Skip to content

Gibberish output during training/inference with Qwen2.5-Omni-7B #13

Description

@WendySong31

Hi,
I am trainin on 4 H100 GPUs and encountering an issue were the model outputs gibberish.
The model produced a valid output then a repetitive chatacter loop, and the pattern repeats
--- DEBUG: RAW MODEL OUTPUT START ---
C
--- DEBUG: RAW MODEL OUTPUT START ---
< to for ( with, and, (2 ( for C, and, ( B ( A C, or (, at. ( p. , or. and ( and D ( , , and, for (4, C in A L., de (4. in, and in A '3 (.5 2.8 and and A ( ( ( in, and,, as B. ( ( (3. (. ( of (, ( C. ( C, and, in,, to ( de ( C ( at (4 in for. 2 C,, for ( (1 for B to to as, ( ( ( ( ( ( ( in ( ( at by (11 A [05 and p and in26, and ( (, (. ( a, ( and C A ( ( ( to to,2 ( ( 3 and on while ( [2. at, ( ( and for ( gunshot ( (, A, by - ( -, (4 (, [ (, as in, ( ( or A, to ( A and B as ( as (0 in ', for (, (1 I in cl - (1 ( (2 as and, I ( with., B ( '. a " ( (5 (,, and ( (, as (4 in (. P. as in ( ( (1. P to. P of and to, and ( ( for ( (, if ( (,. is A. ( (26 of p, if, ( and. is at (:, a^2, ( or I, of. ( ( and (: and to (5: ( (: at ( ( [0 ( (1 (, ( for (, (2 for, in - rec, for ( A, and if at ( for ( ( (. ( in. and gunshot, with, (,. or ( to by, ' (: (, B ( I ( I (. ( ( at:21 to be ( as as, ( (. or for ( ( of in or at, to (1.3 ( ( (,, ( A ',5 2,, if ( '2., 1 and to.: , as, in ( ( ( in and , is. (, A. ( in ( ( ( B and of. and for ( I and, and, C (, for in ( ( ( and by, or while (, and,, in, 27 (5. in ( (, ( of in as and for de ( if as,, 4 A ( for ( at, as de: and ( in, and the, I p, as is A, (, ( ( ( as p and (. A and (5 ( (4 in at,., as, of B to (3 ( ( (, (,1 at. ( at if. (, I ( ( and,, and, in (. for ( and ( (, (1 (, ( ( (,. in, D ( [., by ( B 5, 11 A is (. at a valid C, (1., 2 at at or. (以下 ( and for (: B in for the (1 (: ( ( ( and in, and in,, , C. to - to at at, for, in to,, to ( ( (, ( de,, in at a specific with ( (, B. (, ( and, ( at ( (8: ( A of,, - (: B. ( ( ( ( (8,, at ( (. and,, ( of as it ( - ( (1 . C,,. 1 A is and (1 in to 8,, in. 3: I and ( ( (, to ( ( L and, at, ( ( ( ( and..3 for ( ' P. in, 2: ( ( in for (,,.1. is (, to ( [ at by F, B, ( I cl, with, C, ( 1 , - (2., p to.. ( ( and ( with [ (1. of D

I was not able to unistall Flash Attention 2 and used SDPA instead with BF16.

Does the model require attn_implementation="flash_attention_2" to maintain numerical stability?
Should I be applying a specific repetition_penalty or temperature setting in the generation_config to prevent this degeneration?

Thank you for any help provided.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions