[Attention] Use FA3 for MLA on Hopper (#12807)

Signed-off-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>

vllm/attention/backends/flash_attn.py

@@ -14,19 +14,16 @@ from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
                                               AttentionMetadataBuilder,
                                               AttentionType)
 from vllm.attention.backends.utils import (
-    PAD_SLOT_ID, CommonAttentionState, compute_slot_mapping,
-    compute_slot_mapping_start_idx, get_num_prefill_decode_query_kv_tokens,
-    get_seq_len_block_table_args, is_all_cross_attn_metadata_set,
-    is_all_encoder_attn_metadata_set, is_block_tables_empty)
-from vllm.envs import VLLM_FLASH_ATTN_VERSION
+    PAD_SLOT_ID, VLLM_FLASH_ATTN_VERSION, CommonAttentionState,
+    compute_slot_mapping, compute_slot_mapping_start_idx,
+    get_num_prefill_decode_query_kv_tokens, get_seq_len_block_table_args,
+    is_all_cross_attn_metadata_set, is_all_encoder_attn_metadata_set,
+    is_block_tables_empty)
 from vllm.logger import init_logger
 from vllm.multimodal import MultiModalPlaceholderMap
-from vllm.platforms import current_platform
 from vllm.utils import async_tensor_h2d, make_tensor_with_pad
-from vllm.vllm_flash_attn import (fa_version_unsupported_reason,
-                                  flash_attn_varlen_func,
-                                  flash_attn_with_kvcache,
-                                  is_fa_version_supported)
+from vllm.vllm_flash_attn import (flash_attn_varlen_func,
+                                  flash_attn_with_kvcache)
 
 if TYPE_CHECKING:
     from vllm.worker.model_runner import (ModelInputForGPUBuilder,
@@ -644,25 +641,6 @@ class FlashAttentionImpl(AttentionImpl):
                 f"Supported head sizes are: {support_head_sizes}.")
         self.attn_type = attn_type
 
-        # if hopper default to FA3, otherwise stick to FA2 for now
-        # TODO(lucas): profile FA3 on ampere to see if it makes sense to
-        #  use FA3 as default for both
-        if current_platform.get_device_capability()[0] >= 9:
-            self.fa_version = 3 if is_fa_version_supported(3) else 2
-        else:
-            self.fa_version = 2
-
-        if VLLM_FLASH_ATTN_VERSION is not None:
-            assert VLLM_FLASH_ATTN_VERSION in [2, 3]
-            self.fa_version = VLLM_FLASH_ATTN_VERSION
-
-        if not is_fa_version_supported(self.fa_version):
-            logger.error("Cannot use FA version %d is not supported due to %s",
-                         self.fa_version,
-                         fa_version_unsupported_reason(self.fa_version))
-
-        assert is_fa_version_supported(self.fa_version)
-
     def forward(
         self,
         layer: AttentionLayer,
@@ -781,7 +759,7 @@ class FlashAttentionImpl(AttentionImpl):
                     alibi_slopes=alibi_slopes,
                     softcap=logits_soft_cap,
                     out=prefill_output,
-                    fa_version=self.fa_version,
+                    fa_version=VLLM_FLASH_ATTN_VERSION,
                 )
             else:
                 # prefix-enabled attention
@@ -804,7 +782,7 @@ class FlashAttentionImpl(AttentionImpl):
                     block_table=prefill_meta.block_tables,
                     softcap=logits_soft_cap,
                     out=prefill_output,
-                    fa_version=self.fa_version,
+                    fa_version=VLLM_FLASH_ATTN_VERSION,
                 )
 
         if decode_meta := attn_metadata.decode_metadata:
@@ -833,7 +811,7 @@ class FlashAttentionImpl(AttentionImpl):
                     softcap=logits_soft_cap,
                     block_table=decode_meta.block_tables,
                     out=decode_output,
-                    fa_version=self.fa_version,
+                    fa_version=VLLM_FLASH_ATTN_VERSION,
                 )
             else:
                 # Use flash_attn_with_kvcache for normal decoding.
@@ -854,7 +832,7 @@ class FlashAttentionImpl(AttentionImpl):
                     alibi_slopes=alibi_slopes,
                     softcap=logits_soft_cap,
                     out=decode_output.unsqueeze(1),
-                    fa_version=self.fa_version,
+                    fa_version=VLLM_FLASH_ATTN_VERSION,
                 )
         return output
 

vllm/attention/backends/mla/utils.py

@@ -12,6 +12,7 @@ from vllm import envs
 from vllm.attention.backends.abstract import (AttentionLayer,
                                               AttentionMetadata,
                                               MLAAttentionImpl, T)
+from vllm.attention.backends.utils import VLLM_FLASH_ATTN_VERSION
 from vllm.distributed import (get_tensor_model_parallel_world_size,
                               tensor_model_parallel_all_reduce)
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
@@ -533,6 +534,7 @@ class MLACommonImpl(MLAAttentionImpl[T], Generic[T]):
             max_seqlen_k=max_prefill_seq_len,
             softmax_scale=self.scale,
             causal=True,
+            fa_version=VLLM_FLASH_ATTN_VERSION,
         )
         attn_output = attn_output\
             .view(-1, self.num_heads, q.shape[-1])[..., :v.shape[-1]]\

vllm/attention/backends/utils.py

@@ -8,12 +8,17 @@ from typing import TYPE_CHECKING, Any, Dict, List, Tuple, Type, TypeVar, Union
 import numpy as np
 import torch
 
+from vllm import envs
 from vllm.attention import (AttentionMetadata, AttentionMetadataBuilder,
                             AttentionState)
 from vllm.attention.backends.abstract import AttentionType
+from vllm.logger import logging
 from vllm.multimodal import MultiModalPlaceholderMap
+from vllm.platforms import current_platform
 from vllm.utils import async_tensor_h2d, make_tensor_with_pad
 
+logger = logging.getLogger(__name__)
+
 if TYPE_CHECKING:
     from vllm.worker.model_runner_base import ModelRunnerBase
 
@@ -580,3 +585,32 @@ def get_num_prefill_decode_query_kv_tokens(
 
     return (num_prefill_query_tokens, num_prefill_kv_tokens,
             num_decode_query_tokens)
+
+
+try:
+    from vllm.vllm_flash_attn.flash_attn_interface import (
+        fa_version_unsupported_reason, is_fa_version_supported)
+
+    def flash_attn_version():
+        # if hopper default to FA3, otherwise stick to FA2 for now
+        # TODO(lucas): profile FA3 on ampere to see if it makes sense to
+        #  use FA3 as default for both
+        if current_platform.get_device_capability()[0] >= 9:
+            fa_version = 3 if is_fa_version_supported(3) else 2
+        else:
+            fa_version = 2
+
+        if envs.VLLM_FLASH_ATTN_VERSION is not None:
+            assert envs.VLLM_FLASH_ATTN_VERSION in [2, 3]
+            fa_version = envs.VLLM_FLASH_ATTN_VERSION
+
+        if not is_fa_version_supported(fa_version):
+            logger.error("Cannot use FA version %d is not supported due to %s",
+                         fa_version, fa_version_unsupported_reason(fa_version))
+
+        assert is_fa_version_supported(fa_version)
+        return fa_version
+
+    VLLM_FLASH_ATTN_VERSION = flash_attn_version()
+except ImportError:
+    VLLM_FLASH_ATTN_VERSION = None

vllm/v1/attention/backends/flash_attn.py

@@ -10,13 +10,10 @@ import triton.language as tl
 
 from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
                                               AttentionMetadata, AttentionType)
-from vllm.envs import VLLM_FLASH_ATTN_VERSION
+from vllm.attention.backends.utils import VLLM_FLASH_ATTN_VERSION
 from vllm.logger import init_logger
-from vllm.platforms import current_platform
 from vllm.utils import cdiv
-from vllm.vllm_flash_attn import (fa_version_unsupported_reason,
-                                  flash_attn_varlen_func,
-                                  is_fa_version_supported)
+from vllm.vllm_flash_attn import flash_attn_varlen_func
 
 logger = init_logger(__name__)
 
@@ -136,25 +133,6 @@ class FlashAttentionImpl(AttentionImpl):
                                       "are not implemented for "
                                       "FlashAttentionImpl")
 
-        # if hopper default to FA3, otherwise stick to FA2 for now
-        # TODO(lucas): profile FA3 on ampere to see if it makes sense to
-        #  use FA3 as default for both
-        if current_platform.get_device_capability()[0] >= 9:
-            self.fa_version = 3 if is_fa_version_supported(3) else 2
-        else:
-            self.fa_version = 2
-
-        if VLLM_FLASH_ATTN_VERSION is not None:
-            assert VLLM_FLASH_ATTN_VERSION in [2, 3]
-            self.fa_version = VLLM_FLASH_ATTN_VERSION
-
-        if not is_fa_version_supported(self.fa_version):
-            logger.error("Cannot use FA version %d is not supported due to %s",
-                         self.fa_version,
-                         fa_version_unsupported_reason(self.fa_version))
-
-        assert is_fa_version_supported(self.fa_version)
-
     def forward(
         self,
         layer: torch.nn.Module,
@@ -227,7 +205,7 @@ class FlashAttentionImpl(AttentionImpl):
                 window_size=self.sliding_window,
                 block_table=attn_metadata.block_table,
                 softcap=self.logits_soft_cap,
-                fa_version=self.fa_version,
+                fa_version=VLLM_FLASH_ATTN_VERSION,
             )
             return output
 
@@ -249,7 +227,7 @@ class FlashAttentionImpl(AttentionImpl):
             logits_soft_cap=self.logits_soft_cap,
             block_table=attn_metadata.block_table,
             common_prefix_len=attn_metadata.common_prefix_len,
-            fa_version=self.fa_version,
+            fa_version=VLLM_FLASH_ATTN_VERSION,
         )
         return output