[Question]: Usage of Graph Algorithms and Any Slowdowns?

[nv-rliu]

Hi there,

I'm interested in understanding if nncf depends on any graph algorithms from its usage of NetworkX? If so,

• What algorithms are used for what purpose?
• What graph sizes are they being used with?
• Have users experienced any slowdowns or issues with algorithms provided by NetworkX? (Speed, algorithm availability, etc)

Furthermore, would users be interested in accelerated nx algorithms via a GPU backend? This would involve zero code change.

Any insight into this topic would be greatly appreciated! Thank you.

[alexsu52]

Hello @nv-rliu,

Thank you for reaching us.

NNCF implements pattern matching:

nncf/nncf/common/graph/graph_matching.py

Line 129 in be26567

def find_subgraphs_matching_pattern(

includes: generate weakly connected components, graph isomorphism testing. The graph size is thousands nodes not tens of thousands nodes. NNCF is interested in accelerating this part and we are open to contributions. Do you have any estimates of what acceleration can be achieved for such graph sizes using GPU?

Some details: Taking into account the quantization pipeline, the time spent working with the graph is significantly less than the rest of the pipeline, which includes collecting statistics and manipulating weights.

[nv-rliu]

Hi @alexsu52

Hmm, while those graphs are certainly possible to be accelerated via GPU, I believe that for the <tens-of-thousands of nodes sizes, it would be difficult to foresee any meaningful speedups. GPU acceleration tends to be beneficial after >hundreds-of-thousands of nodes.

From the latter part of your answer, it may appear that the graph-work is not a bottleneck in your workflow. I appreciate you taking the time to inform me about your project. Please feel free to reach out in the future if there are any collaboration opportunities with nx-cugraph!