add sort-by-length

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

tritonbench/operators/ragged_attention/hstu.py

@@ -48,6 +48,7 @@ def __init__(
         num_buckets,
         sparsity,
         target_size,
+        sort_by_length,
         requires_grad,
         persistent_kernel: bool = False,
     ) -> None:
@@ -58,6 +59,7 @@ def __init__(
         self.num_buckets = num_buckets
         self.sparsity = sparsity
         self.target_size = target_size
+        self.sort_by_length = sort_by_length
         self.all_ts_weights = torch.nn.Parameter(
             torch.randn(
                 (self.num_buckets + 1,),
@@ -175,7 +177,7 @@ def forward(
                 kwargs["ATTN_BIAS_TYPE"],  # relative_bias_type
                 kwargs["MAX_ATTN_LEN"],  # max_attn_len
                 kwargs["CONTEXTUAL_SEQ_LEN"],  # contextual_seq_len
-                kwargs["sort_by_length_indices"],  # sort_by_length
+                self.sort_by_length,
             )
 
         return out
@@ -213,7 +215,7 @@ def generate_sparse_seq_len(
 
 
 def get_test_inputs(
-    batch_size, num_heads, max_seq_len, sparsity, target_size, requires_grad
+    batch_size, num_heads, max_seq_len, sparsity, target_size, sort_by_length, requires_grad
 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
     timestamp_deltas: torch.Tensor = torch.randint(
         86400,

tritonbench/operators/ragged_attention/operator.py

@@ -24,6 +24,7 @@ def parse_op_args(args: List[str]):
     parser.add_argument("--num-buckets", type=int, default=2048)
     parser.add_argument("--sparsity", type=float, default=0.8)
     parser.add_argument("--target-size", type=int, default=20)
+    parser.add_argument("--sort-by-length", type=bool, default=False)
     return parser.parse_args(args)
 
 
@@ -41,6 +42,7 @@ def __init__(
         self.num_buckets = args.num_buckets
         self.sparsity = args.sparsity
         self.target_size = args.target_size
+        self.sort_by_length = args.sort_by_length
         # set a default number of inputs
         self._num_inputs = 10 if self._num_inputs is None else self._num_inputs
         self.requires_grad = not (self.mode == Mode.FWD_NO_GRAD)
@@ -54,6 +56,7 @@ def hstu_triton_ragged_attention(self, qkv, seq_offsets, timestamps, num_targets
             self.num_buckets,
             self.sparsity,
             self.target_size,
+            self.sort_by_length,
             self.requires_grad,
             persistent_kernel=False,
         )
@@ -69,18 +72,19 @@ def hstu_triton_ragged_attention_persistent(self, qkv, seq_offsets, timestamps,
             self.num_buckets,
             self.sparsity,
             self.target_size,
+            self.sort_by_length,
             self.requires_grad,
             persistent_kernel=True,
         )
         return lambda: attn(qkv, seq_offsets, timestamps, num_targets)
 
     def get_x_val(self, example_inputs):
-        return (self.batch_size, self.num_heads, self.max_seq_len, self.num_buckets, self.sparsity, self.target_size)
+        return (self.batch_size, self.num_heads, self.max_seq_len, self.num_buckets, self.sparsity, self.target_size, self.sort_by_length)
 
     def get_input_iter(self):
         for _input_id in range(self._num_inputs):
             inputs = get_test_inputs(
-                self.batch_size, self.num_heads, self.max_seq_len, self.sparsity, self.target_size, self.requires_grad
+                self.batch_size, self.num_heads, self.max_seq_len, self.sparsity, self.target_size, self.sort_by_length, self.requires_grad
             )
             yield inputs