Gemma3 27B crashes on 8 nodes with Sequence Parallel

## 8 nodes full w/ FSDP (v1 and v2 dtensor backends)

Gemma3 27B will crash with 8 nodes and sequence_parallel=True using this as a base config

https://github.com/NVIDIA-NeMo/RL/blob/ae89e126bedd57020dcb3b5173e393bac287d634/examples/configs/recipes/llm/grpo-gemma3-27b-it-16n8g-fsdp2tp8sp-actckpt-long.yaml

repro: `uv run tests/test_suites/llm/grpo-gemma3-27b-it-16n8g-fsdp2tp8sp-actckpt-long.sh cluster.num_nodes=8`
###
```
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1805, in inner
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/code_snapshots/grpo-gemma3-27b-it-16n8g-fsdp2tp8sp-actckpt-long/3rdparty/Automodel-workspace/Automodel/nemo_automodel/components/_transformers/auto_model.py", line 80, in wrapper
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/transformers/models/gemma3/modeling_gemma3.py", line 1085, in forward
^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/transformers/utils/generic.py", line 943, in wrapper
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/transformers/models/gemma3/modeling_gemma3.py", line 939, in forward
^^^^^^^^^^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/transformers/utils/generic.py", line 943, in wrapper
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/transformers/models/gemma3/modeling_gemma3.py", line 566, in forward
^^^^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/transformers/modeling_layers.py", line 83, in __call__
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
^^^^^^^^^^^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/transformers/models/gemma3/modeling_gemma3.py", line 385, in forward
hidden_states, self_attn_weights = self.self_attn(
^^^^^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/transformers/models/gemma3/modeling_gemma3.py", line 330, in forward
attn_output, attn_weights = attention_interface(
^^^^^^^^^^^^^^^^^^^^
File "/opt/ray_venvs/nemo_rl.models.policy.dtensor_policy_worker_v2.DTensorPolicyWorkerV2/lib/python3.12/site-packages/transformers/integrations/sdpa_attention.py", line 66, in sdpa_attention_forward
attn_output = torch.nn.functional.scaled_dot_product_attention(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: The expanded size of the tensor (3800) must match the existing size (475) at non-singleton dimension 3. Target sizes: [2, 4, 3800, 3800]. Tensor sizes: [2, 1, 475, 475]
```

https://wandb.ai/nvidia/nemo-rl/runs/8prq9eot

Also filed against Automodel: https://github.com/NVIDIA-NeMo/Automodel/issues/426

WAR is to just set `sequence_parallel=False`