alihan/llama-cpp-python
1
0
Fork 0
mirror of https://github.com/abetlen/llama-cpp-python.git synced 2023-09-07 17:34:22 +03:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'main' of github.com:abetlen/llama_cpp_python into better-server-params-and-fields

Browse Source
This commit is contained in:
Andrei Betlen
2023-05-07 01:54:00 -04:00
parent 3008a954c1 397ae97f64
commit d8fddcce73
13 changed files with 341 additions and 142 deletions
Download Patch File Download Diff File Expand all files Collapse all files

62
llama_cpp/llama.py
View File

@@ -33,12 +33,10 @@ class LlamaCache:
return k
return None
def __getitem__(
self, key: Sequence[llama_cpp.llama_token]
) -> Optional["LlamaState"]:
def __getitem__(self, key: Sequence[llama_cpp.llama_token]) -> "LlamaState":
_key = self._find_key(tuple(key))
if _key is None:
return None
raise KeyError(f"Key not found: {key}")
return self.cache_state[_key]
def __contains__(self, key: Sequence[llama_cpp.llama_token]) -> bool:
@@ -53,8 +51,8 @@ class LlamaState:
def __init__(
self,
eval_tokens: Deque[llama_cpp.llama_token],
eval_logits: Deque[List[llama_cpp.c_float]],
llama_state,
eval_logits: Deque[List[float]],
llama_state, # type: llama_cpp.Array[llama_cpp.c_uint8]
llama_state_size: llama_cpp.c_size_t,
):
self.eval_tokens = eval_tokens
@@ -129,7 +127,7 @@ class Llama:
self.last_n_tokens_size = last_n_tokens_size
self.n_batch = min(n_ctx, n_batch)
self.eval_tokens: Deque[llama_cpp.llama_token] = deque(maxlen=n_ctx)
self.eval_logits: Deque[List[llama_cpp.c_float]] = deque(
self.eval_logits: Deque[List[float]] = deque(
maxlen=n_ctx if logits_all else 1
)
@@ -247,7 +245,7 @@ class Llama:
n_vocab = llama_cpp.llama_n_vocab(self.ctx)
cols = int(n_vocab)
logits_view = llama_cpp.llama_get_logits(self.ctx)
logits: List[List[llama_cpp.c_float]] = [
logits: List[List[float]] = [
[logits_view[i * cols + j] for j in range(cols)] for i in range(rows)
]
self.eval_logits.extend(logits)
@@ -289,7 +287,7 @@ class Llama:
candidates=llama_cpp.ctypes.pointer(candidates),
penalty=repeat_penalty,
)
if temp == 0.0:
if float(temp.value) == 0.0:
return llama_cpp.llama_sample_token_greedy(
ctx=self.ctx,
candidates=llama_cpp.ctypes.pointer(candidates),
@@ -299,21 +297,25 @@ class Llama:
ctx=self.ctx,
candidates=llama_cpp.ctypes.pointer(candidates),
k=top_k,
min_keep=llama_cpp.c_size_t(1),
)
llama_cpp.llama_sample_tail_free(
ctx=self.ctx,
candidates=llama_cpp.ctypes.pointer(candidates),
z=llama_cpp.c_float(1.0),
min_keep=llama_cpp.c_size_t(1),
)
llama_cpp.llama_sample_typical(
ctx=self.ctx,
candidates=llama_cpp.ctypes.pointer(candidates),
p=llama_cpp.c_float(1.0),
min_keep=llama_cpp.c_size_t(1),
)
llama_cpp.llama_sample_top_p(
ctx=self.ctx,
candidates=llama_cpp.ctypes.pointer(candidates),
p=top_p,
min_keep=llama_cpp.c_size_t(1),
)
llama_cpp.llama_sample_temperature(
ctx=self.ctx,
@@ -390,18 +392,28 @@ class Llama:
"""
assert self.ctx is not None
if (
reset
and len(self.eval_tokens) > 0
and tuple(self.eval_tokens) == tuple(tokens[: len(self.eval_tokens)])
):
if self.verbose:
print("Llama.generate: cache hit", file=sys.stderr)
reset = False
tokens = tokens[len(self.eval_tokens) :]
if reset and len(self.eval_tokens) > 0:
longest_prefix = 0
for a, b in zip(self.eval_tokens, tokens[:-1]):
if a == b:
longest_prefix += 1
else:
break
if longest_prefix > 0:
if self.verbose:
print("Llama.generate: prefix-match hit", file=sys.stderr)
reset = False
tokens = tokens[longest_prefix:]
for _ in range(len(self.eval_tokens) - longest_prefix):
self.eval_tokens.pop()
try:
self.eval_logits.pop()
except IndexError:
pass
if reset:
self.reset()
while True:
self.eval(tokens)
token = self.sample(
@@ -639,7 +651,10 @@ class Llama:
self.detokenize([token]).decode("utf-8", errors="ignore")
for token in all_tokens
]
all_logprobs = [Llama._logits_to_logprobs(row) for row in self.eval_logits]
all_logprobs = [
Llama.logits_to_logprobs(list(map(float, row)))
for row in self.eval_logits
]
for token, token_str, logprobs_token in zip(
all_tokens, all_token_strs, all_logprobs
):
@@ -958,7 +973,10 @@ class Llama:
llama_state_compact = (llama_cpp.c_uint8 * int(n_bytes))()
llama_cpp.ctypes.memmove(llama_state_compact, llama_state, int(n_bytes))
if self.verbose:
print(f"Llama.save_state: saving {n_bytes} bytes of llama state", file=sys.stderr)
print(
f"Llama.save_state: saving {n_bytes} bytes of llama state",
file=sys.stderr,
)
return LlamaState(
eval_tokens=self.eval_tokens.copy(),
eval_logits=self.eval_logits.copy(),
@@ -985,7 +1003,7 @@ class Llama:
return llama_cpp.llama_token_bos()
@staticmethod
def logits_to_logprobs(logits: List[llama_cpp.c_float]) -> List[llama_cpp.c_float]:
def logits_to_logprobs(logits: List[float]) -> List[float]:
exps = [math.exp(float(x)) for x in logits]
sum_exps = sum(exps)
return [llama_cpp.c_float(math.log(x / sum_exps)) for x in exps]
return [math.log(x / sum_exps) for x in exps]

132
llama_cpp/llama_cpp.py
View File

@@ -8,6 +8,7 @@ from ctypes import (
c_void_p,
c_bool,
POINTER,
_Pointer, # type: ignore
Structure,
Array,
c_uint8,
@@ -17,7 +18,7 @@ import pathlib
# Load the library
def _load_shared_library(lib_base_name):
def _load_shared_library(lib_base_name: str):
# Determine the file extension based on the platform
if sys.platform.startswith("linux"):
lib_ext = ".so"
@@ -67,11 +68,11 @@ _lib_base_name = "llama"
_lib = _load_shared_library(_lib_base_name)
# C types
LLAMA_FILE_VERSION = ctypes.c_int(1)
LLAMA_FILE_VERSION = c_int(1)
LLAMA_FILE_MAGIC = b"ggjt"
LLAMA_FILE_MAGIC_UNVERSIONED = b"ggml"
LLAMA_SESSION_MAGIC = b"ggsn"
LLAMA_SESSION_VERSION = ctypes.c_int(1)
LLAMA_SESSION_VERSION = c_int(1)
llama_context_p = c_void_p
@@ -127,18 +128,23 @@ class llama_context_params(Structure):
llama_context_params_p = POINTER(llama_context_params)
LLAMA_FTYPE_ALL_F32 = ctypes.c_int(0)
LLAMA_FTYPE_MOSTLY_F16 = ctypes.c_int(1) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_0 = ctypes.c_int(2) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1 = ctypes.c_int(3) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = ctypes.c_int(
LLAMA_FTYPE_ALL_F32 = c_int(0)
LLAMA_FTYPE_MOSTLY_F16 = c_int(1) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_0 = c_int(2) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1 = c_int(3) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = c_int(
4
) # tok_embeddings.weight and output.weight are F16
LLAMA_FTYPE_MOSTLY_Q4_2 = ctypes.c_int(5) # except 1d tensors
# LLAMA_FTYPE_MOSTYL_Q4_3 = ctypes.c_int(6) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q8_0 = ctypes.c_int(7) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_0 = ctypes.c_int(8) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_1 = ctypes.c_int(9) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_2 = c_int(5) # except 1d tensors
# LLAMA_FTYPE_MOSTYL_Q4_3 = c_int(6) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q8_0 = c_int(7) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_0 = c_int(8) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_1 = c_int(9) # except 1d tensors
# Misc
c_float_p = POINTER(c_float)
c_uint8_p = POINTER(c_uint8)
c_size_t_p = POINTER(c_size_t)
# Functions
@@ -210,8 +216,8 @@ _lib.llama_model_quantize.restype = c_int
# Returns 0 on success
def llama_apply_lora_from_file(
ctx: llama_context_p,
path_lora: ctypes.c_char_p,
path_base_model: ctypes.c_char_p,
path_lora: c_char_p,
path_base_model: c_char_p,
n_threads: c_int,
) -> c_int:
return _lib.llama_apply_lora_from_file(ctx, path_lora, path_base_model, n_threads)
@@ -252,21 +258,25 @@ _lib.llama_get_state_size.restype = c_size_t
# Copies the state to the specified destination address.
# Destination needs to have allocated enough memory.
# Returns the number of bytes copied
def llama_copy_state_data(ctx: llama_context_p, dest) -> c_size_t:
def llama_copy_state_data(
ctx: llama_context_p, dest # type: Array[c_uint8]
) -> c_size_t:
return _lib.llama_copy_state_data(ctx, dest)
_lib.llama_copy_state_data.argtypes = [llama_context_p, POINTER(c_uint8)]
_lib.llama_copy_state_data.argtypes = [llama_context_p, c_uint8_p]
_lib.llama_copy_state_data.restype = c_size_t
# Set the state reading from the specified address
# Returns the number of bytes read
def llama_set_state_data(ctx: llama_context_p, src) -> c_size_t:
def llama_set_state_data(
ctx: llama_context_p, src # type: Array[c_uint8]
) -> c_size_t:
return _lib.llama_set_state_data(ctx, src)
_lib.llama_set_state_data.argtypes = [llama_context_p, POINTER(c_uint8)]
_lib.llama_set_state_data.argtypes = [llama_context_p, c_uint8_p]
_lib.llama_set_state_data.restype = c_size_t
@@ -274,9 +284,9 @@ _lib.llama_set_state_data.restype = c_size_t
def llama_load_session_file(
ctx: llama_context_p,
path_session: bytes,
tokens_out,
tokens_out, # type: Array[llama_token]
n_token_capacity: c_size_t,
n_token_count_out,
n_token_count_out, # type: _Pointer[c_size_t]
) -> c_size_t:
return _lib.llama_load_session_file(
ctx, path_session, tokens_out, n_token_capacity, n_token_count_out
@@ -288,13 +298,16 @@ _lib.llama_load_session_file.argtypes = [
c_char_p,
llama_token_p,
c_size_t,
POINTER(c_size_t),
c_size_t_p,
]
_lib.llama_load_session_file.restype = c_size_t
def llama_save_session_file(
ctx: llama_context_p, path_session: bytes, tokens, n_token_count: c_size_t
ctx: llama_context_p,
path_session: bytes,
tokens, # type: Array[llama_token]
n_token_count: c_size_t,
) -> c_size_t:
return _lib.llama_save_session_file(ctx, path_session, tokens, n_token_count)
@@ -374,22 +387,22 @@ _lib.llama_n_embd.restype = c_int
# Can be mutated in order to change the probabilities of the next token
# Rows: n_tokens
# Cols: n_vocab
def llama_get_logits(ctx: llama_context_p):
def llama_get_logits(ctx: llama_context_p): # type: (...) -> Array[float] # type: ignore
return _lib.llama_get_logits(ctx)
_lib.llama_get_logits.argtypes = [llama_context_p]
_lib.llama_get_logits.restype = POINTER(c_float)
_lib.llama_get_logits.restype = c_float_p
# Get the embeddings for the input
# shape: [n_embd] (1-dimensional)
def llama_get_embeddings(ctx: llama_context_p):
def llama_get_embeddings(ctx: llama_context_p): # type: (...) -> Array[float] # type: ignore
return _lib.llama_get_embeddings(ctx)
_lib.llama_get_embeddings.argtypes = [llama_context_p]
_lib.llama_get_embeddings.restype = POINTER(c_float)
_lib.llama_get_embeddings.restype = c_float_p
# Token Id -> String. Uses the vocabulary in the provided context
@@ -433,8 +446,8 @@ _lib.llama_token_nl.restype = llama_token
# @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
def llama_sample_repetition_penalty(
ctx: llama_context_p,
candidates,
last_tokens_data,
candidates, # type: _Pointer[llama_token_data_array]
last_tokens_data, # type: Array[llama_token]
last_tokens_size: c_int,
penalty: c_float,
):
@@ -456,8 +469,8 @@ _lib.llama_sample_repetition_penalty.restype = None
# @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
def llama_sample_frequency_and_presence_penalties(
ctx: llama_context_p,
candidates,
last_tokens_data,
candidates, # type: _Pointer[llama_token_data_array]
last_tokens_data, # type: Array[llama_token]
last_tokens_size: c_int,
alpha_frequency: c_float,
alpha_presence: c_float,
@@ -484,7 +497,9 @@ _lib.llama_sample_frequency_and_presence_penalties.restype = None
# @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
def llama_sample_softmax(ctx: llama_context_p, candidates):
def llama_sample_softmax(
ctx: llama_context_p, candidates # type: _Pointer[llama_token_data]
):
return _lib.llama_sample_softmax(ctx, candidates)
@@ -497,7 +512,10 @@ _lib.llama_sample_softmax.restype = None
# @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
def llama_sample_top_k(
ctx: llama_context_p, candidates, k: c_int, min_keep: c_size_t = c_size_t(1)
ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array]
k: c_int,
min_keep: c_size_t,
):
return _lib.llama_sample_top_k(ctx, candidates, k, min_keep)
@@ -513,7 +531,10 @@ _lib.llama_sample_top_k.restype = None
# @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
def llama_sample_top_p(
ctx: llama_context_p, candidates, p: c_float, min_keep: c_size_t = c_size_t(1)
ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array]
p: c_float,
min_keep: c_size_t,
):
return _lib.llama_sample_top_p(ctx, candidates, p, min_keep)
@@ -529,7 +550,10 @@ _lib.llama_sample_top_p.restype = None
# @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
def llama_sample_tail_free(
ctx: llama_context_p, candidates, z: c_float, min_keep: c_size_t = c_size_t(1)
ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array]
z: c_float,
min_keep: c_size_t,
):
return _lib.llama_sample_tail_free(ctx, candidates, z, min_keep)
@@ -545,7 +569,10 @@ _lib.llama_sample_tail_free.restype = None
# @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
def llama_sample_typical(
ctx: llama_context_p, candidates, p: c_float, min_keep: c_size_t = c_size_t(1)
ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array]
p: c_float,
min_keep: c_size_t,
):
return _lib.llama_sample_typical(ctx, candidates, p, min_keep)
@@ -559,7 +586,11 @@ _lib.llama_sample_typical.argtypes = [
_lib.llama_sample_typical.restype = None
def llama_sample_temperature(ctx: llama_context_p, candidates, temp: c_float):
def llama_sample_temperature(
ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array]
temp: c_float,
):
return _lib.llama_sample_temperature(ctx, candidates, temp)
@@ -578,7 +609,12 @@ _lib.llama_sample_temperature.restype = None
# @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
# @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
def llama_sample_token_mirostat(
ctx: llama_context_p, candidates, tau: c_float, eta: c_float, m: c_int, mu
ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array]
tau: c_float,
eta: c_float,
m: c_int,
mu, # type: _Pointer[c_float]
) -> llama_token:
return _lib.llama_sample_token_mirostat(ctx, candidates, tau, eta, m, mu)
@@ -589,7 +625,7 @@ _lib.llama_sample_token_mirostat.argtypes = [
c_float,
c_float,
c_int,
POINTER(c_float),
c_float_p,
]
_lib.llama_sample_token_mirostat.restype = llama_token
@@ -600,7 +636,11 @@ _lib.llama_sample_token_mirostat.restype = llama_token
# @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
# @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
def llama_sample_token_mirostat_v2(
ctx: llama_context_p, candidates, tau: c_float, eta: c_float, mu
ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array]
tau: c_float,
eta: c_float,
mu, # type: _Pointer[c_float]
) -> llama_token:
return _lib.llama_sample_token_mirostat_v2(ctx, candidates, tau, eta, mu)
@@ -610,13 +650,16 @@ _lib.llama_sample_token_mirostat_v2.argtypes = [
llama_token_data_array_p,
c_float,
c_float,
POINTER(c_float),
c_float_p,
]
_lib.llama_sample_token_mirostat_v2.restype = llama_token
# @details Selects the token with the highest probability.
def llama_sample_token_greedy(ctx: llama_context_p, candidates) -> llama_token:
def llama_sample_token_greedy(
ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array]
) -> llama_token:
return _lib.llama_sample_token_greedy(ctx, candidates)
@@ -628,7 +671,10 @@ _lib.llama_sample_token_greedy.restype = llama_token
# @details Randomly selects a token from the candidates based on their probabilities.
def llama_sample_token(ctx: llama_context_p, candidates) -> llama_token:
def llama_sample_token(
ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array]
) -> llama_token:
return _lib.llama_sample_token(ctx, candidates)

18
llama_cpp/server/__main__.py
View File

@@ -22,12 +22,26 @@ Then visit http://localhost:8000/docs to see the interactive API docs.
"""
import os
import argparse
import uvicorn
from llama_cpp.server.app import create_app
from llama_cpp.server.app import create_app, Settings
if __name__ == "__main__":
app = create_app()
parser = argparse.ArgumentParser()
for name, field in Settings.__fields__.items():
parser.add_argument(
f"--{name}",
dest=name,
type=field.type_,
default=field.default,
help=field.field_info.description,
)
args = parser.parse_args()
settings = Settings(**vars(args))
app = create_app(settings=settings)
uvicorn.run(
app, host=os.getenv("HOST", "localhost"), port=int(os.getenv("PORT", 8000))