Fix : BLOOM tie_word_embeddings in GGUF (#35812)

* fix bloom ggml

* fix falcon output

* make style

Author: MekkCyber

src/transformers/modeling_gguf_pytorch_utils.py

@@ -400,7 +400,7 @@ def load_gguf_checkpoint(gguf_checkpoint_path, return_tensors=False, model_to_lo
 
     # Handle tie_word_embeddings, if lm_head.weight is not present in tensors,
     # tie_word_embeddings is true otherwise false
-    exceptions = ["falcon"]
+    exceptions = ["falcon", "bloom"]
     parsed_parameters["config"]["tie_word_embeddings"] = (
         all("output.weight" != tensor.name for tensor in reader.tensors) or architecture in exceptions
     )

tests/quantization/ggml/test_ggml.py

@@ -633,7 +633,7 @@ def test_falcon7b_q2_k(self):
         text = tokenizer(self.example_text, return_tensors="pt")["input_ids"].to(torch_device)
         out = model.generate(text, max_new_tokens=16)
 
-        EXPECTED_TEXT = 'Hello,\nI am trying to use the "get_post_meta"'
+        EXPECTED_TEXT = "Hello All,\nI am new to this forum.\nI am using the "
         self.assertEqual(tokenizer.decode(out[0], skip_special_tokens=True), EXPECTED_TEXT)
 
     @unittest.skip("The test causes a torch.OutOfMemoryError on the CI but it passes with enough memory")