Add get_input_embeddings method to nemo_automodel.components.models.gpt_oss.model.GptOssForCausalLM

[torsli]

Describe the bug

Unlike other component models, GptOssForCausalLM does not have the get_input_embeddings() method and related methods.

Steps/Code to reproduce bug

Call model.get_input_embeddings() on an instance of GptOssForCausalLM.

Expected behavior

Method should exist.

Additional context

I use this method in benchmarking scripts. I can write a workaround in the short term but it would be better to have this for all special models here.

[SwekeR-463]

Is this up for grabs? With some little more context would try to contribute

[akoumpa]

Thank you so much for your interest @SwekeR-463, please go for it!

[torsli]

For context: models like https://github.com/NVIDIA-NeMo/Automodel/blob/main/nemo_automodel/components/models/llama/model.py#L427 have getters and setters for input and output embeddings, but not all models have them. Adding these methods would be a great way to get started contributing!

[SwekeR-463]

but not all models have them

So have to add the embedding part for all models which dont have the embedding part alongside GptOss?

[SwekeR-463]

@akoumpa, @torsli have created the pr by applying the changes to all models missing the embedding utils. Waiting for review.

[akoumpa]

Thanks a lot @SwekeR-463 for the PR & fix!

[qiaochuz-nv]

QA RCCA Analysis

Fix PR: #1288 (+165/-0, 15 model files)
Root Cause: GptOssForCausalLM and 14 other model classes missing get_input_embeddings()/set_input_embeddings() methods.

Scope: Added 2-line accessor methods to 15 model classes: GPT-OSS, DeepSeek V3/V32, GLM4 MoE, Mistral3, NemotronV3, Qwen2, Qwen3 MoE, etc.

Coverage Assessment:

Dimension	Bug Config	Nearest Test	Match
GPT-OSS model	Required	gpt_oss_20b_finetune_test, gpt_oss_20b_peft_test	Yes
get_input_embeddings	API method	tied_weights_keys_unit_test (uses tied weight internals)	Indirect
HF tied weights	Uses method	All models with lm_head tying	Yes

Status: COVERED — get_input_embeddings() is called by HF internals during tied weight handling. Our tied_weights_keys_unit_test.sh tests tied weight migration logic, and GPT-OSS functional tests exercise the full training path including tied weights. If the method were removed, these tests would fail.

Conclusion: No new test needed.