[Bugfix] fix qwen tokenizer config when converting to nemo format

[chrjxj]

What does this PR do ?

This PR fixes qwen tokenizer config when converting qwen2 from HF to nemo format;

Background: when running Qwen SFT, _build_tokenizer() will return Qwen2Tokenizer or Qwen2TokenizerFast based on cfg.tokenizer .

without this update, _build_tokenizer will use use_fast default value ( False) and create an Qwen2Tokenizer instance (wrapped by NeMo/nemo/collections/common/tokenizers/huggingface/auto_tokenizer.py). Qwen2Tokenizer instance don't have vocab attributes, which results in errors during SFT data loading.

Above changes is tested within nvcr.io/nvidia/nemo:24.07; passed.

Collection: LLM, Qwen2

Changelog

• line 86 to 92, in NeMo/scripts/checkpoint_converters/convert_qwen2_hf_to_nemo.py

Usage

• won't affect usage

GitHub Actions CI

The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.

The GitHub Actions CI will run automatically when the "Run CICD" label is added to the PR.
To re-run CI remove and add the label again.
To run CI on an untrusted fork, a NeMo user with write access must first click "Approve and run".

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?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

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

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

• Related to # (issue)

[ntajbakhsh]

@cuichenx can you review this PR please?

[cuichenx]

LGTM! thanks for the fix

[zhaoyang-star]

Thanks for your bugfix. I tested your code and found that I can not get the right response when do inference using the generated .nemo file. Is the .nemo file right? Could you please have a look? Thanks. @chrjxj @cuichenx

1. HF -> nemo

Convert Qwen2.5-7B to Qwen2.5-7B.nemo. After tar xvf, the Qwen2.5-7B.nemo is as following:

root@inp16075348439349544319-3-1:/mnt/tenant-home_speed/nemo_models# tar xvf Qwen2.5-7B.nemo 
./
./model_config.yaml
./model_weights/
./model_weights/.metadata
./model_weights/__0_0.distcp
./model_weights/__0_1.distcp
./model_weights/common.pt
./model_weights/metadata.json

2. Inference using Qwen2.5-7B.nemo

Then I tried to do inference using Qwen2.5-7B.nemo.

python3 megatron_gpt_eval.py \
            gpt_model_file=Qwen2.5-7B.nemo \
            inference.greedy=True \
            inference.add_BOS=True \
            trainer.devices=1 \
            trainer.num_nodes=1 \
            tensor_model_parallel_size=1 \
            pipeline_model_parallel_size=1 \
            prompts='["who are you?", "What is the captial of China?"]'

The response seems wrong. Part of the output:

[NeMo I 2024-12-04 21:57:07 nlp_overrides:1386] Model MegatronGPTModel was successfully restored from Qwen2.5-7B.nemo.
prompt=========:['who are you?', 'What is the captial of China?']
setting number of microbatches to constant 1
***************************
{'sentences': ['who are you?1000000000000000000000000000000000', 'What is the captial of China?100000000000000000000000000000'], 'tokens': [['<|im_start|>', 'who', 'Ġare', 'Ġyou', '?', '1', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0'], ['<|im_start|>', 'What', 'Ġis', 'Ġthe', 'Ġcapt', 'ial', 'Ġof', 'ĠChina', '?', '1', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0']], 'logprob': None, 'full_logprob': None, 'token_ids': [[151644, 14623, 525, 498, 30, 16, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15], [151644, 3838, 374, 279, 6427, 530, 315, 5616, 30, 16, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15]], 'offsets': [[0, 0, 3, 7, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45], [0, 0, 4, 7, 11, 16, 19, 22, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58]]}
***************************
[NeMo I 2024-12-04 21:57:14 megatron_gpt_model:1717] Pipeline model parallel rank: 0, Tensor model parallel rank: 0, Number of model parameters on device: 7.62e+09. Number of precise model parameters on device: 7615616512.
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1,2,3,4,5,6,7]
    
Predicting DataLoader 0:   0%|                                                                                                                                          | 0/1 [00:00<?, ?it/s]setting number of microbatches to constant 1
Predicting DataLoader 0: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:01<00:00,  0.51it/s]
***************************
[{'sentences': ['who are you?1000000000000000000000000000000000', 'What is the captial of China?100000000000000000000000000000'], 'tokens': [['<|im_start|>', 'who', 'Ġare', 'Ġyou', '?', '1', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0'], ['<|im_start|>', 'What', 'Ġis', 'Ġthe', 'Ġcapt', 'ial', 'Ġof', 'ĠChina', '?', '1', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0']], 'logprob': None, 'full_logprob': None, 'token_ids': [[151644, 14623, 525, 498, 30, 16, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15], [151644, 3838, 374, 279, 6427, 530, 315, 5616, 30, 16, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15]], 'offsets': [[0, 0, 3, 7, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45], [0, 0, 4, 7, 11, 16, 19, 22, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58]]}]
***************************

[github-actions]

This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.

[snowmanwwg]

So what is preventing this to be merged? @cuichenx @chrjxj

[github-actions]

This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.