Closed medduk9871 closed 6 months ago
- predicted_indices[0] => torch.Size([n])
The format is correct but we need the logits and not the most likely token. Can you please output "next_token_probs" instead of torch.max..?
is it enough just for next 1 token?
Because this is outputs['logits'] shape
torch.Size([1, 636, 32128])
- predicted_indices[0] => torch.Size([n])
The format is correct but we need the logits and not the most likely token. Can you please output "next_token_probs" instead of torch.max..?
is it enough just for next 1 token? Because this is outputs['logits'] shape torch.Size([1, 636, 32128])
Then I just misunderstood the torch.max method. As long as you have the logits for the new token it's what specdec needs :)
- predicted_indices[0] => torch.Size([n])
The format is correct but we need the logits and not the most likely token. Can you please output "next_token_probs" instead of torch.max..?
is it enough just for next 1 token? Because this is outputs['logits'] shape torch.Size([1, 636, 32128])
Then I just misunderstood the torch.max method. As long as you have the logits for the new token it's what specdec needs :)
- predicted_indices[0] => torch.Size([n])
The format is correct but we need the logits and not the most likely token. Can you please output "next_token_probs" instead of torch.max..?
is it enough just for next 1 token? Because this is outputs['logits'] shape torch.Size([1, 636, 32128])
Then I just misunderstood the torch.max method. As long as you have the logits for the new token it's what specdec needs :)
you understood well, the info in the previous comment is not related with max, So I should revise the output. My question is if the output should be just logits for 1 token or something like this tensor torch.Size([1, 636, 32128])?