Enable SDPA without kv cache (#8950)

Summary:

Sdpa custom op has been decoupled from kv cache by kimishpatel. Update the llama definition so that the sdpa op is applied both with and without kv cache.

Reviewed By: kimishpatel

Differential Revision: D70593177

Author: Martin Yuan

examples/models/llama/attention.py

@@ -133,7 +133,7 @@ def __init__(
 
     def forward(
         self,
-        input_pos: torch.Tensor,
+        input_pos: Optional[torch.Tensor],
         q: torch.Tensor,  # Already have rotary embeddings. (bs, n_local_heads, seqlen, head_dim)
         k: torch.Tensor,  # Already have rotary embeddings. (bs, n_local_kv_heads, seqlen, head_dim)
         v: torch.Tensor,  # (bs, n_local_kv_heads, seqlen, head_dim)
@@ -218,13 +218,17 @@ def __init__(self, args: ModelArgs, layer_id: int, rope: Rope):
                 self.head_dim,
                 args.enable_dynamic_shape,
             )
-            self.SDPA = SDPA(
-                dim=self.n_local_heads * self.head_dim,
-                head_dim=self.head_dim,
-                n_rep=self.n_rep,
-                max_context_len=self.max_context_len,
-                enable_dynamic_shape=args.enable_dynamic_shape,
-            )
+        else:
+            # Use a constant state to avoid export error
+            self.zero_pos = torch.tensor([0])
+
+        self.SDPA = SDPA(
+            dim=self.n_local_heads * self.head_dim,
+            head_dim=self.head_dim,
+            n_rep=self.n_rep,
+            max_context_len=self.max_context_len,
+            enable_dynamic_shape=args.enable_dynamic_shape,
+        )
 
     def forward(
         self,
@@ -258,20 +262,8 @@ def forward(
             assert input_pos is not None
             k, v = self.kv_cache.update(input_pos, k, v)
             output = self.SDPA(input_pos, q, k, v, bsz, seqlen, self.mask)
-            return self.wo(output), None
-
-        # grouped multiquery attention: expand out keys and values
-        k = k.repeat_interleave(self.n_rep, dim=1)
-        v = v.repeat_interleave(self.n_rep, dim=1)
-
-        assert hasattr(self, "mask")
-
-        mask = self.mask[:seqlen, :seqlen]
-
-        output = F.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0)
-
-        output = output.transpose(1, 2).contiguous().view(bsz, seqlen, -1)
-
-        output = self.wo(output)
-
-        return output, None
+        else:
+            mask = self.mask[:seqlen, :seqlen]
+            # No kv cache. Pass 0 input_pos
+            output = self.SDPA(self.zero_pos, q, k, v, bsz, seqlen, mask)
+        return self.wo(output), None

examples/models/llama/tests/test_attention_sma.py

@@ -0,0 +1,88 @@
+import unittest
+
+import torch
+from executorch.examples.models.llama.attention import (
+    AttentionMHA,
+    KVCache,
+    ModelArgs,
+    Rope,
+    SDPA,
+)
+
+
+class TestAttentionMHA(unittest.TestCase):
+
+    def create_mock_args(self):
+        return ModelArgs(
+            use_kv_cache=True,
+            n_heads=8,
+            n_kv_heads=4,
+            head_dim=64,
+            max_batch_size=2,
+            max_context_len=16,
+            dim=512,
+            attention_qkv_bias=False,
+            enable_dynamic_shape=False,
+        )
+
+    def test_attentionmha_init(self):
+        args = self.create_mock_args()
+        rope = Rope(args)
+        attn = AttentionMHA(args, layer_id=0, rope=rope)
+
+        self.assertEqual(attn.n_heads, 8)
+        self.assertEqual(attn.n_kv_heads, 4)
+        self.assertEqual(attn.n_local_heads, 8)
+        self.assertEqual(attn.n_local_kv_heads, 4)
+        self.assertEqual(attn.head_dim, 64)
+        self.assertEqual(attn.dim, 512)
+        self.assertEqual(attn.mask.shape, (16, 16))  # Causal mask shape check
+        self.assertTrue(attn.use_kv_cache)
+
+        if attn.use_kv_cache:
+            self.assertIsInstance(attn.kv_cache, KVCache)
+        self.assertIsInstance(attn.SDPA, SDPA)
+
+    def test_attentionmha_forward(self):
+        args = self.create_mock_args()
+        rope = Rope(args)
+        attn = AttentionMHA(args, layer_id=0, rope=rope)
+
+        bsz, seqlen, dim = 2, 4, args.dim
+        x = torch.randn(bsz, seqlen, dim)
+        freqs_cos = torch.randn(seqlen, args.head_dim // 2)
+        freqs_sin = torch.randn(seqlen, args.head_dim // 2)
+        input_pos = torch.tensor([0, 1, 2, 3])
+
+        output, _ = attn.forward(x, freqs_cos, freqs_sin, input_pos=input_pos)
+
+        self.assertEqual(output.shape, (bsz, seqlen, dim))
+
+    def test_attentionmha_forward_no_kv_cache(self):
+        args = self.create_mock_args()
+        args.use_kv_cache = False  # Disable KV cache for this test
+        rope = Rope(args)
+        attn = AttentionMHA(args, layer_id=0, rope=rope)
+
+        bsz, seqlen, dim = 2, 4, args.dim
+        x = torch.randn(bsz, seqlen, dim)
+        freqs_cos = torch.randn(seqlen, args.head_dim // 2)
+        freqs_sin = torch.randn(seqlen, args.head_dim // 2)
+
+        output, _ = attn.forward(x, freqs_cos, freqs_sin)
+
+        self.assertEqual(output.shape, (bsz, seqlen, dim))
+
+    def test_attentionmha_invalid_kv_cache(self):
+        args = self.create_mock_args()
+        rope = Rope(args)
+        attn = AttentionMHA(args, layer_id=0, rope=rope)
+
+        bsz, seqlen, dim = 2, 4, args.dim
+        x = torch.randn(bsz, seqlen, dim)
+        freqs_cos = torch.randn(seqlen, args.head_dim // 2)
+        freqs_sin = torch.randn(seqlen, args.head_dim // 2)
+
+        # No input_pos provided, should raise assertion error
+        with self.assertRaises(AssertionError):
+            attn.forward(x, freqs_cos, freqs_sin)