fix: tile subsampling for long audio to avoid ggml 2^31 tensor overflow on GPU#19

Merged

mudler merged 7 commits into

worktree-long-audio-tiling

Jun 7, 2026

Commits on Jun 7, 2026

feat(subsampling): add subsample_len spatial-length helper
mudler
committed
feat(subsampling): tiled long-audio path (parity vs forward)

mudler
and
claude
committed
refactor(subsampling): cache valid_out_len in forward_tiled; document tiling test invariant
mudler
committed
feat(encoder): forward_batch_tiled from pre-subsampled features

mudler
and
claude
committed
feat(model): tile subsampling for long audio above safe threshold

mudler
and
claude
committed
feat(model): tile single-clip transcribe for long audio (CLI/path C-API)
mudler
committed
fix(ggml-cuda): grid-stride pad kernel for dims > 65535 (long-audio attention)
mudler
committed