naver-hyperclovax

/

HyperCLOVAX-SEED-Text-Instruct-0.5B

like 81

Follow

HyperCLOVA X 903

Summary

generation_config.json is not included in this repository, which causes
GenerationConfig.from_pretrained() to fail with a file-not-found error.
While model.generate() still works via fallback from the model config (model.config),
this breaks explicit generation config loading workflows and some serving frameworks.

Reproduction

from transformers import GenerationConfig
config = GenerationConfig.from_pretrained(
    "naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-0.5B"
)
# Error: does not appear to have a file named generation_config.json

Environment

• transformers: latest (installed from main branch)
• torch: 2.10.0
• Platform: macOS, Apple M4 Pro (MPS)

Impact

1. GenerationConfig.from_pretrained() raises an error
2. Workflows relying on explicit GenerationConfig.from_pretrained() load/save cycle break
3. Some serving stacks (e.g., DeepSpeed-MII) may assume generation_config.json exists
4. Inconsistent with other models of similar scale (Qwen2.5-0.5B, Llama-3.2-1B, EXAONE-3.5-2.4B all include it)

Token IDs from tokenizer (for reference)

bos_token_id: 100257  (<|endoftext|>)
eos_token_id: 100275  (<|endofturn|>)
pad_token_id: 100257  (<|endoftext|>)

Suggestion

Would it be possible to add a minimal generation_config.json?
A minimal version would be:

{
  "_from_model_config": true,
  "bos_token_id": 100257,
  "eos_token_id": 100275,
  "pad_token_id": 100257
}

I noticed that the Vision-Instruct-3B repo had generation_config.json removed in a recent commit,
so this may be intentional. If so, could you share the reasoning?

Happy to submit a PR if this would be welcome.

Notes

• model.generate() works fine via model config fallback — this is not a model inference bug
• This applies to the Text-Instruct series; I was unable to verify 1.5B due to gated access