From f227e918f95d1f68d17560de9e3dd1cac529b084 Mon Sep 17 00:00:00 2001
From: Ettore Di Giacinto <mudler@users.noreply.github.com>
Date: Wed, 25 Oct 2023 20:56:25 +0200
Subject: [PATCH] feat(llama.cpp): Bump llama.cpp, adapt grpc server (#1211)

* feat(llama.cpp): Bump llama.cpp, adapt grpc server

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* ci: fixups

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

---------

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---
 .github/workflows/test-gpu.yml    |  5 ++++-
 Makefile                          |  2 +-
 backend/cpp/llama/grpc-server.cpp | 16 ++++++++--------
 3 files changed, 13 insertions(+), 10 deletions(-)

diff --git a/.github/workflows/test-gpu.yml b/.github/workflows/test-gpu.yml
index 3d126bbf7869..34d3cafb79b8 100644
--- a/.github/workflows/test-gpu.yml
+++ b/.github/workflows/test-gpu.yml
@@ -40,6 +40,8 @@ jobs:
           if [ ! -e /run/systemd/system ]; then
             sudo mkdir /run/systemd/system
           fi
+          sudo mkdir -p /host/tests/${{ github.head_ref || github.ref }}
+          sudo chmod -R 777 /host/tests/${{ github.head_ref || github.ref }}
           make \
             TEST_DIR="/host/tests/${{ github.head_ref || github.ref }}" \
             BUILD_TYPE=cublas \
@@ -57,4 +59,5 @@ jobs:
           make \
             TEST_DIR="/host/tests/${{ github.head_ref || github.ref }}" \
             teardown-e2e || true
-          docker system prune -f -a --volumes || true
\ No newline at end of file
+          sudo rm -rf /host/tests/${{ github.head_ref || github.ref }}
+          docker system prune -f -a --volumes || true
diff --git a/Makefile b/Makefile
index f944b00252f3..9d8d3e957db4 100644
--- a/Makefile
+++ b/Makefile
@@ -8,7 +8,7 @@ GOLLAMA_VERSION?=aeba71ee842819da681ea537e78846dc75949ac0
 
 GOLLAMA_STABLE_VERSION?=50cee7712066d9e38306eccadcfbb44ea87df4b7
 
-CPPLLAMA_VERSION?=96981f37b1e3f450d9e63e571514217bf60f0a7f
+CPPLLAMA_VERSION?=9d02956443e5c1ded29b7b5ed8a21bc01ba6f563
 
 # gpt4all version
 GPT4ALL_REPO?=https://github.com/nomic-ai/gpt4all
diff --git a/backend/cpp/llama/grpc-server.cpp b/backend/cpp/llama/grpc-server.cpp
index 50d59ab1f34e..c177fa73b3cf 100644
--- a/backend/cpp/llama/grpc-server.cpp
+++ b/backend/cpp/llama/grpc-server.cpp
@@ -275,11 +275,11 @@ struct llama_server_context
         if (suff_rm_leading_spc  && suffix_tokens[0] == space_token) {
             suffix_tokens.erase(suffix_tokens.begin());
         }
-        prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(ctx));
-        prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(ctx)); // always add BOS
-        prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(ctx));
+        prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(model));
+        prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS
+        prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(model));
         prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
-        prefix_tokens.push_back(llama_token_middle(ctx));
+        prefix_tokens.push_back(llama_token_middle(model));
 
         auto prompt_tokens = prefix_tokens;
 
@@ -419,7 +419,7 @@ struct llama_server_context
         if (params.n_predict == 0)
         {
             has_next_token = false;
-            result.tok = llama_token_eos(ctx);
+            result.tok = llama_token_eos(model);
             return result;
         }
 
@@ -453,7 +453,7 @@ struct llama_server_context
         // decrement remaining sampling budget
         --n_remain;
 
-        if (!embd.empty() && embd.back() == llama_token_eos(ctx))
+        if (!embd.empty() && embd.back() == llama_token_eos(model))
         {
             // stopping_word = llama_token_to_piece(ctx, embd.back());
             has_next_token = false;
@@ -594,7 +594,7 @@ static void parse_options_completion(bool streaming,const backend::PredictOption
 
     if (predict->ignoreeos())
     {
-        llama.params.sparams.logit_bias[llama_token_eos(llama.ctx)] = -INFINITY;
+        llama.params.sparams.logit_bias[llama_token_eos(llama.model)] = -INFINITY;
     }
 
     // const auto &logit_bias = body.find("logit_bias");
@@ -676,7 +676,7 @@ static void params_parse(const backend::ModelOptions* request,
 }
 
 static bool is_at_eob(llama_server_context &server_context, const llama_token *tokens, const size_t n_tokens) {
-    return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.ctx);
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.model);
 }
 
 // Function matching type llama_beam_search_callback_fn_t.