[bananighosh]

Description of changes:
On the Neuron SDK 2.33 DLAMI, importing spmd_types before torch triggers the torch_neuronx backend autoload, which loads torch.distributed.fsdp; _fsdp_param_group.py then re-imports the partially initialized spmd_types and fails with "has no attribute 'register_local_autograd_function'".

The Qwen3, DeepSeek-V3 and GPT-OSS recipes import spmd_types at module top, so their config_registry fails to import and ConfigManager reports "Config function ... not found" (e.g. the Qwen3-32B command in docs/dense-tutorials.md).

Import torch in torchtitan_neuron/init.py so it is loaded before any recipe module.

Tested on trn3-dev1.48xlarge, DLAMI ami-0dea3e9e3e3a16119 (torch 2.14.0, torch-neuronx 2.14.3.0.1, spmd_types 0.2.1): the four model config_registry modules import, and the dense-tutorials Qwen3-32B eager command runs 10/10 steps (rc=0).

error:

W1107 21:06:24.372000 2218974 torch/distributed/run.py:982] *****************************************
W1107 21:06:24.372000 2218974 torch/distributed/run.py:982] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
W1107 21:06:24.372000 2218974 torch/distributed/run.py:982] *****************************************
[titan] 2026-11-07 21:06:28,423 - root - INFO - torchtitan version: 0.2.2+neuron.1 (0.0.0 means __version__ is not defined correctly).
Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "/home/ubuntu/nl/exp/torchtitan-neuron/torchtitan/train.py", line 79, in <module>
    main()
  File "/home/ubuntu/nl/exp/torchtitan-neuron/torchtitan/train.py", line 29, in main
    config = config_manager.parse_args()
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/nl/exp/torchtitan-neuron/torchtitan/config/manager.py", line 37, in parse_args
    loaded_config, args = self._load_config(args)
                          ^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/nl/exp/torchtitan-neuron/torchtitan/config/manager.py", line 153, in _load_config
    raise ValueError(
ValueError: Config function 'qwen3_32b_tp4_fsdp16_fullac_eager' not found in torchtitan_neuron.models.qwen3. Available config functions: ['core_model_registry', 'model_registry', 'replace', 'validate_converter_order']
W1107 21:06:29.340000 2218974 torch/distributed/elastic/multiprocessing/api.py:1028] Sending process 2219042 closing signal SIGTERM
W1107 21:06:29.341000 2218974 torch/distributed/elastic/multiprocessing/api.py:1028] Sending process 2219043 closing signal SIGTERM

Issue #, if available: - N/A

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.