Uh oh!
There was an error while loading. Please reload this page.
[Collage] SubGraphs - #11981
Conversation
mbs-octoml
commented
Jun 30, 2022
@mbaret, thanks. |
cfc5aff to
300dc7dComparembs-octoml
commented
Jul 1, 2022
green and ready for review |
d0d4237 to
d71779dCompareSee https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md. (Special thanks to Matthew Barrett for authoring partition_rule_test.cc and suggesting a PR partitioning strategy.) Collage uses a small 'combinator library' of PartitionRule to decribe how candidate partitions can be extracted from a model for measurement and comparison. This introduces most of that machinery, however we defer the all important 'CombinerPartitionRule' for the next PR. Thus the rules at this stage can only express the sorts of DFPattern-based rules we find in most BYOC integrations, and cannot describe rules more traditionally associated with operator fusion. Based on apache#11981.
See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md. (Special thanks to Matthew Barrett for authoring partition_rule_test.cc and suggesting a PR partitioning strategy.) Collage uses a small 'combinator library' of PartitionRule to decribe how candidate partitions can be extracted from a model for measurement and comparison. This introduces most of that machinery, however we defer the all important 'CombinerPartitionRule' for the next PR. Thus the rules at this stage can only express the sorts of DFPattern-based rules we find in most BYOC integrations, and cannot describe rules more traditionally associated with operator fusion. Based on apache#11981.
d71779d to
70fc25aCompareUh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
SebastianBoblest
commented
Jul 6, 2022
I read this out of curiosity, I cannot review this on a technical level. |
mbaret
left a comment
There was a problem hiding this comment.
Largely looks good to me, couple of points around naming. It's tacitly acknowledged that more of this infrastructure will get exercised by tests in later PRs.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md. Collage works in units of 'sub-graphs', which are potential partitions of the overall Relay model. This PR introduces SubGraph (an arbitrary partitioning, without any implication about how it is to be represented), it's companion SubSubGraph (implying a representation as a function), and some supporting odds 'n ends.
- make 'Compiler' name explicit
70fc25a to
33e8c94Comparembs-octoml
commented
Jul 11, 2022
Thanks @SebastianBoblest, every little bit helps. Thanks and PTAL @mbaret. |
See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md. (Special thanks to Matthew Barrett for authoring partition_rule_test.cc and suggesting a PR partitioning strategy.) Collage uses a small 'combinator library' of PartitionRule to decribe how candidate partitions can be extracted from a model for measurement and comparison. This introduces most of that machinery, however we defer the all important 'CombinerPartitionRule' for the next PR. Thus the rules at this stage can only express the sorts of DFPattern-based rules we find in most BYOC integrations, and cannot describe rules more traditionally associated with operator fusion. Based on apache#11981.
* [Collage] PartitionRule (though without CombinePartitionRule) See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md. (Special thanks to Matthew Barrett for authoring partition_rule_test.cc and suggesting a PR partitioning strategy.) Collage uses a small 'combinator library' of PartitionRule to decribe how candidate partitions can be extracted from a model for measurement and comparison. This introduces most of that machinery, however we defer the all important 'CombinerPartitionRule' for the next PR. Thus the rules at this stage can only express the sorts of DFPattern-based rules we find in most BYOC integrations, and cannot describe rules more traditionally associated with operator fusion. Based on #11981. * - Backport improvements to partiton_rule_test.cc * - Oops
* [Collage] SubGraphs See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md. Collage works in units of 'sub-graphs', which are potential partitions of the overall Relay model. This PR introduces SubGraph (an arbitrary partitioning, without any implication about how it is to be represented), it's companion SubSubGraph (implying a representation as a function), and some supporting odds 'n ends. * - make Integer <-> size_t conversion explicit - make 'Compiler' name explicit * - fix namespace ambiguity * - review comments
…#11993) * [Collage] PartitionRule (though without CombinePartitionRule) See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md. (Special thanks to Matthew Barrett for authoring partition_rule_test.cc and suggesting a PR partitioning strategy.) Collage uses a small 'combinator library' of PartitionRule to decribe how candidate partitions can be extracted from a model for measurement and comparison. This introduces most of that machinery, however we defer the all important 'CombinerPartitionRule' for the next PR. Thus the rules at this stage can only express the sorts of DFPattern-based rules we find in most BYOC integrations, and cannot describe rules more traditionally associated with operator fusion. Based on apache#11981. * - Backport improvements to partiton_rule_test.cc * - Oops
* [Collage] SubGraphs See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md. Collage works in units of 'sub-graphs', which are potential partitions of the overall Relay model. This PR introduces SubGraph (an arbitrary partitioning, without any implication about how it is to be represented), it's companion SubSubGraph (implying a representation as a function), and some supporting odds 'n ends. * - make Integer <-> size_t conversion explicit - make 'Compiler' name explicit * - fix namespace ambiguity * - review comments
…#11993) * [Collage] PartitionRule (though without CombinePartitionRule) See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md. (Special thanks to Matthew Barrett for authoring partition_rule_test.cc and suggesting a PR partitioning strategy.) Collage uses a small 'combinator library' of PartitionRule to decribe how candidate partitions can be extracted from a model for measurement and comparison. This introduces most of that machinery, however we defer the all important 'CombinerPartitionRule' for the next PR. Thus the rules at this stage can only express the sorts of DFPattern-based rules we find in most BYOC integrations, and cannot describe rules more traditionally associated with operator fusion. Based on apache#11981. * - Backport improvements to partiton_rule_test.cc * - Oops
* [Collage] SubGraphs See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md. Collage works in units of 'sub-graphs', which are potential partitions of the overall Relay model. This PR introduces SubGraph (an arbitrary partitioning, without any implication about how it is to be represented), it's companion SubSubGraph (implying a representation as a function), and some supporting odds 'n ends. * - make Integer <-> size_t conversion explicit - make 'Compiler' name explicit * - fix namespace ambiguity * - review comments
…#11993) * [Collage] PartitionRule (though without CombinePartitionRule) See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md. (Special thanks to Matthew Barrett for authoring partition_rule_test.cc and suggesting a PR partitioning strategy.) Collage uses a small 'combinator library' of PartitionRule to decribe how candidate partitions can be extracted from a model for measurement and comparison. This introduces most of that machinery, however we defer the all important 'CombinerPartitionRule' for the next PR. Thus the rules at this stage can only express the sorts of DFPattern-based rules we find in most BYOC integrations, and cannot describe rules more traditionally associated with operator fusion. Based on apache#11981. * - Backport improvements to partiton_rule_test.cc * - Oops
* [Collage] SubGraphs See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md. Collage works in units of 'sub-graphs', which are potential partitions of the overall Relay model. This PR introduces SubGraph (an arbitrary partitioning, without any implication about how it is to be represented), it's companion SubSubGraph (implying a representation as a function), and some supporting odds 'n ends. * - make Integer <-> size_t conversion explicit - make 'Compiler' name explicit * - fix namespace ambiguity * - review comments
…#11993) * [Collage] PartitionRule (though without CombinePartitionRule) See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md. (Special thanks to Matthew Barrett for authoring partition_rule_test.cc and suggesting a PR partitioning strategy.) Collage uses a small 'combinator library' of PartitionRule to decribe how candidate partitions can be extracted from a model for measurement and comparison. This introduces most of that machinery, however we defer the all important 'CombinerPartitionRule' for the next PR. Thus the rules at this stage can only express the sorts of DFPattern-based rules we find in most BYOC integrations, and cannot describe rules more traditionally associated with operator fusion. Based on apache#11981. * - Backport improvements to partiton_rule_test.cc * - Oops
See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md.
Collage works in units of 'sub-graphs', which are potential partitions of the
overall Relay model. This PR introduces SubGraph (an arbitrary partitioning, without
any implication about how it is to be represented), it's companion SubSubGraph
(implying a representation as a function), and some supporting odds 'n ends.