[v3] NoBadWordsLogitsProcessor breaks when the size of bad_word_ids exceeds the size of the input ids

[tarekziade]

System Info

v3

Environment/Platform

• Website/web-app
• Browser extension
• Server-side (e.g., Node.js, Deno, Bun)
• Desktop app (e.g., Electron)
• Other (e.g., VSCode extension)

Description

this block breaks for me with mozilla/distilvit where we have a long list of bad words

                for (let i = 1; i <= bad_word_ids.length - 1 && bad_word_ids.length < input_ids[i].length; ++i) {

                    // NOTE: We use != instead of !== to compare bigint and number
                    // @ts-ignore
                    if (bad_word_ids.at(-i - 1) != input_ids[i].at(-i)) {
                        // We have found a mismatch
                        mark = false;
                        break;
                    }
                }

with:

Uncaught (in promise) TypeError: can't access property "length", input_ids[i] is undefined

my input_ids has just one entry when I use the regular pipeline function with dtype=q4 and devide=wasm
The code is a bit confusing because i is reused in both the inner and outer loop.

I fixed it by changing the inner loop with :

// using `j` to avoid confusion with the outer loop
for (let j = 1; j <= bad_word_ids.length - 1; ++j) {
    // Ensure input_ids[i] is defined and has sufficient length before accessing
    if (input_ids[i] && bad_word_ids.length < input_ids[i].length) {
        // NOTE: We use != instead of !== to compare bigint and number
        if (bad_word_ids.at(-j - 1) != input_ids[i].at(-j)) {
            // We have found a mismatch
            mark = false;
            break;
        }
    } else {
        mark = false;
        break;
    }
}

I am not sure if that is the correct fix. Happy to do a PR

Reproduction

.

[xenova]

Whoops my bad - a silly issue of reusing variables (input_ids[i].at(-i) should have been a dead giveaway). Probably some bad replace all 🤷

Fixed by d21c87c.

Let me know if you face any other issues 👍