Merge pull request #230 from tushar2407/version0.1.6

Version0.1.6

README.md

@@ -170,8 +170,8 @@ model = BaseModel.load("x/distilgpt2_lora_finetuned_alpaca")
 - [x] INT4 LLaMA LoRA fine-tuning with INT4 generation
 - [x] Support for a `Generic model` wrapper
 - [x] Support for `Falcon-7B` model
+- [X] INT4 low-precision fine-tuning support
 - [ ] Evaluation of LLM models
-- [ ] INT4 low-precision fine-tuning support
 - [ ] INT3, INT2, INT1 low-precision fine-tuning support
 - [ ] Support for Stable Diffusion
 

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "xturing"
-version = "0.1.5"
+version = "0.1.6"
 description = "Fine-tuning, evaluation and data generation for LLMs"
 
 authors = [

src/xturing/__about__.py

@@ -1 +1 @@
-__version__ = "0.1.5"
+__version__ = "0.1.6"

src/xturing/config/finetuning_config.yaml

@@ -14,80 +14,87 @@ defaults:
   optimizer_name: adamw
   output_dir: saved_model
 
-llama:
+bloom:
   learning_rate: 5e-5
   weight_decay: 0.01
   num_train_epochs: 3
-  optimizer_name: cpu_adam
-
-llama_lora:
-  learning_rate: 1e-4
-  weight_decay: 0.01
-  num_train_epochs: 3
-  batch_size: 1
 
-llama_lora_int8:
+bloom_lora:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
-  batch_size: 8
-  max_length: 256
+  batch_size: 4
 
-llama_lora_int4:
+bloom_lora_int8:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
   batch_size: 8
   max_length: 256
 
-gptj:
+cerebras:
   learning_rate: 5e-5
   weight_decay: 0.01
   num_train_epochs: 3
-  optimizer_name: cpu_adam
 
-gptj_lora:
+cerebras_lora:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
-  batch_size: 1
+  batch_size: 4
 
-gptj_lora_int8:
+cerebras_lora_int8:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
   batch_size: 8
   max_length: 256
 
-gpt2:
+distilgpt2:
   learning_rate: 1e-3
   weight_decay: 0.01
   num_train_epochs: 3
   batch_size: 8
 
-gpt2_lora:
+distilgpt2_lora:
   learning_rate: 3e-3
   weight_decay: 0.01
   num_train_epochs: 3
   batch_size: 16
 
-gpt2_lora_int8:
-  learning_rate: 3e-3
+falcon:
+  learning_rate: 5e-5
   weight_decay: 0.01
   num_train_epochs: 3
-  batch_size: 16
+  batch_size: 1
+  max_length: 256
 
-distilgpt2:
-  learning_rate: 1e-3
+falcon_int8:
+  learning_rate: 1e-4
+  weight_decay: 0.01
+  num_train_epochs: 3
+  batch_size: 1
+  max_length: 256
+
+falcon_lora:
+  learning_rate: 1e-4
+  weight_decay: 0.01
+  num_train_epochs: 3
+  batch_size: 1
+
+falcon_lora_int8:
+  learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
   batch_size: 8
+  max_length: 256
 
-distilgpt2_lora:
-  learning_rate: 3e-3
+falcon_lora_kbit:
+  learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
-  batch_size: 16
+  batch_size: 8
+  max_length: 256
 
 galactica:
   learning_rate: 5e-5
@@ -108,109 +115,130 @@ galactica_lora_int8:
   batch_size: 8
   max_length: 256
 
-opt:
-  learning_rate: 5e-5
-  weight_decay: 0.01
-  num_train_epochs: 3
-
-opt_lora:
+generic:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
-  batch_size: 1
+  batch_size: 8
+  max_length: 256
 
-opt_lora_int8:
+generic_int8:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
   batch_size: 8
   max_length: 256
 
-cerebras:
-  learning_rate: 5e-5
+generic_lora:
+  learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
+  batch_size: 8
+  max_length: 256
 
-cerebras_lora:
+generic_lora_int8:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
-  batch_size: 4
+  batch_size: 8
+  max_length: 256
 
-cerebras_lora_int8:
+generic_lora_kbit:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
   batch_size: 8
   max_length: 256
 
-bloom:
+gptj:
   learning_rate: 5e-5
   weight_decay: 0.01
   num_train_epochs: 3
+  optimizer_name: cpu_adam
 
-bloom_lora:
+gptj_lora:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
-  batch_size: 4
+  batch_size: 1
 
-bloom_lora_int8:
+gptj_lora_int8:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
   batch_size: 8
   max_length: 256
 
-generic:
-  learning_rate: 1e-4
+gpt2:
+  learning_rate: 1e-3
   weight_decay: 0.01
   num_train_epochs: 3
   batch_size: 8
-  max_length: 256
 
-generic_int8:
-  learning_rate: 1e-4
+gpt2_lora:
+  learning_rate: 3e-3
   weight_decay: 0.01
   num_train_epochs: 3
-  batch_size: 8
-  max_length: 256
+  batch_size: 16
 
-generic_int8_lora:
+gpt2_lora_int8:
+  learning_rate: 3e-3
+  weight_decay: 0.01
+  num_train_epochs: 3
+  batch_size: 16
+
+llama:
+  learning_rate: 5e-5
+  weight_decay: 0.01
+  num_train_epochs: 3
+  optimizer_name: cpu_adam
+
+llama_lora:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
-  batch_size: 8
-  max_length: 256
+  batch_size: 1
 
-generic_lora:
+llama_lora_int8:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
   batch_size: 8
   max_length: 256
-
-falcon:
-  learning_rate: 5e-5
-  weight_decay: 0.01
+
+llama_lora_kbit:
+  learning_rate: 3e-4
   num_train_epochs: 3
   batch_size: 1
   max_length: 256
+  lora_r: 32
+  lora_alpha: 128
+  lora_groupsize: 128
+  lora_dropout: 0.05
+  seed: 0
+  cache: False
+  seqlen: 2048
+  kl_weight: 1.0
+  ce_weight: 200.0
+  save_freq: 1
+  trainable_kl_weight: False
+  trainable_ce_weight: False
+  weight_decay: 1e-5
+  intra_save_freq: 200
+  groupsize: 128
 
-falcon_int8:
-  learning_rate: 1e-4
+opt:
+  learning_rate: 5e-5
   weight_decay: 0.01
   num_train_epochs: 3
-  batch_size: 1
-  max_length: 256
 
-falcon_lora:
+opt_lora:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3
   batch_size: 1
 
-falcon_lora_int8:
+opt_lora_int8:
   learning_rate: 1e-4
   weight_decay: 0.01
   num_train_epochs: 3