Improve .llm code coverage

[xnuohz]

Issue

Close #10514, #10529

CodeCov

Before

After

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 87.53%. Comparing base (c211214) to head (a905acb).
⚠️ Report is 145 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #10516      +/-   ##
==========================================
+ Coverage   86.11%   87.53%   +1.41%     
==========================================
  Files         496      510      +14     
  Lines       33655    35960    +2305     
==========================================
+ Hits        28981    31476    +2495     
+ Misses       4674     4484     -190     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[xnuohz]

@puririshi98 @akihironitta .llm test coverage has been successfully uploaded to codecov. ready for review and merge.

[puririshi98]

lgtm, just plz address my one comment

[puririshi98]

1 more

[xnuohz]

@puririshi98 I forgot to force reload MoleculeDataset when llm switched to Qwen, so the text in the dataset was generated by TinyLlama but trained it with Qwen.
See the training log from scratch below.

master branch from scratch

Setting up 'TinyLlama/TinyLlama-1.1B-Chat-v0.1' with configuration: {'revision': 'main', 'max_memory': {0: '21GiB'}, 'low_cpu_mem_usage': True, 'device_map': 'auto', 'torch_dtype': torch.bfloat16}
Some weights of RobertaModel were not initialized from the model checkpoint at DeepChem/ChemBERTa-77M-MTR and are newly initialized: ['pooler.dense.bias', 'pooler.dense.weight']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
Total Preparation Time: 3359.429658s
Training beginning...
Epoch: 1|3:   0%|                                                                                      | 0/1682 [00:00<?, ?it/s]/workspace/pytorch_geometric/torch_geometric/llm/models/molecule_gpt.py:158: UserWarning: HuggingFace model TinyLlama/TinyLlama-1.1B-Chat-v0.1 is not using a chat template, using Llama 2 style prompting. Please consider using a more recent model and initialize the LLM with `sys_prompt`.
  ) = self.llm._get_embeds(instructions, additional_text_context, xs,
Epoch: 1|3: 100%|███████████████████████████████████████████████████████████████████████████| 1682/1682 [04:12<00:00,  6.67it/s]
Epoch: 1|3, Train loss: 1.020544, Val loss: 0.994869
Epoch: 2|3: 100%|███████████████████████████████████████████████████████████████████████████| 1682/1682 [04:11<00:00,  6.69it/s]
Epoch: 2|3, Train loss: 0.816425, Val loss: 0.960044
Epoch: 3|3: 100%|███████████████████████████████████████████████████████████████████████████| 1682/1682 [04:10<00:00,  6.70it/s]
Epoch: 3|3, Train loss: 0.795707, Val loss: 0.943275
Total Training Time: 778.760076s
Test loss: 0.957731
Total Time: 4144.875072s

this pr from scratch

Setting up 'Qwen/Qwen3-0.6B' with configuration: {'revision': 'main', 'max_memory': {0: '22GiB'}, 'low_cpu_mem_usage': True, 'device_map': 'auto', 'torch_dtype': torch.bfloat16}
Some weights of RobertaModel were not initialized from the model checkpoint at DeepChem/ChemBERTa-77M-MTR and are newly initialized: ['pooler.dense.bias', 'pooler.dense.weight']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
Total Preparation Time: 5493.621828s
Training beginning...
Epoch: 1|3: 100%|███████████████████████████████████████████████████████████████████████████| 1682/1682 [03:21<00:00,  8.37it/s]
Epoch: 1|3, Train loss: 0.581263, Val loss: 0.575342
Epoch: 2|3: 100%|███████████████████████████████████████████████████████████████████████████| 1682/1682 [03:20<00:00,  8.38it/s]
Epoch: 2|3, Train loss: 0.435040, Val loss: 0.553491
Epoch: 3|3: 100%|███████████████████████████████████████████████████████████████████████████| 1682/1682 [03:20<00:00,  8.38it/s]
Epoch: 3|3, Train loss: 0.405096, Val loss: 0.549858
Total Training Time: 622.549687s
Test loss: 0.585572
Total Time: 6122.272548s

[puririshi98]

okay since the loss has dropped 2x on the test branch i think this is safe to merge just need CI to be green

[xnuohz]

@puririshi98 ready to merge. the same nightly PyTorch CI error as the master branch.

[puririshi98]

@puririshi98 ready to merge. the same nightly PyTorch CI error as the master branch.

hopefully @akihironitta can help w this

[xnuohz]

@puririshi98 all CI passed

[xnuohz]

Hi @puririshi98 what would be a good time to revisit and merge this PR?

[akihironitta]

Could we only run rag tests in the workflow being deleted? I'd like to avoid this PR from almost tripling the CI time.

[xnuohz]

@akihironitta rag test can't sync to codecov, can you help with this?