From 67be2ce1015d070b3b2cd488bcb041eefb61de72 Mon Sep 17 00:00:00 2001
From: slaren <slarengh@gmail.com>
Date: Sun, 3 Mar 2024 14:26:18 +0100
Subject: [PATCH 1/3] cuda : fix data race in soft max (#5853)

---
 ggml-cuda.cu | 1 +
 1 file changed, 1 insertion(+)

diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 7ed97430f..04c6cb1b8 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -6904,6 +6904,7 @@ static __global__ void soft_max_f32(const float * x, const float * mask, const f
     // find the sum of exps in the block
     tmp = warp_reduce_sum(tmp);
     if (block_size > WARP_SIZE) {
+        __syncthreads();
         if (warp_id == 0) {
             buf_iw[lane_id] = 0.0f;
         }

From 5a51cc1bb4592f0d71f9af89cd08b11a066ba447 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?DAN=E2=84=A2?= <dranger003@gmail.com>
Date: Mon, 4 Mar 2024 02:57:20 -0500
Subject: [PATCH 2/3] main : support special tokens as reverse/anti prompt
 (#5847)

* Support special tokens as reverse/anti prompt.

* Tokenize antiprompts only once.

* main : minor

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---
 examples/main/main.cpp | 20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)

diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index 34e84d0d4..47059e582 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -511,6 +511,14 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
 
+    // tokenized antiprompts
+    std::vector<std::vector<llama_token>> antiprompt_ids;
+
+    antiprompt_ids.reserve(params.antiprompt.size());
+    for (const std::string & antiprompt : params.antiprompt) {
+        antiprompt_ids.emplace_back(::llama_tokenize(ctx, antiprompt, false, true));
+    }
+
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
 
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
@@ -769,6 +777,18 @@ int main(int argc, char ** argv) {
                     }
                 }
 
+                // check for reverse prompt using special tokens
+                llama_token last_token = llama_sampling_last(ctx_sampling);
+                for (std::vector<llama_token> ids : antiprompt_ids) {
+                    if (ids.size() == 1 && last_token == ids[0]) {
+                        if (params.interactive) {
+                            is_interacting = true;
+                        }
+                        is_antiprompt = true;
+                        break;
+                    }
+                }
+
                 if (is_antiprompt) {
                     LOG("found antiprompt: %s\n", last_output.c_str());
                 }

From 82f3e668adafba647de703f835991e91a96b5ac4 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?DAN=E2=84=A2?= <dranger003@gmail.com>
Date: Mon, 4 Mar 2024 03:08:19 -0500
Subject: [PATCH 3/3] common : use LLAMA_DEFAULT_SEED (#5855)

---
 common/common.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/common/common.h b/common/common.h
index d3682b7ad..b2868833b 100644
--- a/common/common.h
+++ b/common/common.h
@@ -43,7 +43,7 @@ extern char const *LLAMA_BUILD_TARGET;
 int32_t get_num_physical_cores();
 
 struct gpt_params {
-    uint32_t seed                 = -1;    // RNG seed
+    uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
     int32_t n_threads             = get_num_physical_cores();
     int32_t n_threads_draft       = -1;