From 558c7c3c6db2e755939245df574a7d3f1629d6e6 Mon Sep 17 00:00:00 2001
From: antirez <antirez@gmail.com>
Date: Wed, 27 Dec 2023 18:54:36 +0100
Subject: [PATCH] Clarify the need for FP16 implementation.

---
 fp16.c | 13 +++++++++----
 fp16.h |  2 ++
 2 files changed, 11 insertions(+), 4 deletions(-)

diff --git a/fp16.c b/fp16.c
index 9a62719..fa90605 100644
--- a/fp16.c
+++ b/fp16.c
@@ -1,11 +1,16 @@
-#include <stdint.h>
-#include <math.h>
-
-/* This code comes originally from:
+/* Conversion from floats to FP16 and the other way around.
+ * This is useful as in GGUF files we have both FP16 tensors
+ * and quantized blocks where half-precisions floats are used
+ * to store the scaling factor (delta) and other parameters.
+ *
+ * This code comes originally from:
  * https://github.com/Maratyszcza/FP16/blob/master/include/fp16/fp16.h
  *
  * The original code is MIT licensed. */
 
+#include <stdint.h>
+#include <math.h>
+
 static inline float fp32_from_bits(uint32_t w) {
     union {
         uint32_t as_bits;
diff --git a/fp16.h b/fp16.h
index 32543b6..ad26ea7 100644
--- a/fp16.h
+++ b/fp16.h
@@ -1,3 +1,5 @@
+/* See fp16.c. */
+
 #ifdef FP16_h
 #define FP16_h
 float from_half(uint16_t h);