llama : disable Direct IO by default (#19109)

* llama : disable Direct IO by default

* cont : override mmap if supported

Author: ggerganov

common/arg.cpp

@@ -2198,18 +2198,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"--mmap"},
         {"--no-mmap"},
-        string_format("whether to memory-map model. Explicitly enabling mmap disables direct-io. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: %s)", params.use_mmap ? "enabled" : "disabled"),
+        string_format("whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: %s)", params.use_mmap ? "enabled" : "disabled"),
         [](common_params & params, bool value) {
             params.use_mmap = value;
-            if (value) {
-                params.use_direct_io = false;  // disable direct io when mmap is explicitly enabled
-            }
         }
     ).set_env("LLAMA_ARG_MMAP"));
     add_opt(common_arg(
         {"-dio", "--direct-io"},
         {"-ndio", "--no-direct-io"},
-        string_format("use DirectIO if available. Takes precedence over --mmap (default: %s)", params.use_direct_io ? "enabled" : "disabled"),
+        string_format("use DirectIO if available. (default: %s)", params.use_direct_io ? "enabled" : "disabled"),
         [](common_params & params, bool value) {
             params.use_direct_io = value;
         }

common/common.h

@@ -438,7 +438,7 @@ struct common_params {
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool use_mmap          = true;  // enable mmap to use filesystem cache
-    bool use_direct_io     = true;  // read from disk without buffering for faster model loading
+    bool use_direct_io     = false; // read from disk without buffering
     bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation

include/llama.h

@@ -309,7 +309,7 @@ extern "C" {
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool vocab_only;      // only load the vocabulary, no weights
         bool use_mmap;        // use mmap if possible
-        bool use_direct_io;   // use direct io, takes precedence over use_mmap
+        bool use_direct_io;   // use direct io, takes precedence over use_mmap when supported
         bool use_mlock;       // force system to keep model in RAM
         bool check_tensors;   // validate model tensor data
         bool use_extra_bufts; // use extra buffer types (used for weight repacking)

src/llama-model-loader.cpp

@@ -541,15 +541,15 @@ llama_model_loader::llama_model_loader(
 
     if (use_mmap && use_direct_io) {
         if (files.back()->has_direct_io()) {
-            // Disable mmap, as DirectIO is available
-            use_mmap = false;
             LLAMA_LOG_WARN("%s: direct I/O is enabled, disabling mmap\n", __func__);
+            use_mmap = false;
         } else {
-            // Disable DirectIO and reopen file using std::fopen for mmap
+            LLAMA_LOG_WARN("%s: direct I/O is not available, using mmap\n", __func__);
             use_direct_io = false;
+
+            // reopen file using std::fopen for mmap
             files.pop_back();
             files.emplace_back(new llama_file(fname.c_str(), "rb", false));
-            LLAMA_LOG_WARN("%s: direct I/O is not available, using mmap\n", __func__);
         }
     }
 

src/llama-model.cpp

@@ -8125,7 +8125,7 @@ llama_model_params llama_model_default_params() {
         /*.kv_overrides                =*/ nullptr,
         /*.vocab_only                  =*/ false,
         /*.use_mmap                    =*/ true,
-        /*.use_direct_io               =*/ true,
+        /*.use_direct_io               =*/ false,
         /*.use_mlock                   =*/ false,
         /*.check_tensors               =*/ false,
         /*.use_extra_bufts             =*/ true,

src/llama-quant.cpp

@@ -545,7 +545,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     }
 
     std::vector<std::string> splits = {};
-    llama_model_loader ml(fname_inp, splits, use_mmap, /*use_direct_io*/ true, /*check_tensors*/ true, /*no_alloc*/ false, kv_overrides, nullptr);
+    llama_model_loader ml(fname_inp, splits, use_mmap, /*use_direct_io*/ false, /*check_tensors*/ true, /*no_alloc*/ false, kv_overrides, nullptr);
     ml.init_mappings(false); // no prefetching
 
     llama_model model(llama_model_default_params());