[bnb] Fine-tuning HF 8-bit models (#21290)

* force `memory_efficient_backward=True`

* enhancements

- trainer support
- add new flag

* some changes

- internal changes in `Trainer`
- small refactor

* make quality

* Fixes

- add new testing util
- add new test
- change test in Trainer

* fix CI test

* educate users on how to ft 8bit models

* more checks

* fix `logger` error

* Apply suggestions from code review

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* adapt from review

* fix

* add comment

* use return instead

---------

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

src/transformers/modeling_utils.py

@@ -73,6 +73,7 @@ from .utils import (
     logging,
     replace_return_docstrings,
 )
+from .utils.import_utils import importlib_metadata
 from .utils.versions import require_version_core
 
 
@@ -2439,6 +2440,11 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                 model, threshold=load_in_8bit_threshold, modules_to_not_convert=modules_to_not_convert
             )
 
+            # training in 8-bit is only available in 0.37.0+
+            model._is_int8_training_enabled = version.parse(
+                importlib_metadata.version("bitsandbytes")
+            ) >= version.parse("0.37.0")
+
         if isinstance(device_map, str):
             if model._no_split_modules is None:
                 raise ValueError(f"{model.__class__.__name__} does not support `device_map='{device_map}'` yet.")

src/transformers/trainer.py

@@ -368,10 +368,18 @@ class Trainer:
 
         # At this stage the model is already loaded
         if getattr(model, "is_loaded_in_8bit", False):
-            raise ValueError(
+            if getattr(model, "_is_int8_training_enabled", False):
-                "The model you want to train is loaded in 8-bit precision. "
+                logger.info(
-                "Training an 8-bit model is not supported yet. "
+                    "The model is loaded in 8-bit precision. To train this model you need to add additional modules"
-            )
+                    " inside the model such as adapters using `peft` library and freeze the model weights. Please"
+                    " check "
+                    " the examples in https://github.com/huggingface/peft for more details."
+                )
+            else:
+                raise ValueError(
+                    "The model you want to train is loaded in 8-bit precision.  if you want to fine-tune an 8-bit"
+                    " model, please make sure that you have installed `bitsandbytes>=0.37.0`. "
+                )
 
         # Setup Sharded DDP training
         self.sharded_ddp = None
@@ -458,7 +466,7 @@ class Trainer:
         self.eval_dataset = eval_dataset
         self.tokenizer = tokenizer
 
-        if self.place_model_on_device:
+        if self.place_model_on_device and not getattr(model, "is_loaded_in_8bit", False):
             self._move_model_to_device(model, args.device)
 
         # Force n_gpu to 1 to avoid DataParallel as MP will manage the GPUs

tests/mixed_int8/test_mixed_int8.py

@@ -16,6 +16,8 @@ import gc
 import tempfile
 import unittest
 
+from packaging import version
+
 from transformers import (
     AutoModel,
     AutoModelForCausalLM,
@@ -33,10 +35,30 @@ from transformers.testing_utils import (
     require_torch_multi_gpu,
     slow,
 )
+from transformers.utils.versions import importlib_metadata
 
 
 if is_torch_available():
     import torch
+    import torch.nn as nn
+
+    class LoRALayer(nn.Module):
+        """Wraps a linear layer with LoRA-like adapter - Used for testing purposes only"""
+
+        def __init__(self, module: nn.Module, rank: int):
+            super().__init__()
+            self.module = module
+            self.adapter = nn.Sequential(
+                nn.Linear(module.in_features, rank, bias=False),
+                nn.Linear(rank, module.out_features, bias=False),
+            )
+            small_std = (2.0 / (5 * min(module.in_features, module.out_features))) ** 0.5
+            nn.init.normal_(self.adapter[0].weight, std=small_std)
+            nn.init.zeros_(self.adapter[1].weight)
+            self.adapter.to(module.weight.device)
+
+        def forward(self, input, *args, **kwargs):
+            return self.module(input, *args, **kwargs) + self.adapter(input)
 
 
 @require_bitsandbytes
@@ -335,3 +357,44 @@ class MixedInt8TestMultiGpu(BaseMixedInt8Test):
         # Second real batch
         output_parallel = model_parallel.generate(input_ids=encoded_input["input_ids"].to(0), max_new_tokens=10)
         self.assertEqual(self.tokenizer.decode(output_parallel[0], skip_special_tokens=True), self.EXPECTED_OUTPUT)
+
+
+class MixedInt8TestTraining(BaseMixedInt8Test):
+    def setUp(self):
+        self.model_name = "facebook/opt-350m"
+        super().setUp()
+
+    def test_training(self):
+        if version.parse(importlib_metadata.version("bitsandbytes")) < version.parse("0.37.0"):
+            return
+
+        # Step 1: freeze all parameters
+        model = AutoModelForCausalLM.from_pretrained(self.model_name, load_in_8bit=True, device_map="auto")
+
+        for param in model.parameters():
+            param.requires_grad = False  # freeze the model - train adapters later
+            if param.ndim == 1:
+                # cast the small parameters (e.g. layernorm) to fp32 for stability
+                param.data = param.data.to(torch.float32)
+
+        # Step 2: add adapters
+        for _, module in model.named_modules():
+            if "OPTAttention" in repr(type(module)):
+                module.q_proj = LoRALayer(module.q_proj, rank=16)
+                module.k_proj = LoRALayer(module.k_proj, rank=16)
+                module.v_proj = LoRALayer(module.v_proj, rank=16)
+
+        # Step 3: dummy batch
+        batch = self.tokenizer("Test batch ", return_tensors="pt").to(0)
+
+        # Step 4: Check if the gradient is not None
+        with torch.cuda.amp.autocast():
+            out = model.forward(**batch)
+            out.logits.norm().backward()
+
+        for module in model.modules():
+            if isinstance(module, LoRALayer):
+                self.assertTrue(module.adapter[1].weight.grad is not None)
+                self.assertTrue(module.adapter[1].weight.grad.norm().item() > 0)
+            elif isinstance(module, nn.Embedding):
+                self.assertTrue(module.weight.grad is None)