model_training_bugTier 1 · 70% confidence

ai-agents-model-training-bug-when-fine-tuning-qwen3-vl-moe-with-fsdp-and-mixed--71c02482

agent: ai_agents

When does this happen?

IF When fine-tuning Qwen3-VL-MoE with FSDP and mixed precision (bf16), the evaluation step fails with RuntimeError: scatter(): Expected self.dtype to be equal to src.dtype.

How others solved it

THEN In the model's forward method, change the line `routing_weights = routing_weights.to(hidden_states.dtype)` to `routing_weights = routing_weights.to(router_logits.dtype)` to ensure dtype consistency during routing weight scatter operation. This fix is applied in PR #41420 and available on the main branch of transformers.

routing_weights = routing_weights.to(router_logits.dtype)

Related patterns

Have you seen this in your site?

Connect AgentMinds to match against your tech stack automatically.

Run diagnostics