From 099f7077a146917a78e136b4bbcf19ab134961d5 Mon Sep 17 00:00:00 2001
From: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
Date: Sat, 26 Oct 2024 14:58:54 -0700
Subject: [PATCH] Fix deepseek deseret regex (#7369)

On windows compiled with gcc the c++ regex library failed to handle
the characters
---
 llama/llama-vocab.cpp                         |  2 +-
 .../0012-fix-deepseek-deseret-regex.patch     | 66 +++++++++++++++++++
 llama/unicode.cpp                             | 21 ++++++
 3 files changed, 88 insertions(+), 1 deletion(-)
 create mode 100644 llama/patches/0012-fix-deepseek-deseret-regex.patch

diff --git a/llama/llama-vocab.cpp b/llama/llama-vocab.cpp
index 6932de44..eff70883 100644
--- a/llama/llama-vocab.cpp
+++ b/llama/llama-vocab.cpp
@@ -415,7 +415,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
             case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM:
                 regex_exprs = {
                     "[\r\n]",
-                    "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿDŽ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿff-stﬓ-ﬗＡ-Ｚａ-ｚ𐐀-𐑏𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+",
+                    "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿDŽ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿff-stﬓ-ﬗＡ-Ｚａ-ｚ\U00010400-\U0001044f𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+",
                     "\\s?[!-/:-~！-／：-～‘-‟　-。]+",
                     "\\s+$",
                     "[一-龥ࠀ-一가-퟿]+",
diff --git a/llama/patches/0012-fix-deepseek-deseret-regex.patch b/llama/patches/0012-fix-deepseek-deseret-regex.patch
new file mode 100644
index 00000000..c42cf355
--- /dev/null
+++ b/llama/patches/0012-fix-deepseek-deseret-regex.patch
@@ -0,0 +1,66 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Daniel Hiltgen <daniel@ollama.com>
+Date: Fri, 25 Oct 2024 16:25:18 -0700
+Subject: [PATCH] fix deepseek deseret regex
+
+On windows compiled with gcc the c++ regex library failed to handle
+the characters
+---
+ src/llama-vocab.cpp |  2 +-
+ src/unicode.cpp     | 21 +++++++++++++++++++++
+ 2 files changed, 22 insertions(+), 1 deletion(-)
+
+diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
+index d2f34ddd..3ef6af19 100644
+--- a/src/llama-vocab.cpp
++++ b/src/llama-vocab.cpp
+@@ -389,7 +389,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
+             case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM:
+                 regex_exprs = {
+                     "[\r\n]",
+-                    "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿDŽ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿff-stﬓ-ﬗＡ-Ｚａ-ｚ𐐀-𐑏𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+",
++                    "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿDŽ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿff-stﬓ-ﬗＡ-Ｚａ-ｚ\U00010400-\U0001044f𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+",
+                     "\\s?[!-/:-~！-／：-～‘-‟　-。]+",
+                     "\\s+$",
+                     "[一-龥ࠀ-一가-퟿]+",
+diff --git a/src/unicode.cpp b/src/unicode.cpp
+index f4e941cd..9d78ff16 100644
+--- a/src/unicode.cpp
++++ b/src/unicode.cpp
+@@ -2,6 +2,11 @@
+ #define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
+ #endif
+ 
++#if defined(_WIN32)
++#define WIN32_LEAN_AND_MEAN
++#include <windows.h>
++#endif
++
+ #include "unicode.h"
+ #include "unicode-data.h"
+ 
+@@ -201,8 +206,24 @@ static std::unordered_map<std::string, uint8_t> unicode_utf8_to_byte_map() {
+ }
+ 
+ static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
++#ifdef _WIN32
++    int wlen = MultiByteToWideChar(CP_UTF8, 0, s.c_str(), -1, NULL, 0);
++    if (!wlen) {
++        throw std::invalid_argument("failed to convert regex");
++    }
++    wchar_t * wbuf = (wchar_t *) malloc(wlen * sizeof(wchar_t));
++    wlen = MultiByteToWideChar(CP_UTF8, 0, s.c_str(), -1, wbuf, wlen);
++    if (!wlen) {
++        free(wbuf);
++        throw std::invalid_argument("failed to convert regex");
++    }
++    std::wstring ret = std::wstring(wbuf);
++    free(wbuf);
++    return ret;
++#else
+     std::wstring_convert<std::codecvt_utf8<wchar_t>> conv;
+     return conv.from_bytes(s);
++#endif
+ }
+ 
+ static std::vector<std::string> unicode_byte_encoding_process(const std::vector<std::string> & bpe_words) {
diff --git a/llama/unicode.cpp b/llama/unicode.cpp
index d7af3686..4a8946d2 100644
--- a/llama/unicode.cpp
+++ b/llama/unicode.cpp
@@ -28,6 +28,11 @@
 #define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
 #endif
 
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#endif
+
 #include "unicode.h"
 #include "unicode-data.h"
 
@@ -227,8 +232,24 @@ static std::unordered_map<std::string, uint8_t> unicode_utf8_to_byte_map() {
 }
 
 static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
+#ifdef _WIN32
+    int wlen = MultiByteToWideChar(CP_UTF8, 0, s.c_str(), -1, NULL, 0);
+    if (!wlen) {
+        throw std::invalid_argument("failed to convert regex");
+    }
+    wchar_t * wbuf = (wchar_t *) malloc(wlen * sizeof(wchar_t));
+    wlen = MultiByteToWideChar(CP_UTF8, 0, s.c_str(), -1, wbuf, wlen);
+    if (!wlen) {
+        free(wbuf);
+        throw std::invalid_argument("failed to convert regex");
+    }
+    std::wstring ret = std::wstring(wbuf);
+    free(wbuf);
+    return ret;
+#else
     std::wstring_convert<std::codecvt_utf8<wchar_t>> conv;
     return conv.from_bytes(s);
+#endif
 }
 
 static std::vector<std::string> unicode_byte_encoding_process(const std::vector<std::string> & bpe_words) {