SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Fix : BLOOM tie_word_embeddings in GGUF (#35812)

Browse Source 
* fix bloom ggml

* fix falcon output

* make style
This commit is contained in:
Mohamed Mekkouri
2025-01-21 15:35:54 +01:00
committed by GitHub
parent 678bd7f1ce
commit dbd8474125
2 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/transformers/modeling_gguf_pytorch_utils.py
View File

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

2
tests/quantization/ggml/test_ggml.py
View File

@@ -633,7 +633,7 @@ class GgufIntegrationTests(unittest.TestCase):
text = tokenizer(self.example_text, return_tensors="pt")["input_ids"].to(torch_device) text = tokenizer(self.example_text, return_tensors="pt")["input_ids"].to(torch_device)
out = model.generate(text, max_new_tokens=16) 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) 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") @unittest.skip("The test causes a torch.OutOfMemoryError on the CI but it passes with enough memory")