fix: avoid duplicate gradient norm reduction in pure DDP

[taking-lying-flat]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Pure DDP backward synchronizes complete gradients on every rank, but passing _dp_group into gradient clipping sums their squared norms again. With two ranks, a normalized gradient of 0.8 is reported as 1.131371 and incorrectly clipped to 0.707106 when max_grad_norm=1.

Use the existing local clipping path when the model is wrapped in DistributedDataParallel and its mesh contains only data parallelism. Keep the DP group for token-count gathering and retain gradient normalization, scaler handling, and training-state updates. Other parallel configurations keep their existing norm-reduction group.

Experiment results

Local validation with PyTorch 2.13.0 and CPU/Gloo, using an out-of-tree harness with real DDP backward and the TransformersModel.clip_grad_norm method body:

• Reproduced the original two-rank error against the base commit.
• Passed 36 DDP cases across 1, 2, and 4 ranks, covering L1/L2/infinity norms, gradients above and below the clipping threshold, unequal token counts, gradient accumulation, and SGD updates.
• Verified skipped accumulation steps, explicit Tensor-shard reductions, DTensor shard norms/clipping, and preservation of hybrid/EP group arguments.
• uvx pre-commit run --all-files passed.

For the two-rank 0.8 case, the corrected norm and clipped gradient are both 0.800000.