Add permute_duplicate_pooled_embeddings op #1912
Closed
+141
−11
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
✅ Deploy Preview for pytorch-fbgemm-docs canceled.
|
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 5, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Differential Revision: D48090591 fbshipit-source-id: a398bc495e3461abc429974dc6d2212e0178e59d
AGZain
force-pushed
the
export-D48090591
branch
from
8f33247 to
1a13a72
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 5, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Differential Revision: D48090591 fbshipit-source-id: 99a9d88a2f1ada30dd41c24b4df97745c5e108d8
AGZain
force-pushed
the
export-D48090591
branch
from
1a13a72 to
29835eb
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 7, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Differential Revision: D48090591 fbshipit-source-id: 9ca8cfa7b066c84d20e3618b073bcd66627c32ee
AGZain
force-pushed
the
export-D48090591
branch
from
29835eb to
ed253e2
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 7, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Differential Revision: D48090591 fbshipit-source-id: 03f3792cc2309e5dae5af4532887e29ce0e07cd4
AGZain
force-pushed
the
export-D48090591
branch
from
ed253e2 to
a1132db
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 7, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Differential Revision: D48090591 fbshipit-source-id: e66734ca9e6e105dbe0031025239ae585063c255
AGZain
force-pushed
the
export-D48090591
branch
from
a1132db to
a9dd610
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
1 similar comment
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 11, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Differential Revision: D48090591 fbshipit-source-id: 922e8c472c3ffb42f23c36dd38c67a3dffa8e119
AGZain
force-pushed
the
export-D48090591
branch
from
a9dd610 to
5fb9fd4
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 14, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Differential Revision: D48090591 fbshipit-source-id: 96edf038b8c991b0331e05a0ec5a84d9e35bcf91
AGZain
force-pushed
the
export-D48090591
branch
from
5fb9fd4 to
c592d35
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 14, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Differential Revision: D48090591 fbshipit-source-id: f8dd2cce92b768c2c1269d071d32fa795e6bf29e
AGZain
force-pushed
the
export-D48090591
branch
from
c592d35 to
1751abe
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 14, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Differential Revision: D48090591 fbshipit-source-id: 87fbb41662af0c8bb6cba0b6ecc8aa6f06b3d30c
AGZain
force-pushed
the
export-D48090591
branch
from
52be70d to
d077c44
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 18, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Differential Revision: D48090591 fbshipit-source-id: daa9b037786e44123afad875b05ab68792fbd2ab
AGZain
force-pushed
the
export-D48090591
branch
from
d077c44 to
7f8305b
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 18, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Reviewed By: sryap Differential Revision: D48090591 fbshipit-source-id: e649a33d6b5ff70733e37bd96e9419328c3c0624
AGZain
force-pushed
the
export-D48090591
branch
from
7f8305b to
276fefa
Compare
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 21, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Reviewed By: sryap Differential Revision: D48090591 fbshipit-source-id: e12c85ab1ce9066a8cc4eb7bd29759d5c4a5be6c
AGZain
force-pushed
the
export-D48090591
branch
from
276fefa to
34a7fba
Compare
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 21, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Reviewed By: sryap Differential Revision: D48090591 fbshipit-source-id: d484cd779edb37f9c3a4bb0af1725aa7960c03cd
AGZain
force-pushed
the
export-D48090591
branch
from
34a7fba to
a9ada05
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 22, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Reviewed By: sryap Differential Revision: D48090591 fbshipit-source-id: 083b519187afdb7ae9a75be68c7e5725b6968194
AGZain
force-pushed
the
export-D48090591
branch
from
a9ada05 to
6d2ce9e
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 22, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Reviewed By: sryap Differential Revision: D48090591 fbshipit-source-id: bb202c905728d01171a642ae0722d0aa97f25af7
AGZain
force-pushed
the
export-D48090591
branch
from
6d2ce9e to
40b569e
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 22, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Reviewed By: sryap Differential Revision: D48090591 fbshipit-source-id: 4096408033988ab17e0ca3c76a661e4bce6d8fde
AGZain
force-pushed
the
export-D48090591
branch
from
40b569e to
66cb344
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
added a commit
to AGZain/FBGEMM
that referenced
this pull request
Aug 23, 2023
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Reviewed By: sryap Differential Revision: D48090591 fbshipit-source-id: ebb5a15ae0a5cf9b0b03b6b5be78360838e05849
AGZain
force-pushed
the
export-D48090591
branch
from
66cb344 to
11a8d49
Compare
Summary: Pull Request resolved: pytorch#1912 # Background Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation. # Details The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list. Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] Offset_dims: [0, 2, 5, 6, 10] Permute: [3, 0, 2, 1, 3] Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9] # Backwards Compatibility The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu. # Documentation [Link](https://docs.google.com/document/d/1rp31nhcRuVmh8RVbWcktL3Hq-0ccc84AYL4nVegDywI) Reviewed By: sryap Differential Revision: D48090591 fbshipit-source-id: 54fce7365a7608a175b0eb0de9484a829fe2cd34
|
This pull request was exported from Phabricator. Differential Revision: D48090591 |
AGZain
force-pushed
the
export-D48090591
branch
from
11a8d49 to
ab8791d
Compare
|
This pull request has been merged in 9d221b7. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
Background
Currently permute_pooled_embs_gpu does not support duplicates in a permutation, this poses a problem with passing the same embeddings to multiple modules. This doc proposes a solution to allow duplicate subsets in the resultant permutation.
Details
The required implementation of permute_duplicate_pooled_embs_gpu should support a subset being repeated. This is represented by having duplicates in the permute list. This also results in the output list size being greater than the input list.
Input: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
Offset_dims: [0, 2, 5, 6, 10]
Permute: [3, 0, 2, 1, 3]
Output: [6, 7, 8, 9, 0, 1, 5, 2, 3, 4, 6, 7, 8, 9]
Backwards Compatibility
The permute_duplicate_pooled_embs_gpu is backwards compatible with the permute_pooled_embs_gpu (non duplicate permutations) use case, hence any use of permute_pooled_embs_gpu in the inference path can be replaced with permute_duplicate_pooled_embs_gpu.
Documentation
Link
Differential Revision: D48090591