for debug update

siliconflow · doombeaker · Feb 2, 2024 · Feb 4, 2024 · Feb 4, 2024 · Feb 5, 2024
commit 4e7cd9b2aa51045f27e40fb5625cca9d39c1e32a
diff --git a/examples/mywarmupmodule.py b/examples/mywarmupmodule.py
@@ -11,63 +11,59 @@ def __init__(self):
         self.groupnorm = nn.GroupNorm(3, 6)
         self.layer_norm = nn.LayerNorm([24, 6, 24])
 
-    def forward(self, x):
-        # # Perform operations
-        # #x = torch.expand_dims(x, dim=1)
-        # #x = torch.broadcast_dim_like(x, self.weight)
-        # x = torch.arange(10)
-        # #x = torch.scalar_mul(x, 0.5)
-        # #x = torch.scalar_div(x, 2.0)
-        # x = torch.exp(x)
-        # #x = torch.broadcast_mul(x, self.weight)
-        # x = torch.sin(x)
-        # x = torch.cos(x)
-        # #x = torch.clip(x, -1, 0)
-        # #x = torch.cat([x, x], dim=1)
-        # #x = torch.slice(x, 1, 0, 5)
-        # #x = torch.cast(x, dtype=torch.float16)
-        # #x = torch.fused_matmul_bias(x, self.weight, self.bias)
-        # #x = torch.silu(x)
-        # x = torch.flatten(x)
-        # #x = torch.unflatten(x, sizes=[2, 3, 4])
-        # #x = torch.add_n([x, x])
-
+    def forward(self, x, a, b, w):
         x = self.conv2d(x)
         x = self.groupnorm(x)
-        #x = torch.broadcast_add(x, self.bias)
         x = torch.transpose(x, 1, 2)
         x = self.layer_norm(x)
-        #x = torch.broadcast_matmul(x, x)
         x = torch.reshape(x, (2, -1))
-        x = torch.empty(3, 3)
-
-        a = torch.randn(2, 4, 8)
-        b = torch.randn(2, 8, 3)
-        x = torch.bmm(a, b)
+        x = torch.sin(x)
+        x = torch.cos(x)
+        tup_list = [[None, None, None], [0, 4, 1],]
+        x = torch.slice(x, slice_tup_list=tup_list)
+        x = x.reshape(2, -1, 4)
+        x = x + a
+        x = torch.bmm(x, b)
         x = torch.softmax(x, dim=1)
-        x = torch.narrow(x, 1, 0, 2)
-        #x = torch.gelu(x)
-        #x = torch.reshape_like(x, self.weight)
-        x = torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')
-        #x = torch.grouped_matmul_bias(x, self.weight, self.bias)
-        #x = torch.timestep_embedding(x, 4)
-        #x = torch.fused_multi_head_attention_inference(x, self.weight, self.bias)
-        #x = torch.fused_glu(x)
-
+        x = x*1.0
+        x = x/1.0
+        print(f"### x:{x.shape}, a:{a.shape}, b:{b.shape}, w:{w.shape}")
+        x = torch.flatten(x, start_dim=1)
+        a = torch.flatten(x, start_dim=1)
+        print(f"### x:{x.shape}, a:{a.shape}, b:{b.shape}, w:{w.shape}")
+        x = torch.cat([x,a])
+        x = torch.clip(x, min=-1, max=1)
+        x = torch.exp(x)
+        x = torch.flatten(x, start_dim=0)
+        tup_list = [[0, 32, 1],]
+        x = torch.slice(x, slice_tup_list=tup_list)
+        x = torch._C.unflatten(x, 0, (4, 8))
+        a_1 = torch.empty(a.shape).cuda()
+        a = a + a_1
+        x = torch._C.reshape_like(x, a)
+        x = torch.cast(x, torch.float16)
+        print(f"### x:{x.shape} {x.dtype}, a:{a.shape}, b:{b.shape}, w:{w.shape}")
+        x = torch.narrow(x, 0, 0, 2)
         return x
 class SimpleGraph(nn.Graph):
     def __init__(self, model):
         super().__init__()
         self.model = model
 
-    def build(self, input):
-        return self.model(input)
-        
-if __name__ == "__main__":
+    def build(self, *input):
+        return self.model(*input)
+
+def global_warmup():
     m = SimpleModule()
     m.eval()
     m = m.cuda()
+    a = torch.randn(2, 4, 4).cuda()
+    b = torch.randn(2, 4, 4).cuda()
     x = torch.randn(1,3,28,28).cuda()
+    w = torch.randn(2,4).cuda()
     g = SimpleGraph(m)
-    y = g(x)
-    print(y.shape)
+    y = g(x,a,b,w)
+    print(y.shape)
+
+if __name__ == "__main__":
+    global_warmup()
diff --git a/judge_op_wanted.sh b/judge_op_wanted.sh
@@ -0,0 +1,6 @@
+set -e
+echo "build simple graph"
+python examples/mywarmupmodule.py
+
+echo "====================="
+python extract_op_type_name.py