Liger GRPO model testing: Shape mismatch during training
Action Required
Failure to resolve the shape mismatch will prevent the Liger GRPO model from training correctly, delaying model development and potentially impacting performance.
AI Impact Summary
This event describes a technical issue encountered during testing of the Liger GRPO model, specifically related to a shape mismatch during the forward pass within a deepspeed zero3 training setup using Qwen/Qwen2.5-0.5B-Instruct. The issue stems from an incompatibility between the model's expected input shape and the actual shape provided during computation, likely due to differences in data processing or model architecture. This requires investigation and potential code adjustments to resolve the shape mismatch and ensure correct model execution.
Affected Systems
- Date
- 25 May 2025
- Change type
- capability
- Severity
- medium