Fix a bug in sparse optimizer. (#3779)

* fix a bug.

* Update pylintrc

* fix.
Co-authored-by: Quan (Andy) Gan <coin2028@hotmail.com>

python/dgl/distributed/optim/pytorch/sparse_optim.py

@@ -48,11 +48,16 @@ class DistSparseGradOptimizer(abc.ABC):
             for emb in self._params:
                 name = emb._tensor.name
                 kvstore = emb._tensor.kvstore
-                trace = emb._trace
                 trainers_per_server = self._world_size // kvstore.num_servers
 
-                idics = [t[0] for t in trace]
-                grads = [t[1].grad.data for t in trace]
+                idics = []
+                grads = []
+                for trace in emb._trace:
+                    if trace[1].grad is not None:
+                        idics.append(trace[0])
+                        grads.append(trace[1].grad.data)
+                    else:
+                        assert len(trace[0]) == 0
                 # If the sparse embedding is not used in the previous forward step
                 # The idx and grad will be empty, initialize them as empty tensors to
                 # avoid crashing the optimizer step logic.

tests/lint/pylintrc

@@ -200,6 +200,9 @@ function-naming-style=snake_case
 # f - files
 # i, j, k - loop variables
 # u, v, e - nodes and edges
+# s, d - source and destination
+# t - time
+# r - relation type
 # n, m - general integers representing quantity
 # w, x, y, z - general math variables
 # g, G - graphs
@@ -210,7 +213,7 @@ function-naming-style=snake_case
 # op - operators
 # ty - type
 # A, B, C, W - for tensor operators like matmul
-good-names=f,i,j,k,u,v,e,n,m,w,x,y,z,g,G,hg,sg,fn,ex,Run,_,us,vs,gs,es,op,ty,A,B,C,W,a,b,N,D1,D2,R
+good-names=f,i,j,k,u,v,e,n,m,w,x,y,z,s,d,t,r,g,G,hg,sg,fn,ex,Run,_,us,vs,gs,es,op,ty,A,B,C,W,a,b,N,D1,D2,R
 
 # Include a hint for the correct naming format with invalid-name.
 include-naming-hint=no