ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-27 01:49:28 +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 7e956a32ce
commit 5536e99d42
21 changed files with 850 additions and 78 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
src/llama-hparams.cpp
View File

@@ -405,22 +405,59 @@ void llm_load_hparams(
}
} break;
case LLM_ARCH_QWEN3:
case LLM_ARCH_QWEN3:
{
ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type, false);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
case 28: model.type = hparams.n_embd == 1024 ? e_model::MODEL_0_6B : e_model::MODEL_1_7B; break;
case 36: model.type = hparams.n_embd == 2560 ? e_model::MODEL_4B : e_model::MODEL_8B; break;
case 40: model.type = e_model::MODEL_14B; break;
case 64: model.type = e_model::MODEL_32B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_QWEN3VL:
{
ml.get_key(LLM_KV_NUM_DEEPSTACK_LAYERS, hparams.n_deepstack_layers, false);
ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
case 28: model.type = e_model::MODEL_1_7B; break;
case 36: model.type = hparams.n_embd == 2560 ? e_model::MODEL_4B : e_model::MODEL_8B; break;
case 64: model.type = e_model::MODEL_32B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
// since vision model stacks deepstack features along feature dim
// we also create a fake "n_embd" for text model to be the main embd + deepstack embds
hparams.n_embd *= hparams.n_deepstack_layers + 1;
} break;
case LLM_ARCH_QWEN3MOE:
{
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
case 48: model.type = e_model::MODEL_30B_A3B; break;
case 94: model.type = e_model::MODEL_235B_A22B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_QWEN3VLMOE:
{
ml.get_key(LLM_KV_NUM_DEEPSTACK_LAYERS, hparams.n_deepstack_layers, false);
ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true);
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
case 48: model.type = e_model::MODEL_30B_A3B; break;
case 94: model.type = e_model::MODEL_235B_A22B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
// since vision model stacks deepstack features along feature dim
// we also create a fake "n_embd" for text model to be the main embd + deepstack embds
hparams.n_embd *= hparams.n_deepstack_layers + 1;
} break;
case LLM_ARCH_PHI2:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);