Linear to Conv2d transform for static attention (pytorch#9025)

Summary:
Pull Request resolved: pytorch#9025

This is needed by some backends such as HTP.

Differential Revision: D70726317

Author: sxu

examples/models/llama/static_attention.py

@@ -209,6 +209,7 @@ def __init__(self, config: ModelArgs, layer_id: int, rope: Rope):
         self.head_dim = config.head_dim
         self.inv_scale = 1.0 / (float(self.head_dim) ** 0.5)
         self.attention_qkv_bias = config.attention_qkv_bias
+        self.use_conv2d = False
 
         self.wqs = nn.ModuleList(
             [
@@ -253,9 +254,25 @@ def forward(
         in_cache_state = kwargs.get("in_cache_state")
         out_cache_state = kwargs.get("out_cache_state")
 
+        bsz, seq_len, dim = x.shape
+        if self.use_conv2d:
+            x = x.reshape(bsz, seq_len, 1, dim).transpose(1, 3)
+
         new_qs = [self.wqs[i](x) for i in range(self.n_heads)]
         new_ks = [self.wks[i](x) for i in range(self.n_kv_heads)]
         new_vs = [self.wvs[i](x) for i in range(self.n_kv_heads)]
+
+        if self.use_conv2d:
+
+            def from_conv2ds(ts):
+                return [
+                    t.reshape(bsz, self.head_dim, seq_len).transpose(1, 2) for t in ts
+                ]
+
+            new_qs = from_conv2ds(new_qs)
+            new_ks = from_conv2ds(new_ks)
+            new_vs = from_conv2ds(new_vs)
+
         new_qs = [self.rope(q, freqs_cos, freqs_sin) for q in new_qs]
         new_ks = [self.rope(k, freqs_cos, freqs_sin) for k in new_ks]
 
@@ -281,7 +298,14 @@ def forward(
             heads.append(attn @ all_vs[kv_idx])
 
         y = torch.cat(heads, dim=-1)
-        y = self.wo(y)
+        if self.use_conv2d:
+            y = (
+                self.wo(y.reshape(bsz, seq_len, 1, -1).transpose(1, 3))
+                .transpose(1, 3)
+                .reshape(bsz, seq_len, -1)
+            )
+        else:
+            y = self.wo(y)
         return y, {"out_cache_state": out_cache_state}
 
     def load_weights_from_attention_mha(self, other: AttentionMHA):
@@ -299,3 +323,44 @@ def load_weights_from_attention_mha(self, other: AttentionMHA):
             )
 
         self.wo.weight.data.copy_(other.wo.weight)
+
+    def linear_to_conv2d(self):
+        def transfer_weight(linear, conv2d):
+            conv2d.weight.data.copy_(linear.weight[:, :, None, None])
+            return conv2d
+
+        self.wqs = nn.ModuleList(
+            [
+                transfer_weight(
+                    linear,
+                    nn.Conv2d(self.dim, self.head_dim, 1, bias=self.attention_qkv_bias),
+                )
+                for linear in self.wqs
+            ]
+        )
+        self.wks = nn.ModuleList(
+            [
+                transfer_weight(
+                    linear,
+                    nn.Conv2d(self.dim, self.head_dim, 1, bias=self.attention_qkv_bias),
+                )
+                for linear in self.wks
+            ]
+        )
+        self.wvs = nn.ModuleList(
+            [
+                transfer_weight(
+                    linear,
+                    nn.Conv2d(self.dim, self.head_dim, 1, bias=self.attention_qkv_bias),
+                )
+                for linear in self.wvs
+            ]
+        )
+        self.wo = transfer_weight(
+            self.wo,
+            nn.Conv2d(
+                self.n_heads * self.head_dim, self.dim, 1, bias=self.attention_qkv_bias
+            ),
+        )
+
+        self.use_conv2d = True

examples/models/llama/tests/test_static_attention.py

@@ -17,32 +17,38 @@ def setUp(self):
         torch.manual_seed(42)
 
     def test_without_cache(self):
-        config = ModelArgs(
-            dim=64,
-            n_heads=4,
-            n_kv_heads=2,
-            max_seq_len=8,
-        )
-        layer_id = 0
-        rope = Rope(config)
-        attn_mha = AttentionMHA(config, layer_id, rope).eval()
-        static_attn = StaticAttention(config, layer_id, rope).eval()
-        static_attn.load_weights_from_attention_mha(attn_mha)
+        def test(use_conv2d):
+            config = ModelArgs(
+                dim=64,
+                n_heads=4,
+                n_kv_heads=2,
+                max_seq_len=8,
+            )
+            layer_id = 0
+            rope = Rope(config)
+            attn_mha = AttentionMHA(config, layer_id, rope).eval()
+            static_attn = StaticAttention(config, layer_id, rope).eval()
+            static_attn.load_weights_from_attention_mha(attn_mha)
+            if use_conv2d:
+                static_attn.linear_to_conv2d()
+
+            x = torch.rand(1, config.max_seq_len, config.dim)
+            freqs_cos, freqs_sin = rope.get_freqs(None, config.max_seq_len)
+            expected, _ = attn_mha(x, freqs_cos, freqs_sin)
+            mask = torch.triu(
+                torch.full((1, config.max_seq_len, config.max_seq_len), float("-inf")),
+                diagonal=1,
+            )
+            y, _ = static_attn(
+                x,
+                freqs_cos,
+                freqs_sin,
+                mask=mask,
+            )
+            self.assertTrue(torch.isclose(y, expected, rtol=1e-3).all())
 
-        x = torch.rand(1, config.max_seq_len, config.dim)
-        freqs_cos, freqs_sin = rope.get_freqs(None, config.max_seq_len)
-        expected, _ = attn_mha(x, freqs_cos, freqs_sin)
-        mask = torch.triu(
-            torch.full((1, config.max_seq_len, config.max_seq_len), float("-inf")),
-            diagonal=1,
-        )
-        y, _ = static_attn(
-            x,
-            freqs_cos,
-            freqs_sin,
-            mask=mask,
-        )
-        self.assertTrue(torch.isclose(y, expected, rtol=1e-3).all())
+        test(True)
+        test(False)
 
     def test_hf_rope_without_cache(self):
         config = ModelArgs(