notesum.ai
Collect (0)
Published at December 6
Share
Code generation and runtime techniques for enabling data-efficient deep learning training on GPUs
cs.DC
cs.NE
Released Date: December 6, 2024
Authors:
Kun Wu
Arxiv:
http://arxiv.org/pdf/2412.04747v1
Methods of graph variables
node iterator
g.dst_nodes()
,
g.src_nodes()
edge iterator
g.edges()
weight slicing, e.g.,
W[e.etype]
neighbor iterator
n.incoming_edges()
,
n.outgoing_edges()
Attributes
nodes
e.src
,
e.dst
types
e.etype
,
n.ntype
input data, e.g.,
n.feature
produced data, e.g.,
e["att"]
Operators
GEMM-eligible computation, e.g.,
linear()
,
outer_prod()
GEMM-ineligible computation, e.g.,
dot_prod()
manipulation, e.g.,
reshape()
,
concat()
