From ff2c64a9f4190cd5da3c59f65d8417299c72a336 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Thu, 25 Apr 2024 15:51:46 +0300
Subject: [PATCH] tests : remove TMP_ATTN_BENCH

ggml-ci
---
 tests/test-backend-ops.cpp | 70 --------------------------------------
 1 file changed, 70 deletions(-)

diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index d044a6ea0..b27c1291e 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -15,9 +15,6 @@
 #include <thread>
 #include <vector>
 
-// TODO: remove before merging
-//#define TMP_ATTN_BENCH
-
 static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float max = 1.0f) {
     // static RNG initialization (revisit if n_threads stops being constant)
     static const size_t n_threads = std::thread::hardware_concurrency();
@@ -574,19 +571,9 @@ struct test_case {
         // duplicate the op
         size_t target_size = ggml_backend_is_cpu(backend) ? 1ULL << 33 : 1ULL << 35; // 8 GB CPU, 32 GB GPU
         int n_runs = std::min((size_t)gf->size - gf->n_nodes, target_size / op_size(out)) + 1;
-#ifndef TMP_ATTN_BENCH
         for (int i = 1; i < n_runs; i++) {
             gf->nodes[gf->n_nodes++] = out;
         }
-#else
-        int n_nodes = gf->n_nodes;
-        n_runs = 1000;
-        for (int i = 1; i < n_runs; i++) {
-            for (int j = 0; j < n_nodes; j++) {
-                gf->nodes[gf->n_nodes++] = gf->nodes[j];
-            }
-        }
-#endif
 
         // calculate memory
         size_t mem = n_runs * op_size(out);
@@ -1522,50 +1509,6 @@ struct test_flash_attn_ext : public test_case {
     }
 };
 
-#ifdef TMP_ATTN_BENCH
-// ATTN
-struct test_attn : public test_case {
-    const int64_t hs; // head size
-    const int64_t nh; // num heads
-    const int64_t kv; // kv size
-    const int64_t nb; // batch size
-
-    std::string op_desc(ggml_tensor * t) override {
-        return "ATTN";
-
-        GGML_UNUSED(t);
-    }
-
-    std::string vars() override {
-        return VARS_TO_STR4(hs, nh, kv, nb);
-    }
-
-    double max_nmse_err() override {
-        return 5e-4;
-    }
-
-    test_attn(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8)
-        : hs(hs), nh(nh), kv(kv), nb(nb) {}
-
-    ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * q = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, hs, nb, nh, 1);
-        ggml_tensor * k = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, hs, kv, nh, 1);
-        ggml_tensor * v = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, hs, nh, 1); // transposed
-        ggml_tensor * mask = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, nb, 1, 1);
-
-        struct ggml_tensor * cur;
-
-        cur = ggml_mul_mat     (ctx, k, q);
-        cur = ggml_soft_max_ext(ctx, cur, mask, nullptr, 1.0f/sqrtf(hs), 0.0f);
-        cur = ggml_mul_mat     (ctx, v, cur);
-        cur = ggml_permute     (ctx, cur, 0, 2, 1, 3);
-        cur = ggml_cont_2d     (ctx, cur, hs*nh, nb);
-
-        return cur;
-    }
-};
-#endif
-
 enum llm_norm_type {
     LLM_NORM,
     LLM_NORM_RMS,
@@ -2230,18 +2173,6 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_timestep_embedding());
     test_cases.emplace_back(new test_leaky_relu());
 
-#ifdef TMP_ATTN_BENCH
-    for (int hs : { 128, 256, 64, 80, }) {
-        for (int nh : { 32, }) {
-            for (int kv : { 512, 1024, 2048, 4096, }) {
-                for (int nb : { 1, 2, 4, 8, 512, 1024, 2048, }) {
-                    test_cases.emplace_back(new test_attn          (hs, nh, kv, nb));
-                    test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb));
-                }
-            }
-        }
-    }
-#else
     for (int hs : { 64, 80, 128, 256, }) {
         for (int nh : { 32, }) {
             for (int kv : { 512, 1024, }) {
@@ -2251,7 +2182,6 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
             }
         }
     }
-#endif
 
     // these tests are disabled to save execution time, but they can be handy for debugging
 #if 0