Llama 3 streaming repeats the previous request's first token

TL;DR

Upgrade past 0.25.2 and the repeated prefix tokens go away. The client-side fix shipped in replicate-python 0.25.2.

Steps

1. If your streamed Llama 3 responses start by echoing the first tokens of previous turns, you are on a client version with the streaming bug, not a model problem. The maintainer reproduced it in the client alone after ruling out stop sequences, and the fix is in replicate-python 0.25.2. Upgrade past 0.25.2 and the repeated prefix tokens go away. If you still see it, check the dashboard predictions to confirm whether the repetition is in the model output or the client rendering.

When to use

You are seeing this: Streaming meta/meta-llama-3-70b-instruct produced a strange artifact: each new response in a chat started by repeating earlier turns' first tokens, first "Hi", then "I", then "How", growing with the conversation. Use this skill when you run into "Llama 3 streaming repeats the previous request's first token".

When not to use

If your error message or symptom does not match what is described above, this is probably not your fix. Search for your exact error text instead of forcing this one to fit.

Versions

Versions mentioned in the source: replicate-python 0.25.2. If you are on something much newer or older, the details may have shifted.

Why this happens

The original report does not dig into a root cause. It documents the symptom and the fix that resolved it.

Question asked in the thread: Streaming meta/meta-llama-3-70b-instruct produced a strange artifact: each new response in a chat started by repeating earlier turns' first tokens, first "Hi", then "I", then "How", growing with the conversation. Maintainer mattt first suspected the model's stop sequences and pushed a new model build, then reproduced the issue in the client in isolation. The client-side fix shipped in replicate-python 0.25.2.

Verified answer: If your streamed Llama 3 responses start by echoing the first tokens of previous turns, you are on a client version with the streaming bug, not a model problem. The maintainer reproduced it in the client alone after ruling out stop sequences, and the fix is in replicate-python 0.25.2. Upgrade past 0.25.2 and the repeated prefix tokens go away. If you still see it, check the dashboard predictions to confirm whether the repetition is in the model output or the client rendering. Source: https://github.com/replicate/replicate-python/issues/287

Thread: https://vectle.com/threads/thrfGDqRWh0CuLeam-ZeAdMw