Fix the incorrect permutation of gguf (#31788)

* Fix the incorrect permutation of gguf

* rename num_kv_heads

Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>

* add typing to num_kv_heads

Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>

* rename variables

* refactor permute function name

* update the expected text of the llama3 q4 test

---------

Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>

src/transformers/modeling_gguf_pytorch_utils.py

@@ -14,6 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from typing import Optional
+
 import numpy as np
 from tqdm import tqdm
 
@@ -147,10 +149,11 @@ def load_gguf_checkpoint(gguf_checkpoint_path, return_tensors=False):
 
             if architecture == "llama" and (".attn_k." in name or ".attn_q." in name):
                 num_heads = parsed_parameters["config"]["num_attention_heads"]
-                tmp_shape = (int(shape[-1] // num_heads // 2), num_heads, 2, shape[0])
+                num_kv_heads = parsed_parameters["config"]["num_key_value_heads"]
-                weights = weights.reshape(tmp_shape)
+                if ".attn_q." in name:
-                weights = weights.transpose(0, 2, 1, 3)
+                    weights = reverse_permute_weights(weights, num_heads, num_heads)
-                weights = weights.reshape(shape[::-1])
+                elif ".attn_k." in name:
+                    weights = reverse_permute_weights(weights, num_heads, num_kv_heads)
 
             for tensor_name in tensor_key_mapping:
                 if tensor_name in name:
@@ -163,3 +166,14 @@ def load_gguf_checkpoint(gguf_checkpoint_path, return_tensors=False):
         logger.info(f"Some keys of the GGUF file were not considered: {reader_keys}")
 
     return parsed_parameters
+
+
+def reverse_permute_weights(weights: np.ndarray, n_head: int, num_kv_heads: Optional[int] = None) -> np.ndarray:
+    # Original permutation implementation
+    # https://github.com/ggerganov/llama.cpp/blob/a38b884c6c4b0c256583acfaaabdf556c62fabea/convert_hf_to_gguf.py#L1402-L1408
+    if num_kv_heads is not None and n_head != num_kv_heads:
+        n_head = num_kv_heads
+
+    dim = weights.shape[0] // n_head // 2
+    w = weights.reshape(n_head, dim, 2, *weights.shape[1:])
+    return w.swapaxes(2, 1).reshape(weights.shape)

tests/quantization/ggml/test_ggml.py

@@ -188,8 +188,7 @@ class GgufIntegrationTests(unittest.TestCase):
         text = tokenizer(self.example_text, return_tensors="pt").to(torch_device)
         out = model.generate(**text, max_new_tokens=10)
 
-        EXPECTED_TEXT = "Hello, I am new to this forum. I am"
+        EXPECTED_TEXT = "Hello, I am interested in [The Park]\nThe"
-
         self.assertEqual(tokenizer.decode(out[0], skip_special_tokens=True), EXPECTED_TEXT)
 
     def test_tokenization_xnli(self):