use the enable_gqa param in torch.nn.functional.scaled_dot_product_at… (#39412)

* use the enable_gqa param in torch.nn.functional.scaled_dot_product_attention

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* ci failure fix

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* add check

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* fix ci failure

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* refine code, extend to cuda

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* refine code

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* fix review comments

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* refine the PR

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

---------

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
Co-authored-by: Cyril Vallez <cyril.vallez@huggingface.co>

src/transformers/integrations/sdpa_attention.py

@@ -3,11 +3,15 @@ from typing import Optional
 import torch
 
 from ..utils import logging
+from ..utils.import_utils import is_torch_greater_or_equal
 
 
 logger = logging.get_logger(__name__)
 
 
+_is_torch_greater_or_equal_than_2_5 = is_torch_greater_or_equal("2.5", accept_dev=True)
+
+
 def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
     """
     This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
@@ -20,6 +24,14 @@ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
     return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
 
 
+def use_gqa_in_sdpa(attention_mask: Optional[torch.Tensor], key: torch.Tensor) -> bool:
+    # GQA can only be used under the following conditions
+    # 1. torch version >= 2.5
+    # 2. attention_mask is None (otherwise it will fall back to the math kernel)
+    # 3. key is not a torch.fx.Proxy (otherwise it will fail with a tracing error)
+    return _is_torch_greater_or_equal_than_2_5 and attention_mask is None and not isinstance(key, torch.fx.Proxy)
+
+
 def sdpa_attention_forward(
     module: torch.nn.Module,
     query: torch.Tensor,
@@ -36,10 +48,13 @@ def sdpa_attention_forward(
             "`sdpa` attention does not support `output_attentions=True` or `head_mask`."
             " Please set your attention to `eager` if you want any of these features."
         )
-
+    sdpa_kwargs = {}
     if hasattr(module, "num_key_value_groups"):
+        if not use_gqa_in_sdpa(attention_mask, key):
             key = repeat_kv(key, module.num_key_value_groups)
             value = repeat_kv(value, module.num_key_value_groups)
+        else:
+            sdpa_kwargs = {"enable_gqa": True}
 
     if attention_mask is not None and attention_mask.ndim == 4:
         attention_mask = attention_mask[:, :, :, : key.shape[-2]]
@@ -71,6 +86,7 @@ def sdpa_attention_forward(
         dropout_p=dropout,
         scale=scaling,
         is_causal=is_causal,
+        **sdpa_kwargs,
     )
     attn_output = attn_output.transpose(1, 2).contiguous()
 

tests/utils/test_cache_utils.py

@@ -588,12 +588,12 @@ class CacheExportIntegrationTest(unittest.TestCase):
             past_key_values=past_key_values_eager,
             use_cache=True,
         )
-        self.assertTrue(torch.allclose(res.logits, res_eager.logits))
+        self.assertTrue(torch.allclose(res.logits, res_eager.logits, atol=1e-5))
         for k1, k2 in zip(res.past_key_values.key_cache, res_eager.past_key_values.key_cache):
-            self.assertTrue(torch.allclose(k1, k2))
+            self.assertTrue(torch.allclose(k1, k2, atol=1e-5))
 
         for v1, v2 in zip(res.past_key_values.value_cache, res_eager.past_key_values.value_cache):
-            self.assertTrue(torch.allclose(v1, v2))
+            self.assertTrue(torch.allclose(v1, v2, atol=1e-5))
 
     def test_dynamic_cache_exportability_multiple_run(self):
         # When exporting with DynamicCache, you should export two graphs:
@@ -686,10 +686,10 @@ class CacheExportIntegrationTest(unittest.TestCase):
         )
 
         for k1, k2 in zip(res_export_2.past_key_values.key_cache, res_eager_2.past_key_values.key_cache):
-            self.assertTrue(torch.allclose(k1, k2))
+            self.assertTrue(torch.allclose(k1, k2, atol=1e-5))
 
         for v1, v2 in zip(res_export_2.past_key_values.value_cache, res_eager_2.past_key_values.value_cache):
-            self.assertTrue(torch.allclose(v1, v2))
+            self.assertTrue(torch.allclose(v1, v2, atol=1e-5))
 
     @unittest.skip("Runs on my machine locally, passed, no idea why it does not online")
     def test_static_cache_exportability(self):