add a test for caching

meshgpt_pytorch/meshgpt_pytorch.py

@@ -1157,6 +1157,7 @@ def __init__(
         self.conditioner = None
 
         cross_attn_dim_context = None
+        dim_text = None
 
         if condition_on_text:
             self.conditioner = TextEmbeddingReturner(

meshgpt_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.5.1'
+__version__ = '1.5.2'

tests/test_meshgpt.py

@@ -16,7 +16,7 @@ def test_readme(adaptive_rmsnorm):
     # mock inputs
 
     vertices = torch.randn((2, 121, 3))            # (batch, num vertices, coor (3))
-    faces = torch.randint(0, 121, (2, 64, 3))      # (batch, num faces, vertices (3))
+    faces = torch.randint(0, 121, (2, 2, 3))      # (batch, num faces, vertices (3))
 
     # forward in the faces
 
@@ -33,7 +33,7 @@ def test_readme(adaptive_rmsnorm):
     transformer = MeshTransformer(
         autoencoder,
         dim = 512,
-        max_seq_len = 768,
+        max_seq_len = 60,
         num_sos_tokens = 1,
         fine_cross_attend_text = True,
         text_cond_with_film = False,
@@ -51,3 +51,21 @@ def test_readme(adaptive_rmsnorm):
     loss.backward()
 
     faces_coordinates, face_mask = transformer.generate(texts = ['a small chair'], cond_scale = 3.)
+
+def test_cache():
+    # test that the output for generation with and without kv (and optional gateloop) cache is equivalent
+
+    autoencoder = MeshAutoencoder(
+        num_discrete_coors = 128
+    )
+
+    transformer = MeshTransformer(
+        autoencoder,
+        dim = 512,
+        max_seq_len = 12
+    )
+
+    uncached_faces_coors, _ = transformer.generate(cache_kv = False, temperature = 0)
+    cached_faces_coors, _ = transformer.generate(cache_kv = True, temperature = 0)
+
+    assert torch.allclose(uncached_faces_coors, cached_faces_coors)