saved as a single protobuf. Saving with external data: model.data
model = onnx.load("tmp.onnx")
save_onnx(model, model_path = "tmp.onnx")
# The ONNX model is too large to be saved as a single protobuf. Saving with external data: model.data
# Attempting to save external data for a model: tmp.onnx to a directory: that already contains external data file: model.data. The external data file will be overwritten.
# du -sh model.data -> 5.3G

This has been also successfully tested with the sparseml.transformers.export pathway.

(venv) damian@nmgpuserver1:~/sparsegpt@opt-2.7b@c4@opt2.7b.W8A8linear.fp16_matmul_resi_relu_ln_emb_head@SQ1@PTQ1@ID21988/deployment$ du -sh *
4.0K	config.json
448K	merges.txt
**3.4G	model.data**
996K	model.onnx
4.0K	special_tokens_map.json
4.0K	tokenizer_config.json
976K	vocab.json