llama.swiftui : add bench functionality (ggml-org#4483)

* llama.swiftui : add bench button

* llama.swiftui : initial bench functionality

* force to use n_gpu_layers on simulator

* add download buttons & expose llamaState.loadModel

* update project.pbxproj

* comment #Preview & fix editorconfig check

* gitignore : xcode stuff

* llama.swiftui : UX improvements

* llama.swiftui : avoid data copy via "downloadTask"

* llama.swiftui : remove model from project

* llama : remove "mostly" from model infos

* llama.swiftui : improve bench

---------

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

Author: ggerganov

.editorconfig

@@ -23,3 +23,6 @@ insert_final_newline = unset
 
 [examples/server/public/*]
 indent_size = 2
+
+[examples/llama.swiftui/llama.swiftui.xcodeproj/*]
+indent_style = tab

examples/llama.swiftui/.gitignore

@@ -1 +1,2 @@
 xcuserdata
+xcshareddata

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -6,16 +6,34 @@ enum LlamaError: Error {
     case couldNotInitializeContext
 }
 
+func llama_batch_clear(_ batch: inout llama_batch) {
+    batch.n_tokens = 0
+}
+
+func llama_batch_add(_ batch: inout llama_batch, _ id: llama_token, _ pos: llama_pos, _ seq_ids: [llama_seq_id], _ logits: Bool) {
+    batch.token   [Int(batch.n_tokens)] = id
+    batch.pos     [Int(batch.n_tokens)] = pos
+    batch.n_seq_id[Int(batch.n_tokens)] = Int32(seq_ids.count)
+    for i in 0..<seq_ids.count {
+        batch.seq_id[Int(batch.n_tokens)]![Int(i)] = seq_ids[i]
+    }
+    batch.logits  [Int(batch.n_tokens)] = logits ? 1 : 0
+
+    batch.n_tokens += 1
+}
+
 actor LlamaContext {
     private var model: OpaquePointer
     private var context: OpaquePointer
     private var batch: llama_batch
     private var tokens_list: [llama_token]
+
     /// This variable is used to store temporarily invalid cchars
     private var temporary_invalid_cchars: [CChar]
 
-    var n_len: Int32 = 512
+    var n_len: Int32 = 64
     var n_cur: Int32 = 0
+
     var n_decode: Int32 = 0
 
     init(model: OpaquePointer, context: OpaquePointer) {
@@ -27,25 +45,34 @@ actor LlamaContext {
     }
 
     deinit {
+        llama_batch_free(batch)
         llama_free(context)
         llama_free_model(model)
         llama_backend_free()
     }
 
-    static func createContext(path: String) throws -> LlamaContext {
+    static func create_context(path: String) throws -> LlamaContext {
         llama_backend_init(false)
-        let model_params = llama_model_default_params()
+        var model_params = llama_model_default_params()
 
+#if targetEnvironment(simulator)
+        model_params.n_gpu_layers = 0
+        print("Running on simulator, force use n_gpu_layers = 0")
+#endif
         let model = llama_load_model_from_file(path, model_params)
         guard let model else {
             print("Could not load model at \(path)")
             throw LlamaError.couldNotInitializeContext
         }
+
+        let n_threads = max(1, min(8, ProcessInfo.processInfo.processorCount - 2))
+        print("Using \(n_threads) threads")
+
         var ctx_params = llama_context_default_params()
-        ctx_params.seed = 1234
+        ctx_params.seed  = 1234
         ctx_params.n_ctx = 2048
-        ctx_params.n_threads = 8
-        ctx_params.n_threads_batch = 8
+        ctx_params.n_threads       = UInt32(n_threads)
+        ctx_params.n_threads_batch = UInt32(n_threads)
 
         let context = llama_new_context_with_model(model, ctx_params)
         guard let context else {
@@ -56,6 +83,26 @@ actor LlamaContext {
         return LlamaContext(model: model, context: context)
     }
 
+    func model_info() -> String {
+        let result = UnsafeMutablePointer<Int8>.allocate(capacity: 256)
+        result.initialize(repeating: Int8(0), count: 256)
+        defer {
+            result.deallocate()
+        }
+
+        // TODO: this is probably very stupid way to get the string from C
+
+        let nChars = llama_model_desc(model, result, 256)
+        let bufferPointer = UnsafeBufferPointer(start: result, count: Int(nChars))
+
+        var SwiftString = ""
+        for char in bufferPointer {
+            SwiftString.append(Character(UnicodeScalar(UInt8(char))))
+        }
+
+        return SwiftString
+    }
+
     func get_n_tokens() -> Int32 {
         return batch.n_tokens;
     }
@@ -79,16 +126,11 @@ actor LlamaContext {
             print(String(cString: token_to_piece(token: id) + [0]))
         }
 
-        // batch = llama_batch_init(512, 0) // done in init()
-        batch.n_tokens = Int32(tokens_list.count)
+        llama_batch_clear(&batch)
 
-        for i1 in 0..<batch.n_tokens {
+        for i1 in 0..<tokens_list.count {
             let i = Int(i1)
-            batch.token[i] = tokens_list[i]
-            batch.pos[i] = i1
-            batch.n_seq_id[Int(i)] = 1
-            batch.seq_id[Int(i)]![0] = 0
-            batch.logits[i] = 0
+            llama_batch_add(&batch, tokens_list[i], Int32(i), [0], false)
         }
         batch.logits[Int(batch.n_tokens) - 1] = 1 // true
 
@@ -141,18 +183,11 @@ actor LlamaContext {
         print(new_token_str)
         // tokens_list.append(new_token_id)
 
-        batch.n_tokens = 0
-
-        batch.token[Int(batch.n_tokens)] = new_token_id
-        batch.pos[Int(batch.n_tokens)] = n_cur
-        batch.n_seq_id[Int(batch.n_tokens)] = 1
-        batch.seq_id[Int(batch.n_tokens)]![0] = 0
-        batch.logits[Int(batch.n_tokens)] = 1 // true
-        batch.n_tokens += 1
+        llama_batch_clear(&batch)
+        llama_batch_add(&batch, new_token_id, n_cur, [0], true)
 
         n_decode += 1
-
-        n_cur += 1
+        n_cur    += 1
 
         if llama_decode(context, batch) != 0 {
             print("failed to evaluate llama!")
@@ -161,14 +196,111 @@ actor LlamaContext {
         return new_token_str
     }
 
+    func bench(pp: Int, tg: Int, pl: Int, nr: Int = 1) -> String {
+        var pp_avg: Double = 0
+        var tg_avg: Double = 0
+
+        var pp_std: Double = 0
+        var tg_std: Double = 0
+
+        for r in 0..<nr {
+            // bench prompt processing
+
+            llama_batch_clear(&batch)
+
+            let n_tokens = pp
+
+            for i in 0..<n_tokens {
+                llama_batch_add(&batch, 0, Int32(i), [0], false)
+            }
+            batch.logits[Int(batch.n_tokens) - 1] = 1 // true
+
+            llama_kv_cache_clear(context)
+
+            let t_pp_start = ggml_time_us()
+
+            if llama_decode(context, batch) != 0 {
+                print("llama_decode() failed during prompt")
+            }
+
+            let t_pp_end = ggml_time_us()
+
+            // bench text generation
+
+            llama_kv_cache_clear(context)
+
+            let t_tg_start = ggml_time_us()
+
+            for i in 0..<tg {
+                llama_batch_clear(&batch)
+
+                for j in 0..<pl {
+                    llama_batch_add(&batch, 0, Int32(i), [Int32(j)], true)
+                }
+
+                if llama_decode(context, batch) != 0 {
+                    print("llama_decode() failed during text generation")
+                }
+            }
+
+            let t_tg_end = ggml_time_us()
+
+            llama_kv_cache_clear(context)
+
+            let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0
+            let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0
+
+            let speed_pp = Double(pp)    / t_pp
+            let speed_tg = Double(pl*tg) / t_tg
+
+            pp_avg += speed_pp
+            tg_avg += speed_tg
+
+            pp_std += speed_pp * speed_pp
+            tg_std += speed_tg * speed_tg
+
+            print("pp \(speed_pp) t/s, tg \(speed_tg) t/s")
+        }
+
+        pp_avg /= Double(nr)
+        tg_avg /= Double(nr)
+
+        if nr > 1 {
+            pp_std = sqrt(pp_std / Double(nr - 1) - pp_avg * pp_avg * Double(nr) / Double(nr - 1))
+            tg_std = sqrt(tg_std / Double(nr - 1) - tg_avg * tg_avg * Double(nr) / Double(nr - 1))
+        } else {
+            pp_std = 0
+            tg_std = 0
+        }
+
+        let model_desc     = model_info();
+        let model_size     = String(format: "%.2f GiB", Double(llama_model_size(model)) / 1024.0 / 1024.0 / 1024.0);
+        let model_n_params = String(format: "%.2f B", Double(llama_model_n_params(model)) / 1e9);
+        let backend        = "Metal";
+        let pp_avg_str     = String(format: "%.2f", pp_avg);
+        let tg_avg_str     = String(format: "%.2f", tg_avg);
+        let pp_std_str     = String(format: "%.2f", pp_std);
+        let tg_std_str     = String(format: "%.2f", tg_std);
+
+        var result = ""
+
+        result += String("| model | size | params | backend | test | t/s |\n")
+        result += String("| --- | --- | --- | --- | --- | --- |\n")
+        result += String("| \(model_desc) | \(model_size) | \(model_n_params) | \(backend) | pp \(pp) | \(pp_avg_str) ± \(pp_std_str) |\n")
+        result += String("| \(model_desc) | \(model_size) | \(model_n_params) | \(backend) | tg \(tg) | \(tg_avg_str) ± \(tg_std_str) |\n")
+
+        return result;
+    }
+
     func clear() {
         tokens_list.removeAll()
         temporary_invalid_cchars.removeAll()
+        llama_kv_cache_clear(context)
     }
 
     private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
         let utf8Count = text.utf8.count
-        let n_tokens = utf8Count + (add_bos ? 1 : 0)
+        let n_tokens = utf8Count + (add_bos ? 1 : 0) + 1
         let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
         let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, false)