fix: position-ids in qwen3-next

[akoumpa]

What does this PR do ?

torchrun --nproc-per-node 8 nemo_automodel/recipes/llm/benchmark.py --config examples/benchmark/configs/qwen3_next_te_deepep.yaml --model.config.num_hidden_layers 8 --distributed.ep_size 8

What it does: Benchmarks Qwen3-Next-80B-A3B-Instruct with TE + DeepEP, 8 GPUs. Crashes on the very first forward pass.

Traceback:

benchmark.py:246      → _forward_backward_step(...)
train_ft.py:1343      → out = model(**batch)
qwen3_next/model.py:296 → self.model(...)
qwen3_next/model.py:202 → layer(...)
qwen3_next/model.py:77  → self.linear_attn(
    ..., position_ids=position_ids)   ← CRASH

TypeError: Qwen3NextGatedDeltaNet.forward() got an unexpected keyword argument 'position_ids'
Root cause: In nemo_automodel/components/models/qwen3_next/model.py:77, the decoder layer passes position_ids to self.linear_attn(). But Qwen3NextGatedDeltaNet.forward() doesn't accept that parameter — gated delta-nets encode position implicitly and have a different signature than standard attention.

So Qwen3NextGatedDeltaNet.forward() accepts:

(self, hidden_states, cache_params=None, cache_position=None, attention_mask=None)

Changelog

• Add specific line by line info of high level changes in this PR.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?

If you haven't finished some of the above items you can still open "Draft" PR.

Additional Information

• Related to # (issue)

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[akoumpa]

/ok to test 66a2804

[akoumpa]

/ok to test db1f660

[akoumpa]

/ok to test ce0d6ab