all : be more strict about converting float to double (#458)

* Be more strict about converting float to double

* Test equivalence of round, SILU implementations

Test module is commented out in CMakeLists.txt because the tests may
take a long time, depending on how much the compiler optimizes.

* Fix softmax in perplexity.cpp

* all : prefer float over double where appropriate

* perplexity : add <cmath>

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

4 files changed, 23 insertions, 18 deletions

diff --git a/examples/common.cpp b/examples/common.cpp
index 880ebe9..af3ad9e 100644
--- a/examples/common.cpp
+++ b/examples/common.cpp
@@ -215,13 +215,13 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "                        prompt file to start generation.\n");
     fprintf(stderr, "  -n N, --n_predict N   number of tokens to predict (default: %d, -1 = infinity)\n", params.n_predict);
     fprintf(stderr, "  --top_k N             top-k sampling (default: %d)\n", params.top_k);
-    fprintf(stderr, "  --top_p N             top-p sampling (default: %.1f)\n", params.top_p);
+    fprintf(stderr, "  --top_p N             top-p sampling (default: %.1f)\n", (double)params.top_p);
     fprintf(stderr, "  --repeat_last_n N     last n tokens to consider for penalize (default: %d)\n", params.repeat_last_n);
-    fprintf(stderr, "  --repeat_penalty N    penalize repeat sequence of tokens (default: %.1f)\n", params.repeat_penalty);
+    fprintf(stderr, "  --repeat_penalty N    penalize repeat sequence of tokens (default: %.1f)\n", (double)params.repeat_penalty);
     fprintf(stderr, "  -c N, --ctx_size N    size of the prompt context (default: %d)\n", params.n_ctx);
     fprintf(stderr, "  --ignore-eos          ignore end of stream token and continue generating\n");
     fprintf(stderr, "  --memory_f32          use f32 instead of f16 for memory key+value\n");
-    fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", params.temp);
+    fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
     fprintf(stderr, "  --n_parts N           number of model parts (default: -1 = determine from dimensions)\n");
     fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
     fprintf(stderr, "  --perplexity          compute perplexity over the prompt\n");
diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index d5ab2cf..3130aef 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -209,7 +209,8 @@ int main(int argc, char ** argv) {
             fprintf(stderr, "Input prefix: '%s'\n", params.input_prefix.c_str());
         }
     }
-    fprintf(stderr, "sampling: temp = %f, top_k = %d, top_p = %f, repeat_last_n = %i, repeat_penalty = %f\n", params.temp, params.top_k, params.top_p, params.repeat_last_n, params.repeat_penalty);
+    fprintf(stderr, "sampling: temp = %f, top_k = %d, top_p = %f, repeat_last_n = %i, repeat_penalty = %f\n",
+        params.temp, params.top_k, params.top_p, params.repeat_last_n, params.repeat_penalty);
     fprintf(stderr, "generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
     fprintf(stderr, "\n\n");
 
@@ -274,10 +275,10 @@ int main(int argc, char ** argv) {
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
             // out of user input, sample next token
-            const float top_k          = params.top_k;
-            const float top_p          = params.top_p;
-            const float temp           = params.temp;
-            const float repeat_penalty = params.repeat_penalty;
+            const int32_t top_k          = params.top_k;
+            const float   top_p          = params.top_p;
+            const float   temp           = params.temp;
+            const float   repeat_penalty = params.repeat_penalty;
 
             llama_token id = 0;
 
diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp
index 75d526d..07ed0a8 100644
--- a/examples/perplexity/perplexity.cpp
+++ b/examples/perplexity/perplexity.cpp
@@ -1,15 +1,17 @@
 #include "common.h"
 #include "llama.h"
 
-std::vector<double> softmax(const std::vector<float>& logits) {
-    std::vector<double> probs(logits.size());
+#include <cmath>
+
+std::vector<float> softmax(const std::vector<float>& logits) {
+    std::vector<float> probs(logits.size());
     float max_logit = logits[0];
     for (float v : logits) max_logit = std::max(max_logit, v);
     double sum_exp = 0.0;
     for (size_t i = 0; i < logits.size(); i++) {
         // Subtract the maximum logit value from the current logit value for numerical stability
-        float logit = logits[i] - max_logit;
-        double exp_logit = std::exp(logit);
+        const float logit = logits[i] - max_logit;
+        const float exp_logit = expf(logit);
         sum_exp += exp_logit;
         probs[i] = exp_logit;
     }
@@ -24,14 +26,16 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
     auto tokens = ::llama_tokenize(ctx, params.prompt, true);
 
     int count = 0;
-    double nll = 0.0;
     int seq_count = tokens.size() / params.n_ctx;
 
+    double nll = 0.0;
+
     fprintf(stderr, "%s : calculating perplexity over %d chunks\n", __func__, seq_count);
 
     for (int i = 0; i < seq_count; ++i) {
         int start = i * params.n_ctx;
-        int end = start + params.n_ctx - 1;
+        int end = start + params.n_ctx - 1; // TODO: this is not optimal, e.g. it makes the batch 511 instead of 512
+                                            //       it is better to always be power of 2 for better performance
         std::vector<llama_token> embd(tokens.begin() + start, tokens.begin() + end);
         auto start_t = std::chrono::high_resolution_clock::now();
         if (llama_eval(ctx, embd.data(), embd.size(), 0, params.n_threads)) {
@@ -40,7 +44,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
         }
         auto end_t = std::chrono::high_resolution_clock::now();
         if (i == 0) {
-            double seconds = std::chrono::duration<double>(end_t - start_t).count();
+            const float seconds = std::chrono::duration<float>(end_t - start_t).count();
             printf("%.2f seconds per pass - ETA %.2f hours\n", seconds, (seconds * seq_count) / (60.0*60.0));
         }
         // We get the logits for all the tokens in the context window (params.n_ctx)
@@ -63,7 +67,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
             std::vector<float> tok_logits(
                 logits + j * n_vocab,
                 logits + (j + 1) * n_vocab);
-            double prob = softmax(tok_logits)[tokens[start + j + 1]];
+            const float prob = softmax(tok_logits)[tokens[start + j + 1]];
             nll += -std::log(prob);
             ++count;
         }
diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp
index 3888ff5..b444328 100644
--- a/examples/quantize/quantize.cpp
+++ b/examples/quantize/quantize.cpp
@@ -50,8 +50,8 @@ int main(int argc, char ** argv) {
         const int64_t t_main_end_us = ggml_time_us();
 
         printf("\n");
-        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
-        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0);
     }
 
     return 0;