note: clip_is_mrope was moved to mtmd_decode_use_mrope upstream and no longer syncs since https://github.com/ggml-org/llama.cpp/pull/18793

Merge commit 'c1e79e610fd28f2c3923539fee9313734bbf8cfa' into concedo_experimental

# Conflicts:
#	.github/workflows/build.yml
#	.github/workflows/release.yml
#	CMakeLists.txt
#	CONTRIBUTING.md
#	MIT_LICENSE_GGML_SDCPP_LLAMACPP_ONLY.md
#	README.md
#	SECURITY.md
#	ci/run.sh
#	common/CMakeLists.txt
#	common/arg.cpp
#	docs/ops.md
#	docs/ops/BLAS.csv
#	docs/ops/zDNN.csv
#	docs/preset.md
#	examples/batched/batched.cpp
#	examples/debug/debug.cpp
#	ggml/src/ggml-blas/CMakeLists.txt
#	ggml/src/ggml-opencl/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	licenses/LICENSE-curl
#	licenses/LICENSE-httplib
#	scripts/pr2wt.sh
#	scripts/sync_vendor.py
#	tests/CMakeLists.txt
#	tests/test-backend-ops.cpp
#	tools/cli/README.md
#	tools/completion/README.md
#	tools/llama-bench/llama-bench.cpp
#	tools/server/README.md
#	vendor/cpp-httplib/LICENSE
This commit is contained in:
Concedo
2026-01-13 23:31:14 +08:00
45 changed files with 859 additions and 661 deletions
+1 -1
View File
@@ -4205,7 +4205,7 @@ bool clip_is_glm(const struct clip_ctx * ctx) {
return ctx->proj_type() == PROJECTOR_TYPE_GLM_EDGE;
}
bool clip_is_mrope(const struct clip_ctx * ctx) {
bool clip_is_mrope(const struct clip_ctx * ctx) { //kcpp: this was removed in https://github.com/ggml-org/llama.cpp/pull/18793 and moved to mtmd_decode_use_mrope
switch (ctx->proj_type()) {
case PROJECTOR_TYPE_QWEN2VL:
case PROJECTOR_TYPE_QWEN25VL:
+11 -9
View File
@@ -146,8 +146,6 @@ struct mtmd_context {
bool tok_row_end_trail = false;
bool ov_img_first = false;
bool use_mrope = false; // for Qwen2VL, we need to use M-RoPE
// string template for slice image delimiters with row/col (idefics3)
std::string sli_img_start_tmpl;
@@ -217,7 +215,6 @@ struct mtmd_context {
void init_vision() {
GGML_ASSERT(ctx_v != nullptr);
use_mrope = clip_is_mrope(ctx_v);
projector_type proj = clip_get_projector_type(ctx_v);
int minicpmv_version = clip_is_minicpmv(ctx_v);
@@ -627,7 +624,7 @@ struct mtmd_tokenizer {
}
mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
if (ctx->use_mrope) {
if (mtmd_decode_use_mrope(ctx)) {
// for Qwen2VL, we need this information for M-RoPE decoding positions
image_tokens->nx = clip_n_output_tokens_x(ctx->ctx_v, batch_f32.entries[0].get());
image_tokens->ny = clip_n_output_tokens_y(ctx->ctx_v, batch_f32.entries[0].get());
@@ -863,10 +860,7 @@ float * mtmd_get_output_embd(mtmd_context * ctx) {
bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
switch (ctx->proj_type_v()) {
case PROJECTOR_TYPE_QWEN2VL:
case PROJECTOR_TYPE_QWEN25VL:
case PROJECTOR_TYPE_QWEN3VL:
case PROJECTOR_TYPE_YOUTUVL:
case PROJECTOR_TYPE_GEMMA3:
return true;
default:
return false;
@@ -874,7 +868,15 @@ bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
}
bool mtmd_decode_use_mrope(mtmd_context * ctx) {
return ctx->use_mrope;
switch (ctx->proj_type_v()) {
case PROJECTOR_TYPE_QWEN2VL:
case PROJECTOR_TYPE_QWEN25VL:
case PROJECTOR_TYPE_QWEN3VL:
case PROJECTOR_TYPE_GLM4V:
return true;
default:
return false;
}
}
bool mtmd_support_vision(mtmd_context * ctx) {