From 6915fa3441e3ab0026d996e3c6c100930b1e5dda Mon Sep 17 00:00:00 2001
From: Morgan Funtowicz <funtowiczmo@gmail.com>
Date: Sat, 9 Nov 2024 22:19:38 +0100
Subject: [PATCH] feat(backend): remove reinterpret_cast converting from
 uint32_t to llama_token(int32_t)

---
 backends/llamacpp/csrc/ffi.hpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/backends/llamacpp/csrc/ffi.hpp b/backends/llamacpp/csrc/ffi.hpp
index 51a524cb..70669b7c 100644
--- a/backends/llamacpp/csrc/ffi.hpp
+++ b/backends/llamacpp/csrc/ffi.hpp
@@ -56,8 +56,8 @@ namespace huggingface::tgi::backends::llamacpp {
             };
 
             // Ask the compiler to create view over Rust slice transmuting from uint32_t* to llama_token*
-            auto input_tokens_v =
-                    std::span(reinterpret_cast<const llama_token *>(input_tokens.data()), input_tokens.size());
+            auto input_tokens_v = std::vector<llama_token>(input_tokens.size());
+            std::memcpy(input_tokens_v.data(), input_tokens.data(), input_tokens.size());
 
             const auto generation_context = generation_context_t {generation_params, sampling_params, input_tokens_v};
             if(const auto result = worker_.generate(generation_context, context_forwarding_callback); result.has_value()) [[likely]] {