feat: enable flash attention by default (#82)

Co-authored-by: vansangpfiev <[email protected]>

README.md

@@ -145,6 +145,6 @@ Table of parameters
 |`model_type` | String | Model type we want to use: llm or embedding, default value is llm|
 |`model_alias`| String | Used as model_id if specified in request, mandatory in loadmodel|
 |`model`      | String | Used as model_id if specified in request, mandatory in chat/embedding request|
-|`flash_attn` | Boolean| To enable Flash Attention, default is false|
+|`flash_attn` | Boolean| To enable Flash Attention, default is true|
 |`cache_type` | String| KV cache type: f16, q8_0, q4_0, default is f16|
-|`use_mmap` | Boolean| To enable mmap, default is true|
+|`use_mmap` | Boolean| To enable mmap, default is true|

src/llama_engine.cc

@@ -346,14 +346,12 @@ bool LlamaEngine::LoadModelImpl(std::shared_ptr<Json::Value> json_body) {
     params.cache_type_v = params.cache_type_k;
     LOG_DEBUG << "cache_type: " << params.cache_type_k;
 
-    // Check for backward compatible
-    auto fa0 = json_body->get("flash-attn", false).asBool();
-    auto fa1 = json_body->get("flash_attn", false).asBool();
+    auto fa = json_body->get("flash_attn", true).asBool();
     auto force_enable_fa = params.cache_type_k != kTypeF16;
     if (force_enable_fa) {
       LOG_DEBUG << "Using KV cache quantization, force enable Flash Attention";
     }
-    params.flash_attn = fa0 || fa1 || force_enable_fa;
+    params.flash_attn = fa || force_enable_fa;
     if (params.flash_attn) {
       LOG_DEBUG << "Enabled Flash Attention";
     }