ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-28 10:21:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Port of Qwen3-VL support from mainline (#883)

Browse Source 
* Port of Qwen3-VL for latest ik_llama.cpp

- convert_hf_to_gguf.py - Not touched, use llama.cpp to convert model instead
- sysl and metal support for imrope not added
- Vulkan support for imrope not tested
- Code not tested

* Bugfix n_embd was declared multiple times

https://github.com/ikawrakow/ik_llama.cpp/pull/883#issuecomment-3471179655

* Fix n_embd issue with qwen3vl

* model.output tensor not required

https://github.com/ikawrakow/ik_llama.cpp/pull/883#discussion_r2480388389

* Improved logic for qkv combined tensors

59ceaf8fcb (r2480395800)
59ceaf8fcb (r2480398187)

* Fix n_embd for merge_qkv() + cleaner code

https://github.com/ikawrakow/ik_llama.cpp/pull/883#discussion_r2481227395

* Revert TENSOR_NOT_REQUIRED
This commit is contained in:
Thireus ☠
2025-11-04 17:20:54 +00:00
committed by GitHub
parent efcb5f9d9e
commit 86597623a5
21 changed files with 850 additions and 78 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/llama.cpp
View File

@@ -1165,6 +1165,10 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train);
LLAMA_LOG_INFO("%s: n_ctx_orig_yarn = %u\n", __func__, hparams.n_ctx_orig_yarn);
LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown");
// MRoPE (Multi-axis Rotary Position Embedding) sections
if (const auto & s = hparams.rope_sections; s[0] || s[1] || s[2] || s[3]) {
LLAMA_LOG_INFO("%s: mrope sections = [%d, %d, %d, %d]\n", __func__, s[0], s[1], s[2], s[3]);
}
LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv);
LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner);
LLAMA_LOG_INFO("%s: ssm_d_state = %u\n", __func__, hparams.ssm_d_state);
@@ -1230,7 +1234,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp);
}
if (model.arch == LLM_ARCH_QWEN3MOE || model.arch == LLM_ARCH_OPENAI_MOE) {
if (model.arch == LLM_ARCH_QWEN3MOE || model.arch == LLM_ARCH_OPENAI_MOE || model.arch == LLM_ARCH_QWEN3VLMOE) {
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
}
@@ -2054,7 +2058,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
auto tim1 = ggml_time_us();
#endif
const int64_t n_tokens = batch.n_tokens;
const int n_pos_per_embd = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
const int n_pos_per_embd = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE ? 4 : 1;
if (batch.token && n_pos_per_embd == 4) {
std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd);
for (int i = 0; i < n_tokens; ++i) {
@@ -4668,6 +4672,10 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
case LLM_ARCH_QWEN2VL:
return LLAMA_ROPE_TYPE_MROPE;
case LLM_ARCH_QWEN3VL:
case LLM_ARCH_QWEN3VLMOE:
return LLAMA_ROPE_TYPE_IMROPE;
// all model arches should be listed explicitly here
case LLM_ARCH_UNKNOWN:
GGML_ABORT("unknown architecture");