[pull] master from ggerganov:master

CMakeLists.txt

@@ -50,9 +50,6 @@ endif()
 # option list
 #
 
-# general
-option(LLAMA_CCACHE "llama: use ccache if available" ON)
-
 # debug
 option(LLAMA_ALL_WARNINGS           "llama: enable all compiler warnings"                   ON)
 option(LLAMA_ALL_WARNINGS_3RD_PARTY "llama: enable all compiler warnings in 3rd party libs" OFF)
@@ -77,7 +74,6 @@ option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 
 # override ggml options
-set(GGML_CCACHE             ${LLAMA_CCACHE})
 set(GGML_SANITIZE_THREAD    ${LLAMA_SANITIZE_THREAD})
 set(GGML_SANITIZE_ADDRESS   ${LLAMA_SANITIZE_ADDRESS})
 set(GGML_SANITIZE_UNDEFINED ${LLAMA_SANITIZE_UNDEFINED})

Makefile

@@ -64,10 +64,14 @@ TEST_TARGETS = \
 	tests/test-tokenizer-1-spm
 
 # Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned
-LEGACY_TARGETS = main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
+LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
 	simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama \
 	retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm
 
+# Legacy build targets that were renamed in #7809, but we want to build binaries that for them that output a deprecation warning if people try to use them.
+#  We don't want to clutter things too much, so we only build replacements for the most commonly used binaries.
+LEGACY_TARGETS_BUILD = main quantize perplexity embedding server finetune
+
 # Deprecation aliases
 ifdef LLAMA_CUBLAS
 $(error LLAMA_CUBLAS is removed. Use GGML_CUDA instead.)
@@ -193,7 +197,7 @@ ifdef GGML_RPC
 	BUILD_TARGETS += rpc-server
 endif
 
-default: $(BUILD_TARGETS)
+default: $(BUILD_TARGETS) $(LEGACY_TARGETS_BUILD)
 
 test: $(TEST_TARGETS)
 	@failures=0; \
@@ -228,7 +232,7 @@ test: $(TEST_TARGETS)
 	fi
 	@echo 'All tests passed.'
 
-all: $(BUILD_TARGETS) $(TEST_TARGETS)
+all: $(BUILD_TARGETS) $(TEST_TARGETS) $(LEGACY_TARGETS_BUILD)
 
 ifdef RISCV_CROSS_COMPILE
 CC	:= riscv64-unknown-linux-gnu-gcc
@@ -245,17 +249,22 @@ MK_CFLAGS    = -std=c11   -fPIC
 MK_CXXFLAGS  = -std=c++11 -fPIC
 MK_NVCCFLAGS = -std=c++11
 
-ifndef LLAMA_NO_CCACHE
+ifdef LLAMA_NO_CCACHE
+GGML_NO_CCACHE := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifndef GGML_NO_CCACHE
 CCACHE := $(shell which ccache)
 ifdef CCACHE
 export CCACHE_SLOPPINESS = time_macros
-$(info I ccache found, compilation results will be cached. Disable with LLAMA_NO_CCACHE.)
+$(info I ccache found, compilation results will be cached. Disable with GGML_NO_CCACHE.)
 CC    := $(CCACHE) $(CC)
 CXX   := $(CCACHE) $(CXX)
 else
 $(info I ccache not found. Consider installing it for faster compilation.)
 endif # CCACHE
-endif # LLAMA_NO_CCACHE
+endif # GGML_NO_CCACHE
 
 # clock_gettime came in POSIX.1b (1993)
 # CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
@@ -545,7 +554,7 @@ endif # GGML_BLIS
 
 ifndef GGML_NO_LLAMAFILE
 	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
-	OBJ_GGML    += ggml/src/sgemm.o
+	OBJ_GGML    += ggml/src/llamafile/sgemm.o
 endif
 
 ifdef GGML_RPC
@@ -826,7 +835,8 @@ OBJ_GGML += \
 	ggml/src/ggml.o \
 	ggml/src/ggml-alloc.o \
 	ggml/src/ggml-backend.o \
-	ggml/src/ggml-quants.o
+	ggml/src/ggml-quants.o \
+	ggml/src/ggml-aarch64.o
 
 OBJ_LLAMA = \
 	src/llama.o \
@@ -926,6 +936,7 @@ $(info   - LLAMA_NO_LLAMAFILE)
 $(info   - LLAMA_NO_ACCELERATE)
 $(info   - LLAMA_NO_OPENMP)
 $(info   - LLAMA_NO_METAL)
+$(info   - LLAMA_NO_CCACHE)
 $(info )
 endif
 
@@ -959,15 +970,22 @@ ggml/src/ggml-quants.o: \
 	ggml/src/ggml-common.h
 	$(CC) $(CFLAGS)    -c $< -o $@
 
+ggml/src/ggml-aarch64.o: \
+	ggml/src/ggml-aarch64.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-aarch64.h \
+	ggml/src/ggml-common.h
+	$(CC) $(CFLAGS)    -c $< -o $@
+
 ggml/src/ggml-blas.o: \
 	ggml/src/ggml-blas.cpp \
 	ggml/include/ggml-blas.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 ifndef GGML_NO_LLAMAFILE
-ggml/src/sgemm.o: \
-	ggml/src/sgemm.cpp \
-	ggml/src/sgemm.h \
+ggml/src/llamafile/sgemm.o: \
+	ggml/src/llamafile/sgemm.cpp \
+	ggml/src/llamafile/sgemm.h \
 	ggml/include/ggml.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif # GGML_NO_LLAMAFILE
@@ -1092,7 +1110,7 @@ clean:
 	rm -vrf ggml/src/ggml-cuda/template-instances/*.o
 	rm -rvf $(BUILD_TARGETS)
 	rm -rvf $(TEST_TARGETS)
-	rm -rvf $(LEGACY_TARGETS)
+	rm -rvf $(LEGACY_TARGETS_CLEAN)
 	find examples pocs -type f -name "*.o" -delete
 
 #
@@ -1488,3 +1506,69 @@ llama-q8dot: pocs/vdot/q8dot.cpp ggml/src/ggml.o \
 	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+#
+# Deprecated binaries that we want to keep around long enough for people to migrate to the new filenames, then these can be removed.
+#
+# Mark legacy binary targets as .PHONY so that they are always checked.
+.PHONY: main quantize perplexity embedding server finetune
+
+main: examples/deprecation-warning/deprecation-warning.cpp
+ifneq (,$(wildcard main))
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	@echo "#########"
+	@echo "WARNING: The 'main' binary is deprecated. Please use 'llama-cli' instead."
+	@echo "  Remove the 'main' binary to remove this warning."
+	@echo "#########"
+endif
+
+quantize: examples/deprecation-warning/deprecation-warning.cpp
+ifneq (,$(wildcard quantize))
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	@echo "#########"
+	@echo "WARNING: The 'quantize' binary is deprecated. Please use 'llama-quantize' instead."
+	@echo "  Remove the 'quantize' binary to remove this warning."
+	@echo "#########"
+endif
+
+perplexity: examples/deprecation-warning/deprecation-warning.cpp
+ifneq (,$(wildcard perplexity))
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	@echo "#########"
+	@echo "WARNING: The 'perplexity' binary is deprecated. Please use 'llama-perplexity' instead."
+	@echo "  Remove the 'perplexity' binary to remove this warning."
+	@echo "#########"
+endif
+
+embedding: examples/deprecation-warning/deprecation-warning.cpp
+ifneq (,$(wildcard embedding))
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	@echo "#########"
+	@echo "WARNING: The 'embedding' binary is deprecated. Please use 'llama-embedding' instead."
+	@echo "  Remove the 'embedding' binary to remove this warning."
+	@echo "#########"
+endif
+
+server: examples/deprecation-warning/deprecation-warning.cpp
+ifneq (,$(wildcard server))
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	@echo "#########"
+	@echo "WARNING: The 'server' binary is deprecated. Please use 'llama-server' instead."
+	@echo "  Remove the 'server' binary to remove this warning."
+	@echo "#########"
+endif
+
+finetune: examples/deprecation-warning/deprecation-warning.cpp
+ifneq (,$(wildcard finetune))
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	@echo "#########"
+	@echo "WARNING: The 'finetune' binary is deprecated. Please use 'llama-finetune' instead."
+	@echo "  Remove the 'finetune' binary to remove this warning."
+	@echo "#########"
+endif

Package.swift

@@ -10,6 +10,7 @@ var sources = [
     "ggml/src/ggml-alloc.c",
     "ggml/src/ggml-backend.c",
     "ggml/src/ggml-quants.c",
+    "ggml/src/ggml-aarch64.c",
 ]
 
 var resources: [Resource] = []

README.md

@@ -453,7 +453,7 @@ To learn more how to measure perplexity using llama.cpp, [read this documentatio
 - [How to build](./docs/build.md)
 - [Running on Docker](./docs/docker.md)
 - [Build on Android](./docs/android.md)
-- [Performance troubleshooting](./docs/token_generation_performance_tips.md)
+- [Performance troubleshooting](./docs/development/token_generation_performance_tips.md)
 - [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
 
 **Seminal papers and background on the models**

common/common.cpp

@@ -1,3 +1,7 @@
+#if defined(_MSC_VER)
+#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
+#endif
+
 #include "common.h"
 // Change JSON_ASSERT from assert() to GGML_ASSERT:
 #define JSON_ASSERT GGML_ASSERT

convert_hf_to_gguf.py

@@ -1356,7 +1356,7 @@ class LlamaModel(Model):
 
     def set_vocab(self):
         try:
-            self. _set_vocab_sentencepiece()
+            self._set_vocab_sentencepiece()
         except FileNotFoundError:
             try:
                 self._set_vocab_llama_hf()
@@ -2144,6 +2144,9 @@ def set_vocab(self):
                 toktype = SentencePieceTokenTypes.UNUSED
             elif tokenizer.IsByte(token_id):
                 toktype = SentencePieceTokenTypes.BYTE
+            # take care of ununsed raw token
+            if piece.startswith('[UNUSED'):
+                toktype = SentencePieceTokenTypes.UNKNOWN
 
             tokens.append(text)
             scores.append(score)
@@ -2159,6 +2162,47 @@ def set_vocab(self):
                     scores.append(-1000.0)
                     toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
 
+        chat_eos_token = '<|im_end|>'
+        chat_eos_token_id = None
+
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {})
+                for token_id, foken_data in added_tokens_decoder.items():
+                    token_id = int(token_id)
+                    token = foken_data["content"]
+                    if token == chat_eos_token:
+                        chat_eos_token_id = token_id
+                    token = token.encode("utf-8")
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
+                        assert(tokens[token_id] == token)
+                    tokens[token_id] = token
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+                    if foken_data.get("special"):
+                        toktypes[token_id] = SentencePieceTokenTypes.CONTROL
+
+        tokenizer_file = self.dir_model / 'tokenizer.json'
+        if tokenizer_file.is_file():
+            with open(tokenizer_file, "r", encoding="utf-8") as f:
+                tokenizer_json = json.load(f)
+                added_tokens = tokenizer_json.get("added_tokens", [])
+                for foken_data in added_tokens:
+                    token_id = int(foken_data["id"])
+                    token = foken_data["content"]
+                    if token == chat_eos_token:
+                        chat_eos_token_id = token_id
+                    token = token.encode("utf-8")
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
+                        assert(tokens[token_id] == token)
+                    tokens[token_id] = token
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+                    if foken_data.get("special"):
+                        toktypes[token_id] = SentencePieceTokenTypes.CONTROL
+
         self.gguf_writer.add_tokenizer_model("llama")
         self.gguf_writer.add_tokenizer_pre("default")
         self.gguf_writer.add_token_list(tokens)
@@ -2168,28 +2212,16 @@ def set_vocab(self):
 
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         old_eos = special_vocab.special_token_ids["eos"]
-        if "chat" in os.path.basename(self.dir_model.absolute()):
+        if chat_eos_token_id is not None:
             # For the chat model, we replace the eos with '<|im_end|>'.
             # TODO: this is a hack, should be fixed
             #       https://github.com/ggerganov/llama.cpp/pull/6745#issuecomment-2067687048
-            special_vocab.special_token_ids["eos"] = self._try_get_sft_eos(tokenizer)
-            logger.warning(f"Replace eos:{old_eos} with a special token:{special_vocab.special_token_ids['eos']} \
-in chat mode so that the conversation can end normally.")
+            special_vocab.special_token_ids["eos"] = chat_eos_token_id
+            logger.warning(f"Replace eos:{old_eos} with a special token:{chat_eos_token_id}"
+                           " in chat mode so that the conversation can end normally.")
 
         special_vocab.add_to_gguf(self.gguf_writer)
 
-    def _try_get_sft_eos(self, tokenizer):
-        unused_145_list = tokenizer.Encode('[UNUSED_TOKEN_145]')
-        im_end_list = tokenizer.Encode('<|im_end|>')
-        eos_token = None
-        assert (len(unused_145_list) == 1) ^ (len(im_end_list) == 1)
-        if len(unused_145_list) == 1:
-            eos_token = unused_145_list[0]
-        if len(im_end_list) == 1:
-            eos_token = im_end_list[0]
-        assert eos_token
-        return eos_token
-
     def _hf_permute_qk(self, weights, n_head: int, n_head_kv: int):
         if n_head_kv is not None and n_head != n_head_kv:
             n_head = n_head_kv
@@ -2208,6 +2240,10 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
         self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"])
         self.gguf_writer.add_file_type(self.ftype)
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         num_heads = self.hparams["num_attention_heads"]

docs/build.md

@@ -28,6 +28,7 @@ In order to build llama.cpp you have four different options.
         ```
 
   - Notes:
+    - For `Q4_0_4_4` quantization type build, add the `GGML_NO_LLAMAFILE=1` flag. For example, use `make GGML_NO_LLAMAFILE=1`.
     - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel.
     - For faster repeated compilation, install [ccache](https://ccache.dev/).
     - For debug builds, run `make LLAMA_DEBUG=1`
@@ -41,6 +42,7 @@ In order to build llama.cpp you have four different options.
 
   **Notes**:
 
+    - For `Q4_0_4_4` quantization type build, add the `-DGGML_LLAMAFILE=OFF` cmake option. For example, use `cmake -B build -DGGML_LLAMAFILE=OFF`.
     - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
     - For faster repeated compilation, install [ccache](https://ccache.dev/).
     - For debug builds, there are two cases: