baalajimaestro/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity
1521 commits 1 branch 0 tags 2.1 MiB
Commit graph

304 commits

Author SHA1 Message Date
Andrei Betlen
e6d6260a91 fix: Update from_pretrained defaults to match hf_hub_download 2024-02-22 00:10:23 -05:00
Andrei
7f51b6071f
feat(low-level-api): Improve API static type-safety and performance (#1205) 2024-02-21 16:25:38 -05:00
Andrei
0f8aa4ab5c
feat: Pull models directly from huggingface (#1206) 
* Add from_pretrained method to Llama class

* Update docs

* Merge filename and pattern
 2024-02-21 16:25:10 -05:00
Andrei Betlen
53f6f5f415 fix: self.numa missing 2024-02-17 01:02:33 -05:00
Andrei Betlen
fdce078cb9 feat: Update llama.cpp 2024-02-17 00:37:51 -05:00
Andrei Betlen
0ce66bc080 fix: create_embedding broken response for input type str 2024-02-15 16:09:48 -05:00
Douglas Hanley
7bb91f025f
fix: Incorporate embedding pooling layer fixes (#1194) 
* remove division by token count

* truncate to n_batch, not n_ctx
 2024-02-15 15:16:30 -05:00
Douglas Hanley
d7a67917ba
feat: Support batch embeddings (#1186) 
* handle batched embeddings

* fix normalization issue

* fix type hints, ensure no breaking changes to embed

* Clear kv cache / reset internal state after embedding complete

---------

Co-authored-by: Andrei <abetlen@gmail.com>
 2024-02-14 04:26:09 -05:00
Andrew Lapp
d6be5333e1
fix: sample idx off-by-one error for logit_processors (#1179) 
* fix sample_idx off-by-one error

* self._scores is indexed differently, only modify the index within self._input_ids

---------

Co-authored-by: Andrew Lapp <andrew@rew.la>
Co-authored-by: Andrei <abetlen@gmail.com>
 2024-02-13 12:26:07 -05:00
Andrei Betlen
cb791716b4 fix: Always set logits_all = True when using speculative decoding 2024-02-12 16:19:05 -05:00
Andrei
153a0049d9
feat: Generic chatml Function Calling (#957) 
* Add demo notebook

* Add initial chat handler

* Update OpenAI types

* Add generic chatml function calling (wip)

* Update chatml generic function calling.

* Progress on auto-tool calls

* fix streaming functions

* Remove print statements

* fix: Suppress output from llama.cpp init and grammar creation

* Add OpenAI v1 python api compatible chat completion function

* Support non-streaming multi-tool calls

* Format

* Include function_call in response.
 2024-02-12 15:56:07 -05:00
Andrei Betlen
4abb8c9386 Merge branch 'main' of github.com:abetlen/llama_cpp_python into main 2024-02-09 13:32:31 -05:00
Andrei Betlen
e16f06e6eb fix: revert _create_completions. 2024-02-09 02:02:13 -05:00
Andrei Betlen
b5fca911b5 feat: Move tokenizer to own module 2024-02-08 01:08:18 -05:00
Jeffrey Fong
901827013b
feat: Integrate functionary v1.4 and v2 models + add custom tokenizer support to Llama class (#1078) 
* convert functionary-v1 chat handler to use hf autotokenizer

* add hf_tokenizer + inteegrate functionary-v1.4 prompt template

* integrate functionary v2 prompt template

* update readme

* set up parallel function calling wip

* set up parallel function calling

* Update README.md

* Update README.md

* refactor tokenizers

* include old functionary handler for backward compatibility

* add hf_tokenizer_path in server ModelSettings

* convert functionary-v1 chat handler to use hf autotokenizer

* add hf_tokenizer + inteegrate functionary-v1.4 prompt template

* integrate functionary v2 prompt template

* update readme

* set up parallel function calling wip

* resolve merge conflict

* Update README.md

* Update README.md

* refactor tokenizers

* include old functionary handler for backward compatibility

* add hf_tokenizer_path in server ModelSettings

* Cleanup PR, fix breaking changes

* Use hf_pretrained_model_name_or_path for tokenizer

* fix hf tokenizer in streaming

* update README

* refactor offset mapping

---------

Co-authored-by: Andrei <abetlen@gmail.com>
 2024-02-07 20:07:03 -05:00
Andrei Betlen
59760c85ed fix: Use llama_log_callback to avoid suppress_stdout_stderr 2024-02-05 21:52:12 -05:00
Andrei
fb762a6041
Add speculative decoding (#1120) 
* Add draft model param to llama class, implement basic prompt lookup decoding draft model

* Use samplingcontext for sampling

* Use 1d array

* Use draft model for sampling

* Fix dumb mistake

* Allow for later extensions to the LlamaDraftModel api

* Cleanup

* Adaptive candidate prediction

* Update implementation to match hf transformers

* Tuning

* Fix bug where last token was not used for ngram prediction

* Remove heuristic for num_pred_tokens (no benefit)

* fix: n_candidates bug.

* Add draft_model_num_pred_tokens server setting

* Cleanup

* Update README
 2024-01-31 14:08:14 -05:00
Andrei
da003d8768
Automatically set chat format from gguf (#1110) 
* Use jinja formatter to load chat format from gguf

* Fix off-by-one error in metadata loader

* Implement chat format auto-detection
 2024-01-29 14:22:23 -05:00
Andrei Betlen
9677a1f2c8 fix: Check order 2024-01-23 22:28:03 -05:00
Andrei Betlen
4d6b2f7b91 fix: format 2024-01-23 22:08:27 -05:00
Phil H
fe5d6ea648
fix: GGUF metadata KV overrides, re #1011 (#1116) 
* kv overrides another attempt

* add sentinel element, simplify array population

* ensure sentinel element is zeroed
 2024-01-23 22:00:38 -05:00
Andrei Betlen
5a34c57e54 feat: Expose gguf model metadata in metadata property 2024-01-19 10:46:03 -05:00
Andrei Betlen
3babe3512c Fix mirostat sampling 2024-01-19 08:31:59 -05:00
Andrei Betlen
48c3b77e6f Offload KQV by default 2024-01-18 11:08:57 -05:00
Andrei Betlen
7b46bb5a78 Re-order classes in llama.py 2024-01-17 09:16:13 -05:00
Andrei Betlen
cc4630e66f Move helper classes to _internals submodule 2024-01-17 09:14:00 -05:00
Andrei Betlen
3b92419132 Move cache classes to llama_cache submodule. 2024-01-17 09:09:12 -05:00
Andrei Betlen
84615adbc6 Add split_mode option. Closes #1085 2024-01-15 12:49:20 -05:00
Phil H
76aafa6149
Implement GGUF metadata KV overrides (#1011) 
* Implement GGUF metadata overrides

* whitespace fix

* Fix kv overrides.

* Fix pointer and pickle

* Match llama.cpp kv_overrides cli argument

---------

Co-authored-by: Andrei <abetlen@gmail.com>
 2024-01-15 12:29:29 -05:00
Stephen Hankinson
df3be58d6c
Add ability to pass in penalize_nl param (#1068) 2024-01-10 02:46:27 -05:00
Andrei Betlen
d9a1d90fd7 Fix typo 2023-12-22 15:12:27 -05:00
swg
4b01a873ef
server: Support none defaulting to infinity for completions (#111) 
* Support defaulting to infinity or -1 for chat completions

* Check if completion_tokens is none in error handler.

* fix: max_tokens in create completion should match openai spec

* Fix __call__

---------

Co-authored-by: Andrei Betlen <abetlen@gmail.com>
 2023-12-22 14:05:13 -05:00
twaka
2f03fb0231
fix text_offset of multi-token characters (#1037) 
* fix text_offsets for bytes tokens

* fix
 2023-12-22 00:03:29 -05:00
Andrei Betlen
a05b4da80a fix: float32 is not JSON serializable when streaming logits. 2023-12-18 18:40:36 -05:00
Andrei Betlen
095c650006 Add offload_kqv option to llama and server 2023-12-18 15:36:09 -05:00
Andrei Betlen
472b344ae3 Remove unnused import 2023-12-18 15:32:40 -05:00
kddubey
6b2e0e05b4
perf: Don't convert logprobs arrays to lists (#1021) 2023-12-18 14:28:12 -05:00
Brandon Roberts
62944df142
Bugfix: Remove f16_kv, add offload_kqv field (#1019) 
F16_KV appears to have been removed here: af99c6fbfc

This addresses two issues:

 - #995 which just requests to add the KV cache offloading param
 - #1006 a NULL ptr exception when using the embeddings (introduced by
   leaving f16_kv in the fields struct)
 2023-12-18 14:27:11 -05:00
Daniele Morotti
f1c631dc53
Bug fixed with n_ctx=0 (#1015) 
If the n_ctx is set to 0 the code should use the maximum context length of the selected model, but it didn't work. There was a problem with the initialization of this parameter and a related problem with 'n_batch'.
 2023-12-16 18:59:50 -05:00
kddubey
5a8944672f
Fix logits_to_logprobs for 2-D and 3-D logits (#1002) 
* Fix logits_to_logprobs for 2-D and 3-D logits

* Set dtype to single

* Test size
 2023-12-16 18:59:26 -05:00
Tanner Hobson
ef22e478db
Replace logits_to_logprobs implementation with numpy equivalent to llama.cpp (#991) 
See #990. This change makes the logits_to_logprobs function equivalent to the version in the llama.cpp repository. It uses numpy so it's much faster than the previous version.
 2023-12-11 20:46:27 -05:00
Andrei Betlen
ec26f364cc Remove f16_kv 2023-12-11 10:25:37 -05:00
kddubey
b069d06346
Fix #891 (#952) 2023-11-29 05:39:52 -05:00
Andrei Betlen
6308f21d5e docs: Update Llama docs 2023-11-26 15:56:40 -05:00
Andrei Betlen
4026166e68 docs: Update completion and chat_completion parameter docstrings 2023-11-24 03:24:19 -05:00
Andrei Betlen
b6bb7ac76a docs: Add Llama class example 2023-11-22 23:10:04 -05:00
Andrei Betlen
7a3f87846b Format 2023-11-21 04:02:20 -05:00
Andrei Betlen
422ebc89ce Fix: Add logit_bias to all completion api methods 2023-11-21 04:01:36 -05:00
Andrei Betlen
07e47f55ba Add support for logit_bias outside of server api. Closes #827 2023-11-21 03:59:46 -05:00
TK-Master
b8438f70b5
Added support for min_p (#921) 
* Added support for min_p

My small contribution to this great project.

Ref: https://github.com/ggerganov/llama.cpp/pull/3841

Closes: https://github.com/abetlen/llama-cpp-python/issues/911

* Fix for negative temp (sample_softmax)
 2023-11-20 23:21:33 -05:00