Skip to content

[BUG] Do not use generation of hf model in inferencers #920

Open

Description

@Kamichanw

Describe the bug
As I mentioned in this issue, the default value of top_p and temperature is not guaranteed to be 1. Therefore, the code below will get a modified logits, i.e., a distribution processed depending on generation_config from hf end.

if self.use_accelerator:
outputs = self.backend_model.generate(
input_ids=inputs,
pad_token_id=self.tokenizer.pad_token_id,
*args,
**kwargs
)
else:
if self.device == "gpu":
outputs = self.ds_engine.module.generate(
input_ids=inputs,
synced_gpus=True,
pad_token_id=self.tokenizer.pad_token_id,
*args,
**kwargs
)
elif self.device == "cpu":
outputs = self.backend_model.generate(
input_ids=inputs,
synced_gpus=True,
pad_token_id=self.tokenizer.pad_token_id,
*args,
**kwargs
)

Much worse, you applied top_p and temperature again in score_to_prob, resulting unexpected distribution:

for _ in range(num_new_tokens):
pred = self.predict_next_token(model=model, input_ids=sequence, num_new_tokens=1) # predict next one token
prob = self.score_to_prob(pred.scores[0], temperature=temperature)
sampled = self.sample(prob=prob, num_samples=1)
new_tokens.append(sampled)
sequence = torch.cat([sequence, sampled['sampled_token']], dim=1)

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    pendingSomething isn't working

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions