update(llama.cpp): with Llama v3.1 model support

Author: grencez

dep/cmake_fetchcontent/llama_cpp.cmake

@@ -1,7 +1,7 @@
 FetchContent_Declare(
   LlamaCpp
   GIT_REPOSITORY "https://github.com/ggerganov/llama.cpp.git"
-  GIT_TAG "50e05353e88d50b644688caa91f5955e8bdb9eb9"
+  GIT_TAG "b5e95468b1676e1e5c9d80d1eeeb26f542a38f42"
 )
 
 set(GGML_OPENMP FALSE CACHE BOOL "OpenMP off for compatibility.")

example/prompt/README.md

@@ -12,6 +12,7 @@ In order of interest:
   - [assistant_alpaca](assistant_alpaca/): Alpaca prompt format.
   - [assistant_chatml](assistant_chatml/): ChatML prompt format that typically requires special `<|im_start|>` and `<|im_end|>` tokens but is configured with fallbacks.
   - [assistant_gemma](assistant_gemma/): Gemma prompt format that requires special `<start_of_turn>` and `<end_of_turn>` tokens.
+  - [assistant_llama](assistant_llama/): Llama 3 prompt format that requires special `<|start_header_id|>`, `<|end_header_id|>`, and `<|eot_id|>` tokens.
   - [assistant_mistral](assistant_mistral/): Mistral propmt format that requires special `[INST]` and `[/INST]` tokens.
   - [assistant_vicuna](assistant_vicuna/): Vicuna prompt format.
 - [assistant_coprocess](assistant_coprocess/): A simple assistant that can be controlled as a coprocess.

example/prompt/assistant_llama/README.md

@@ -0,0 +1,4 @@
+# Llama Assistant
+
+This example should be run with Llama 3 models that are tuned to behave like an instruction-following assistant chatbot.
+Most importantly, the model must have special `<|start_header_id|>`, `<|end_header_id|>`, and `<|eot_id|>` tokens.

example/prompt/assistant_llama/priming.txt

@@ -0,0 +1,6 @@
+<|start_header_id|>system<|end_header_id|>
+
+Cutting Knowledge Date: December 2023
+Today Date: July 2024
+
+You are a helpful assistant.<|eot_id|>

example/prompt/assistant_llama/rolling.txt

@@ -0,0 +1,3 @@
+<|start_header_id|>user<|end_header_id|>
+
+Hello!<|eot_id|>

example/prompt/assistant_llama/setting.sxpb

@@ -0,0 +1,37 @@
+
+; Newlines are included after <|eot_id|> to make the chat easier to read,
+; but Llama 3 instruct prompt format does not actually include them.
+(chat_prefixes (())
+ (m
+  (prefix "<|start_header_id|>user<|end_header_id|>\n\n")
+  (suffix "<|eot_id|>\n")
+ )
+ (m
+  (prefix "<|start_header_id|>assistant<|end_header_id|>\n\n")
+  (suffix "<|eot_id|>\n")
+ )
+)
+
+(substitution
+ (special_tokens (())
+  (() (alias "<|start_header_id|>"))
+  (() (alias "<|end_header_id|>"))
+  (() (alias "<|eot_id|>"))
+ )
+)
+
+(x_priming "priming.txt")
+(x_rolling "rolling.txt")
+(o_rolling "../../../bld/example/prompt/assistant_llama.txt")
+
+; No starting space.
+(startspace_on +false)
+; No token penalization.
+(repeat_window 0)
+
+; 10 reasonably-long sentences at a time.
+(sentence_limit 10)
+(sentence_token_limit 100)
+
+; Limit context to avoid blowing up RAM on large context models.
+(model_token_limit 8000)

test/example/prompt/CMakeLists.txt

@@ -17,6 +17,8 @@ add_test(NAME example_prompt_assistant_coprocess_parse_test COMMAND example_prom
   "${PROJECT_SOURCE_DIR}/example/prompt/assistant_coprocess/setting.sxpb")
 add_test(NAME example_prompt_assistant_gemma_parse_test COMMAND example_prompt_parse_test
   "${PROJECT_SOURCE_DIR}/example/prompt/assistant_gemma/setting.sxpb")
+add_test(NAME example_prompt_assistant_llama_parse_test COMMAND example_prompt_parse_test
+  "${PROJECT_SOURCE_DIR}/example/prompt/assistant_llama/setting.sxpb")
 add_test(NAME example_prompt_assistant_mistral_parse_test COMMAND example_prompt_parse_test
   "${PROJECT_SOURCE_DIR}/example/prompt/assistant_mistral/setting.sxpb")
 add_test(NAME example_prompt_assistant_plain_parse_test COMMAND example_prompt_parse_test