openadapt-ml 0.14.1: multiple broken internal imports prevent serve, train, and demo commands

[ultimus247]

Environment

• OS: Windows 2019 Server
• Python: 3.14.3
• openadapt-ml: 0.14.1 (also reproduced on main branch)
• GPU: None (CPU mode)

Description

Several commands fail with misleading "openadapt-ml not installed" errors due to internal import mismatches within openadapt-ml. The except ImportError blocks in cli.py mask the real errors.

Bugs Found

1. cli.py imports non-existent serve_dashboard

openadapt/cli.py serve function imports serve_dashboard from openadapt_ml.cloud.local, but the actual function is cmd_serve.

2. cmd_serve imports non-existent functions from trainer.py

cloud/local.py:cmd_serve() imports regenerate_local_dashboard and update_current_symlink_to_latest from openadapt_ml.training.trainer — neither exists. Only get_current_job_directory is available.

3. train.py uses wrong parameter name for capture_to_episode

scripts/train.py line ~41 calls capture_to_episode(capture_path, goal=goal) but the actual signature is capture_to_episode(capture_path, instruction=None).

4. demo_policy.py imports non-existent generate_synthetic_sessions

scripts/demo_policy.py imports generate_synthetic_sessions from openadapt_ml.ingest.synthetic, but this function doesn't exist.

5. Missing configs/ directory

The configs/ directory (containing qwen3vl_capture.yaml, qwen3vl_capture_4bit.yaml, etc.) is not included in the pip package, causing training to fail with "Config not found".

Steps to Reproduce

pip install openadapt[all]
openadapt serve --port 8080    # "Error: openadapt-ml not installed"
openadapt train start --capture MyCapture --model qwen3vl-2b  # Same error

Workarounds

Detailed workarounds involve:

• Installing from main branch: pip install --force-reinstall git+https://github.com/OpenAdaptAI/openadapt-ml.git@main
• Monkey-patching missing regenerate_local_dashboard and update_current_symlink_to_latest functions for serve command
• Patching goal=goal to instruction=goal in scripts/train.py
• Manually creating config YAML files with model, lora, and training settings
• Installing missing dependency: pip install datasets

[ultimus247]

Additional bugs found during continued testing

After filing this issue, we discovered three more bugs while getting training to work:

6. Missing package dependencies

datasets, trl, and huggingface_hub are required for training but are not listed in openadapt-ml's package requirements. Users must manually install them:

pip install datasets trl huggingface_hub

7. CPU training fails — missing use_cpu in SFTConfig

_run_sft_training() in training/trl_trainer.py does not pass use_cpu=True, bf16=False, or fp16=False to SFTConfig in the non-Unsloth (else) branch. On CPU-only machines this causes:

ValueError: Your setup doesn't support bf16/gpu. You need to assign use_cpu if you want to train the model on CPU.

Fix: add use_cpu=True, bf16=False, fp16=False to the SFTConfig() call in the else branch (~line 412).

8. Default model name doesn't exist on HuggingFace

The hardcoded default config path references Qwen/Qwen2.5-VL-2B-Instruct (in cloud/local.py), but this model does not exist on HuggingFace. The correct model ID is Qwen/Qwen2-VL-2B-Instruct (without the .5). The smallest available Qwen VL models are:

• Qwen/Qwen2-VL-2B-Instruct
• Qwen/Qwen2.5-VL-3B-Instruct

[adamczerw]

Hi @ultimus247, after dealing with those bugs were you able to successfully work with this package?

[ultimus247]

Yes I finally got it working using claude code for a few hours to massage everything into place.

…

On Tue, Apr 21, 2026 at 11:19 AM Adam ***@***.***> wrote: *adamczerw* left a comment (OpenAdaptAI/OpenAdapt#999) <#999 (comment)> Hi @ultimus247 <https://github.com/ultimus247>, after dealing with those bugs were you able to successfully work with this package? — Reply to this email directly, view it on GitHub <#999?email_source=notifications&email_token=AAT6SGEGLDY5PZLN3G3FIN34W6URLA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMRZGA2DQNZUGMY2M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4290487431>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAT6SGGSHTHJSBOXS4QEUKD4W6URLAVCNFSM6AAAAACWW6FO2CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DEOJQGQ4DONBTGE> . You are receiving this because you were mentioned.Message ID: ***@***.***>

-- David Thomson