matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-14 09:47:02 +00:00
Code Issues Packages Projects Releases Wiki Activity

chore: update mllama to use ollama engine (#10637)

Browse Source
This commit is contained in:
Michael Yang
2025-05-13 17:36:02 -07:00
committed by GitHub
parent 0478d440f0
commit 23125648b8
67 changed files with 785 additions and 4354 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
llama/llama.cpp/src/llama-kv-cache.cpp vendored
View File

@@ -100,16 +100,8 @@ llama_kv_cache_unified::llama_kv_cache_unified(
throw std::runtime_error("failed to create ggml context for kv cache");
}
ggml_tensor * k, *v;
// for cross attention layers
if (model.arch == LLM_ARCH_MLLAMA && hparams.cross_attention_layers(i)) {
k = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_k, 6404, hparams.n_head_kv(i));
v = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_v, 6404, hparams.n_head_kv(i));
} else {
k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size);
v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size);
}
ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size);
ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size);
ggml_format_name(k, "cache_k_l%d", i);
ggml_format_name(v, "cache_v_l%d", i);
k_l.push_back(k);
@@ -459,7 +451,7 @@ void llama_kv_cache_unified::set_full() {
llama_sbatch llama_kv_cache_unified::sbatch_init(
const llama_batch & batch,
bool logits_all) {
return llama_sbatch(batch, batch.n_embd, true, logits_all);
return llama_sbatch(batch, hparams.n_embd, true, logits_all);
}
llama_ubatch llama_kv_cache_unified::ubatch_next(