FIX [quantization / ESM] Fix ESM 8bit / 4bit with bitsandbytes (#29329)

* fix ESM 8bit

* Apply suggestions from code review

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* fixup

---------

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

src/transformers/models/esm/modeling_esm.py

@@ -377,7 +377,7 @@ class EsmSelfAttention(nn.Module):
         if head_mask is not None:
             attention_probs = attention_probs * head_mask
 
-        context_layer = torch.matmul(attention_probs, value_layer)
+        context_layer = torch.matmul(attention_probs.to(value_layer.dtype), value_layer)
 
         context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
         new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)

src/transformers/quantizers/quantizer_bnb_4bit.py

@@ -121,7 +121,7 @@ class Bnb4BitHfQuantizer(HfQuantizer):
         import bitsandbytes as bnb
 
         module, tensor_name = get_module_from_name(model, param_name)
-        if isinstance(module._parameters[tensor_name], bnb.nn.Params4bit):
+        if isinstance(module._parameters.get(tensor_name, None), bnb.nn.Params4bit):
             # Add here check for loaded components' dtypes once serialization is implemented
             return True
         elif isinstance(module, bnb.nn.Linear4bit) and tensor_name == "bias":

src/transformers/quantizers/quantizer_bnb_8bit.py

@@ -139,7 +139,7 @@ class Bnb8BitHfQuantizer(HfQuantizer):
         import bitsandbytes as bnb
 
         module, tensor_name = get_module_from_name(model, param_name)
-        if isinstance(module._parameters[tensor_name], bnb.nn.Int8Params):
+        if isinstance(module._parameters.get(tensor_name, None), bnb.nn.Int8Params):
             if self.pre_quantized:
                 if param_name.replace("weight", "SCB") not in state_dict.keys():
                     raise ValueError("Missing quantization component `SCB`")

tests/models/esm/test_modeling_esm.py

@@ -18,7 +18,7 @@
 import unittest
 
 from transformers import EsmConfig, is_torch_available
-from transformers.testing_utils import TestCasePlus, require_torch, slow, torch_device
+from transformers.testing_utils import TestCasePlus, require_bitsandbytes, require_torch, slow, torch_device
 
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
@@ -303,9 +303,9 @@ class EsmModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
         pass
 
 
+@slow
 @require_torch
 class EsmModelIntegrationTest(TestCasePlus):
-    @slow
     def test_inference_masked_lm(self):
         with torch.no_grad():
             model = EsmForMaskedLM.from_pretrained("facebook/esm2_t6_8M_UR50D")
@@ -323,7 +323,6 @@ class EsmModelIntegrationTest(TestCasePlus):
             )
             self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
 
-    @slow
     def test_inference_no_head(self):
         with torch.no_grad():
             model = EsmModel.from_pretrained("facebook/esm2_t6_8M_UR50D")
@@ -336,3 +335,18 @@ class EsmModelIntegrationTest(TestCasePlus):
                 [[[0.1444, 0.5413, 0.3248], [0.3034, 0.0053, 0.3108], [0.3228, -0.2499, 0.3415]]]
             )
             self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
+
+    @require_bitsandbytes
+    def test_inference_bitsandbytes(self):
+        model = EsmForMaskedLM.from_pretrained("facebook/esm2_t36_3B_UR50D", load_in_8bit=True)
+
+        input_ids = torch.tensor([[0, 6, 4, 13, 5, 4, 16, 12, 11, 7, 2]])
+        # Just test if inference works
+        with torch.no_grad():
+            _ = model(input_ids)[0]
+
+        model = EsmForMaskedLM.from_pretrained("facebook/esm2_t36_3B_UR50D", load_in_4bit=True)
+
+        input_ids = torch.tensor([[0, 6, 4, 13, 5, 4, 16, 12, 11, 7, 2]])
+        # Just test if inference works
+        _ = model(input_ids)[0]