Initial release

.gitignore

@@ -0,0 +1,21 @@
+*.o
+*.a
+.cache/
+.vs/
+.vscode/
+.DS_Store
+
+build/
+build-em/
+build-debug/
+build-release/
+build-static/
+build-no-accel/
+build-sanitize-addr/
+build-sanitize-thread/
+
+/main
+/quantize
+
+arm_neon.h
+compile_commands.json

Makefile

@@ -0,0 +1,203 @@
+ifndef UNAME_S
+UNAME_S := $(shell uname -s)
+endif
+
+ifndef UNAME_P
+UNAME_P := $(shell uname -p)
+endif
+
+ifndef UNAME_M
+UNAME_M := $(shell uname -m)
+endif
+
+CCV := $(shell $(CC) --version | head -n 1)
+CXXV := $(shell $(CXX) --version | head -n 1)
+
+# Mac OS + Arm can report x86_64
+# ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
+ifeq ($(UNAME_S),Darwin)
+	ifneq ($(UNAME_P),arm)
+		SYSCTL_M := $(shell sysctl -n hw.optional.arm64)
+		ifeq ($(SYSCTL_M),1)
+			# UNAME_P := arm
+			# UNAME_M := arm64
+			warn := $(warning Your arch is announced as x86_64, but it seems to actually be ARM64. Not fixing that can lead to bad performance. For more info see: https://github.com/ggerganov/whisper.cpp/issues/66\#issuecomment-1282546789)
+		endif
+	endif
+endif
+
+#
+# Compile flags
+#
+
+CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
+CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
+LDFLAGS  =
+
+# OS specific
+# TODO: support Windows
+ifeq ($(UNAME_S),Linux)
+	CFLAGS   += -pthread
+	CXXFLAGS += -pthread
+endif
+ifeq ($(UNAME_S),Darwin)
+	CFLAGS   += -pthread
+	CXXFLAGS += -pthread
+endif
+ifeq ($(UNAME_S),FreeBSD)
+	CFLAGS   += -pthread
+	CXXFLAGS += -pthread
+endif
+ifeq ($(UNAME_S),Haiku)
+	CFLAGS   += -pthread
+	CXXFLAGS += -pthread
+endif
+
+# Architecture specific
+# TODO: probably these flags need to be tweaked on some architectures
+#       feel free to update the Makefile for your architecture and send a pull request or issue
+ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
+	ifeq ($(UNAME_S),Darwin)
+		CFLAGS += -mf16c
+		AVX1_M := $(shell sysctl machdep.cpu.features)
+		ifneq (,$(findstring FMA,$(AVX1_M)))
+			CFLAGS += -mfma
+		endif
+		ifneq (,$(findstring AVX1.0,$(AVX1_M)))
+			CFLAGS += -mavx
+		endif
+		AVX2_M := $(shell sysctl machdep.cpu.leaf7_features)
+		ifneq (,$(findstring AVX2,$(AVX2_M)))
+			CFLAGS += -mavx2
+		endif
+	else ifeq ($(UNAME_S),Linux)
+		AVX1_M := $(shell grep "avx " /proc/cpuinfo)
+		ifneq (,$(findstring avx,$(AVX1_M)))
+			CFLAGS += -mavx
+		endif
+		AVX2_M := $(shell grep "avx2 " /proc/cpuinfo)
+		ifneq (,$(findstring avx2,$(AVX2_M)))
+			CFLAGS += -mavx2
+		endif
+		FMA_M := $(shell grep "fma " /proc/cpuinfo)
+		ifneq (,$(findstring fma,$(FMA_M)))
+			CFLAGS += -mfma
+		endif
+		F16C_M := $(shell grep "f16c " /proc/cpuinfo)
+		ifneq (,$(findstring f16c,$(F16C_M)))
+			CFLAGS += -mf16c
+		endif
+		SSE3_M := $(shell grep "sse3 " /proc/cpuinfo)
+		ifneq (,$(findstring sse3,$(SSE3_M)))
+			CFLAGS += -msse3
+		endif
+	else ifeq ($(UNAME_S),Haiku)
+		AVX1_M := $(shell sysinfo -cpu | grep "AVX ")
+		ifneq (,$(findstring avx,$(AVX1_M)))
+			CFLAGS += -mavx
+		endif
+		AVX2_M := $(shell sysinfo -cpu | grep "AVX2 ")
+		ifneq (,$(findstring avx2,$(AVX2_M)))
+			CFLAGS += -mavx2
+		endif
+		FMA_M := $(shell sysinfo -cpu | grep "FMA ")
+		ifneq (,$(findstring fma,$(FMA_M)))
+			CFLAGS += -mfma
+		endif
+		F16C_M := $(shell sysinfo -cpu | grep "F16C ")
+		ifneq (,$(findstring f16c,$(F16C_M)))
+			CFLAGS += -mf16c
+		endif
+	else
+		CFLAGS += -mfma -mf16c -mavx -mavx2
+	endif
+endif
+ifeq ($(UNAME_M),amd64)
+	CFLAGS += -mavx -mavx2 -mfma -mf16c
+endif
+ifneq ($(filter ppc64%,$(UNAME_M)),)
+	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
+	ifneq (,$(findstring POWER9,$(POWER9_M)))
+		CFLAGS += -mpower9-vector
+	endif
+	# Require c++23's std::byteswap for big-endian support.
+	ifeq ($(UNAME_M),ppc64)
+		CXXFLAGS += -std=c++23 -DGGML_BIG_ENDIAN
+	endif
+endif
+ifndef WHISPER_NO_ACCELERATE
+	# Mac M1 - include Accelerate framework
+	ifeq ($(UNAME_S),Darwin)
+		CFLAGS  += -DGGML_USE_ACCELERATE
+		LDFLAGS += -framework Accelerate
+	endif
+endif
+ifdef WHISPER_OPENBLAS
+	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
+	LDFLAGS += -lopenblas
+endif
+ifdef WHISPER_GPROF
+	CFLAGS   += -pg
+	CXXFLAGS += -pg
+endif
+ifneq ($(filter aarch64%,$(UNAME_M)),)
+	CFLAGS += -mcpu=native
+	CXXFLAGS += -mcpu=native
+endif
+ifneq ($(filter armv6%,$(UNAME_M)),)
+	# Raspberry Pi 1, 2, 3
+	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
+endif
+ifneq ($(filter armv7%,$(UNAME_M)),)
+	# Raspberry Pi 4
+	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+endif
+ifneq ($(filter armv8%,$(UNAME_M)),)
+	# Raspberry Pi 4
+	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
+endif
+
+#
+# Print build information
+#
+
+$(info I llama.cpp build info: )
+$(info I UNAME_S:  $(UNAME_S))
+$(info I UNAME_P:  $(UNAME_P))
+$(info I UNAME_M:  $(UNAME_M))
+$(info I CFLAGS:   $(CFLAGS))
+$(info I CXXFLAGS: $(CXXFLAGS))
+$(info I LDFLAGS:  $(LDFLAGS))
+$(info I CC:       $(CCV))
+$(info I CXX:      $(CXXV))
+$(info )
+
+default: main quantize
+
+#
+# Build library
+#
+
+ggml.o: ggml.c ggml.h
+	$(CC)  $(CFLAGS)   -c ggml.c -o ggml.o
+
+utils.o: utils.cpp utils.h
+	$(CXX) $(CXXFLAGS) -c utils.cpp -o utils.o
+
+clean:
+	rm -f *.o main quantize
+
+main: main.cpp ggml.o utils.o
+	$(CXX) $(CXXFLAGS) main.cpp ggml.o utils.o -o main $(LDFLAGS)
+	./main -h
+
+quantize: quantize.cpp ggml.o utils.o
+	$(CXX) $(CXXFLAGS) quantize.cpp ggml.o utils.o -o quantize $(LDFLAGS)
+
+#
+# Tests
+#
+
+.PHONY: tests
+tests:
+	bash ./tests/run-tests.sh

convert-pth-to-ggml.py

@@ -0,0 +1,136 @@
+# Convert a LLaMA model checkpoint to a ggml compatible file
+#
+# Load the model using Torch
+# Iterate over all variables and write them to a binary file.
+#
+# For each variable, write the following:
+#   - Number of dimensions (int)
+#   - Name length (int)
+#   - Dimensions (int[n_dims])
+#   - Name (char[name_length])
+#   - Data (float[n_dims])
+#
+# By default, the bigger matrices are converted to 16-bit floats.
+# This can be disabled by adding the "use-f32" CLI argument.
+#
+# At the start of the ggml file we write the model parameters
+# and vocabulary.
+#
+
+import sys
+import json
+import struct
+import numpy as np
+import torch
+
+from sentencepiece import SentencePieceProcessor
+
+if len(sys.argv) < 3:
+    print("Usage: convert-ckpt-to-ggml.py dir-model ftype\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model.bin"
+
+fname_hparams   = sys.argv[1] + "/params.json"
+fname_model     = sys.argv[1] + "/consolidated.00.pth"
+fname_tokenizer = sys.argv[1] + "/../tokenizer.model"
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 2:
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
+
+with open(fname_hparams, "r") as f:
+    hparams = json.load(f)
+
+tokenizer = SentencePieceProcessor(fname_tokenizer)
+
+hparams.update({"vocab_size": tokenizer.vocab_size()})
+
+print(hparams)
+
+model = torch.load(fname_model, map_location="cpu")
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["dim"]))
+fout.write(struct.pack("i", hparams["multiple_of"]))
+fout.write(struct.pack("i", hparams["n_heads"]))
+fout.write(struct.pack("i", hparams["n_layers"]))
+fout.write(struct.pack("i", 64)) # rot
+fout.write(struct.pack("i", ftype))
+
+# Is this correct??
+for i in range(32000):
+    # TODO: this is probably wrong - not sure how this tokenizer works
+    text = tokenizer.decode([29889, i]).encode('utf-8')
+    # remove the first byte (it's always '.')
+    text = text[1:]
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for k, v in model.items():
+    name = k
+    shape = v.shape
+
+    # skip layers.X.attention.inner_attention.rope.freqs
+    if name[-5:] == "freqs":
+        continue
+
+    print("Processing variable: " + name + " with shape: ", shape, " and type: ", v.dtype)
+
+    #data = tf.train.load_variable(dir_model, name).squeeze()
+    data = v.numpy().squeeze()
+    n_dims = len(data.shape);
+
+    # for efficiency - transpose some matrices
+    # "model/h.*/attn/c_attn/w"
+    # "model/h.*/attn/c_proj/w"
+    # "model/h.*/mlp/c_fc/w"
+    # "model/h.*/mlp/c_proj/w"
+    #if name[-14:] == "/attn/c_attn/w" or \
+    #   name[-14:] == "/attn/c_proj/w" or \
+    #   name[-11:] == "/mlp/c_fc/w" or \
+    #   name[-13:] == "/mlp/c_proj/w":
+    #    print("  Transposing")
+    #    data = data.transpose()
+
+    dshape = data.shape
+
+    # default type is fp16
+    ftype_cur = 1
+    if ftype == 0 or n_dims == 1:
+        print("  Converting to float32")
+        data = data.astype(np.float32)
+        ftype_cur = 0
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", dshape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")