Commit Graph

132 Commits

Author SHA1 Message Date
oobabooga
09a7c36e1b Minor improvement while running custom models 2023-03-06 15:36:35 -03:00
oobabooga
24c4c20391 Fix bug in multigpu setups (attempt #2) 2023-03-06 15:23:29 -03:00
oobabooga
d88b7836c6 Fix bug in multigpu setups 2023-03-06 14:58:30 -03:00
oobabooga
5bed607b77 Increase repetition frequency/penalty for RWKV 2023-03-06 14:25:48 -03:00
oobabooga
bf56b6c1fb Load settings.json without the need for --settings settings.json
This is for setting UI defaults
2023-03-06 10:57:45 -03:00
oobabooga
e91f4bc25a Add RWKV tokenizer 2023-03-06 08:45:49 -03:00
oobabooga
c855b828fe Better handle <USER> 2023-03-05 17:01:47 -03:00
oobabooga
2af66a4d4c Fix <USER> in pygmalion replies 2023-03-05 16:08:50 -03:00
oobabooga
a54b91af77 Improve readability 2023-03-05 10:21:15 -03:00
oobabooga
8e706df20e Fix a memory leak when text streaming is on 2023-03-05 10:12:43 -03:00
oobabooga
c33715ad5b Move towards HF LLaMA implementation 2023-03-05 01:20:31 -03:00
oobabooga
bd8aac8fa4 Add LLaMA 8-bit support 2023-03-04 13:28:42 -03:00
oobabooga
c93f1fa99b Count the tokens more conservatively 2023-03-04 03:10:21 -03:00
oobabooga
ed8b35efd2 Add --pin-weight parameter for FlexGen 2023-03-04 01:04:02 -03:00
oobabooga
05e703b4a4 Print the performance information more reliably 2023-03-03 21:24:32 -03:00
oobabooga
5a79863df3 Increase the sequence length, decrease batch size
I have no idea what I am doing
2023-03-03 15:54:13 -03:00
oobabooga
a345a2acd2 Add a tokenizer placeholder 2023-03-03 15:16:55 -03:00
oobabooga
5b354817f6 Make chat minimally work with LLaMA 2023-03-03 15:04:41 -03:00
oobabooga
ea5c5eb3da Add LLaMA support 2023-03-03 14:39:14 -03:00
oobabooga
2bff646130 Stop chat from flashing dark when processing 2023-03-03 13:19:13 -03:00
oobabooga
169209805d Model-aware prompts and presets 2023-03-02 11:25:04 -03:00
oobabooga
7bbe32f618 Don't return a value in an iterator function 2023-03-02 00:48:46 -03:00
oobabooga
ff9f649c0c Remove some unused imports 2023-03-02 00:36:20 -03:00
oobabooga
1a05860ca3 Ensure proper no-streaming with generation_attempts > 1 2023-03-02 00:10:10 -03:00
oobabooga
a2a3e8f797 Add --rwkv-strategy parameter 2023-03-01 20:02:48 -03:00
oobabooga
449116a510 Fix RWKV paths on Windows (attempt) 2023-03-01 19:17:16 -03:00
oobabooga
955cf431e8 Minor consistency fix 2023-03-01 19:11:26 -03:00
oobabooga
f3da6dcc8f
Merge pull request #149 from oobabooga/RWKV
Add RWKV support
2023-03-01 16:57:45 -03:00
oobabooga
831ac7ed3f Add top_p 2023-03-01 16:45:48 -03:00
oobabooga
7c4d5ca8cc Improve the text generation call a bit 2023-03-01 16:40:25 -03:00
oobabooga
2f16ce309a Rename a variable 2023-03-01 12:33:09 -03:00
oobabooga
9e9cfc4b31 Parameters 2023-03-01 12:19:37 -03:00
oobabooga
0f6708c471 Sort the imports 2023-03-01 12:18:17 -03:00
oobabooga
e735806c51 Add a generate() function for RWKV 2023-03-01 12:16:11 -03:00
oobabooga
659bb76722 Add RWKVModel class 2023-03-01 12:08:55 -03:00
oobabooga
9c86a1cd4a Add RWKV pip package 2023-03-01 11:42:49 -03:00
oobabooga
6837d4d72a Load the model by name 2023-02-28 02:52:29 -03:00
oobabooga
a1429d1607 Add default extensions to the settings 2023-02-28 02:20:11 -03:00
oobabooga
19ccb2aaf5 Handle <USER> and <BOT> 2023-02-28 01:05:43 -03:00
oobabooga
626da6c731 Handle {{user}} and {{char}} in example dialogue 2023-02-28 00:59:05 -03:00
oobabooga
e861e68e38 Move the chat example dialogue to the prompt 2023-02-28 00:50:46 -03:00
oobabooga
f871971de1 Trying to get the chat to work 2023-02-28 00:25:30 -03:00
oobabooga
67ee7bead7 Add cpu, bf16 options 2023-02-28 00:09:11 -03:00
oobabooga
ebd698905c Add streaming to RWKV 2023-02-28 00:04:04 -03:00
oobabooga
70e522732c Move RWKV loader into a separate file 2023-02-27 23:50:16 -03:00
oobabooga
ebc64a408c RWKV support prototype 2023-02-27 23:03:35 -03:00
oobabooga
021bd55886 Better format the prompt when generation attempts > 1 2023-02-27 21:37:03 -03:00
oobabooga
43b6ab8673 Store thumbnails as files instead of base64 strings
This improves the UI responsiveness for large histories.
2023-02-27 13:41:00 -03:00
oobabooga
f24b6e78a3 Fix clear history 2023-02-26 23:58:04 -03:00
oobabooga
8e3e8a070f Make FlexGen work with the newest API 2023-02-26 16:53:41 -03:00