tmp

Author: orca-zhang

src/llama-kv-cache.cpp

@@ -32,7 +32,7 @@ bool llama_kv_cache_init(
 
     cache.recurrent = llama_model_is_recurrent(&model);
     cache.v_trans   = !cache.recurrent && !cparams.flash_attn;
-    cache.can_shift = !cache.recurrent; // not supported due to MLA
+    cache.can_shift = !cache.recurrent && model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA
 
     LLAMA_LOG_INFO("%s: kv_size = %d, offload = %d, type_k = '%s', type_v = '%s', n_layer = %d, can_shift = %d\n",
             __func__, kv_size, offload, ggml_type_name(type_k), ggml_type_name(type_v), n_layer, cache.can_shift);