Custom SDPA in attention

Author: jackzhxng

extension/llm/modules/attention.py

@@ -10,6 +10,7 @@
 import torch
 import torchtune.modules.attention as TorchTuneAttention
 from executorch.extension.llm.modules.kv_cache import KVCache as InferenceKVCache
+from executorch.extension.llm.custom_ops import custom_ops
 from torch import nn
 from torchtune.modules.attention_utils import _MaskType, _sdpa_or_flex_attention
 from torchtune.modules.kv_cache import KVCache
@@ -146,6 +147,7 @@ def __init__(
         # Use flex attention if supported and we are sample packing
         self._attention_call = _sdpa_or_flex_attention()
         self._sdpa = SDPA(
+            max_seq_len=self.max_seq_len,
             num_kv_heads=self.num_kv_heads,
             num_heads=self.num_heads,
             head_dim=self.head_dim,
@@ -310,7 +312,7 @@ def false_fn(y):
             self.kv_cache.v_cache.copy_(v)
             self.kv_cache.cache_pos.copy_(cache_pos)
 
-        output = self._sdpa(q, k, v, b, s_x, mask=mask)
+        output = self._sdpa(q, k, v, b, s_x, mask=mask, input_pos=input_pos)
         return self.output_proj(output)
 
 
@@ -322,6 +324,7 @@ class SDPA(nn.Module):
 
     def __init__(
         self,
+        max_seq_len: int,
         num_kv_heads: int,
         num_heads: int,
         head_dim: int,
@@ -331,6 +334,7 @@ def __init__(
         kv_cache,
     ) -> None:
         super().__init__()
+        self.max_seq_len = max_seq_len
         self.num_kv_heads = num_kv_heads
         self.num_heads = num_heads
         self.head_dim = head_dim
@@ -348,32 +352,48 @@ def forward(
         bsz: int,
         seq_len: int,
         mask: Optional[_MaskType] = None,
+        # Below args are only used for ET custom sdpa op.
+        input_pos: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
-        # View + expand + reshape bring num_kv_heads to num_heads for k and v
-        # to match q.
-
-        # [bsz, n_h, s, h_d]
-        q = q.transpose(1, 2)
-        k = k.transpose(1, 2)
-        v = v.transpose(1, 2)
-
-        # Expand the key and value tensors to have the same shape
-        # as the query tensor by copying values across the relevant dim
-        if self.num_heads != self.num_kv_heads:
-            expand_shape = (-1, -1, self.q_per_kv, -1, -1)
-            k = k.unsqueeze(2).expand(expand_shape).flatten(1, 2)
-            v = v.unsqueeze(2).expand(expand_shape).flatten(1, 2)
-
-        output = self._attention_fn(
+        start_pos = input_pos[0][-1].item() - seq_len + 1
+        torch._check_is_size(start_pos)
+        torch._check(start_pos <= self.max_seq_len)
+        output = torch.ops.llama.custom_sdpa(
             q,
             k,
             v,
-            mask=mask,
-            dropout_p=self.attn_dropout,
-            is_causal=self.kv_cache is None and mask is None and self.is_causal,
+            start_pos,
+            None,  # Attention mask
+            0,  # dropout probability. Ignored by the code
+            True,  # is_causal TODO: flip to false if kv cache is enabled???
         )
-        # Reshape the output to be the same shape as the input
-        return output.transpose(1, 2).contiguous().view(bsz, seq_len, -1)
+        return output.view(bsz, seq_len, -1)
+        
+        # # View + expand + reshape bring num_kv_heads to num_heads for k and v
+        # # to match q.
+
+        # # [bsz, n_h, s, h_d]
+        # q = q.transpose(1, 2)
+        # k = k.transpose(1, 2)
+        # v = v.transpose(1, 2)
+
+        # # Expand the key and value tensors to have the same shape
+        # # as the query tensor by copying values across the relevant dim
+        # if self.num_heads != self.num_kv_heads:
+        #     expand_shape = (-1, -1, self.q_per_kv, -1, -1)
+        #     k = k.unsqueeze(2).expand(expand_shape).flatten(1, 2)
+        #     v = v.unsqueeze(2).expand(expand_shape).flatten(1, 2)
+
+        # output = self._attention_fn(
+        #     q,
+        #     k,
+        #     v,
+        #     mask=mask,
+        #     dropout_p=self.attn_dropout,
+        #     is_causal=self.kv_cache is None and mask is None and self.is_causal,
+        # )
+        # # Reshape the output to be the same shape as the input
+        # return output.transpose(1, 2).contiguous().view(bsz, seq_len, -1)
 
 
 def _replace_mha_with_inference_mha(module: torch.nn.Module) -> None: