Free cached GPU memory before AR validation to avoid OOM

With LoRA co-training the model carries extra parameters and optimizer
states (LoRA A/B + Adam moments), reducing the headroom available for
the validation forward passes. Call torch.cuda.empty_cache() before
validate_ar() to release unused cached allocations without affecting
any live tensors (parameters, optimizer states, gradients).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Ye Yu <yeyu@nvidia.com>

Author: yeyu-nvidia

examples/speculative_decoding/eagle_utils.py

@@ -260,6 +260,7 @@ def on_step_end(self, args, state, control, **kwargs):
             return control
         if state.global_step % self.ar_validate_steps == 0 and state.global_step > 0:
             print_rank_0("Running AR validation...")
+            torch.cuda.empty_cache()
             try:
                 ars = validate_ar(
                     model=kwargs["model"],