rm Chameleon's slow tokenizer

src/transformers/models/auto/tokenization_auto.py

@@ -109,10 +109,7 @@
             ("canine", ("CanineTokenizer", None)),
             (
                 "chameleon",
-                (
-                    "LlamaTokenizer" if is_sentencepiece_available() else None,
-                    "LlamaTokenizerFast" if is_tokenizers_available() else None,
-                ),
+                (None, "LlamaTokenizerFast" if is_tokenizers_available() else None),
             ),
             ("chinese_clip", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
             (

src/transformers/models/chameleon/convert_chameleon_weights_to_hf.py

@@ -24,7 +24,7 @@
 
 from transformers import (
     ChameleonConfig,
-    ChameleonForCausalLM,
+    ChameleonForConditionalGeneration,
     ChameleonImageProcessor,
     ChameleonProcessor,
 )
@@ -49,10 +49,10 @@
 Thereafter, models can be loaded via:
 
 ```py
-from transformers import ChameleonForCausalLM, LlamaTokenizer
+from transformers import ChameleonForConditionalGeneration, LlamaTokenizerFast
 
-model = ChameleonForCausalLM.from_pretrained("/output/path")
-tokenizer = LlamaTokenizer.from_pretrained("/output/path")
+model = ChameleonForConditionalGeneration.from_pretrained("/output/path")
+tokenizer = LlamaTokenizerFast.from_pretrained("/output/path")
 ```
 
 Important note: you need to be able to host the whole model in RAM to execute this script (even if the biggest versions
@@ -372,7 +372,7 @@ def permute(w, n_heads, dim1=dim, dim2=dim):
         vocabulary_map=vocabulary_map,
     )
     with init_empty_weights():
-        model = ChameleonForCausalLM(config)
+        model = ChameleonForConditionalGeneration(config)
 
     model.load_state_dict(state_dict, assign=True, strict=False)
     model.save_pretrained(model_path, safe_serialization=True)
@@ -397,7 +397,7 @@ def permute(w, n_heads, dim1=dim, dim2=dim):
     # taken from https://github.com/facebookresearch/chameleon/blob/7a72f40aa5f462965c8374f25257f55b65b25ff4/data/prompts_for_human_evaluations.jsonl
     print("Loading the checkpoint in a Chameleon model...")
     print("*" * 100)
-    model = ChameleonForCausalLM.from_pretrained(
+    model = ChameleonForConditionalGeneration.from_pretrained(
         model_path, attn_implementation="eager", torch_dtype=torch.bfloat16, device_map="auto"
     )
     processor = ChameleonProcessor.from_pretrained(model_path)

src/transformers/models/chameleon/processing_chameleon.py

@@ -68,7 +68,7 @@ class ChameleonProcessor(ProcessorMixin):
     """
 
     attributes = ["image_processor", "tokenizer"]
-    tokenizer_class = ("LlamaTokenizer", "LlamaTokenizerFast")
+    tokenizer_class = "LlamaTokenizerFast"
     image_processor_class = "ChameleonImageProcessor"
 
     def __init__(self, image_processor, tokenizer, image_seq_length: int = 1024, image_token: str = "<image>"):

tests/models/chameleon/test_processor_chameleon.py

@@ -2,7 +2,6 @@
 import unittest
 
 from transformers import ChameleonProcessor
-from transformers.models.auto.processing_auto import processor_class_from_name
 
 from ...test_processing_common import ProcessorTesterMixin
 
@@ -11,26 +10,6 @@ class ChameleonProcessorTest(ProcessorTesterMixin, unittest.TestCase):
     from_pretrained_id = "leloy/Anole-7b-v0.1-hf"
     processor_class = ChameleonProcessor
 
-    def get_component(self, attribute, **kwargs):
-        assert attribute in self.processor_class.attributes
-        if attribute != "tokenizer":
-            return super().get_component(attribute, **kwargs)
-        # We use the fast tokenizer by default as the slow tokenizer expects the vocab file to be present in the loading directory.
-        # This vocab file is neither in the official repo nor does it get saved when we save the processor in `setUp` below.
-        component_class_name = getattr(self.processor_class, f"{attribute}_class")
-        if isinstance(component_class_name, tuple):
-            if "_fast" in component_class_name[0]:
-                component_class_name = component_class_name[0]
-            else:
-                component_class_name = component_class_name[1]
-
-        component_class = processor_class_from_name(component_class_name)
-        component = component_class.from_pretrained(self.tmpdirname, **kwargs)  # noqa
-        if attribute == "tokenizer" and not component.pad_token:
-            component.pad_token = "[TEST_PAD]"
-
-        return component
-
     def setUp(self):
         self.tmpdirname = tempfile.mkdtemp()
         processor = self.processor_class.from_pretrained(self.from_pretrained_id)