Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-20 01:31:31 +02:00
Code Issues Packages Projects Releases Wiki Activity

fix clamp cold prefix

Browse Source 
Co-authored-by: Pascal <admin@serveurperso.com>
This commit is contained in:
Xuan Son Nguyen
2026-08-01 12:33:52 +02:00
parent ffe530beb8
commit 914fd950a3
1 changed files with 8 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/mtmd/models/qwen3tts-gen.cpp
+8
View File
@@ -488,6 +488,14 @@ ggml_tensor * clip_graph_qwen3tts_gen::code2wav::tfm_layer_forward(ggml_tensor *
ggml_tensor * causal_keep = ggml_step(ctx0, ggml_scale_bias(ctx0, diff, 1.0f, 0.5f)); // diff >= 0
ggml_tensor * in_window = ggml_step(ctx0, ggml_scale_bias(ctx0, diff, -1.0f, (float) W - 0.5f)); // diff < W
ggml_tensor * keep = ggml_mul(ctx0, causal_keep, in_window);
// clamp the cold prefix: key j holds real state only when j >= prefix -
// tfm_pos, everything before is the zero-filled cold start and attending
// to zero keys dilutes the softmax instead of skipping them
ggml_tensor * warm = ggml_step(ctx0, ggml_scale_bias(ctx0, ggml_add(ctx0, pos_k, base),
1.0f, 0.5f - (float) prefix)); // j + pos > prefix - 0.5
keep = ggml_mul(ctx0, keep, warm);
ggml_tensor * mask = ggml_reshape_4d(ctx0, ggml_log(ctx0, keep), total_kv, N, 1, 1); // 0 = keep, -inf = masked
ggml_tensor * q_cur = ggml_reshape_4d(ctx0, q, d_head, n_head, N, 1);