ai-agents-training-configurati-recomputed-tensor-size-mismatch-error-when-using-a-49734b1e

THEN In the model constructor, set use_cache=False when either gradient_checkpointing or fsdp_config.activation_checkpointing is enabled. For example, replace `use_cache=not gradient_checkpointing` with `use_cache=not (gradient_checkpointing or fsdp_config.activation_checkpointing)`. This prevents the cache from interfering with activation checkpointing under FSDP.

```python
model_kwargs = dict(
    attn_implementation=sft_config.attn_implementation,
    torch_dtype=sft_config.torch_dtype,
    use_cache=not (sft_config.gradient_checkpointing or sft_config.fsdp_config.activation_checkpointing)
)
model = AutoModelForCausalLM.from_pretrained(sft_config.model_name_or_path, **model_kwargs)
```