Add batch warmup to sweep-bench

common/common.cpp

@@ -1457,6 +1457,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.warmup = false;
         return true;
     }
+    if (arg == "--warmup-batch" || arg == "-wb") {
+        params.batch_warmup = true;
+        return true;
+    }
     if (arg == "--output-format") {
         CHECK_ARG
         std::string value(argv[i]);

common/common.h

@@ -199,6 +199,7 @@ struct gpt_params {
     bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
     bool no_kv_offload     = false; // disable KV offloading
     bool warmup            = true;  // warmup run
+    bool batch_warmup      = false; // batch warmup run
     bool check_tensors     = false; // validate tensor data
     bool repack_tensors    = false; // repack tensors if interleaved variant is available
     bool use_thp           = false; // use transparent huge pages (linux only)

examples/sweep-bench/sweep-bench.cpp

@@ -107,7 +107,7 @@ int main(int argc, char ** argv) {
     llama_batch batch = llama_batch_init(n_kv_max, 0, 1);
 
     // warm up
-    {
+    if (params.warmup) {
         llama_batch_add(batch, bos, 0, { 0 }, false);
 
         if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
@@ -115,6 +115,22 @@ int main(int argc, char ** argv) {
             return 1;
         }
     }
+    if (params.batch_warmup) {
+        // clean up KV cache after generation
+        llama_kv_cache_seq_rm(ctx, 0, params.n_ubatch, -1);
+
+        // prepare batch of pp size for prompt processing performance measurement
+        llama_batch_clear(batch);
+
+        for (unsigned int i = 0; i < params.n_ubatch; ++i) {
+            llama_batch_add(batch, std::rand() % n_vocab, i, { 0 }, false);
+        }
+
+        if (!decode_helper(ctx, batch, ctx_params.n_ubatch)) {
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            return 1;
+        }
+    }
 
     llama_batch_clear(batch);
     llama_kv_cache_clear(ctx);