fix lora training with alpaca_lora_4bit (#3853)

docs/GPTQ-models-(4-bit-mode).md

@@ -163,22 +163,17 @@ This requires using a monkey patch that is supported by this web UI: https://git
 
 To use it:
 
-1. Clone `johnsmith0031/alpaca_lora_4bit` into the repositories folder:
+1. Install alpaca_lora_4bit using pip
 
 ```
-cd text-generation-webui/repositories
+git clone https://github.com/johnsmith0031/alpaca_lora_4bit.git
-git clone https://github.com/johnsmith0031/alpaca_lora_4bit
+cd alpaca_lora_4bit
+git fetch origin winglian-setup_pip
+git checkout winglian-setup_pip
+pip install .
 ```
 
-⚠️  I have tested it with the following commit specifically: `2f704b93c961bf202937b10aac9322b092afdce0`
+2. Start the UI with the `--monkey-patch` flag:
-
-2. Install https://github.com/sterlind/GPTQ-for-LLaMa with this command:
-
-```
-pip install git+https://github.com/sterlind/GPTQ-for-LLaMa.git@lora_4bit
-```
-
-3. Start the UI with the `--monkey-patch` flag:
 
 ```
 python server.py --model llama-7b-4bit-128g --listen --lora tloen_alpaca-lora-7b --monkey-patch

modules/monkey_patch_gptq_lora.py

@@ -1,39 +1,35 @@
 # Copied from https://github.com/johnsmith0031/alpaca_lora_4bit
 
-import sys
 from pathlib import Path
 
-sys.path.insert(0, str(Path("repositories/alpaca_lora_4bit")))
+import alpaca_lora_4bit.autograd_4bit as autograd_4bit
-
+from alpaca_lora_4bit.amp_wrapper import AMPWrapper
-import autograd_4bit
+from alpaca_lora_4bit.autograd_4bit import (
-from amp_wrapper import AMPWrapper
-from autograd_4bit import (
     Autograd4bitQuantLinear,
     load_llama_model_4bit_low_ram
 )
-from monkeypatch.peft_tuners_lora_monkey_patch import (
+from alpaca_lora_4bit.models import Linear4bitLt
-    Linear4bitLt,
+from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
-    replace_peft_model_with_gptq_lora_model
+    replace_peft_model_with_int4_lora_model
 )
 
 from modules import shared
 from modules.GPTQ_loader import find_quantized_model_file
 
-replace_peft_model_with_gptq_lora_model()
+replace_peft_model_with_int4_lora_model()
 
 
 def load_model_llama(model_name):
     config_path = str(Path(f'{shared.args.model_dir}/{model_name}'))
     model_path = str(find_quantized_model_file(model_name))
     model, tokenizer = load_llama_model_4bit_low_ram(config_path, model_path, groupsize=shared.args.groupsize, is_v1_model=False)
-    for n, m in model.named_modules():
+    for _, m in model.named_modules():
         if isinstance(m, Autograd4bitQuantLinear) or isinstance(m, Linear4bitLt):
             if m.is_v1_model:
                 m.zeros = m.zeros.half()
             m.scales = m.scales.half()
             m.bias = m.bias.half()
 
-    autograd_4bit.use_new = True
     autograd_4bit.auto_switch = True
 
     model.half()

modules/training.py

@@ -269,10 +269,10 @@ def calc_trainable_parameters(model):
 def do_train(lora_name: str, always_override: bool, save_steps: int, micro_batch_size: int, batch_size: int, epochs: int, learning_rate: str, lr_scheduler_type: str, lora_rank: int, lora_alpha: int, lora_dropout: float, cutoff_len: int, dataset: str, eval_dataset: str, format: str, eval_steps: int, raw_text_file: str, overlap_len: int, newline_favor_len: int, higher_rank_limit: bool, warmup_steps: int, optimizer: str, hard_cut_string: str, train_only_after: str, stop_at_loss: float, add_eos_token: bool, min_chars: int, report_to: str):
 
     if shared.args.monkey_patch:
-        from monkeypatch.peft_tuners_lora_monkey_patch import (
+        from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
-            replace_peft_model_with_gptq_lora_model
+            replace_peft_model_with_int4_lora_model
         )
-        replace_peft_model_with_gptq_lora_model()
+        replace_peft_model_with_int4_lora_model()
 
     global WANT_INTERRUPT
     WANT_INTERRUPT = False
@@ -512,11 +512,12 @@ def do_train(lora_name: str, always_override: bool, save_steps: int, micro_batch
         return
 
     if shared.args.monkey_patch:
-        for n, m in lora_model.named_modules():
+        from alpaca_lora_4bit.autograd_4bit import Autograd4bitQuantLinear
-            if '4bit' in str(type(m)):
+        from alpaca_lora_4bit.models import Linear4bitLt
+        for _, m in lora_model.named_modules():
+            if isinstance(m, Autograd4bitQuantLinear) or isinstance(m, Linear4bitLt):
                 if m.is_v1_model:
                     m.zeros = m.zeros.half()
-
                 m.scales = m.scales.half()
 
     class Tracked():