Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

model : add Tencent Hy 4 (hy_v4) preview architecture support (#28127)

Browse Source 
* model: add Tencent Hy 4 (hy_v4) preview architecture support

Adds support for the Tencent Hy 4 model (Hugging Face architecture
HYV4ForCausalLM, GGUF arch hy_v4):

Add HF -> GGUF conversion script (conversion/hy_v4.py) and wire it into the conversion registry
Register hy_v4 GGUF constants, arch enum, and writer support
Implement the hy-v4 model graph, hparams, vocab and context changes
Register the new arch in llama-arch and models registry
Extend arch tests to cover hy_v4

Assisted by Claude Opus 5

* Update convert_hf_to_gguf_update.py

Co-authored-by: fairydreaming <166155368+fairydreaming@users.noreply.github.com>

* Update conversion/base.py

Co-authored-by: fairydreaming <166155368+fairydreaming@users.noreply.github.com>

* convert : move hy_v4 entry to the same place as in convert_hf_to_gguf_update.py

* model : apply changes related to n_ff_exp becoming per-layer in Hy4-preview

* n_layer_all

---------

Co-authored-by: fairydreaming <166155368+fairydreaming@users.noreply.github.com>
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
This commit is contained in:
HongHuang
2026-09-04 14:31:36 +02:00
committed by GitHub
co-authored by fairydreaming Stanisław Szymczyk Sigbjørn Skjæret
parent 5266f24da7
commit 49c0dc82b8
18 changed files with 1124 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-model.cpp
+48 -1
View File
@@ -288,6 +288,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
return new llama_model_hunyuan_dense(params);
case LLM_ARCH_HY_V3:
return new llama_model_hy_v3(params);
case LLM_ARCH_HY_V4:
return new llama_model_hy_v4(params);
case LLM_ARCH_SMOLLM3:
return new llama_model_smollm3(params);
case LLM_ARCH_OPENAI_MOE:
@@ -2053,7 +2055,8 @@ void llama_model::print_info() const {
if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_DEEPSEEK2OCR ||
arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA ||
arch == LLM_ARCH_DOTS3NOTE || arch == LLM_ARCH_MISTRAL4) {
arch == LLM_ARCH_DOTS3NOTE || arch == LLM_ARCH_MISTRAL4 ||
arch == LLM_ARCH_HY_V4) {
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q);
LLAMA_LOG_INFO("%s: n_lora_kv = %d\n", __func__, hparams.n_lora_kv);
@@ -2322,6 +2325,48 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
nullptr);
}
} break;
case LLM_ARCH_HY_V4:
{
if (hparams.indexer_top_k == 0) {
// full-attention checkpoint: no indexer, so no indexer key cache
res = new llama_kv_cache(
*this,
hparams,
params.type_k,
params.type_v,
!cparams.flash_attn,
cparams.offload_kqv,
cparams.kv_unified,
cparams.n_ctx_seq,
cparams.n_seq_max,
1,
hparams.n_swa,
hparams.swa_type,
nullptr,
nullptr,
nullptr,
nullptr);
} else {
// only "full" layers own an indexer, so the shared layers need no indexer cache
llama_kv_cache::layer_filter_cb filter_lid = [&](uint32_t il) { return hparams.is_indexer_full(il); };
res = new llama_kv_cache_dsa(
*this,
params.type_k,
params.type_v,
!cparams.flash_attn,
cparams.offload_kqv,
cparams.kv_unified,
cparams.n_ctx_seq,
cparams.n_seq_max,
1,
hparams.n_swa,
hparams.swa_type,
nullptr,
filter_lid,
nullptr);
}
} break;
case LLM_ARCH_DOTS3NOTE:
{
GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE);
@@ -2881,6 +2926,8 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_DOTS3NOTE:
case LLM_ARCH_NANBEIGE:
case LLM_ARCH_POCKETTTS:
// HY_V4 rotates consecutive pairs, matching the reference implementation
case LLM_ARCH_HY_V4:
return LLAMA_ROPE_TYPE_NORM;
// the pairs of head values are offset by n_rot/2