models: pad on the left with ggml_pad_ext (#29567)

* models: pad on the left with ggml_pad_ext

The Parakeet, LFM2-Audio, Granite Speech and Gemma 4 audio encoders
build a left padding as a right pad followed by a roll, and DFlash2
concatenates a zero filled block in front of the previous tokens.
ggml_pad_ext does both in one node now that every backend supports a
left padding. The Gemma 4 audio embeddings are bit identical.

* models: skip the DFlash2 taps that only read padding

A tap at or past block_size shifts every row out of the block, so its
term is zero. The loop runs min(kernel_size, block_size) taps.
This commit is contained in:
Pascal authored and GitHub committed 2026-09-28 20:56:15 +02:00
1 parent 14ebbd5f2f
commit 57b557cb95
5 files changed
+16 -28

No files matched your search

+7 -10
View File
@@ -446,19 +446,16 @@ static ggml_tensor * build_dflash2_conv(
ggml_tensor * weight_all = ggml_add(ctx0, coeff_all, base_side);
// taps at or past block_size only read the left padding and add nothing
const int64_t n_taps = std::min(kernel_size, block_size);
ggml_tensor * result = nullptr;
for (int64_t tap = 0; tap < kernel_size; ++tap) {
for (int64_t tap = 0; tap < n_taps; ++tap) {
ggml_tensor * values = blocks;
if (tap > 0) {
ggml_tensor * zeros = ggml_fill(ctx0,
ggml_new_tensor_3d(ctx0, hidden->type, hidden_size, std::min(tap, block_size), n_blocks), 0.0f);
if (tap < block_size) {
ggml_tensor * previous = ggml_view_3d(ctx0, blocks, hidden_size, block_size - tap, n_blocks,
blocks->nb[1], blocks->nb[2], 0);
values = ggml_concat(ctx0, zeros, previous, 1);
} else {
values = zeros;
}
ggml_tensor * previous = ggml_view_3d(ctx0, blocks, hidden_size, block_size - tap, n_blocks,
blocks->nb[1], blocks->nb[2], 0);
values = ggml_pad_ext(ctx0, previous, 0, 0, tap, 0, 0, 0, 0, 0);
}
values = ggml_reshape_2d(ctx0, values, hidden_size, n_tokens);