pythonpytorchcoremlcoremltools

"RuntimeError: BlobWriter not loaded" error when exporting a PyTorch model to CoreML. How to fix it?

I get a "RuntimeError: BlobWriter not loaded" error when exporting a PyTorch model to CoreML. How to fix it?

Same issue with Python 3.11 and Python 3.10. Same issue with torch 2.3.1 and 2.2.0. Tested on Windows 10.

Export script:

# -*- coding: utf-8 -*-
"""Core ML Export
pip install transformers torch coremltools nltk
"""
import os
from transformers import AutoModelForTokenClassification, AutoTokenizer
import torch
import torch.nn as nn
import nltk
import coremltools as ct

nltk.download('punkt')

# Load the model and tokenizer
model_path = os.path.join('model')
model = AutoModelForTokenClassification.from_pretrained(model_path, local_files_only=True)
tokenizer = AutoTokenizer.from_pretrained(model_path, local_files_only=True)

# Modify the model's forward method to return a tuple
class ModifiedModel(nn.Module):
    def __init__(self, model):
        super(ModifiedModel, self).__init__()
        self.model = model
        self.device = model.device  # Add the device attribute

    def forward(self, input_ids, attention_mask, token_type_ids=None):
        outputs = self.model(input_ids=input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids)
        return outputs.logits


modified_model = ModifiedModel(model)

# Export to Core ML
def convert_to_coreml(model, tokenizer):
    # Define a dummy input for tracing
    dummy_input = tokenizer("A French fan", return_tensors="pt")
    dummy_input = {k: v.to(model.device) for k, v in dummy_input.items()}

    # Trace the model with the dummy input
    traced_model = torch.jit.trace(model, (
    dummy_input['input_ids'], dummy_input['attention_mask'], dummy_input.get('token_type_ids')))

    # Convert to Core ML
    inputs = [
        ct.TensorType(name="input_ids", shape=dummy_input['input_ids'].shape),
        ct.TensorType(name="attention_mask", shape=dummy_input['attention_mask'].shape)
    ]
    if 'token_type_ids' in dummy_input:
        inputs.append(ct.TensorType(name="token_type_ids", shape=dummy_input['token_type_ids'].shape))

    mlmodel = ct.convert(traced_model, inputs=inputs)

    # Save the Core ML model
    mlmodel.save("model.mlmodel")
    print("Model exported to Core ML successfully")

convert_to_coreml(modified_model, tokenizer)

Error stack:

C:\Users\dernoncourt\anaconda3\envs\coreml\python.exe C:\Users\dernoncourt\PycharmProjects\coding\export_model_to_coreml6_fopr_SE_q.py 
Failed to load _MLModelProxy: No module named 'coremltools.libcoremlpython'
Fail to import BlobReader from libmilstoragepython. No module named 'coremltools.libmilstoragepython'
Fail to import BlobWriter from libmilstoragepython. No module named 'coremltools.libmilstoragepython'
[nltk_data] Downloading package punkt to
[nltk_data]     C:\Users\dernoncourt\AppData\Roaming\nltk_data...
[nltk_data]   Package punkt is already up-to-date!
C:\Users\dernoncourt\anaconda3\envs\coreml\lib\site-packages\transformers\modeling_utils.py:4565: FutureWarning: `_is_quantized_training_enabled` is going to be deprecated in transformers 4.39.0. Please use `model.hf_quantizer.is_trainable` instead
  warnings.warn(
When both 'convert_to' and 'minimum_deployment_target' not specified, 'convert_to' is set to "mlprogram" and 'minimum_deployment_target' is set to ct.target.iOS15 (which is same as ct.target.macOS12). Note: the model will not run on systems older than iOS15/macOS12/watchOS8/tvOS15. In order to make your model run on older system, please set the 'minimum_deployment_target' to iOS14/iOS13. Details please see the link: https://apple.github.io/coremltools/docs-guides/source/target-conversion-formats.html
Model is not in eval mode. Consider calling '.eval()' on your model prior to conversion
Converting PyTorch Frontend ==> MIL Ops:   0%|          | 0/127 [00:00<?, ? ops/s]Core ML embedding (gather) layer does not support any inputs besides the weights and indices. Those given will be ignored.
Converting PyTorch Frontend ==> MIL Ops:  99%|█████████▉| 126/127 [00:00<00:00, 2043.73 ops/s]
Running MIL frontend_pytorch pipeline: 100%|██████████| 5/5 [00:00<00:00, 212.62 passes/s]
Running MIL default pipeline:  37%|███▋      | 29/78 [00:00<00:00, 289.75 passes/s]C:\Users\dernoncourt\anaconda3\envs\coreml\lib\site-packages\coremltools\converters\mil\mil\ops\defs\iOS15\elementwise_unary.py:894: RuntimeWarning: overflow encountered in cast
  return input_var.val.astype(dtype=string_to_nptype(dtype_val))
Running MIL default pipeline: 100%|██████████| 78/78 [00:00<00:00, 137.56 passes/s]
Running MIL backend_mlprogram pipeline: 100%|██████████| 12/12 [00:00<00:00, 315.01 passes/s]
Traceback (most recent call last):
  File "C:\Users\dernoncourt\PycharmProjects\coding\export_model_to_coreml6_fopr_SE_q.py", line 58, in <module>
    convert_to_coreml(modified_model, tokenizer)
  File "C:\Users\dernoncourt\PycharmProjects\coding\export_model_to_coreml6_fopr_SE_q.py", line 51, in convert_to_coreml
    mlmodel = ct.convert(traced_model, inputs=inputs)
  File "C:\Users\dernoncourt\anaconda3\envs\coreml\lib\site-packages\coremltools\converters\_converters_entry.py", line 581, in convert
    mlmodel = mil_convert(
  File "C:\Users\dernoncourt\anaconda3\envs\coreml\lib\site-packages\coremltools\converters\mil\converter.py", line 188, in mil_convert
    return _mil_convert(model, convert_from, convert_to, ConverterRegistry, MLModel, compute_units, **kwargs)
  File "C:\Users\dernoncourt\anaconda3\envs\coreml\lib\site-packages\coremltools\converters\mil\converter.py", line 212, in _mil_convert
    proto, mil_program = mil_convert_to_proto(
  File "C:\Users\dernoncourt\anaconda3\envs\coreml\lib\site-packages\coremltools\converters\mil\converter.py", line 307, in mil_convert_to_proto
    out = backend_converter(prog, **kwargs)
  File "C:\Users\dernoncourt\anaconda3\envs\coreml\lib\site-packages\coremltools\converters\mil\converter.py", line 130, in __call__
    return backend_load(*args, **kwargs)
  File "C:\Users\dernoncourt\anaconda3\envs\coreml\lib\site-packages\coremltools\converters\mil\backend\mil\load.py", line 902, in load
    mil_proto = mil_proto_exporter.export(specification_version)
  File "C:\Users\dernoncourt\anaconda3\envs\coreml\lib\site-packages\coremltools\converters\mil\backend\mil\load.py", line 400, in export
    raise RuntimeError("BlobWriter not loaded")
RuntimeError: BlobWriter not loaded

Process finished with exit code 1
Solution

  • The code works fine on Ubuntu 20.04 (tested with Python 3.10 and torch 2.3.1). One just needs to change:

    mlmodel.save("model.mlmodel")

    to

    mlmodel.save("model.mlpackage")

    Note that the error message was different when using mlmodel.save("model.mlmodel"):

      File "/home/dernonco/code/export_model_to_coreml6_fopr_SE_q.py", line 58, in <module>
    convert_to_coreml(modified_model, tokenizer)
  File "/home/dernonco/code/export_model_to_coreml6_fopr_SE_q.py", line 54, in convert_to_coreml
    mlmodel.save("model.mlmodel")
  File "/opt/conda/envs/py310/lib/python3.10/site-packages/coremltools/models/model.py", line 515, in save
    raise Exception(
Exception: For an ML Program, extension must be .mlpackage (not .mlmodel). Please see https://coremltools.readme.io/docs/unified-conversion-api#target-conversion-formats to see the difference between neuralnetwork and mlprogram model types.

    As for Windows, I guess coremltools doesn't work well there, or need some special treatment.