Skip to content

LLM inference in C/C++

C++ ggml

mtmd: use ggml_rope_set_offset (#27521)

Xuan-Son Nguyenb21e4de74567
diff --git a/tools/mtmd/clip-graph.h b/tools/mtmd/clip-graph.h
index 2cf1b683a7..bbee35bead 100644
--- a/tools/mtmd/clip-graph.h
+++ b/tools/mtmd/clip-graph.h
@@ -137,9 +137,15 @@ struct clip_graph {
int il,
ggml_tensor * sinks = nullptr) const;
- // implementation of the 2D RoPE without adding a new op in ggml
- // this is not efficient (use double the memory), but works on all backends
- // TODO: there was a more efficient which relies on ggml_view and ggml_rope_ext_inplace, but the rope inplace does not work well with non-contiguous tensors ; we should fix that and revert back to the original implementation in https://github.com/ggml-org/llama.cpp/pull/13065
+ // implementation of the 2D RoPE using two ggml_rope_ext calls
+ //
+ // unlike GGML_ROPE_TYPE_VISION which forces NEOX ordering, this rotates adjacent pairs (normal ordering)
+ //
+ // example:
+ // given a single head with size = 8 --> [00000000]
+ // dims [0, 4) rotate with pos_a, dims [4, 8) rotate with pos_b --> [aaaabbbb]
+ // interleave_freq = false --> both halves use the same inv_freq set (like GGML_ROPE_TYPE_VISION)
+ // interleave_freq = true --> first half uses even inv_freq, second half uses odd inv_freq (used by pixtral)
ggml_tensor * build_rope_2d(
ggml_context * ctx0,
ggml_tensor * cur,
diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index 9977ed4909..89ca65a7bb 100644
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -819,8 +819,6 @@ ggml_tensor * clip_graph::build_attn(
}
// implementation of the 2D RoPE without adding a new op in ggml
-// this is not efficient (use double the memory), but works on all backends
-// TODO: there was a more efficient which relies on ggml_view and ggml_rope_ext_inplace, but the rope inplace does not work well with non-contiguous tensors ; we should fix that and revert back to the original implementation in https://github.com/ggml-org/llama.cpp/pull/13065
ggml_tensor * clip_graph::build_rope_2d(
ggml_context * ctx0,
ggml_tensor * cur,
@@ -829,9 +827,7 @@ ggml_tensor * clip_graph::build_rope_2d(
const float freq_base,
const bool interleave_freq
) {
- const int64_t n_dim = cur->ne[0];
- const int64_t n_head = cur->ne[1];
- const int64_t n_pos = cur->ne[2];
+ const int64_t n_dim = cur->ne[0];
// for example, if we have cur tensor of shape (n_dim=8, n_head, n_pos)
// we will have a list of 4 inv_freq: 1e-0, 1e-1, 1e-2, 1e-3
@@ -845,46 +841,30 @@ ggml_tensor * clip_graph::build_rope_2d(
? std::pow(freq_base, (float)-2/n_dim)
: 1.0;
- // first half
- ggml_tensor * first;
- {
- first = ggml_view_3d(ctx0, cur,
- n_dim/2, n_head, n_pos,
- cur->nb[1],
- cur->nb[2],
- 0);
- first = ggml_rope_ext(
- ctx0,
- first,
- pos_a, // positions
- nullptr, // freq factors
- n_dim/2, // n_dims
- 0, 0, freq_base,
- 1.0f, 0.0f, 1.0f, 0.0f, 0.0f
- );
- }
-
- // second half
- ggml_tensor * second;
- {
- second = ggml_view_3d(ctx0, cur,
- n_dim/2, n_head, n_pos,
- cur->nb[1],
- cur->nb[2],
- n_dim/2 * ggml_element_size(cur));
- second = ggml_rope_ext(
- ctx0,
- second,
- pos_b, // positions
- nullptr, // freq factors
- n_dim/2, // n_dims
- 0, 0, freq_base,
- freq_scale_odd,
- 0.0f, 1.0f, 0.0f, 0.0f
- );
- }
+ // first half, dims [0, n_dim/2)
+ cur = ggml_rope_ext(
+ ctx0,
+ cur,
+ pos_a, // positions
+ nullptr, // freq factors
+ n_dim/2, // n_dims
+ 0, 0, freq_base,
+ 1.0f, 0.0f, 1.0f, 0.0f, 0.0f
+ );
+
+ // second half, dims [n_dim/2, n_dim)
+ cur = ggml_rope_ext(
+ ctx0,
+ cur,
+ pos_b, // positions
+ nullptr, // freq factors
+ n_dim/2, // n_dims
+ 0, 0, freq_base,
+ freq_scale_odd,
+ 0.0f, 1.0f, 0.0f, 0.0f
+ );
+ cur = ggml_rope_set_offset(cur, n_dim/2);
- cur = ggml_concat(ctx0, first, second, 0);
return cur;
}
diff --git a/tools/mtmd/models/gemma4v.cpp b/tools/mtmd/models/gemma4v.cpp
index 87cbd43fc5..4484389470 100644
--- a/tools/mtmd/models/gemma4v.cpp
+++ b/tools/mtmd/models/gemma4v.cpp
@@ -44,51 +44,31 @@ ggml_cgraph * clip_graph_gemma4v::build() {
// similar to build_rope_2d, but use neox ordering
auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
- const int64_t n_dim = cur->ne[0];
- const int64_t n_head = cur->ne[1];
- const int64_t n_pos = cur->ne[2];
-
- // first half
- ggml_tensor * first;
- {
- 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,
- first,
- pos_x, // positions
- nullptr, // freq factors
- n_dim/2, // n_dims
- GGML_ROPE_TYPE_NEOX, 0, hparams.rope_theta,
- 1.0f, 0.0f, 1.0f, 0.0f, 0.0f
- );
- }
-
- // second half
- ggml_tensor * second;
- {
- 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,
- second,
- pos_y, // positions
- nullptr, // freq factors
- n_dim/2, // n_dims
- GGML_ROPE_TYPE_NEOX, 0, hparams.rope_theta,
- 1.0f, 0.0f, 1.0f, 0.0f, 0.0f
- );
- }
-
- cur = ggml_concat(ctx0, first, second, 0);
+ const int64_t n_dim = cur->ne[0];
+
+ // first half, dims [0, n_dim/2)
+ cur = ggml_rope_ext(
+ ctx0,
+ cur,
+ pos_x, // positions
+ nullptr, // freq factors
+ n_dim/2, // n_dims
+ GGML_ROPE_TYPE_NEOX, 0, hparams.rope_theta,
+ 1.0f, 0.0f, 1.0f, 0.0f, 0.0f
+ );
+
+ // second half, dims [n_dim/2, n_dim)
+ cur = ggml_rope_ext(
+ ctx0,
+ cur,
+ pos_y, // positions
+ nullptr, // freq factors
+ n_dim/2, // n_dims
+ GGML_ROPE_TYPE_NEOX, 0, hparams.rope_theta,
+ 1.0f, 0.0f, 1.0f, 0.0f, 0.0f
+ );
+ cur = ggml_rope_set_offset(cur, n_dim/2);
+
return cur;
};
diff --git a/tools/mtmd/models/minimax-m3.cpp b/tools/mtmd/models/minimax-m3.cpp
index 447621754e..256e531057 100644
--- a/tools/mtmd/models/minimax-m3.cpp
+++ b/tools/mtmd/models/minimax-m3.cpp
@@ -2,30 +2,22 @@
ggml_tensor * clip_graph_minimax_m3::apply_rope(
ggml_tensor * x, ggml_tensor * pos_h, ggml_tensor * pos_w) {
- const int64_t Hn = x->ne[1];
- const int64_t P = x->ne[2];
- const size_t es = ggml_element_size(x);
- const int dh = (int) x->ne[0];
- const int axd = 2 * ((2 * (dh / 2) / 3) / 2);
+ const int dh = (int) x->ne[0];
+ const int axd = 2 * ((2 * (dh / 2) / 3) / 2);
- GGML_ASSERT(x->nb[0] == es);
GGML_ASSERT(3 * axd <= dh);
const float th = hparams.rope_theta;
// layout of x is [t, h, w, pad]
// t is unrotated, h and w are rotated, pad is unrotated
- // note: everything from n_dims onward untouched, so w and pad are rotated in one call.
- auto sl = [&](int off, int n) {
- return ggml_cont(ctx0, ggml_view_3d(ctx0, x, n, Hn, P, x->nb[1], x->nb[2], (size_t) off * es));
- };
- ggml_tensor * t = sl(0, axd);
- ggml_tensor * h = sl(axd, axd);
- ggml_tensor * w = sl(2 * axd, dh - 2 * axd); // w + pad
-
- h = ggml_rope_ext(ctx0, h, pos_h, nullptr, axd, GGML_ROPE_TYPE_NEOX, 0, th, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
- w = ggml_rope_ext(ctx0, w, pos_w, nullptr, axd, GGML_ROPE_TYPE_NEOX, 0, th, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
- return ggml_concat(ctx0, ggml_concat(ctx0, t, h, 0), w, 0);
+ x = ggml_rope_ext(ctx0, x, pos_h, nullptr, axd, GGML_ROPE_TYPE_NEOX, 0, th, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+ x = ggml_rope_set_offset(x, axd);
+
+ x = ggml_rope_ext(ctx0, x, pos_w, nullptr, axd, GGML_ROPE_TYPE_NEOX, 0, th, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+ x = ggml_rope_set_offset(x, 2 * axd);
+
+ return x;
}
ggml_cgraph * clip_graph_minimax_m3::build() {