gate attn impls

.github/workflows/pr-gpu.yaml

@@ -31,6 +31,11 @@ jobs:
           container: mosaicml/pytorch:2.1.0_cu121-python3.10-ubuntu20.04
           markers: 'gpu'
           pytest_command: 'coverage run -m pytest'
+        # TODO: After the PR with the flash attention 2 images goes in, add the new unit test suite
+        # - name: 'gpu-2.1.0-flash2'
+        #   container: # UPDATE AFTER THIS PR GOES TO PRODUCTION IMAGE
+        #   markers: 'gpu'
+        #   pytest_command: 'coverage run -m pytest'
     name: ${{ matrix.name }}
     if: github.repository_owner == 'mosaicml'
     with:

llmfoundry/models/layers/attention.py

@@ -17,17 +17,19 @@
 from llmfoundry.models.layers.norm import NORM_CLASS_REGISTRY
 
 
-def raise_if_flash_attn_v2():
-    flash_attn_version = None
-    # This only needs to be in a try except so that huggingface does not try to import it
+def is_flash_v2_installed():
     try:
-        from flash_attn import __version__ as flash_attn_version
-        if version.parse(flash_attn_version) >= version.parse('2.0.0'):
-            raise RuntimeError(
-                'flash-attn==2.0.0+ is not supported. Please use flash-attn==1.0.9.'
-            )
+        import flash_attn as flash_attn
+    except:
+        return False
+    return version.parse(flash_attn.__version__) >= version.parse('2.0.0')
+
+def is_flash_v1_installed():
+    try:
+        import flash_attn as flash_attn
     except:
-        pass
+        return False
+    return version.parse(flash_attn.__version__) < version.parse('2.0.0')
 
 
 def _reset_is_causal(num_query_tokens: int, num_key_tokens: int,
@@ -291,18 +293,35 @@ def flash_attn_fn(
 
     reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
 
-    output_unpad = flash_attn_interface.flash_attn_unpadded_func(
-        query_unpad,
-        key_unpad,
-        value_unpad,
-        cu_seqlens_q,
-        cu_seqlens_k,
-        max_seqlen_q,
-        max_seqlen_k,
-        dropout_p,
-        softmax_scale=softmax_scale,
-        causal=reset_is_causal,
-        return_attn_probs=needs_weights)
+    if is_flash_v1_installed():
+        output_unpad = flash_attn_interface.flash_attn_unpadded_func(
+            query_unpad,
+            key_unpad,
+            value_unpad,
+            cu_seqlens_q,
+            cu_seqlens_k,
+            max_seqlen_q,
+            max_seqlen_k,
+            dropout_p,
+            softmax_scale=softmax_scale,
+            causal=reset_is_causal,
+            return_attn_probs=needs_weights)
+    elif is_flash_v2_installed():
+        output_unpad = flash_attn_interface.flash_attn_func(
+            query_unpad,
+            key_unpad,
+            value_unpad,
+            cu_seqlens_q,
+            cu_seqlens_k,
+            max_seqlen_q,
+            max_seqlen_k,
+            dropout_p,
+            softmax_scale=softmax_scale,
+            causal=reset_is_causal,
+            return_attn_probs=needs_weights)
+    else:
+        raise RuntimeError(
+            'flash-attn==1.0.9 or flash-attn==2.3.2 is required.')
 
     output = bert_padding.pad_input(
         rearrange(output_unpad, 'nnz h d -> nnz (h d)'), indices_q, batch_size,