model: mistral le chonk (#31376)

* llama: extend DSA and sdpark to new model

* fix ci
This commit is contained in:
Xuan-Son Nguyen authored and GitHub committed 2026-10-29 15:43:33 +01:00
1 parent 0ac1a7fd35
commit 78feec062d
23 files changed
+1290 -35

No files matched your search

+35 -5
View File
@@ -218,6 +218,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
return new llama_model_glm_dsa(params);
case LLM_ARCH_MISTRAL4:
return new llama_model_mistral4(params);
case LLM_ARCH_MISTRAL_LARGE4:
return new llama_model_mistral_large4(params);
case LLM_ARCH_CHATGLM:
return new llama_model_chatglm(params);
case LLM_ARCH_GLM4:
@@ -1467,6 +1469,7 @@ void llama_model_base::load_hparams(llama_model_loader & ml) {
std::fill(hparams.is_swa_impl.begin(), hparams.is_swa_impl.end(), 0);
std::fill(hparams.is_recr_impl.begin(), hparams.is_recr_impl.end(), llm_arch_is_recurrent(ml.get_arch()) ? 1 : 0);
std::fill(hparams.is_indexer_full_impl.begin(), hparams.is_indexer_full_impl.end(), 0);
std::iota(hparams.kv_source_impl.begin(), hparams.kv_source_impl.end(), 0);
std::fill(hparams.xielu_alpha_n.begin(), hparams.xielu_alpha_n.end(), 0.0f);
std::fill(hparams.xielu_alpha_p.begin(), hparams.xielu_alpha_p.end(), 0.0f);
@@ -2629,7 +2632,7 @@ 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_HY_V4) {
arch == LLM_ARCH_MISTRAL_LARGE4 || 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);
@@ -2908,6 +2911,32 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
nullptr);
}
} break;
case LLM_ARCH_MISTRAL_LARGE4:
{
// 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); };
// layers without own KV read the MLA cache of their source layer
llama_kv_cache::layer_reuse_cb reuse = [&](int32_t il) -> int32_t {
return hparams.has_kv(il) ? -1 : (int32_t) hparams.kv_source(il);
};
res = new llama_kv_cache_dsa(
*this,
params.no_alloc,
params.type_k,
params.type_v,
!cparams.flash_attn,
cparams.offload_kqv,
cparams.kv_unified,
cparams.n_ctx_seq,
cparams.n_seq_max,
hparams.n_swa,
hparams.swa_type,
nullptr,
filter_lid,
reuse);
} break;
case LLM_ARCH_GLM5_NEXT:
{
// KDA layers are recurrent, the DSA layers use a K-only MLA cache plus an indexer cache.
@@ -3091,8 +3120,8 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
} break;
case LLM_ARCH_DFLASH:
{
// DSV4 DSpark stages store a single MLA-style K per position (window = the draft ring)
if (hparams.dsv4_hc_mult > 0) {
// DSV4 and ML4 DSpark stages store a single MLA-style K per position (window = the draft ring)
if (hparams.dsv4_hc_mult > 0 || hparams.n_lora_kv > 0) {
GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE);
res = new llama_kv_cache_iswa(
@@ -3571,6 +3600,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_MISTRAL3:
case LLM_ARCH_EAGLE3:
case LLM_ARCH_MISTRAL4:
case LLM_ARCH_MISTRAL_LARGE4:
case LLM_ARCH_LLAMA_EMBED:
case LLM_ARCH_MAINCODER:
case LLM_ARCH_GLM_DSA:
@@ -3675,8 +3705,8 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
if (const auto & s = model->hparams.rope_sections; s[0] || s[1] || s[2] || s[3]) {
return LLAMA_ROPE_TYPE_MROPE;
}
// DSV4 DSpark drafters use DeepSeek-V4's normal RoPE; legacy DFlash backbones are NeoX
return model->hparams.dsv4_hc_mult > 0 ? LLAMA_ROPE_TYPE_NORM : LLAMA_ROPE_TYPE_NEOX;
// DSV4 and ML4 DSpark drafters use normal RoPE; legacy DFlash backbones are NeoX
return model->hparams.dsv4_hc_mult > 0 || model->hparams.n_lora_kv > 0 ? LLAMA_ROPE_TYPE_NORM : LLAMA_ROPE_TYPE_NEOX;
case LLM_ARCH_SENSENOVA_U1:
case LLM_ARCH_QWEN2VL: