[Question] Must GLM-5 be converted to dist format offline? (AutoBridge registry mismatch)

Your Question

Hi,

We are trying to train GLM-5 (DeepseekV32ForCausalLM) using --megatron-to-hf-mode bridge to load HF safetensors directly. However, it crashes with: ValueError: Model architecture 'DeepseekV32ForCausalLM' is not yet supported.

It seems slime_plugins/mbridge/deepseek_v32.py registers the bridge using its own @register_model, but model_provider.py relies strictly on Megatron's internal MegatronModelBridge.REGISTRY, causing a disconnect.

We temporarily bypassed this by monkey-patching it in model_provider.py:

from slime_plugins.mbridge.deepseek_v32 import DeepseekV32Bridge
MegatronModelBridge.register_bridge(
source="DeepseekV32ForCausalLM", target=GPTModel
)(DeepseekV32Bridge)

Is online bridge mode supposed to be supported for GLM-5, and this is just a missing registry wrapper?
Or is the official workflow to strictly use mbridge to convert HF weights to a Megatron distributed (dist) checkpoint offline, then pass it via --load to bypass AutoBridge altogether?

What I've Tried

mbridge

Environment (if relevant)

• slime version:
• Python version:
• PyTorch version:
• CUDA/ROCm version:
• GPU type and count:
• OS:

Additional Context

No response

Pre-submission Checklist

• I have read the CONTRIBUTING.md and understand the collaboration scope.
• I have read the documentation and FAQ and my question is not answered there.
• I have searched for existing issues and my question has not been asked before.

Start by reproducing the DeepseekV32ForCausalLM error in model_provider.py, then compare the registration in slime_plugins/mbridge/deepseek_v32.py with MegatronModelBridge.REGISTRY and register_bridge. Determine whether online bridge loading is intended for GLM-5 or whether offline dist conversion is required; the issue is resolved when the supported workflow or missing registry integration is established.