Skip to content

[Bug] Numeric error happens only at few lines and wg1#10

Open
120L021326 wants to merge 1 commit intomainfrom
mla/reg
Open

[Bug] Numeric error happens only at few lines and wg1#10
120L021326 wants to merge 1 commit intomainfrom
mla/reg

Conversation

@120L021326
Copy link
Copy Markdown
Collaborator

No description provided.

@120L021326 120L021326 changed the title 11 rQP bug Nov 5, 2025
@120L021326
Copy link
Copy Markdown
Collaborator Author

kernel name: mla_decode_latent_sharedZ_ws_dim2_2K_txl_change

@deciding deciding changed the title rQP bug [Bug] Numeric error happens only at few lines and wg1 Nov 7, 2025
@deciding
Copy link
Copy Markdown
Owner

deciding commented Nov 7, 2025

  1. "An mbarrier object is used to synchronize threads and asynchronous memory operations."
    use named barrier otherwise
    see cp.async (can be in syntax or separate), wgmma.async, st.async
    ref
  2. if the compiler not aware of a smem change before wgmma, we must manually add fence
    "...fence.proxy.async operation to make the generic proxy operations visible to the async proxy."
    ref

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

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants