Skip to content

fix: run diarization attention with mul_mat softmax on ROCm - #112

Open
staryj wants to merge 1 commit into
mudler:masterfrom
staryj:fix/rocm-diarization-attention
Open

staryj wants to merge 1 commit into
mudler:masterfrom
staryj:fix/rocm-diarization-attention

Conversation

@staryj

@staryj staryj commented Nov 2, 2026

Copy link
Copy Markdown

On RocM (Radeon R9700) diarization fails with error: "SIGSEGV: segmentation violation"

The ggml flash-attn tile kernel faults on gfx1201 for the Nemotron streaming chunk, so diarization uses the same softmax attention as the FastConformer encoder.

The ggml flash-attn tile kernel faults on gfx1201 for the Nemotron streaming chunk, so diarization uses the same softmax attention as the FastConformer encoder.

Assisted-by: Grok:grok-4.7 [Cursor]
Co-authored-by: Cursor <cursoragent@cursor.com>
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.

1 participant