Upgrading to python3.13? [testing needed]

[Aaron-Zhao123]

• removes versioning constraint on coco-tb
• end of support for setencepiece
• adapt loading with pytorch==2.6

testing script

print("This is a save and load test")
import chop.passes as passes

from pathlib import Path
from transformers import AutoModelForSequenceClassification
from chop import MaseGraph

model = AutoModelForSequenceClassification.from_pretrained("prajjwal1/bert-tiny")
model.config.problem_type = "single_label_classification"
mg = MaseGraph(model, hf_input_names=["input_ids", "attention_mask", "labels"])

mg, _ = passes.init_metadata_analysis_pass(mg)
mg, _ = passes.add_common_metadata_analysis_pass(mg)
# test export and load
mg.export(f"{Path.home()}/tutorial_1")
new_mg = MaseGraph.from_checkpoint(f"{Path.home()}/tutorial_1")

Huggingface is also a bit problematic, because it seems in the new version, their model.config.problem_type is changed in their upstream, this would cause transforming to fx fail on some old models, for instance, this bert-tiny, the workaround is to reassign that problem type on the fly.

[Aaron-Zhao123]

@firemountain154B Can you do a quick test on the hardware side to make sure Cocotb works? I had a slight worry here -- I thought we locked cocotb to 1.8 for a reason...