Fix multimer kernel bug (#103)

* fix multimer bug

* fix bug

* remove print

* fix index

* fix layernorm length

* update test

* improve template code

* remove print

* polish layernorm

fastfold/model/fastnn/embedders_multimer.py

@@ -300,6 +300,7 @@ class TemplateEmbedderMultimer(nn.Module):
     ):
         template_embeds = []
         n_templ = batch["template_aatype"].shape[templ_dim]
+        template_pair_embeddings = torch.zeros((z.shape[0], z.shape[1], 64), dtype=z.dtype, device=z.device)
         for i in range(n_templ):
             idx = batch["template_aatype"].new_tensor(i)
             single_template_feats = tensor_tree_map(
@@ -336,7 +337,7 @@ class TemplateEmbedderMultimer(nn.Module):
             rigid_vec = rigid[..., None].inverse().apply_to_point(points)
             unit_vector = rigid_vec.normalized()
 
-            pair_act = self.template_pair_embedder(
+            pair_embedding = self.template_pair_embedder(
                 template_dgram,
                 aatype_one_hot,
                 z,
@@ -345,8 +346,24 @@ class TemplateEmbedderMultimer(nn.Module):
                 multichain_mask_2d,
                 unit_vector,
             )
-
-            single_template_embeds["template_pair_embedding"] = pair_act
+            
+            if not inplace:
+                # [*, S_t, N, N, C_z]
+                template_pair_embeddings = template_pair_embeddings + self.template_pair_stack(
+                    pair_embedding, 
+                    padding_mask_2d.unsqueeze(-3).to(dtype=z.dtype), 
+                    chunk_size=chunk_size,
+                    _mask_trans=False,
+                ).squeeze(0)
+            else:
+                # [*, S_t, N, N, C_z]
+                template_pair_embeddings += self.template_pair_stack.inplace(
+                    [pair_embedding], 
+                    padding_mask_2d.unsqueeze(-3).to(dtype=z.dtype), 
+                    chunk_size=chunk_size,
+                    _mask_trans=False,
+                )[0].squeeze(0)
+            
             single_template_embeds.update(
                 self.template_single_embedder(
                     single_template_feats,
@@ -361,27 +378,10 @@ class TemplateEmbedderMultimer(nn.Module):
             template_embeds,
         )
 
-        if not inplace:
-            # [*, S_t, N, N, C_z]
-            template_embeds["template_pair_embedding"] = self.template_pair_stack(
-                template_embeds["template_pair_embedding"], 
-                padding_mask_2d.unsqueeze(-3).to(dtype=z.dtype), 
-                chunk_size=chunk_size,
-                _mask_trans=False,
-            )
-        else:
-            template_embeds["template_pair_embedding"] = [template_embeds["template_pair_embedding"]]
-            # [*, S_t, N, N, C_z]
-            template_embeds["template_pair_embedding"] = self.template_pair_stack.inplace(
-                template_embeds["template_pair_embedding"], 
-                padding_mask_2d.unsqueeze(-3).to(dtype=z.dtype), 
-                chunk_size=chunk_size,
-                _mask_trans=False,
-            )[0].to(z.device)
-
         # [*, N, N, C_z]
-        template_embeds["template_pair_embedding"] = torch.sum(template_embeds["template_pair_embedding"], dim=-4) / n_templ
-        template_embeds["template_pair_embedding"] = torch.nn.functional.relu(template_embeds["template_pair_embedding"])
-        template_embeds["template_pair_embedding"] = self.linear_t(template_embeds["template_pair_embedding"])
-
+        template_pair_embeddings = template_pair_embeddings / n_templ
+        template_pair_embeddings = torch.nn.functional.relu(template_pair_embeddings)
+        template_pair_embeddings = self.linear_t(template_pair_embeddings)
+        
+        template_embeds["template_pair_embedding"] = template_pair_embeddings
         return template_embeds

fastfold/model/fastnn/kernel/layer_norm.py

@@ -34,6 +34,24 @@ class FusedLayerNorm(torch.nn.Module):
         torch.nn.init.zeros_(self.bias)
 
     def forward(self, input):
+        if len(input.shape) >= 3 and input.shape[-3] > 4000:
+            out = torch.empty_like(input)
+            # set max chunk_size = dim / 2, to max compute efficiency
+            chunk_size = min(4000 * 4000 // input.shape[-3], (input.shape[-3] + 1) // 2)
+            if len(input.shape) == 3:
+                for i in range(input.shape[-3]):
+                    out[i:i + chunk_size] = self.kernel_forward(input[i:i + chunk_size])
+            elif len(input.shape) == 4:
+                for j in range(input.shape[-4]):               
+                    for i in range(0, input.shape[-3], chunk_size):
+                        out[j, i:i + chunk_size] = self.kernel_forward(input[j, i:i + chunk_size])
+            else:
+                raise RuntimeError("Shape" + input.shape + "not implemented for layernorm yet!")
+            return out
+        else:
+            return self.kernel_forward(input)
+
+    def kernel_forward(self, input):
         if _triton_available:
             return LayerNormTritonFunc.apply(input, self.normalized_shape, self.weight, self.bias,
                                              self.eps)

fastfold/model/fastnn/template.py

@@ -240,26 +240,18 @@ class TemplatePairBlock(nn.Module):
             z = scatter(z, dim=1)
 
         mask = torch.nn.functional.pad(mask, (0, padding_size, 0, padding_size))
-
-        # single_templates = [t.unsqueeze(-4) for t in torch.unbind(z, dim=-4)]
-        # single_templates_masks = [m.unsqueeze(-3) for m in torch.unbind(mask, dim=-3)]
-        for i in range(z.shape[0]):
-            single = z[i].unsqueeze(-4)
-            single_mask = mask[i].unsqueeze(-3)
-
-            single_mask_row = scatter(single_mask, dim=1)
-            single_mask_col = scatter(single_mask, dim=2)
-
-            single = self.TriangleAttentionStartingNode(single, single_mask_row)
-            single = row_to_col(single)
-            single = self.TriangleAttentionEndingNode(single, single_mask_col)
-            single = col_to_row(single)
-            single = self.TriangleMultiplicationOutgoing(single, single_mask_row)
-            single = row_to_col(single)
-            single = self.TriangleMultiplicationIncoming(single, single_mask_col)
-            single = self.PairTransition(single)
-            single = col_to_row(single)
-            z[i] = single
+        single_mask_row = scatter(mask, dim=1)
+        single_mask_col = scatter(mask, dim=2)
+
+        z = self.TriangleAttentionStartingNode(z, single_mask_row)
+        z = row_to_col(z)
+        z = self.TriangleAttentionEndingNode(z, single_mask_col)
+        z = col_to_row(z)
+        z = self.TriangleMultiplicationOutgoing(z, single_mask_row)
+        z = row_to_col(z)
+        z = self.TriangleMultiplicationIncoming(z, single_mask_col)
+        z = self.PairTransition(z)
+        z = col_to_row(z)
 
         # z = torch.cat(single_templates, dim=-4)
         if self.last_block:
@@ -275,10 +267,7 @@ class TemplatePairBlock(nn.Module):
         chunk_size: Optional[int] = None,
         _mask_trans: bool = True,
     ):
-        if isinstance(chunk_size, int) and 1 <= chunk_size <= 4:
-            z[0] = z[0].cpu()
         dap_size = gpc.get_world_size(ParallelMode.TENSOR)
-
         seq_length = mask.size(-1)
         padding_size = (int(seq_length / dap_size) + 1) * dap_size - seq_length
 
@@ -287,31 +276,21 @@ class TemplatePairBlock(nn.Module):
             z[0] = scatter(z[0], dim=1)
 
         mask = torch.nn.functional.pad(mask, (0, padding_size, 0, padding_size))
-
-        # single_templates = [t.unsqueeze(-4) for t in torch.unbind(z, dim=-4)]
-        # single_templates_masks = [m.unsqueeze(-3) for m in torch.unbind(mask, dim=-3)]
-        for i in range(z[0].shape[0]):
-            single = z[0][i].unsqueeze(-4).to(mask.device)
-            single_mask = mask[i].unsqueeze(-3)
-
-            single_mask_row = scatter(single_mask, dim=1)
-            single_mask_col = scatter(single_mask, dim=2)
-
-            single = self.TriangleAttentionStartingNode(single, single_mask_row)
-            single = row_to_col(single)
-            single = self.TriangleAttentionEndingNode(single, single_mask_col)
-            single = col_to_row(single)
-            single = self.TriangleMultiplicationOutgoing(single, single_mask_row)
-            single = row_to_col(single)
-            single = self.TriangleMultiplicationIncoming(single, single_mask_col)
-            single = self.PairTransition(single)
-            single = col_to_row(single)
-            z[0][i] = single.to(z[0].device)
+        single_mask_row = scatter(mask, dim=1)
+        single_mask_col = scatter(mask, dim=2)
+
+        z = self.TriangleAttentionStartingNode.inplace(z, single_mask_row)
+        z[0] = row_to_col(z[0])
+        z = self.TriangleAttentionEndingNode.inplace(z, single_mask_col)
+        z[0] = col_to_row(z[0])
+        z[0] = self.TriangleMultiplicationOutgoing(z[0], single_mask_row)
+        z[0] = row_to_col(z[0])
+        z[0] = self.TriangleMultiplicationIncoming(z[0], single_mask_col)
+        z = self.PairTransition.inplace(z)
+        z[0] = col_to_row(z[0])
 
         # z = torch.cat(single_templates, dim=-4)
         if self.last_block:
-            if isinstance(chunk_size, int) and 1 <= chunk_size <= 4:
-                z[0] = z[0].to(mask.device)
             z[0] = gather(z[0], dim=1)
             z[0] = z[0][:, :-padding_size, :-padding_size, :]
 
@@ -408,15 +387,9 @@ class TemplatePairStack(nn.Module):
             args=(t,),
             blocks_per_ckpt=self.blocks_per_ckpt if self.training else None,
         )
-        if chunk_size is None:
-            chunk_size = t.shape[0]
-        for i in range(0, t.shape[0], chunk_size):
-            if t.shape[1] > 4000:
-                chunk_new = int(4000 * 4000 / t.shape[1])
-                for j in range(0, t.shape[1], chunk_new):
-                    t[i:i + chunk_size, j:j + chunk_new] = self.layer_norm(t[i:i + chunk_size, j:j + chunk_new])
-            else:
-                t[i:i + chunk_size] = self.layer_norm(t[i:i + chunk_size])
+
+        for i in range(0, t.shape[0]):
+            t[i] = self.layer_norm(t[i])
         return t
     
     def inplace(
@@ -453,13 +426,7 @@ class TemplatePairStack(nn.Module):
             args=(t,),
             blocks_per_ckpt=self.blocks_per_ckpt if self.training else None,
         )
-        if chunk_size is None:
-            chunk_size = t[0].shape[0]
-        for i in range(0, t[0].shape[0], chunk_size):
-            if t[0].shape[1] > 4000:
-                chunk_new = int(4000 * 4000 / t[0].shape[1])
-                for j in range(0, t[0].shape[1], chunk_new):
-                    t[0][i:i + chunk_size, j:j + chunk_new] = self.layer_norm(t[0][i:i + chunk_size, j:j + chunk_new].to(mask.device)).to(t[0].device)
-            else:
-                t[0][i:i + chunk_size] = self.layer_norm(t[0][i:i + chunk_size].to(mask.device)).to(t[0].device)
+
+        for i in range(0, t[0].shape[0]):
+            t[0][i] = self.layer_norm(t[0][i].to(mask.device)).to(t[0].device)
         return t

tests/test_fastnn/test_msa_att_col.py

@@ -73,4 +73,4 @@ def _test_msa_att_col(rank, world_size, chunk_size, get_openfold_module_and_data
         m_fast = m_fast[:, :-padding_size, :]
 
     error = torch.max(torch.abs(m_out.cuda() - m_fast))
-    assert error < 5e-5, f"Test m failed at chunk size: {chunk_size}. The position dif is {error}"
+    assert error < 1e-4, f"Test m failed at chunk size: {chunk_size}. The position dif is {error}"

tests/test_fastnn/test_template_embedder.py

@@ -46,7 +46,7 @@ def get_openfold_module_and_data():
 
 
 @pytest.mark.parametrize('world_size', [1, 2])
-@pytest.mark.parametrize('chunk_size', [None, 32])
+@pytest.mark.parametrize('chunk_size', [None, 4]) # should set 4 to test offload
 @pytest.mark.parametrize('inplace', [False, True])
 def test_state_dict(world_size, chunk_size, inplace, get_openfold_module_and_data): 
     run_func = partial(_test_template_embedder, world_size=world_size, chunk_size=chunk_size,