Skip to content

get_theoretical_flops_per_token is not model agnostic, even though the function declaration indicates this #327

Closed
Bug
#346
Closed
get_theoretical_flops_per_token is not model agnostic, even though the function declaration indicates this#327
Bug
#346
Assignees
le1nux
Labels
bugSomething isn't workingenhancementNew feature or request
Milestone
v0.4.0

Description

@le1nux
le1nux
opened on Mar 25, 2025

get_theoretical_flops_per_token(model: FSDPX) is not model agnostic. Internally, we check GPT2 model instance variables such as n_layer or n_embd:

modalities/src/modalities/utils/mfu.py

Lines 116 to 118 in bd3e68b

L = model_module.n_layer
T = model_module.sequence_length
H = model_module.n_embd

We could transform get_theoretical_flops_per_token into a parameterizable component.

Activity

  1. le1nux
    added
    bugSomething isn't working
    enhancementNew feature or request
    on Mar 25, 2025
  2. le1nux
    added this to the v0.4.0 milestone on Mar 25, 2025
  3. le1nux
    assigned
    flxst
    on Mar 25, 2025
  4. le1nux
    added the
    Bug
    issue type on Mar 25, 2025
  5. le1nux
    added a parent issue on Mar 25, 2025
  6. le1nux
    assigned
    mali-git
    and
    le1nux
    and unassigned
    flxst
    and
    mali-git
    on Apr 8, 2025
  7. le1nux
    linked a pull request that will close this issueMFU calculation support with FSDP2 #346on Apr 11, 2025
  8. le1nux
    closed this as completedon Apr 13, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

bugSomething isn't workingenhancementNew feature or request

Type

Bug

Projects

No projects

    Milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions