Add attention_backend to let user choose

bump this into the constructor of BuilderArgs

Author: yanbing-j

torchchat/cli/builder.py

@@ -69,6 +69,7 @@ class BuilderArgs:
     prefill_possible: bool = False
     dynamic_shapes: bool = False
     max_seq_length: Optional[int] = None
+    attention_backend: str = "math"
 
     def __post_init__(self):
         if self.device is None:
@@ -178,6 +179,17 @@ def from_args(cls, args: argparse.Namespace) -> "BuilderArgs":
         pp = getattr(args, "pp", 1)
         tp = getattr(args, "tp", 1)
         chpt_from = getattr(args, "chpt_from", "hf")
+        sdp_backend_dict = {
+            'math': torch.nn.attention.SDPBackend.MATH,
+            'flash_attention': torch.nn.attention.SDPBackend.FLASH_ATTENTION,
+            'efficient_attention': torch.nn.attention.SDPBackend.EFFICIENT_ATTENTION,
+            'cudnn_attention': torch.nn.attention.SDPBackend.CUDNN_ATTENTION,
+        }
+        attention_backend = sdp_backend_dict[args.attention_backend]
+        if args.device == "cpu" and (args.attention_backend == "efficient_attention"
+                                     or args.attention_backend == "cudnn_attention"):
+            print(f"Warning: {args.attention_backend} is not supported on CPU. Using math instead.")
+            attention_backend = torch.nn.attention.SDPBackend.MATH
         return cls(
             checkpoint_dir=checkpoint_dir,
             checkpoint_path=checkpoint_path,
@@ -202,6 +214,7 @@ def from_args(cls, args: argparse.Namespace) -> "BuilderArgs":
             is_chat_model=is_chat_model,
             dynamic_shapes=getattr(args, "dynamic_shapes", False),
             max_seq_length=getattr(args, "max_seq_length", None),
+            attention_backend=attention_backend,
         )
 
     @classmethod

torchchat/cli/cli.py

@@ -179,6 +179,13 @@ def _add_model_config_args(parser, verb: str) -> None:
         choices=["fast", "cpu", "cuda", "mps"],
         help="Hardware device to use. Options: fast, cpu, cuda, mps",
     )
+    model_config_parser.add_argument(
+        "--attention-backend",
+        type=str,
+        default="math",
+        choices=["math", "flash_attention", "efficient_attention", "cudnn_attention"],
+        help="SDPBackend to use. Options: MATH, FLASH_ATTENTION, EFFICIENT_ATTENTION, CUDNN_ATTENTION",
+    )
 
 
 # Add CLI Args representing output paths of exported model files

torchchat/generate.py

@@ -26,6 +26,7 @@
 import torch.distributed as dist
 import torch.multiprocessing as mp
 from torch.distributed.pipelining import PipelineStage, ScheduleGPipe
+from torch._C import _SDPBackend as SDPBackend
 
 from PIL import Image
 
@@ -531,6 +532,7 @@ def decode_n_tokens(
         callback=lambda _: _,
         eos_token_id: int = 2,
         eot_id: Optional[int] = None,
+        attention_backend: SDPBackend = torch.nn.attention.SDPBackend.MATH,
         **sampling_kwargs,
     ):
         new_tokens, new_probs = [], []
@@ -539,7 +541,7 @@ def decode_n_tokens(
             num_new_tokens - 1
         ):  # -1 to save space to run an EoS if dont generate it naturally
             # Actually better for Inductor to codegen attention here
-            with torch.nn.attention.sdpa_kernel([torch.nn.attention.SDPBackend.MATH]):
+            with torch.nn.attention.sdpa_kernel([attention_backend]):
 
                 out_token = cur_token.clone()
                 next_token, next_prob = self.decode_one_token(
@@ -683,6 +685,7 @@ def generate(
         sequential_prefill=True,
         callback=lambda x: x,
         max_seq_length: int,
+        attention_backend: str = "math",
         seed: Optional[int] = None,
         **sampling_kwargs,
     ) -> torch.Tensor:
@@ -799,6 +802,7 @@ def generate(
                     if self.is_llama3_model
                     else None
                 ),
+                attention_backend=attention_backend,
                 **sampling_kwargs,
             ):
                 generated_tokens.append(generated_token.view(-1))
@@ -1186,6 +1190,7 @@ def callback(x, *, done_generating=False):
                     start_pos=start_pos,
                     skip_cache_setup=not is_first_sample,
                     max_seq_length=max_seq_length,
+                    attention_backend=self.builder_args.attention_backend,
                 )
                 for token_tensor, metrics in generator_func:
                     if token_tensor is not None: