The model tends to generate a lot of `<fim_pad>` tokens when using the FIM mode for generation.

[Grey4sh]

[cyente]

you may need to add fim releated tokens into eos_tokens_ids

[cyente]

99 similar issue

[Grey4sh]

I still enconter the same problem after modified config.json and generation_config.json with TGI-2.3.0

cat generation_config.json
{
  "bos_token_id": 151643,
  "eos_token_id": [
    151643,
    151662
  ],

  "max_new_tokens": 2048,
  "transformers_version": "4.45.0.dev0"
}

{
  "architectures": [
    "Qwen2ForCausalLM"
  ],
  "attention_dropout": 0.0,
  "bos_token_id": 151643,
  "eos_token_id": [
    151643,
    151662
  ],

[hcbr]

I still enconter the same problem after modified config.json and generation_config.json with TGI-2.3.0

cat generation_config.json
{
  "bos_token_id": 151643,
  "eos_token_id": [
    151643,
    151662
  ],

  "max_new_tokens": 2048,
  "transformers_version": "4.45.0.dev0"
}

{
  "architectures": [
    "Qwen2ForCausalLM"
  ],
  "attention_dropout": 0.0,
  "bos_token_id": 151643,
  "eos_token_id": [
    151643,
    151662
  ],

I think you need to add "<|fim_pad|>" to special token.

[hcbr]

you can try like this.

[Grey4sh]

you can try like this.

Thank you for your detailed response. However, my current requirement is to deploy server-side inference, and I have tried modifying the local model's config, but it hasn't worked.

[cyente]

This issue is being closed due to no response for more than 1 day.