common : rename kv_split -> kv_unified

ggml-ci

Author: ggerganov

common/arg.cpp

@@ -1465,11 +1465,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_ARG_SWA_FULL"));
     add_opt(common_arg(
-        {"--kv-split", "-kvs"},
-        string_format("use multiple streams when computing the attention (default: %s)\n"
-            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/14363)", params.kv_split ? "true" : "false"),
+        {"--kv-unified", "-kvu"},
+        string_format("use single unified KV buffer for the KV cache of all sequences (default: %s)\n"
+            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/14363)", params.kv_unified ? "true" : "false"),
         [](common_params & params) {
-            params.kv_split = true;
+            params.kv_unified = true;
         }
     ).set_env("LLAMA_ARG_KV_SPLIT"));
     add_opt(common_arg(

common/common.cpp

@@ -1157,7 +1157,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.no_perf           = params.no_perf;
     cparams.op_offload        = !params.no_op_offload;
     cparams.swa_full          = params.swa_full;
-    cparams.kv_unified        = !params.kv_split;
+    cparams.kv_unified        = params.kv_unified;
 
     cparams.type_k = params.cache_type_k;
     cparams.type_v = params.cache_type_v;

common/common.h

@@ -330,7 +330,7 @@ struct common_params {
     bool no_perf           = false; // disable performance metrics
     bool ctx_shift         = true;  // context shift on inifinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
-    bool kv_split          = false; // disable unified KV cache
+    bool kv_unified        = false; // enable unified KV cache
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool use_mmap          = true;  // use mmap for faster loads

src/llama-context.cpp

@@ -98,12 +98,21 @@ llama_context::llama_context(
         LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD);
         cparams.n_batch = GGML_KQ_MASK_PAD;
     }
-
     cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
 
     cparams.op_offload = params.op_offload;
     cparams.kv_unified = params.kv_unified;
 
+    {
+        const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
+        const bool supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) : 0;
+
+        if (!supports_set_rows && !cparams.kv_unified) {
+            LLAMA_LOG_WARN("%s: non-unified KV cache requires ggml_set_rows() - forcing unified KV cache\n", __func__);
+            cparams.kv_unified = true;
+        }
+    }
+
     const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
 
     LLAMA_LOG_INFO("%s: n_seq_max     = %u\n",   __func__, cparams.n_seq_max);
@@ -2195,7 +2204,7 @@ llama_context_params llama_context_default_params() {
         /*.no_perf                     =*/ true,
         /*.op_offload                  =*/ true,
         /*.swa_full                    =*/ true,
-        /*.kv_unified                  =*/ true,
+        /*.kv_unified                  =*/ false,
     };
 
     return result;

src/llama-kv-cache-unified.cpp

@@ -195,9 +195,9 @@ llama_kv_cache_unified::llama_kv_cache_unified(
     const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
     supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) : 0;
 
-    if (!supports_set_rows && !unified) {
-        LLAMA_LOG_WARN("%s: non-unified KV cache requires ggml_set_rows() - forcing LLAMA_SET_ROWS=1\n", __func__);
-        supports_set_rows = 1;
+    if (!supports_set_rows) {
+        // ref: https://github.com/ggml-org/llama.cpp/pull/14363
+        GGML_ASSERT(unified && "cannot use non-unified KV cache without ggml_set_rows() support");
     }
 
     if (!supports_set_rows) {