mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

mtmd : Add Nemotron Nano 12B v2 VL support (#19547)

Browse Source 
* nemotron nano v2 vlm support added

* simplified code; addressed reviews

* pre-downsample position embeddings during GGUF conversion for fixed input size
This commit is contained in:
Anav Prasad
2026-02-14 05:07:00 -08:00
committed by GitHub
parent 1725e316c1
commit 01d8eaa28d
9 changed files with 167 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/mtmd/models/models.h
+5
View File
@@ -42,6 +42,11 @@ struct clip_graph_internvl : clip_graph {
ggml_cgraph * build() override;
};
struct clip_graph_nemotron_v2_vl : clip_graph {
clip_graph_nemotron_v2_vl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct clip_graph_llama4 : clip_graph {
clip_graph_llama4(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
ggml_cgraph * build() override;
tools/mtmd/models/nemotron-v2-vl.cpp
+35
View File
@@ -0,0 +1,35 @@
#include "models.h"
ggml_cgraph * clip_graph_nemotron_v2_vl::build() {
GGML_ASSERT(model.class_embedding != nullptr);
GGML_ASSERT(model.position_embeddings != nullptr);
const int n_registers = model.class_embedding->ne[1];
const int n_pos = n_patches + n_registers;
ggml_tensor * inp = build_inp();
// add position embeddings (pre-downsampled during GGUF conversion for fixed 512x512 input)
inp = ggml_add(ctx0, inp, model.position_embeddings);
cb(inp, "inp_pos", -1);
inp = ggml_concat(ctx0, model.class_embedding, inp, 1);
ggml_tensor * cur = build_vit(inp, n_pos, NORM_TYPE_NORMAL, hparams.ffn_op, nullptr, nullptr);
cur = ggml_view_2d(ctx0, cur,
n_embd, n_patches,
ggml_row_size(cur->type, n_embd),
n_registers * ggml_row_size(cur->type, n_embd));
cur = build_patch_merge_permute(cur, model.hparams.n_merge);
{
cur = build_norm(cur, model.mm_0_w, nullptr, NORM_TYPE_RMS, 1e-6, -1);
cur = build_ffn(cur, model.mm_1_w, nullptr, nullptr, nullptr, model.mm_3_w, nullptr, FFN_RELU_SQR, -1);
}
ggml_build_forward_expand(gf, cur);
return gf;
}