Merge branch 'master' of github.com:ggerganov/llama.cpp into grammar-example

* 'master' of github.com:ggerganov/llama.cpp: (24 commits)
  convert : fix Baichuan2 models by using vocab size in config.json (ggml-org#3299)
  readme : add project status link
  ggml : fix build after ggml-org#3329
  llm : add Refact model (ggml-org#3329)
  sync : ggml (conv 1d + 2d updates, UB fixes) (ggml-org#3468)
  finetune : readme fix typo (ggml-org#3465)
  ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (ggml-org#3453)
  main : consistent prefix/suffix coloring (ggml-org#3425)
  llama : fix session saving/loading (ggml-org#3400)
  llama : expose model's rope_freq_scale in the API (ggml-org#3418)
  metal : alibi for arbitrary number of heads (ggml-org#3426)
  cmake : make LLAMA_NATIVE flag actually use the instructions supported by the processor (ggml-org#3273)
  Work on the BPE tokenizer (ggml-org#3252)
  convert : fix vocab size when not defined in hparams (ggml-org#3421)
  cmake : increase minimum version for add_link_options (ggml-org#3444)
  CLBlast: Add broadcast support for matrix multiplication (ggml-org#3402)
  gguf : add BERT, MPT, and GPT-J arch info (ggml-org#3408)
  gguf : general usability improvements (ggml-org#3409)
  cmake : make CUDA flags more similar to the Makefile (ggml-org#3420)
  finetune : fix ggml-org#3404 (ggml-org#3437)
  ...

Author: joelkuiper

.dockerignore

@@ -1,6 +1,9 @@
 *.o
 *.a
 .cache/
+.git/
+.github/
+.gitignore
 .vs/
 .vscode/
 .DS_Store

.github/workflows/build.yml

@@ -188,7 +188,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF ..
+          cmake ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -265,17 +265,17 @@ jobs:
       matrix:
         include:
           - build: 'noavx'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx2'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'avx'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx512'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'clblast'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
           - build: 'openblas'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
 
     steps:
       - name: Clone
@@ -414,7 +414,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON -DBUILD_SHARED_LIBS=ON
+          cmake .. -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON -DBUILD_SHARED_LIBS=ON
           cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Determine tag name

.gitignore

@@ -40,6 +40,7 @@ models-mnt
 /embedding
 /gguf
 /gguf-llama-simple
+/infill
 /libllama.so
 /llama-bench
 /main
@@ -90,4 +91,5 @@ tests/test-quantize-perf
 tests/test-sampling
 tests/test-tokenizer-0-llama
 tests/test-tokenizer-0-falcon
-tests/test-tokenizer-1
+tests/test-tokenizer-1-llama
+tests/test-tokenizer-1-bpe

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.12) # Don't bump this version for no reason
+cmake_minimum_required(VERSION 3.13)  # for add_link_options
 project("llama.cpp" C CXX)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
@@ -44,7 +44,7 @@ endif()
 
 # general
 option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
-option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      OFF)
+option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      ON)
 option(LLAMA_LTO                        "llama: enable link time optimization"                  OFF)
 
 # debug
@@ -58,15 +58,21 @@ option(LLAMA_SANITIZE_ADDRESS           "llama: enable address sanitizer"
 option(LLAMA_SANITIZE_UNDEFINED         "llama: enable undefined sanitizer"                     OFF)
 
 # instruction set specific
-option(LLAMA_AVX                        "llama: enable AVX"                                     ON)
-option(LLAMA_AVX2                       "llama: enable AVX2"                                    ON)
-option(LLAMA_AVX512                     "llama: enable AVX512"                                  OFF)
-option(LLAMA_AVX512_VBMI                "llama: enable AVX512-VBMI"                             OFF)
-option(LLAMA_AVX512_VNNI                "llama: enable AVX512-VNNI"                             OFF)
-option(LLAMA_FMA                        "llama: enable FMA"                                     ON)
+if (LLAMA_NATIVE)
+    set(INS_ENB OFF)
+else()
+    set(INS_ENB ON)
+endif()
+
+option(LLAMA_AVX                             "llama: enable AVX"                                ${INS_ENB})
+option(LLAMA_AVX2                            "llama: enable AVX2"                               ${INS_ENB})
+option(LLAMA_AVX512                          "llama: enable AVX512"                             OFF)
+option(LLAMA_AVX512_VBMI                     "llama: enable AVX512-VBMI"                        OFF)
+option(LLAMA_AVX512_VNNI                     "llama: enable AVX512-VNNI"                        OFF)
+option(LLAMA_FMA                             "llama: enable FMA"                                ${INS_ENB})
 # in MSVC F16C is implied with AVX2/AVX512
 if (NOT MSVC)
-    option(LLAMA_F16C                   "llama: enable F16C"                                    ON)
+    option(LLAMA_F16C                        "llama: enable F16C"                               ${INS_ENB})
 endif()
 
 # 3rd party libs
@@ -343,8 +349,9 @@ if (LLAMA_MPI)
         set(GGML_SOURCES_MPI ggml-mpi.c ggml-mpi.h)
         add_compile_definitions(GGML_USE_MPI)
         add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
-        set(cxx_flags ${cxx_flags} -Wno-cast-qual)
-        set(c_flags   ${c_flags}   -Wno-cast-qual)
+        if (NOT MSVC)
+            add_compile_options(-Wno-cast-qual)
+        endif()
         set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${MPI_C_LIBRARIES})
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
         # Even if you're only using the C header, C++ programs may bring in MPI
@@ -418,10 +425,11 @@ if (LLAMA_ALL_WARNINGS)
         set(c_flags -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int
             -Werror=implicit-function-declaration)
         set(cxx_flags -Wmissing-declarations -Wmissing-noreturn)
+        set(host_cxx_flags "")
 
         if (CMAKE_C_COMPILER_ID MATCHES "Clang")
             set(warning_flags ${warning_flags} -Wunreachable-code-break -Wunreachable-code-return)
-            set(cxx_flags ${cxx_flags} -Wmissing-prototypes -Wextra-semi)
+            set(host_cxx_flags ${host_cxx_flags} -Wmissing-prototypes -Wextra-semi)
 
             if (
                 (CMAKE_C_COMPILER_ID STREQUAL "Clang"      AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 3.8.0) OR
@@ -431,27 +439,38 @@ if (LLAMA_ALL_WARNINGS)
             endif()
         elseif (CMAKE_C_COMPILER_ID STREQUAL "GNU")
             set(c_flags ${c_flags} -Wdouble-promotion)
-            set(cxx_flags ${cxx_flags} -Wno-array-bounds)
+            set(host_cxx_flags ${host_cxx_flags} -Wno-array-bounds)
 
             if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 7.1.0)
-                set(cxx_flags ${cxx_flags} -Wno-format-truncation)
+                set(host_cxx_flags ${host_cxx_flags} -Wno-format-truncation)
             endif()
             if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 8.1.0)
-                set(cxx_flags ${cxx_flags} -Wextra-semi)
+                set(host_cxx_flags ${host_cxx_flags} -Wextra-semi)
             endif()
         endif()
     else()
         # todo : msvc
     endif()
 
-    add_compile_options(
-            ${warning_flags}
-            "$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
-            "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>"
-    )
+    set(c_flags   ${c_flags}   ${warning_flags})
+    set(cxx_flags ${cxx_flags} ${warning_flags})
+    add_compile_options("$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
+                        "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags} ${host_cxx_flags}>")
 
 endif()
 
+if (NOT MSVC)
+    set(cuda_flags -Wno-pedantic)
+endif()
+set(cuda_flags ${cxx_flags} -use_fast_math ${cuda_flags})
+
+list(JOIN host_cxx_flags " " cuda_host_flags)  # pass host compiler flags as a single argument
+if (NOT cuda_host_flags STREQUAL "")
+    set(cuda_flags ${cuda_flags} -Xcompiler ${cuda_host_flags})
+endif()
+
+add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${cuda_flags}>")
+
 if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
 
@@ -491,9 +510,6 @@ if (NOT MSVC)
     if (LLAMA_GPROF)
         add_compile_options(-pg)
     endif()
-    if (LLAMA_NATIVE)
-        add_compile_options(-march=native)
-    endif()
 endif()
 
 if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64"))
@@ -548,6 +564,9 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE
             add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX>)
         endif()
     else()
+        if (LLAMA_NATIVE)
+            add_compile_options(-march=native)
+        endif()
         if (LLAMA_F16C)
             add_compile_options(-mf16c)
         endif()
@@ -705,6 +724,7 @@ set(LLAMA_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR}
 set(LLAMA_BUILD_NUMBER ${BUILD_NUMBER})
 set(LLAMA_BUILD_COMMIT ${BUILD_COMMIT})
 set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
+get_directory_property(LLAMA_TRANSIENT_DEFINES COMPILE_DEFINITIONS)
 
 configure_package_config_file(
         ${CMAKE_CURRENT_SOURCE_DIR}/scripts/LlamaConfig.cmake.in

Makefile

@@ -1,8 +1,8 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml simple batched save-load-state server embd-input-test gguf llama-bench baby-llama beam-search speculative benchmark-matmult parallel finetune export-lora tests/test-c.o
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml simple batched save-load-state server embd-input-test gguf llama-bench baby-llama beam-search speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o
 
 # Binaries only useful for tests
-TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama
+TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -62,9 +62,11 @@ test: $(TEST_TARGETS)
 		if [ "$$test_target" = "tests/test-tokenizer-0-llama" ]; then \
 			./$$test_target $(CURDIR)/models/ggml-vocab-llama.gguf; \
 		elif [ "$$test_target" = "tests/test-tokenizer-0-falcon" ]; then \
-			continue; \
+			./$$test_target $(CURDIR)/models/ggml-vocab-falcon.gguf; \
 		elif [ "$$test_target" = "tests/test-tokenizer-1-llama" ]; then \
 			continue; \
+		elif [ "$$test_target" = "tests/test-tokenizer-1-bpe" ]; then \
+			continue; \
 		else \
 			echo "Running test $$test_target..."; \
 			./$$test_target; \
@@ -543,6 +545,9 @@ main: examples/main/main.cpp                                  build-info.h ggml.
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
+infill: examples/infill/infill.cpp                            build-info.h ggml.o llama.o common.o console.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
@@ -667,6 +672,9 @@ tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h gg
 tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 

README.md

@@ -5,7 +5,7 @@
 [![Actions Status](https://github.com/ggerganov/llama.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/llama.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 
-[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
 
 Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 

common/common.cpp

@@ -389,6 +389,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.interactive_first = true;
         } else if (arg == "-ins" || arg == "--instruct") {
             params.instruct = true;
+        } else if (arg == "--infill") {
+            params.infill = true;
         } else if (arg == "--multiline-input") {
             params.multiline_input = true;
         } else if (arg == "--simple-io") {
@@ -921,6 +923,7 @@ std::string llama_detokenize_bpe(llama_context * ctx, const std::vector<llama_to
         result += piece;
     }
 
+    // NOTE: the original tokenizer decodes bytes after collecting the pieces.
     return result;
 }
 

common/common.h

@@ -120,6 +120,7 @@ struct gpt_params {
     bool use_mlock         = false; // use mlock to keep model in memory
     bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
+    bool infill            = false; // use infill mode
 };
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);

convert-baichuan-hf-to-gguf.py

@@ -11,11 +11,14 @@
 from pathlib import Path
 from typing import TYPE_CHECKING, Any
 import itertools
-import gguf
 import numpy as np
 import torch
 from sentencepiece import SentencePieceProcessor  # type: ignore[import]
 
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
 
 if TYPE_CHECKING:
     from typing import TypeAlias
@@ -174,8 +177,11 @@ def parse_args() -> argparse.Namespace:
 print("gguf: get sentencepiece tokenizer vocab, scores and token types")
 
 tokenizer = SentencePieceProcessor(str(tokenizer_model_file))
+vocab_size = hparams.get('vocab_size')
+if vocab_size is None:
+    vocab_size = tokenizer.vocab_size()
 
-for i in range(tokenizer.vocab_size()):
+for i in range(vocab_size):
     text: bytes
     score: float