vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

merge conflict

Browse source
This commit is contained in:
kchro3 2023-09-02 23:18:23 -07:00
commit 63cef16956
5 changed files with 146 additions and 96 deletions
Download patch file Download diff file Expand all files Collapse all files

146
Makefile
View file

@ -67,21 +67,21 @@ OPT = -Ofast
else else
OPT = -O3 OPT = -O3
endif endif
CFLAGS = -I. $(OPT) -std=c11 -fPIC MK_CPPFLAGS = -I. -Icommon
CXXFLAGS = -I. -I./common $(OPT) -std=c++11 -fPIC MK_CFLAGS = $(CPPFLAGS) $(OPT) -std=c11 -fPIC
LDFLAGS = MK_CXXFLAGS = $(CPPFLAGS) $(OPT) -std=c++11 -fPIC
MK_LDFLAGS =
ifdef LLAMA_DEBUG ifdef LLAMA_DEBUG
CFLAGS += -O0 -g MK_CFLAGS += -O0 -g
CXXFLAGS += -O0 -g MK_CXXFLAGS += -O0 -g
LDFLAGS += -g MK_LDFLAGS += -g
else else
CFLAGS += -DNDEBUG MK_CPPFLAGS += -DNDEBUG
CXXFLAGS += -DNDEBUG
endif endif
ifdef LLAMA_SERVER_VERBOSE ifdef LLAMA_SERVER_VERBOSE
CXXFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE) MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
endif endif
ifdef LLAMA_DISABLE_LOGS ifdef LLAMA_DISABLE_LOGS
@ -90,9 +90,9 @@ ifdef LLAMA_DISABLE_LOGS
endif # LLAMA_DISABLE_LOGS endif # LLAMA_DISABLE_LOGS
# warnings # warnings
CFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \ MK_CFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \
-Wmissing-prototypes -Werror=implicit-int -Wno-unused-function -Wmissing-prototypes -Werror=implicit-int -Wno-unused-function
CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar MK_CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar
ifeq '' '$(findstring clang++,$(CXX))' ifeq '' '$(findstring clang++,$(CXX))'
# g++ only # g++ only
@ -101,29 +101,9 @@ endif
# OS specific # OS specific
# TODO: support Windows # TODO: support Windows
ifeq ($(UNAME_S),Linux) ifneq '' '$(filter $(UNAME_S),Linux Darwin FreeBSD NetBSD OpenBSD Haiku)'
CFLAGS += -pthread MK_CFLAGS += -pthread
CXXFLAGS += -pthread MK_CXXFLAGS += -pthread
endif
ifeq ($(UNAME_S),Darwin)
CFLAGS += -pthread
CXXFLAGS += -pthread
endif
ifeq ($(UNAME_S),FreeBSD)
CFLAGS += -pthread
CXXFLAGS += -pthread
endif
ifeq ($(UNAME_S),NetBSD)
CFLAGS += -pthread
CXXFLAGS += -pthread
endif
ifeq ($(UNAME_S),OpenBSD)
CFLAGS += -pthread
CXXFLAGS += -pthread
endif
ifeq ($(UNAME_S),Haiku)
CFLAGS += -pthread
CXXFLAGS += -pthread
endif endif
# detect Windows # detect Windows
@ -149,12 +129,11 @@ ifeq ($(_WIN32),1)
endif endif
ifdef LLAMA_GPROF ifdef LLAMA_GPROF
CFLAGS += -pg MK_CFLAGS += -pg
CXXFLAGS += -pg MK_CXXFLAGS += -pg
endif endif
ifdef LLAMA_PERF ifdef LLAMA_PERF
CFLAGS += -DGGML_PERF MK_CPPFLAGS += -DGGML_PERF
CXXFLAGS += -DGGML_PERF
endif endif
# Architecture specific # Architecture specific
@ -165,16 +144,16 @@ ifndef RISCV
ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64)) ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
# Use all CPU extensions that are available: # Use all CPU extensions that are available:
CFLAGS += -march=native -mtune=native MK_CFLAGS += -march=native -mtune=native
CXXFLAGS += -march=native -mtune=native MK_CXXFLAGS += -march=native -mtune=native
# Usage AVX-only # Usage AVX-only
#CFLAGS += -mfma -mf16c -mavx #MK_CFLAGS += -mfma -mf16c -mavx
#CXXFLAGS += -mfma -mf16c -mavx #MK_CXXFLAGS += -mfma -mf16c -mavx
# Usage SSSE3-only (Not is SSE3!) # Usage SSSE3-only (Not is SSE3!)
#CFLAGS += -mssse3 #MK_CFLAGS += -mssse3
#CXXFLAGS += -mssse3 #MK_CXXFLAGS += -mssse3
endif endif
# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves. # The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
@ -188,34 +167,33 @@ endif
ifneq ($(filter aarch64%,$(UNAME_M)),) ifneq ($(filter aarch64%,$(UNAME_M)),)
# Apple M1, M2, etc. # Apple M1, M2, etc.
# Raspberry Pi 3, 4, Zero 2 (64-bit) # Raspberry Pi 3, 4, Zero 2 (64-bit)
CFLAGS += -mcpu=native MK_CFLAGS += -mcpu=native
CXXFLAGS += -mcpu=native MK_CXXFLAGS += -mcpu=native
endif endif
ifneq ($(filter armv6%,$(UNAME_M)),) ifneq ($(filter armv6%,$(UNAME_M)),)
# Raspberry Pi 1, Zero # Raspberry Pi 1, Zero
CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access MK_CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
endif endif
ifneq ($(filter armv7%,$(UNAME_M)),) ifneq ($(filter armv7%,$(UNAME_M)),)
# Raspberry Pi 2 # Raspberry Pi 2
CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations MK_CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
endif endif
ifneq ($(filter armv8%,$(UNAME_M)),) ifneq ($(filter armv8%,$(UNAME_M)),)
# Raspberry Pi 3, 4, Zero 2 (32-bit) # Raspberry Pi 3, 4, Zero 2 (32-bit)
CFLAGS += -mfp16-format=ieee -mno-unaligned-access MK_CFLAGS += -mfp16-format=ieee -mno-unaligned-access
MK_CXXFLAGS += -mfp16-format=ieee -mno-unaligned-access
endif endif
ifneq ($(filter ppc64%,$(UNAME_M)),) ifneq ($(filter ppc64%,$(UNAME_M)),)
POWER9_M := $(shell grep "POWER9" /proc/cpuinfo) POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
ifneq (,$(findstring POWER9,$(POWER9_M))) ifneq (,$(findstring POWER9,$(POWER9_M)))
CFLAGS += -mcpu=power9 MK_CFLAGS += -mcpu=power9
CXXFLAGS += -mcpu=power9 MK_CXXFLAGS += -mcpu=power9
endif
# Require c++23's std::byteswap for big-endian support.
ifeq ($(UNAME_M),ppc64)
CXXFLAGS += -std=c++23 -DGGML_BIG_ENDIAN
endif endif
endif endif
@ -225,12 +203,10 @@ else
endif endif
ifndef LLAMA_NO_K_QUANTS ifndef LLAMA_NO_K_QUANTS
CFLAGS += -DGGML_USE_K_QUANTS MK_CPPFLAGS += -DGGML_USE_K_QUANTS
CXXFLAGS += -DGGML_USE_K_QUANTS
OBJS += k_quants.o OBJS += k_quants.o
ifdef LLAMA_QKK_64 ifdef LLAMA_QKK_64
CFLAGS += -DGGML_QKK_64 MK_CPPFLAGS += -DGGML_QKK_64
CXXFLAGS += -DGGML_QKK_64
endif endif
endif endif
@ -238,31 +214,32 @@ ifndef LLAMA_NO_ACCELERATE
# Mac M1 - include Accelerate framework. # Mac M1 - include Accelerate framework.
# `-framework Accelerate` works on Mac Intel as well, with negliable performance boost (as of the predict time). # `-framework Accelerate` works on Mac Intel as well, with negliable performance boost (as of the predict time).
ifeq ($(UNAME_S),Darwin) ifeq ($(UNAME_S),Darwin)
CFLAGS += -DGGML_USE_ACCELERATE MK_CPPFLAGS += -DGGML_USE_ACCELERATE
LDFLAGS += -framework Accelerate MK_LDFLAGS += -framework Accelerate
endif endif
endif # LLAMA_NO_ACCELERATE endif # LLAMA_NO_ACCELERATE
ifdef LLAMA_MPI ifdef LLAMA_MPI
CFLAGS += -DGGML_USE_MPI -Wno-cast-qual MK_CPPFLAGS += -DGGML_USE_MPI
CXXFLAGS += -DGGML_USE_MPI -Wno-cast-qual MK_CFLAGS += -Wno-cast-qual
MK_CXXFLAGS += -Wno-cast-qual
OBJS += ggml-mpi.o OBJS += ggml-mpi.o
endif # LLAMA_MPI endif # LLAMA_MPI
ifdef LLAMA_OPENBLAS ifdef LLAMA_OPENBLAS
CFLAGS += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags openblas) MK_CPPFLAGS += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags-only-I openblas)
LDFLAGS += $(shell pkg-config --libs openblas) MK_CFLAGS += $(shell pkg-config --cflags-only-other openblas)
MK_LDFLAGS += $(shell pkg-config --libs openblas)
endif # LLAMA_OPENBLAS endif # LLAMA_OPENBLAS
ifdef LLAMA_BLIS ifdef LLAMA_BLIS
CFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis MK_CPPFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
LDFLAGS += -lblis -L/usr/local/lib MK_LDFLAGS += -lblis -L/usr/local/lib
endif # LLAMA_BLIS endif # LLAMA_BLIS
ifdef LLAMA_CUBLAS ifdef LLAMA_CUBLAS
CFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include MK_CPPFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
CXXFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include MK_LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
OBJS += ggml-cuda.o OBJS += ggml-cuda.o
NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
ifdef LLAMA_CUDA_NVCC ifdef LLAMA_CUDA_NVCC
@ -313,14 +290,15 @@ endif # LLAMA_CUBLAS
ifdef LLAMA_CLBLAST ifdef LLAMA_CLBLAST
CFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags clblast OpenCL) MK_CPPFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags-only-I clblast OpenCL)
CXXFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags clblast OpenCL) MK_CFLAGS += $(shell pkg-config --cflags-only-other clblast OpenCL)
MK_CXXFLAGS += $(shell pkg-config --cflags-only-other clblast OpenCL)
# Mac provides OpenCL as a framework # Mac provides OpenCL as a framework
ifeq ($(UNAME_S),Darwin) ifeq ($(UNAME_S),Darwin)
LDFLAGS += -lclblast -framework OpenCL MK_LDFLAGS += -lclblast -framework OpenCL
else else
LDFLAGS += $(shell pkg-config --libs clblast OpenCL) MK_LDFLAGS += $(shell pkg-config --libs clblast OpenCL)
endif endif
OBJS += ggml-opencl.o OBJS += ggml-opencl.o
@ -335,10 +313,9 @@ ifdef LLAMA_HIPBLAS
LLAMA_CUDA_DMMV_X ?= 32 LLAMA_CUDA_DMMV_X ?= 32
LLAMA_CUDA_MMV_Y ?= 1 LLAMA_CUDA_MMV_Y ?= 1
LLAMA_CUDA_KQUANTS_ITER ?= 2 LLAMA_CUDA_KQUANTS_ITER ?= 2
CFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
CXXFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas
LDFLAGS += -lhipblas -lamdhip64 -lrocblas
HIPFLAGS += $(addprefix --offload-arch=,$(GPU_TARGETS)) HIPFLAGS += $(addprefix --offload-arch=,$(GPU_TARGETS))
HIPFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X) HIPFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
HIPFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y) HIPFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
@ -353,9 +330,8 @@ ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
endif # LLAMA_HIPBLAS endif # LLAMA_HIPBLAS
ifdef LLAMA_METAL ifdef LLAMA_METAL
CFLAGS += -DGGML_USE_METAL #-DGGML_METAL_NDEBUG MK_CPPFLAGS += -DGGML_USE_METAL #-DGGML_METAL_NDEBUG
CXXFLAGS += -DGGML_USE_METAL MK_LDFLAGS += -framework Foundation -framework Metal -framework MetalKit
LDFLAGS += -framework Foundation -framework Metal -framework MetalKit
OBJS += ggml-metal.o OBJS += ggml-metal.o
endif # LLAMA_METAL endif # LLAMA_METAL
@ -369,11 +345,17 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
$(CC) $(CFLAGS) -c $< -o $@ $(CC) $(CFLAGS) -c $< -o $@
endif # LLAMA_MPI endif # LLAMA_MPI
ifdef LLAMA_NO_K_QUANTS ifndef LLAMA_NO_K_QUANTS
k_quants.o: k_quants.c k_quants.h k_quants.o: k_quants.c k_quants.h
$(CC) $(CFLAGS) -c $< -o $@ $(CC) $(CFLAGS) -c $< -o $@
endif # LLAMA_NO_K_QUANTS endif # LLAMA_NO_K_QUANTS
# combine build flags with cmdline overrides
override CPPFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS)
override CFLAGS := $(MK_CFLAGS) $(CFLAGS)
override CXXFLAGS := $(MK_CXXFLAGS) $(CXXFLAGS)
override LDFLAGS := $(MK_LDFLAGS) $(LDFLAGS)
# #
# Print build information # Print build information
# #

1
README.md
View file

@ -120,6 +120,7 @@ as the main playground for developing new features for the [ggml](https://github
- [nat/openplayground](https://github.com/nat/openplayground) - [nat/openplayground](https://github.com/nat/openplayground)
- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) - [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui)
- [withcatai/catai](https://github.com/withcatai/catai)
--- ---

30
convert.py
View file

@ -323,15 +323,27 @@ class BpeVocab:
self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read()) self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
added_tokens: dict[str, int] added_tokens: dict[str, int]
if fname_added_tokens is not None: if fname_added_tokens is not None:
# FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
added_tokens = json.load(open(fname_added_tokens, encoding="utf-8")) added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
else: else:
# Fall back to trying to find the added tokens in tokenizer.json
tokenizer_json_file = fname_tokenizer.parent / 'tokenizer.json'
if not tokenizer_json_file.is_file():
added_tokens = {} added_tokens = {}
else:
tokenizer_json = json.load(open(tokenizer_json_file, encoding="utf-8"))
added_tokens = dict(
(item['content'], item['id'])
for item in tokenizer_json.get('added_tokens', [])
# Added tokens here can be duplicates of the main vocabulary.
if item['content'] not in self.bpe_tokenizer )
vocab_size: int = len(self.bpe_tokenizer) vocab_size: int = len(self.bpe_tokenizer)
expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
actual_ids = sorted(added_tokens.values()) actual_ids = sorted(added_tokens.values())
if expected_ids != actual_ids: if expected_ids != actual_ids:
raise Exception(f"Expected added token IDs to be sequential and start at {len(added_tokens)}; got {actual_ids}") expected_end_id = vocab_size + len(actual_ids) - 1
raise Exception(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}")
items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1]) items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
self.added_tokens_list = [text for (text, idx) in items] self.added_tokens_list = [text for (text, idx) in items]
@ -345,10 +357,22 @@ class BpeVocab:
from transformers.models.gpt2 import tokenization_gpt2 # type: ignore[import] from transformers.models.gpt2 import tokenization_gpt2 # type: ignore[import]
byte_encoder = tokenization_gpt2.bytes_to_unicode() byte_encoder = tokenization_gpt2.bytes_to_unicode()
byte_decoder = {v: k for k, v in byte_encoder.items()} byte_decoder = {v: k for k, v in byte_encoder.items()}
score = 0.0
for i, item in enumerate(tokenizer): for i, item in enumerate(tokenizer):
text: bytes = item.encode("utf-8") text: bytes = item.encode("utf-8")
score: float = -i # FIXME: These shouldn't be hardcoded, but it's probably better than the current behavior?
yield text, score, gguf.TokenType.USER_DEFINED if i <= 258 and text.startswith(b'<') and text.endswith(b'>'):
if i == 0 and text == b'<unk>':
toktype = gguf.TokenType.UNKNOWN
elif i == 1 or i == 2:
toktype = gguf.TokenType.CONTROL
elif i >= 3 and text.startswith(b'<0x'):
toktype = gguf.TokenType.BYTE
else:
toktype = gguf.TokenType.NORMAL
else:
toktype = gguf.TokenType.NORMAL
yield text, score, toktype
def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
for text in self.added_tokens_list: for text in self.added_tokens_list:

13
examples/gptneox-wip/gptneox-main.cpp
View file

@ -660,9 +660,10 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
ggml_tensor * gpt_neox_ff( ggml_tensor * gpt_neox_ff(
const gpt_neox_block &block, const gpt_neox_block &block,
ggml_context * ctx0, ggml_context * ctx0,
ggml_tensor * inp) { ggml_tensor * inp,
const gpt_neox_hparams &hparams) {
ggml_tensor * cur = ggml_norm(ctx0, inp); ggml_tensor * cur = ggml_norm(ctx0, inp, hparams.norm_eps);
cur = ggml_add(ctx0, ggml_mul(ctx0, ggml_repeat(ctx0, block.ln_2_g, cur), cur), ggml_repeat(ctx0, block.ln_2_b, cur)); cur = ggml_add(ctx0, ggml_mul(ctx0, ggml_repeat(ctx0, block.ln_2_g, cur), cur), ggml_repeat(ctx0, block.ln_2_b, cur));
cur = ggml_mul_mat(ctx0, block.c_mlp_fc_w, cur); cur = ggml_mul_mat(ctx0, block.c_mlp_fc_w, cur);
@ -753,7 +754,7 @@ bool gpt_neox_eval(
// self-attention // self-attention
{ {
{ {
cur = ggml_norm(ctx0, inpL); cur = ggml_norm(ctx0, inpL, hparams.norm_eps);
cur = ggml_add(ctx0, cur = ggml_add(ctx0,
ggml_mul(ctx0, ggml_repeat(ctx0, model.blocks[il].ln_1_g, cur), cur), ggml_mul(ctx0, ggml_repeat(ctx0, model.blocks[il].ln_1_g, cur), cur),
@ -844,7 +845,7 @@ bool gpt_neox_eval(
if (hparams.par_res == 0) { if (hparams.par_res == 0) {
struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpL); struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpL);
cur = gpt_neox_ff(model.blocks[il], ctx0, inpFF); cur = gpt_neox_ff(model.blocks[il], ctx0, inpFF, hparams);
// input for next layer // input for next layer
inpL = ggml_add(ctx0, cur, inpFF); inpL = ggml_add(ctx0, cur, inpFF);
@ -853,7 +854,7 @@ bool gpt_neox_eval(
// this is independent of the self-attention result, so it could be done in parallel to the self-attention // this is independent of the self-attention result, so it could be done in parallel to the self-attention
// note here we pass inpL instead of cur // note here we pass inpL instead of cur
cur = gpt_neox_ff(model.blocks[il], ctx0, inpL); cur = gpt_neox_ff(model.blocks[il], ctx0, inpL, hparams);
// layer input + FF // layer input + FF
cur = ggml_add(ctx0, cur, inpFF); cur = ggml_add(ctx0, cur, inpFF);
@ -867,7 +868,7 @@ bool gpt_neox_eval(
// norm // norm
{ {
inpL = ggml_norm(ctx0, inpL); inpL = ggml_norm(ctx0, inpL, hparams.norm_eps);
// inpL = ln_f_g*inpL + ln_f_b // inpL = ln_f_g*inpL + ln_f_b
inpL = ggml_add(ctx0, inpL = ggml_add(ctx0,

42
llama.cpp
View file

@ -325,6 +325,44 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
}, },
}, },
{
LLM_ARCH_GPT2,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
},
},
{
LLM_ARCH_GPTJ,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
},
},
{
LLM_ARCH_GPTNEOX,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_MPT,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
},
},
{
LLM_ARCH_UNKNOWN,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
},
},
}; };
static llm_arch llm_arch_from_string(const std::string & name) { static llm_arch llm_arch_from_string(const std::string & name) {
@ -1605,10 +1643,14 @@ static void llm_load_hparams(
GGUF_GET_KEY(ctx, hparams.n_rot, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ROPE_DIMENSION_COUNT)); GGUF_GET_KEY(ctx, hparams.n_rot, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ROPE_DIMENSION_COUNT));
if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) {
if (hparams.n_rot != hparams.n_embd / hparams.n_head) { if (hparams.n_rot != hparams.n_embd / hparams.n_head) {
throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd / hparams.n_head)); throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd / hparams.n_head));
} }
} }
// gpt-neox n_rot = rotary_pct * (n_embd / n_head)
// gpt-j n_rot = rotary_dim
}
// arch-specific KVs // arch-specific KVs
switch (model.arch) { switch (model.arch) {