[Bug]: Anima LoRA training does not progress past caching on Apple Silicon / MPS

[Atsushi888]

What happened?

Description

Anima LoRA training does not progress past the caching stage on my Apple Silicon Mac.

OneTrainer successfully starts the training process and enumerates the dataset, but the training does not proceed beyond the caching stage.

The UI remains at:

• step: 0%
• epoch: 0%

and the terminal does not show further training progress.

I have attached a screen recording showing the behavior, together with screenshots of my OneTrainer settings.

---

Environment

• Hardware: MacBook Air (Apple Silicon)
• Memory: 32 GB
• OS: macOS
• Training backend tested: MPS
• Model type: Anima
• Training method: LoRA
• Base model: circlestone-labs/Anima-Base-v1.0-Diffusers

---

Steps to reproduce

1. Launch OneTrainer.
2. Select Anima as the model type.
3. Load circlestone-labs/Anima-Base-v1.0-Diffusers.
4. Configure a LoRA training concept.
5. Start training.
6. Dataset enumeration completes.
7. OneTrainer enters the caching stage.
8. Training remains at 0% and does not proceed.

The attached video shows OneTrainer and the terminal simultaneously so the UI state and console output can be compared.

---

Actual behavior

Dataset enumeration completes successfully.

For example, the terminal reaches messages such as:

enumerating sample paths: 100%

OneTrainer then enters:

Starting epoch/caching

However, progress remains at:

step: 0%
epoch: 0%

and training does not begin.

The application itself does not immediately crash; the training process simply stops making progress at this stage.

---

Previous related error

During earlier attempts while troubleshooting this issue, I also encountered the following exception in Anima text encoding:

UnboundLocalError: cannot access local variable 't5_ids' where it is not associated with a value

The traceback pointed to the Anima encode_text() path and the use of t5_ids.

I am including this because it may be related to the current caching problem, although I cannot confirm that both symptoms have the same cause.

---

What I tested

I tried changing device, dtype, and offloading-related settings while investigating the issue.

In particular, I tested the Apple MPS path and also tried CPU execution as part of isolating the problem.

The dataset itself is detected successfully, and sample enumeration completes.

The failure/stall occurs later, during or immediately after caching/text processing.

Because of this, I suspect the issue may involve the Anima text-encoding/caching path on Apple Silicon, possibly involving MPS compatibility.

This is only a hypothesis; I am attaching the complete settings and reproduction video so that the actual cause can be investigated.

---

Expected behavior

After dataset enumeration and caching, OneTrainer should begin LoRA training and the step/epoch counters should advance.

If the selected Anima configuration is unsupported on Apple Silicon/MPS, OneTrainer should ideally report an explicit error instead of remaining indefinitely at 0%.

---

Attachments

I have included:

• Screen recording of the reproduction
• General settings
• Model settings
• Data settings
• Concept settings
• Training settings
• LoRA settings
• Other OneTrainer configuration screenshots

The screen recording shows both OneTrainer and the terminal during the reproduction.

Please let me know if a config export, complete terminal log, or additional diagnostic information would be useful.

debug_report.log

rec_OneTrainer.mp4

What did you expect would happen?

No response

Relevant log output

Generate and upload debug_report.log

No response

[dxqb]

During earlier attempts while troubleshooting this issue, I also encountered the following exception in Anima text encoding:

UnboundLocalError: cannot access local variable 't5_ids' where it is not associated with a value

The traceback pointed to the Anima encode_text() path and the use of t5_ids.

can you please create a config in the UI, then reproduce the problem, and then post your config here?
I am asking this because I think this code path cannot be reached from a normal config. The toggle that would have to be switched is not visible in the UI.

and since you are using an AI agent to report that bug, my theory is that your AI agent has also manually edited your config

[Atsushi888]

Thanks for the response.
I found the configuration file used for this test.
This configuration was created and saved through the OneTrainer UI and has not been manually edited.
I have attached it here as seika_uniform_test3.
Please let me know if you need any additional files or another reproduction test.

seika_uniform_test3.json

[O-J1]

Thanks for the response. I found the configuration file used for this test. This configuration was created and saved through the OneTrainer UI and has not been manually edited. I have attached it here as seika_uniform_test3. Please let me know if you need any additional files or another reproduction test.

seika_uniform_test3.json

Please for completions sake humour dxqb and recreate the issue by:

1. Load completely blank preset (load the default anima prest)
2. Select basic settings that trigger, dont add anything else
3. Then hit debug and upload the zip file it produces

[dxqb]

Thanks for the response. I found the configuration file used for this test. This configuration was created and saved through the OneTrainer UI and has not been manually edited. I have attached it here as seika_uniform_test3. Please let me know if you need any additional files or another reproduction test.

I was specifically referring to:

During earlier attempts while troubleshooting this issue, I also encountered the following exception in Anima text encoding:
UnboundLocalError: cannot access local variable 't5_ids' where it is not associated with a value
The traceback pointed to the Anima encode_text() path and the use of t5_ids.

which your agent called a "previous issue" but didn't explain how it reached that point.

regarding your config:

seika_uniform_test3.json

you are training on cpu:

    "train_device": "cpu",
    "temp_device": "cpu",

this can be slow :)

[Atsushi888]

Thanks for the response. I found the configuration file used for this test. This configuration was created and saved through the OneTrainer UI and has not been manually edited. I have attached it here as seika_uniform_test3. Please let me know if you need any additional files or another reproduction test.
seika_uniform_test3.json

Please for completions sake humour dxqb and recreate the issue by:

1. Load completely blank preset (load the default anima prest)
2. Select basic settings that trigger, dont add anything else
3. Then hit debug and upload the zip file it produces

Thanks. I reproduced the issue again using the default Anima LoRA preset as requested.
I started from the default Anima LoRA preset and used MPS as the training device. When I started training, OneTrainer failed with the following error:
TypeError: Trying to convert Float8_e4m3fn to the MPS backend but it does not have support for that dtype.
I then clicked the Debug button immediately after reproducing the error.
I have attached:
seika_uniform_test4.json — the configuration saved from this reproduction test
OneTrainer_debug_report.zip — the debug package generated by the Debug button
This configuration was created through the OneTrainer UI and was not manually edited.
Please let me know if you would like me to perform another reproduction test with different settings.

seika_uniform_test4.json

OneTrainer_debug_report.zip

[dxqb]

now your config actually uses mps, not CPU - good.
But Apple doesn't support float8. you have to use bf16.
And Apple doesn't support "Compile transformer blocks". Turn it off.

I'm turning this into a discussion because I don't see a bug. MPS/Apple has been reported working, but we don't have good presets for it because not many users.

[Atsushi888]

Thanks! I changed the settings as you suggested:

• MPS is enabled
• Transformer Data Type: bfloat16
• Text Encoder Data Type: bfloat16
• Train Data Type: bfloat16
• Fallback Train Data Type: bfloat16
• "Compile transformer blocks" is disabled

Training now starts using MPS and caching completes successfully, but it crashes immediately after that.

The final error is:

RuntimeError: Placeholder storage has not been allocated on MPS device!

The traceback appears to point to the Qwen3 text encoder during the embedding operation:

transformers/models/qwen3/modeling_qwen3.py
→ self.embed_tokens(input_ids)
→ torch.nn.functional.embedding

I also see this warning:

Warning: Mixed precision training is untested on device type 'mps'.

So it looks like MPS itself is now being used correctly, but something in the Qwen3 text encoder may not be allocated to the MPS device.

Is there another setting I should change for the text encoder on Apple/MPS?

I'm using Anima LoRA training on Apple Silicon.