INNER CODE UNIT · Python

next_token_logits

FoundationVision/Liquid · evaluation/app.py:241

            next_token_logits = outputs.logits[:, -1:, :]
            
            if guidance_scale>1:
                cond_logits, uncond_logits = torch.split(next_token_logits, len(next_token_logits) // 2, dim=0) 
                cfg_logits = uncond_logits + (cond_logits - uncond_logits) * guidance_scale
                half_next_token, _ = sample(cfg_logits, **sampling_kwargs)
                pred_tokens.append(half_next_token)
                next_token = torch.cat([half_next_token,half_next_token])


            else:
                next_token, next_prob = sample(next_token_logits, **sampling_kwargs)
                pred_tokens.append(next_token)

            # update generated ids, model inputs, and length for next step
            input_ids = torch.cat([input_ids, next_token], dim=-1)
            model_kwargs = vqllm._update_model_kwargs_for_generation(
                outputs,

View source record →

📰 Research Paper
Loading…
⏳ Fetching content…