8000 Merge branch 'main' into binary-wheels · tebjan/llama-cpp-python@79de495 · GitHub
[go: up one dir, main page]
Skip to content

Commit 79de495

Browse files
abetlenabetlen
committed
Merge branch 'main' into binary-wheels
2 parents 22bc1e8 + f96de6d commit 79de495

File tree

15 files changed

+952
-446
lines changed

15 files changed

+952
-446
lines changed

‎CHANGELOG.md‎

Lines changed: 25 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -7,9 +7,33 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
77

88
## [Unreleased]
99

10+
## [0.2.58]
11+
12+
- feat: Update llama.cpp to ggerganov/llama.cpp@ba0c7c70ab5b15f1f2be7fb0dfbe0366dda30d6c
13+
- feat: add support for KV cache quantization options by @Limour-dev in #1307
14+
- feat: Add logprobs support to chat completions by @windspirit95 in #1311
15+
- fix: set LLAMA_METAL_EMBED_LIBRARY=on on MacOS arm64 by @bretello in #1289
16+
- feat: Add tools/functions variables to Jinja2ChatFormatter, add function response formatting for all simple chat formats by @CISC in #1273
17+
- fix: Changed local API doc references to hosted by by @lawfordp2017 in #1317
18+
19+
## [0.2.57]
20+
21+
- feat: Update llama.cpp to ggerganov/llama.cpp@ac9ee6a4ad740bc1ee484ede43e9f92b5af244c1
22+
- fix: set default embedding pooling type to unspecified by @abetlen in 4084aabe867b8ec2aba1b22659e59c9318b0d1f3
23+
- fix: Fix and optimize functionary chat handler by @jeffrey-fong in #1282
24+
- fix: json mode for basic chat formats by @abetlen in 20e6815252d0efd9f015f7adbf108faaf36e3f3c
25+
26+
## [0.2.56]
27+
28+
- feat: Update llama.cpp to ggerganov/llama.cpp@c2101a2e909ac7c08976d414e64e96c90ee5fa9e
29+
- feat(server): Add endpoints for tokenize, detokenize and count tokens by @felipelo in #1136
30+
- feat: Switch embed to llama_get_embeddings_seq by @iamlemec in #1263
31+
- fix: Fixed json strings grammar by blacklisting character control set by @ExtReMLapin in d02a9cf16ff88ad011e2eb1ce29f4d9400f13cd1
32+
- fix: Check for existence of clip model path by @kejcao in #1264
33+
1034
## [0.2.55]
1135

12-
- feat: Update llama.cpp to ggerganov/9731134296af3a6839cd682e51d9c2109a871de5
36+
- feat: Update llama.cpp to ggerganov/llama.cpp@9731134296af3a6839cd682e51d9c2109a871de5
1337
- docs: fix small typo in README: 'model know how' -> 'model knows how' by @boegel in #1244
1438

1539
## [0.2.54]

‎CMakeLists.txt‎

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -17,6 +17,11 @@ if (LLAMA_BUILD)
1717
set(LLAMA_FMA "Off" CACHE BOOL "llama: enable FMA" FORCE)
1818
set(LLAMA_F16C "Off" CACHE BOOL "llama: enable F16C" FORCE)
1919
endif()
20+
21+
if (APPLE AND CMAKE_SYSTEM_PROCESSOR MATCHES "arm64")
22+
set(LLAMA_METAL_EMBED_LIBRARY "On" CACHE BOOL "llama: embed metal library" FORCE)
23+
endif()
24+
2025
add_subdirectory(vendor/llama.cpp)
2126
install(
2227
TARGETS llama
< 6D40 /div>

‎README.md‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -322,7 +322,7 @@ For OpenAI API v1 compatibility, you use the [`create_chat_completion_openai_v1`
322322

323323
### JSON and JSON Schema Mode
324324

325-
To constrain chat responses to only valid JSON or a specific JSON Schema use the `response_format` argument in [`create_chat_completion`](http://localhost:8000/api-reference/#llama_cpp.Llama.create_chat_completion).
325+
To constrain chat responses to only valid JSON or a specific JSON Schema use the `response_format` argument in [`create_chat_completion`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_chat_completion).
326326

327327
#### JSON Mode
328328

@@ -530,7 +530,7 @@ llama = Llama(
530530

531531
### Embeddings
532 9E88 532

533-
To generate text embeddings use [`create_embedding`](http://localhost:8000/api-reference/#llama_cpp.Llama.create_embedding).
533+
To generate text embeddings use [`create_embedding`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_embedding).
534534

535535
```python
536536
import llama_cpp

‎llama_cpp/__init__.py‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,4 +1,4 @@
11
from .llama_cpp import *
22
from .llama import *
33

4-
__version__ = "0.2.55"
4+
__version__ = "0.2.58"

‎llama_cpp/_internals.py‎

Lines changed: 6 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -730,12 +730,14 @@ def sample(
730730
if len(self.prev) > 0:
731731
nl_token = ctx_main.model.token_nl()
732732
nl_logit = logits_array[nl_token]
733-
if self.params.penalty_last_n > 0:
733+
last_tokens = self.prev[-self.params.penalty_last_n:]
734+
last_tokens_size = min(l 6D38 en(last_tokens), self.params.penalty_last_n)
735+
if last_tokens_size > 0:
736+
last_tokens_p = (llama_cpp.llama_token * len(last_tokens))(*last_tokens)
734737
ctx_main.sample_repetition_penalties(
735738
token_data_array,
736-
# TODO: Only create this once
737-
(llama_cpp.llama_token * len(self.prev))(*self.prev),
738-
self.params.penalty_last_n,
739+
last_tokens_p,
740+
last_tokens_size,
739741
self.params.penalty_repeat,
740742
self.params.penalty_freq,
741743
self.params.penalty_present,

‎llama_cpp/llama.py‎

Lines changed: 30 additions & 46 deletions
Original file line numberDiff line numberDiff line change
@@ -79,6 +79,7 @@ def __init__(
7979
n_threads: Optional[int] = None,
8080
n_threads_batch: Optional[int] = None,
8181
rope_scaling_type: Optional[int] = llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED,
82+
pooling_type: int = llama_cpp.LLAMA_POOLING_TYPE_UNSPECIFIED,
8283
rope_freq_base: float = 0.0,
8384
rope_freq_scale: float = 0.0,
8485
yarn_ext_factor: float = -1.0,
@@ -104,6 +105,9 @@ def __init__(
104105
draft_model: Optional[LlamaDraftModel] = None,
105106
# Tokenizer Override
106107
tokenizer: Optional[BaseLlamaTokenizer] = None,
108+
# KV cache quantization
109+
type_k: Optional[int] = None,
110+
type_v: Optional[int] = None,
107111
# Misc
108112
verbose: bool = True,
109113
# Extra Params
@@ -151,6 +155,7 @@ def __init__(
151155
n_threads: Number of threads to use for generation
152156
n_threads_batch: Number of threads to use for batch processing
153157
rope_scaling_type: RoPE scaling type, from `enum llama_rope_scaling_type`. ref: https://github.com/ggerganov/llama.cpp/pull/2054
158+
pooling_type: Pooling type, from `enum llama_pooling_type`.
154159
rope_freq_base: RoPE base frequency, 0 = from model
155160
rope_freq_scale: RoPE frequency scaling factor, 0 = from model
156161
yarn_ext_factor: YaRN extrapolation mix factor, negative = from model
@@ -170,6 +175,8 @@ def __init__(
170175
draft_model: Optional draft model to use for speculative decoding.
171176
tokenizer: Optional tokenizer to override the default tokenizer from llama.cpp.
172177
verbose: Print verbose output to stderr.
178+
type_k: KV cache data type for K (default: f16)
179+
type_v: KV cache data type for V (default: f16)
173180
174181
Raises:
175182
ValueError: If the model path does not exist.
@@ -271,6 +278,7 @@ def __init__(
271278
if rope_scaling_type is not None
272279
else llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED
273280
)
281+
self.context_params.pooling_type = pooling_type
274282
self.context_params.rope_freq_base = (
275283
rope_freq_base if rope_freq_base != 0.0 else 0
276284
)
@@ -293,9 +301,13 @@ def __init__(
293301
self.context_params.logits_all = (
294302
logits_all if draft_model is None else True
295303
) # Must be set to True for speculative decoding
296-
self.context_params.embedding = embedding
304+
self.context_params.embeddings = embedding # TODO: Rename to embeddings
297305
self.context_params.offload_kqv = offload_kqv
298-
306+
# KV cache quantization
307+
if type_k is not None:
308+
self.context_params.type_k = type_k
309+
if type_v is not None:
310+
self.context_params.type_v = type_v
299311
# Sampling Params
300312
self.last_n_tokens_size = last_n_tokens_size
301313

@@ -787,7 +799,7 @@ def embed(
787799
n_embd = self.n_embd()
788800
n_batch = self.n_batch
789801

790-
if self.context_params.embedding == False:
802+
if self.context_params.embeddings == False:
791803
raise RuntimeError(
792804
"Llama model must be created with embedding=True to call this method"
793805
)
@@ -814,9 +826,12 @@ def decode_batch(n_seq: int):
814826

815827
# store embeddings
816828
for i in range(n_seq):
817-
embedding: List[float] = llama_cpp.llama_get_embeddings_ith(
829+
ptr = llama_cpp.llama_get_embeddings_seq(
818830
self._ctx.ctx, i
819-
)[:n_embd]
831+
)
832+
if not ptr:
833+
raise RuntimeError("Failed to get embeddings from sequence pooling type is not set")
834+
embedding: List[float] = ptr[:n_embd]
820835
if normalize:
821836
norm = float(np.linalg.norm(embedding))
822837
embedding = [v / norm for v in embedding]
@@ -1647,6 +1662,7 @@ def create_chat_completion(
16471662
top_k=top_k,
16481663
min_p=min_p,
16491664
typical_p=typical_p,
1665+
logprobs=top_logprobs if logprobs else None,
16501666
stream=stream,
16511667
stop=stop,
16521668
seed=seed,
@@ -1717,6 +1733,7 @@ def __getstate__(self):
17171733
n_threads=self.context_params.n_threads,
17181734
n_threads_batch=self.context_params.n_threads_batch,
17191735
rope_scaling_type=self.context_params.rope_scaling_type,
1736+
pooling_type=self.context_params.pooling_type,
17201737
rope_freq_base=self.context_params.rope_freq_base,
17211738
rope_freq_scale=self.context_params.rope_freq_scale,
17221739
yarn_ext_factor=self.context_params.yarn_ext_factor,
@@ -1725,7 +1742,8 @@ def __getstate__(self):
17251742
yarn_beta_slow=self.context_params.yarn_beta_slow,
17261743
yarn_orig_ctx=self.context_params.yarn_orig_ctx,
17271744
logits_all=self.context_params.logits_all,
1728-
embedding=self.context_params.embedding,
1745+
embedding=self.context_params.embeddings,
1746+
offload_kqv=self.context_params.offload_kqv,
17291747
# Sampling Params
17301748
last_n_tokens_size=self.last_n_tokens_size,
17311749
# LoRA Params
@@ -1737,51 +1755,17 @@ def __getstate__(self):
17371755
# Chat Format Params
17381756
chat_format=self.chat_format,
17391757
chat_handler=self.chat_handler,
1758+
# Speculative Decidng
1759+
draft_model=self.draft_model,
1760+
# KV cache quantization
1761+
type_k=self.context_params.type_k,
1762+
type_v=self.context_params.type_v,
17401763
# Misc
17411764
verbose=self.verbose,
17421765
)
17431766

17441767
def __setstate__(self, state):
1745-
self.__init__(
1746-
model_path=state["model_path"],
1747-
# Model Params
1748-
n_gpu_layers=state["n_gpu_layers"],
1749-
split_mode=state["split_mode"],
1750-
main_gpu=state["main_gpu"],
1751-
tensor_split=state["tensor_split"],
1752-
vocab_only=state["vocab_only"],
1753-
use_mmap=state["use_mmap"],
1754-
use_mlock=state["use_mlock"],
1755-
kv_overrides=state["kv_overrides"],
1756-
# Context Params
1757-
seed=state["seed"],
1758-
n_ctx=state["n_ctx"],
1759-
n_batch=state["n_batch"],
1760-
n_threads=state["n_threads"],
1761-
n_threads_batch=state["n_threads_batch"],
1762-
rope_freq_base=state["rope_freq_base"],
1763-
rope_freq_scale=state["rope_freq_scale"],
1764-
rope_scaling_type=state["rope_scaling_type"],
1765-
yarn_ext_factor=state["yarn_ext_factor"],
1766-
yarn_attn_factor=state["yarn_attn_factor"],
1767-
yarn_beta_fast=state["yarn_beta_fast"],
1768-
yarn_beta_slow=state["yarn_beta_slow"],
1769-
yarn_orig_ctx=state["yarn_orig_ctx"],
1770-
logits_all=state["logits_all"],
1771-
embedding=state["embedding"],
1772-
# Sampling Params
1773-
last_n_tokens_size=state["last_n_tokens_size"],
1774-
# LoRA Params
1775-
lora_base=state["lora_base"],
1776-
lora_path=state["lora_path"],
1777-
# Backend Params
1778-
numa=state["numa"],
1779-
# Chat Format Params
1780-
chat_format=state["chat_format"],
1781-
chat_handler=state["chat_handler"],
1782-
# Misc
1783-
verbose=state["verbose"],
1784-
)
1768+
self.__init__(**state)
17851769

17861770
def save_state(self) -> LlamaState:
17871771
assert self._ctx.ctx is not None

0 commit comments

Comments
 (0)
0