nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 10 Packages Projects Releases Wiki Activity

mtmd: add batching API (#24384)

Browse Source 
* mtmd: add batching API

* wip

* first working version (gemma4v)

* add arg

* nits

* wire up support_batch()

* fix 0.0 output embd

* fix audio

* nits

* refactor a bit

* nits

* fix non-batching case

* fix comment
This commit is contained in:
Xuan-Son Nguyen
2026-06-13 00:10:29 +02:00
committed by GitHub
parent f58bad4137
commit e37abd6b5f
14 changed files with 544 additions and 126 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/mtmd/models/gemma4v.cpp
+11 -9
View File
@@ -10,7 +10,7 @@ ggml_cgraph * clip_graph_gemma4v::build() {
ggml_set_name(inp_raw, "inp_raw_scaled");
ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
inp = ggml_reshape_2d(ctx0, inp, n_patches, n_embd);
inp = ggml_reshape_3d(ctx0, inp, n_patches, n_embd, n_batch);
inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp));
ggml_set_name(inp, "inp");
// note: no patch bias
@@ -51,10 +51,11 @@ ggml_cgraph * clip_graph_gemma4v::build() {
// first half
ggml_tensor * first;
{
first = ggml_view_3d(ctx0, cur,
n_dim/2, n_head, n_pos,
first = ggml_view_4d(ctx0, cur,
n_dim/2, n_head, n_pos, n_batch,
cur->nb[1],
cur->nb[2],
cur->nb[3],
0);
first = ggml_rope_ext(
ctx0,
@@ -70,10 +71,11 @@ ggml_cgraph * clip_graph_gemma4v::build() {
// second half
ggml_tensor * second;
{
second = ggml_view_3d(ctx0, cur,
n_dim/2, n_head, n_pos,
second = ggml_view_4d(ctx0, cur,
n_dim/2, n_head, n_pos, n_batch,
cur->nb[1],
cur->nb[2],
cur->nb[3],
n_dim/2 * ggml_element_size(cur));
second = ggml_rope_ext(
ctx0,
@@ -103,14 +105,14 @@ ggml_cgraph * clip_graph_gemma4v::build() {
const int kernel_size = hparams.n_merge;
GGML_ASSERT(kernel_size > 0);
// [n_embd, n_patches] -> [n_patches_x, n_patches_y, n_embd, 1]
cur = ggml_cont_4d(ctx0, ggml_transpose(ctx0, cur), n_patches_x, n_patches_y, n_embd, 1);
// [n_embd, n_patches] -> [n_patches_x, n_patches_y, n_embd, n_batch]
cur = ggml_cont_4d(ctx0, ggml_transpose(ctx0, cur), n_patches_x, n_patches_y, n_embd, n_batch);
cur = ggml_pool_2d(ctx0, cur, GGML_OP_POOL_AVG,
kernel_size, kernel_size, kernel_size, kernel_size, 0, 0);
const int out_x = n_patches_x / kernel_size;
const int out_y = n_patches_y / kernel_size;
// [out_x, out_y, n_embd, 1] -> [n_embd, out_x * out_y]
cur = ggml_reshape_3d(ctx0, cur, out_x * out_y, n_embd, 1);
// [out_x, out_y, n_embd, n_batch] -> [n_embd, out_x * out_y, n_batch]
cur = ggml_reshape_3d(ctx0, cur, out_x * out_y, n_embd, n_batch);
cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
cur = ggml_scale(ctx0, cur, sqrtf((float)n_embd));
cb(cur, "pooled", -1);