Vulkan lyrics issue (on Intel iGPU)

[StefanoBalocco]

I installed the Vulkan backend on two devices:

1. AMD Ryzen 5 2400G
2. Intel Core Ultra 7 258V

On device 1 I generated a song. Then I take all the same settings (seed included) and use it con device 2.
I generate the same song but some words are missing. Is like there was skipped.

Instead of:
Mi hai legato
con una cintura di sicurezza
hai detto: "È per il tuo bene"
Hai alzato la mano

It say:
Mi hai legato
con una cintura di sicurezza
hai detto: "È per il tuo
la mano

This happens all over the song.

To do a better test, I used the same url changing the backend via an ssh tunnel (so I have the same local storage) and tested also clicking on "edit prompt" of the correct song. Same result.

[StefanoBalocco]

Disabling flash attention (--no-fa) on the Intel iGPU seems to resolve the problem with the missing words

[ServeurpersoCom]

Disabling flash attention (--no-fa) on the Intel iGPU seems to resolve the problem with the missing words

This looks like a flash attention correctness issue in the Intel Vulkan backend (Xe2 / Lunar Lake) rather than an acestep.cpp bug, since the same seed and settings render fine on your AMD Vega path but drop lyric tokens on the Intel one, so --no-fa is the right workaround for now. To really pin it down I'd need to get my hands on that hardware and run cosine similarity tests on the attention output to isolate where the precision breaks, and then fix it upstream in GGML, or it may get fixed in llama.cpp and reach acestep.cpp when I rebase the GGML submodule. I don't have an Intel Xe2 GPU on hand right now, but if I get access to one at some point I'll dig into it and target the bug properly.

[StefanoBalocco]

If you want me to run some specific tests and provide you the output, feel free to provide me the instructions

[Nurb4000]

Ran into a similar issue with Vulcan older ( Alder Lake ) where if i added lyrics it would pretty much go haywire, but in my case its tied to length, not lyrics or structure, so not related to flash. ( i thought i fixed it with adding clamp 16, but my tests were shorter, so i thought it worked, when in reality it didn't matter since it was shorter ). Still tracking down the exact length it freaks out just for reference purposes, but its around the 3 minute mark i think

This may not help this thread and not a huge deal to me in my edge-case ( was mostly testing, not going to 'use' it ), but thought id mention it just in case it helps any with troubleshooting the root post.

[StefanoBalocco]

but in my case its tied to length, or lyrics or structure

Today I tried with another song, using existing long lyrics where I added [verse] [chorus] and other structure elements and ran into the same problem even with flash attention disabled. Didn't have time to try on AMD iGPU to check if the problem is tied to Intel.

but its around the 3 minute mark i think

I make only 180s song, right now.

[Nurb4000]

Switching to CPU only should help. I do upwards of 8 minutes often, but i do have separate dedicated GPU hardware for it.

I make only 180s song, right now.