Skip to content

Performance regression in v0.6.18 caused by synchronization in TRTLLM ragged prefill #4928

Description

@wzhao18

Before submitting

  • I have searched existing issues and did not find a solution or an existing report.

Environment

n/a

Bug description

#3779 introduces checks for inactive rows in trtllm ragged prefill but causes synchronization inside the kernel API. alternatively we can supply CPU tensors q_seq_lens_cpu and kv_seq_lens_cpu to avoid device sync but that introduces non-trivial host overhead.

A potentially fix is to handle the edge case outside the kernel and once per forward pass instead of once every kernel call.

Minimal reproduction

n/a

Error messages / logs


Additional context

No response

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

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions