Datasets:

DanielPFlorian

/

Transformers-Github-Issues

like 0

{ "login": "younesbelkada", "id": 49240599, "node_id": "MDQ6VXNlcjQ5MjQwNTk5", "avatar_url": "https://avatars.githubusercontent.com/u/49240599?v=4", "gravatar_id": "", "url": "https://api.github.com/users/younesbelkada", "html_url": "https://github.com/younesbelkada", "followers_url": "https://api.github.com/users/younesbelkada/followers", "following_url": "https://api.github.com/users/younesbelkada/following{/other_user}", "gists_url": "https://api.github.com/users/younesbelkada/gists{/gist_id}", "starred_url": "https://api.github.com/users/younesbelkada/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/younesbelkada/subscriptions", "organizations_url": "https://api.github.com/users/younesbelkada/orgs", "repos_url": "https://api.github.com/users/younesbelkada/repos", "events_url": "https://api.github.com/users/younesbelkada/events{/privacy}", "received_events_url": "https://api.github.com/users/younesbelkada/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27950). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "This PR enables compatiblity with Mixtral AWQ ! \r\nhttps://github.com/casper-hansen/AutoAWQ/pull/251 being merged, this PR is ready for review 🙏 ", "Thanks a lot for the review @amyeroberts ! \r\nI will merge this as soon as AutoAWQ makes the 0.1.8 Mixtral release cc @casper-hansen just for your information", "Release has been done! Merging!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27950/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27950", "html_url": "https://github.com/huggingface/transformers/pull/27950", "diff_url": "https://github.com/huggingface/transformers/pull/27950.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27950.patch", "merged_at": 1703498817000 }

{ "login": "petergtz", "id": 3618401, "node_id": "MDQ6VXNlcjM2MTg0MDE=", "avatar_url": "https://avatars.githubusercontent.com/u/3618401?v=4", "gravatar_id": "", "url": "https://api.github.com/users/petergtz", "html_url": "https://github.com/petergtz", "followers_url": "https://api.github.com/users/petergtz/followers", "following_url": "https://api.github.com/users/petergtz/following{/other_user}", "gists_url": "https://api.github.com/users/petergtz/gists{/gist_id}", "starred_url": "https://api.github.com/users/petergtz/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/petergtz/subscriptions", "organizations_url": "https://api.github.com/users/petergtz/orgs", "repos_url": "https://api.github.com/users/petergtz/repos", "events_url": "https://api.github.com/users/petergtz/events{/privacy}", "received_events_url": "https://api.github.com/users/petergtz/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27949). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27949/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27949", "html_url": "https://github.com/huggingface/transformers/pull/27949", "diff_url": "https://github.com/huggingface/transformers/pull/27949.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27949.patch", "merged_at": 1702307560000 }

{ "login": "ArthurZucker", "id": 48595927, "node_id": "MDQ6VXNlcjQ4NTk1OTI3", "avatar_url": "https://avatars.githubusercontent.com/u/48595927?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ArthurZucker", "html_url": "https://github.com/ArthurZucker", "followers_url": "https://api.github.com/users/ArthurZucker/followers", "following_url": "https://api.github.com/users/ArthurZucker/following{/other_user}", "gists_url": "https://api.github.com/users/ArthurZucker/gists{/gist_id}", "starred_url": "https://api.github.com/users/ArthurZucker/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ArthurZucker/subscriptions", "organizations_url": "https://api.github.com/users/ArthurZucker/orgs", "repos_url": "https://api.github.com/users/ArthurZucker/repos", "events_url": "https://api.github.com/users/ArthurZucker/events{/privacy}", "received_events_url": "https://api.github.com/users/ArthurZucker/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Actually will compute on GPU!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27948/reactions", "total_count": 7, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 7, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27948", "html_url": "https://github.com/huggingface/transformers/pull/27948", "diff_url": "https://github.com/huggingface/transformers/pull/27948.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27948.patch", "merged_at": 1702380029000 }

{ "login": "muellerzr", "id": 7831895, "node_id": "MDQ6VXNlcjc4MzE4OTU=", "avatar_url": "https://avatars.githubusercontent.com/u/7831895?v=4", "gravatar_id": "", "url": "https://api.github.com/users/muellerzr", "html_url": "https://github.com/muellerzr", "followers_url": "https://api.github.com/users/muellerzr/followers", "following_url": "https://api.github.com/users/muellerzr/following{/other_user}", "gists_url": "https://api.github.com/users/muellerzr/gists{/gist_id}", "starred_url": "https://api.github.com/users/muellerzr/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/muellerzr/subscriptions", "organizations_url": "https://api.github.com/users/muellerzr/orgs", "repos_url": "https://api.github.com/users/muellerzr/repos", "events_url": "https://api.github.com/users/muellerzr/events{/privacy}", "received_events_url": "https://api.github.com/users/muellerzr/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27947/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27947", "html_url": "https://github.com/huggingface/transformers/pull/27947", "diff_url": "https://github.com/huggingface/transformers/pull/27947.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27947.patch", "merged_at": 1702306661000 }

{ "login": "NielsRogge", "id": 48327001, "node_id": "MDQ6VXNlcjQ4MzI3MDAx", "avatar_url": "https://avatars.githubusercontent.com/u/48327001?v=4", "gravatar_id": "", "url": "https://api.github.com/users/NielsRogge", "html_url": "https://github.com/NielsRogge", "followers_url": "https://api.github.com/users/NielsRogge/followers", "following_url": "https://api.github.com/users/NielsRogge/following{/other_user}", "gists_url": "https://api.github.com/users/NielsRogge/gists{/gist_id}", "starred_url": "https://api.github.com/users/NielsRogge/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/NielsRogge/subscriptions", "organizations_url": "https://api.github.com/users/NielsRogge/orgs", "repos_url": "https://api.github.com/users/NielsRogge/repos", "events_url": "https://api.github.com/users/NielsRogge/events{/privacy}", "received_events_url": "https://api.github.com/users/NielsRogge/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27946/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27946", "html_url": "https://github.com/huggingface/transformers/pull/27946", "diff_url": "https://github.com/huggingface/transformers/pull/27946.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27946.patch", "merged_at": 1702306686000 }

{ "login": "CyberTimon", "id": 78795905, "node_id": "MDQ6VXNlcjc4Nzk1OTA1", "avatar_url": "https://avatars.githubusercontent.com/u/78795905?v=4", "gravatar_id": "", "url": "https://api.github.com/users/CyberTimon", "html_url": "https://github.com/CyberTimon", "followers_url": "https://api.github.com/users/CyberTimon/followers", "following_url": "https://api.github.com/users/CyberTimon/following{/other_user}", "gists_url": "https://api.github.com/users/CyberTimon/gists{/gist_id}", "starred_url": "https://api.github.com/users/CyberTimon/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/CyberTimon/subscriptions", "organizations_url": "https://api.github.com/users/CyberTimon/orgs", "repos_url": "https://api.github.com/users/CyberTimon/repos", "events_url": "https://api.github.com/users/CyberTimon/events{/privacy}", "received_events_url": "https://api.github.com/users/CyberTimon/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27945). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27945/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27945", "html_url": "https://github.com/huggingface/transformers/pull/27945", "diff_url": "https://github.com/huggingface/transformers/pull/27945.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27945.patch", "merged_at": 1702306729000 }

{ "login": "NielsRogge", "id": 48327001, "node_id": "MDQ6VXNlcjQ4MzI3MDAx", "avatar_url": "https://avatars.githubusercontent.com/u/48327001?v=4", "gravatar_id": "", "url": "https://api.github.com/users/NielsRogge", "html_url": "https://github.com/NielsRogge", "followers_url": "https://api.github.com/users/NielsRogge/followers", "following_url": "https://api.github.com/users/NielsRogge/following{/other_user}", "gists_url": "https://api.github.com/users/NielsRogge/gists{/gist_id}", "starred_url": "https://api.github.com/users/NielsRogge/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/NielsRogge/subscriptions", "organizations_url": "https://api.github.com/users/NielsRogge/orgs", "repos_url": "https://api.github.com/users/NielsRogge/repos", "events_url": "https://api.github.com/users/NielsRogge/events{/privacy}", "received_events_url": "https://api.github.com/users/NielsRogge/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27944). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27944/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27944", "html_url": "https://github.com/huggingface/transformers/pull/27944", "diff_url": "https://github.com/huggingface/transformers/pull/27944.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27944.patch", "merged_at": 1702317822000 }

{ "login": "vijaye12", "id": 25958261, "node_id": "MDQ6VXNlcjI1OTU4MjYx", "avatar_url": "https://avatars.githubusercontent.com/u/25958261?v=4", "gravatar_id": "", "url": "https://api.github.com/users/vijaye12", "html_url": "https://github.com/vijaye12", "followers_url": "https://api.github.com/users/vijaye12/followers", "following_url": "https://api.github.com/users/vijaye12/following{/other_user}", "gists_url": "https://api.github.com/users/vijaye12/gists{/gist_id}", "starred_url": "https://api.github.com/users/vijaye12/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/vijaye12/subscriptions", "organizations_url": "https://api.github.com/users/vijaye12/orgs", "repos_url": "https://api.github.com/users/vijaye12/repos", "events_url": "https://api.github.com/users/vijaye12/events{/privacy}", "received_events_url": "https://api.github.com/users/vijaye12/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@kashif ", "thanks! @amyeroberts \r\n" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27943/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27943", "html_url": "https://github.com/huggingface/transformers/pull/27943", "diff_url": "https://github.com/huggingface/transformers/pull/27943.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27943.patch", "merged_at": 1702295817000 }

{ "login": "ArthurZucker", "id": 48595927, "node_id": "MDQ6VXNlcjQ4NTk1OTI3", "avatar_url": "https://avatars.githubusercontent.com/u/48595927?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ArthurZucker", "html_url": "https://github.com/ArthurZucker", "followers_url": "https://api.github.com/users/ArthurZucker/followers", "following_url": "https://api.github.com/users/ArthurZucker/following{/other_user}", "gists_url": "https://api.github.com/users/ArthurZucker/gists{/gist_id}", "starred_url": "https://api.github.com/users/ArthurZucker/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ArthurZucker/subscriptions", "organizations_url": "https://api.github.com/users/ArthurZucker/orgs", "repos_url": "https://api.github.com/users/ArthurZucker/repos", "events_url": "https://api.github.com/users/ArthurZucker/events{/privacy}", "received_events_url": "https://api.github.com/users/ArthurZucker/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Are there any other aux losses apart from the LM loss?", "The auxiliary loss can be computed with `output_router_logits = True` automatically, other losses like `z_loss` can be imported from Switch transformers modeling code! Custom losses should be able to use the router_logits returned by the model" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27942/reactions", "total_count": 62, "+1": 26, "-1": 0, "laugh": 0, "hooray": 18, "confused": 0, "heart": 18, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27942", "html_url": "https://github.com/huggingface/transformers/pull/27942", "diff_url": "https://github.com/huggingface/transformers/pull/27942.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27942.patch", "merged_at": 1702295428000 }

{ "login": "R-N", "id": 1442761, "node_id": "MDQ6VXNlcjE0NDI3NjE=", "avatar_url": "https://avatars.githubusercontent.com/u/1442761?v=4", "gravatar_id": "", "url": "https://api.github.com/users/R-N", "html_url": "https://github.com/R-N", "followers_url": "https://api.github.com/users/R-N/followers", "following_url": "https://api.github.com/users/R-N/following{/other_user}", "gists_url": "https://api.github.com/users/R-N/gists{/gist_id}", "starred_url": "https://api.github.com/users/R-N/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/R-N/subscriptions", "organizations_url": "https://api.github.com/users/R-N/orgs", "repos_url": "https://api.github.com/users/R-N/repos", "events_url": "https://api.github.com/users/R-N/events{/privacy}", "received_events_url": "https://api.github.com/users/R-N/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @R-N, \r\n\r\nThanks for raising this issue! The link should now work as v4.36 has been released. ", "> Hi @R-N,\r\n> \r\n> Thanks for raising this issue! The link should now work as v4.36 has been released.\r\n\r\nConfirmed, it works now. Thank you" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27941/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "fxmarty", "id": 9808326, "node_id": "MDQ6VXNlcjk4MDgzMjY=", "avatar_url": "https://avatars.githubusercontent.com/u/9808326?v=4", "gravatar_id": "", "url": "https://api.github.com/users/fxmarty", "html_url": "https://github.com/fxmarty", "followers_url": "https://api.github.com/users/fxmarty/followers", "following_url": "https://api.github.com/users/fxmarty/following{/other_user}", "gists_url": "https://api.github.com/users/fxmarty/gists{/gist_id}", "starred_url": "https://api.github.com/users/fxmarty/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/fxmarty/subscriptions", "organizations_url": "https://api.github.com/users/fxmarty/orgs", "repos_url": "https://api.github.com/users/fxmarty/repos", "events_url": "https://api.github.com/users/fxmarty/events{/privacy}", "received_events_url": "https://api.github.com/users/fxmarty/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27940/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27940", "html_url": "https://github.com/huggingface/transformers/pull/27940", "diff_url": "https://github.com/huggingface/transformers/pull/27940.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27940.patch", "merged_at": 1702288598000 }

{ "login": "jq460494839", "id": 4471203, "node_id": "MDQ6VXNlcjQ0NzEyMDM=", "avatar_url": "https://avatars.githubusercontent.com/u/4471203?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jq460494839", "html_url": "https://github.com/jq460494839", "followers_url": "https://api.github.com/users/jq460494839/followers", "following_url": "https://api.github.com/users/jq460494839/following{/other_user}", "gists_url": "https://api.github.com/users/jq460494839/gists{/gist_id}", "starred_url": "https://api.github.com/users/jq460494839/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jq460494839/subscriptions", "organizations_url": "https://api.github.com/users/jq460494839/orgs", "repos_url": "https://api.github.com/users/jq460494839/repos", "events_url": "https://api.github.com/users/jq460494839/events{/privacy}", "received_events_url": "https://api.github.com/users/jq460494839/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27939/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27939", "html_url": "https://github.com/huggingface/transformers/pull/27939", "diff_url": "https://github.com/huggingface/transformers/pull/27939.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27939.patch", "merged_at": null }

{ "login": "jq460494839", "id": 4471203, "node_id": "MDQ6VXNlcjQ0NzEyMDM=", "avatar_url": "https://avatars.githubusercontent.com/u/4471203?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jq460494839", "html_url": "https://github.com/jq460494839", "followers_url": "https://api.github.com/users/jq460494839/followers", "following_url": "https://api.github.com/users/jq460494839/following{/other_user}", "gists_url": "https://api.github.com/users/jq460494839/gists{/gist_id}", "starred_url": "https://api.github.com/users/jq460494839/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jq460494839/subscriptions", "organizations_url": "https://api.github.com/users/jq460494839/orgs", "repos_url": "https://api.github.com/users/jq460494839/repos", "events_url": "https://api.github.com/users/jq460494839/events{/privacy}", "received_events_url": "https://api.github.com/users/jq460494839/received_events", "type": "User", "site_admin": false }

[]

[]

[ "```python\r\nfrom transformers import CpmAntTokenizer\r\n```\r\n\r\nis what you are looking for", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27938/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "gau-nernst", "id": 26946864, "node_id": "MDQ6VXNlcjI2OTQ2ODY0", "avatar_url": "https://avatars.githubusercontent.com/u/26946864?v=4", "gravatar_id": "", "url": "https://api.github.com/users/gau-nernst", "html_url": "https://github.com/gau-nernst", "followers_url": "https://api.github.com/users/gau-nernst/followers", "following_url": "https://api.github.com/users/gau-nernst/following{/other_user}", "gists_url": "https://api.github.com/users/gau-nernst/gists{/gist_id}", "starred_url": "https://api.github.com/users/gau-nernst/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/gau-nernst/subscriptions", "organizations_url": "https://api.github.com/users/gau-nernst/orgs", "repos_url": "https://api.github.com/users/gau-nernst/repos", "events_url": "https://api.github.com/users/gau-nernst/events{/privacy}", "received_events_url": "https://api.github.com/users/gau-nernst/received_events", "type": "User", "site_admin": false }

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored.", "Left it running as this is model specific rather than transformers bug IMO", "Indeed - feel free to post on the OpenAI repo since it looks like a model regression from large-v2 -> v3: https://github.com/openai/whisper/discussions", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27937/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "ppsmk388", "id": 60417397, "node_id": "MDQ6VXNlcjYwNDE3Mzk3", "avatar_url": "https://avatars.githubusercontent.com/u/60417397?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ppsmk388", "html_url": "https://github.com/ppsmk388", "followers_url": "https://api.github.com/users/ppsmk388/followers", "following_url": "https://api.github.com/users/ppsmk388/following{/other_user}", "gists_url": "https://api.github.com/users/ppsmk388/gists{/gist_id}", "starred_url": "https://api.github.com/users/ppsmk388/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ppsmk388/subscriptions", "organizations_url": "https://api.github.com/users/ppsmk388/orgs", "repos_url": "https://api.github.com/users/ppsmk388/repos", "events_url": "https://api.github.com/users/ppsmk388/events{/privacy}", "received_events_url": "https://api.github.com/users/ppsmk388/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey! 👋 llava is no part of the latest release! You need to install from source or wait until tomorrow! 🤗", "ok, thank you for your reply" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27936/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "steilgedacht", "id": 89748204, "node_id": "MDQ6VXNlcjg5NzQ4MjA0", "avatar_url": "https://avatars.githubusercontent.com/u/89748204?v=4", "gravatar_id": "", "url": "https://api.github.com/users/steilgedacht", "html_url": "https://github.com/steilgedacht", "followers_url": "https://api.github.com/users/steilgedacht/followers", "following_url": "https://api.github.com/users/steilgedacht/following{/other_user}", "gists_url": "https://api.github.com/users/steilgedacht/gists{/gist_id}", "starred_url": "https://api.github.com/users/steilgedacht/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/steilgedacht/subscriptions", "organizations_url": "https://api.github.com/users/steilgedacht/orgs", "repos_url": "https://api.github.com/users/steilgedacht/repos", "events_url": "https://api.github.com/users/steilgedacht/events{/privacy}", "received_events_url": "https://api.github.com/users/steilgedacht/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Can you run `make style`? ", "running `make style` makes no changes, do you have any ideas on what to change?", "Make sure you have `ruff==0.1.5` 🤗 and that you rebased on main! ", "Okay, I have already `ruff=0.1.5` and rebased it, am I doing anything wrong here? 😅\r\n\r\n\r\n\r\n\r\n", "Ah no I have a similar issue. just `make fixup` should help I think otherwise just reverting the changes ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27935/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27935", "html_url": "https://github.com/huggingface/transformers/pull/27935", "diff_url": "https://github.com/huggingface/transformers/pull/27935.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27935.patch", "merged_at": null }

{ "login": "NielsRogge", "id": 48327001, "node_id": "MDQ6VXNlcjQ4MzI3MDAx", "avatar_url": "https://avatars.githubusercontent.com/u/48327001?v=4", "gravatar_id": "", "url": "https://api.github.com/users/NielsRogge", "html_url": "https://github.com/NielsRogge", "followers_url": "https://api.github.com/users/NielsRogge/followers", "following_url": "https://api.github.com/users/NielsRogge/following{/other_user}", "gists_url": "https://api.github.com/users/NielsRogge/gists{/gist_id}", "starred_url": "https://api.github.com/users/NielsRogge/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/NielsRogge/subscriptions", "organizations_url": "https://api.github.com/users/NielsRogge/orgs", "repos_url": "https://api.github.com/users/NielsRogge/repos", "events_url": "https://api.github.com/users/NielsRogge/events{/privacy}", "received_events_url": "https://api.github.com/users/NielsRogge/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27934/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27934", "html_url": "https://github.com/huggingface/transformers/pull/27934", "diff_url": "https://github.com/huggingface/transformers/pull/27934.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27934.patch", "merged_at": 1702282622000 }

{ "login": "lmmx", "id": 2979452, "node_id": "MDQ6VXNlcjI5Nzk0NTI=", "avatar_url": "https://avatars.githubusercontent.com/u/2979452?v=4", "gravatar_id": "", "url": "https://api.github.com/users/lmmx", "html_url": "https://github.com/lmmx", "followers_url": "https://api.github.com/users/lmmx/followers", "following_url": "https://api.github.com/users/lmmx/following{/other_user}", "gists_url": "https://api.github.com/users/lmmx/gists{/gist_id}", "starred_url": "https://api.github.com/users/lmmx/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/lmmx/subscriptions", "organizations_url": "https://api.github.com/users/lmmx/orgs", "repos_url": "https://api.github.com/users/lmmx/repos", "events_url": "https://api.github.com/users/lmmx/events{/privacy}", "received_events_url": "https://api.github.com/users/lmmx/received_events", "type": "User", "site_admin": false }

[ { "id": 2648621985, "node_id": "MDU6TGFiZWwyNjQ4NjIxOTg1", "url": "https://api.github.com/repos/huggingface/transformers/labels/Feature%20request", "name": "Feature request", "color": "FBCA04", "default": false, "description": "Request for a new feature" } ]

[]

[ "please migrate to pydantic v2 🤗", "Please do this!", "Hi @lmmx, thanks for raising this! \r\n\r\nWe can certainly think about having v2 with v1 fallback support. However, I suspect the previous difficulty being raised was that it's hard to manage fallbacks when there's an incompatibility with a third-party library we interface with. Completely agree that pinning to an old version is brittle and it's best we update if and when possible. \r\n\r\nWDYT @ydshieh?", "Hi, thank you for asking!\r\n\r\nFor now, I can (at least) trigger a CI run with `Pydantic v2` and let's see how it goes to decide what to go!", "A PR (and its CI) is opened #28728 🤞 !", "CircleCI is good. I still need to check with docker image and other CI on github", "Docker image is fine\r\n\r\nhttps://github.com/huggingface/transformers/actions/runs/7669130818/job/20902409442\r\n\r\n(the failure at the end is not related to `Pydantic` -> everything is installed successfully)" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27933/reactions", "total_count": 1, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "younesbelkada", "id": 49240599, "node_id": "MDQ6VXNlcjQ5MjQwNTk5", "avatar_url": "https://avatars.githubusercontent.com/u/49240599?v=4", "gravatar_id": "", "url": "https://api.github.com/users/younesbelkada", "html_url": "https://github.com/younesbelkada", "followers_url": "https://api.github.com/users/younesbelkada/followers", "following_url": "https://api.github.com/users/younesbelkada/following{/other_user}", "gists_url": "https://api.github.com/users/younesbelkada/gists{/gist_id}", "starred_url": "https://api.github.com/users/younesbelkada/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/younesbelkada/subscriptions", "organizations_url": "https://api.github.com/users/younesbelkada/orgs", "repos_url": "https://api.github.com/users/younesbelkada/repos", "events_url": "https://api.github.com/users/younesbelkada/events{/privacy}", "received_events_url": "https://api.github.com/users/younesbelkada/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27932). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27932/reactions", "total_count": 2, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 1, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27932", "html_url": "https://github.com/huggingface/transformers/pull/27932", "diff_url": "https://github.com/huggingface/transformers/pull/27932.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27932.patch", "merged_at": 1702460544000 }

{ "login": "ArthurZucker", "id": 48595927, "node_id": "MDQ6VXNlcjQ4NTk1OTI3", "avatar_url": "https://avatars.githubusercontent.com/u/48595927?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ArthurZucker", "html_url": "https://github.com/ArthurZucker", "followers_url": "https://api.github.com/users/ArthurZucker/followers", "following_url": "https://api.github.com/users/ArthurZucker/following{/other_user}", "gists_url": "https://api.github.com/users/ArthurZucker/gists{/gist_id}", "starred_url": "https://api.github.com/users/ArthurZucker/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ArthurZucker/subscriptions", "organizations_url": "https://api.github.com/users/ArthurZucker/orgs", "repos_url": "https://api.github.com/users/ArthurZucker/repos", "events_url": "https://api.github.com/users/ArthurZucker/events{/privacy}", "received_events_url": "https://api.github.com/users/ArthurZucker/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27931). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "If I understand correctly, this PR should close the existing gap between inference with transformers + AutoGPTQ and inference with ExLlama, as the VRAM usage would become much more controlled. I'm rooting for it :)", "Thanks! 🤗", "Exciting PR! ", "1. This was some debug / tests but removed hehe\r\n2. The idea is to be able to save this and track usage + just make the api seamless. You can pass the cache instance but it should be able to be deduced. It's not a generation kwargs for me but a generation_config argument rather than that \r\n3. Attention mask is mostly what is left to adress TBH, because the control flow introduced are non trivial that is why I have a seperate PR that tried to take inspiration for JAX. Still working on it! Jax does update the attention mask with the cache see:\r\n\r\nhttps://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_flax_llama.py#L225\r\n\r\nMain attention challenge is that we have different attention classes, which need different kind of masks. My goal is to only pass the 2d attention mask, and have each attention layers use it. ", "How do I use this PR on npu device, here is my demo with errors.\r\n```python\r\nimport torch\r\nimport torch_npu\r\nfrom transformers import LlamaForCausalLM, LlamaTokenizer, TextStreamer\r\n\r\n\r\n\r\ntokenizer = LlamaTokenizer.from_pretrained(\r\n \"../Llama-2-7b-chat-hf\",\r\n device_map=\"npu:2\"\r\n)\r\nres = tokenizer(\"hello\")\r\n\r\nllama_model = LlamaForCausalLM.from_pretrained(\r\n \"../Llama-2-7b-chat-hf\",\r\n device_map=\"npu:2\"\r\n)\r\nstreamer = TextStreamer(tokenizer)\r\n\r\n\r\nprint(\"load successful\")\r\nwhile True:\r\n ins = input(\"user: \")\r\n res = tokenizer.encode(ins, return_tensors=\"pt\").to(\"npu:2\")\r\n print(res.dtype)\r\n outputs = llama_model.generate(\r\n inputs=res,\r\n streamer=streamer,\r\n max_new_tokens=10,\r\n )\r\n```\r\n\r\nTraceback (most recent call last):\r\n File \"test.py\", line 27, in <module>\r\n streamer=streamer,\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/torch/utils/_contextlib.py\", line 115, in decorate_context\r\n return func(*args, **kwargs)\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/transformers/generation/utils.py\", line 1790, in generate\r\n return self.sample(\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/transformers/generation/utils.py\", line 2887, in sample\r\n outputs = self(\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/transformers/models/llama/modeling_llama.py\", line 1206, in forward\r\n outputs = self.model(\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/transformers/models/llama/modeling_llama.py\", line 1093, in forward\r\n layer_outputs = decoder_layer(\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/transformers/models/llama/modeling_llama.py\", line 820, in forward\r\n hidden_states, self_attn_weights, present_key_value = self.self_attn(\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/usr/local/conda/envs/PyTorch2.1.0/lib/python3.8/site-packages/transformers/models/llama/modeling_llama.py\", line 438, in forward\r\n raise ValueError(\r\nValueError: Attention mask should be of size (1, 1, 2, 2), but is torch.Size([1, 2])", "This PR is still very much in draft mode. A script can be found in the tests, but does not support everything yet! Would recommend you to wait a tad bit, should be in a better state by the end of the week", "cc @oobabooga the PR is more or less ready, would love to get your feedback from using it in `oobabooga`! (this needs a bit of custom work to compile only the decoding step, but should already help!) ", "I had a look at them, the main pain point here is: \r\n- the overhead on a generation is ~2 token /s \r\n- the overhead of not using the `position_ids` tensor in compiled mode is more than 40 tokens / s\r\ns", "Alternative for ROPE that does not break cuda graphs + faster (5%) than vanilla recomputing. Issue is you save these tensor so loading will be wrong for sure.\r\n```python\r\nclass LlamaRotaryEmbedding(nn.Module):\r\n def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):\r\n super().__init__()\r\n\r\n self.dim = dim\r\n self.max_position_embeddings = max_position_embeddings\r\n self.base = base\r\n inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))\r\n self.register_buffer(\"inv_freq\", inv_freq, persistent=False)\r\n # Build here to make `torch.jit.trace` work.\r\n self._set_cos_sin_cache(\r\n seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()\r\n )\r\n\r\n \r\n def _set_cos_sin_cache(self, seq_len, device, dtype):\r\n self.max_seq_len_cached = seq_len\r\n t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)\r\n\r\n freqs = torch.outer(t, self.inv_freq)\r\n # Different from paper, but it uses a different permutation in order to obtain the same calculation\r\n emb = torch.cat((freqs, freqs), dim=-1)\r\n self.register_buffer(\"cos_cached\", emb.cos().to(dtype), persistent=False)\r\n self.register_buffer(\"sin_cached\", emb.sin().to(dtype), persistent=False)\r\n\r\n \r\n def forward(self, x, position_ids, seq_len=None):\r\n # x: [bs, num_attention_heads, seq_len, head_size]\r\n if seq_len > self.max_seq_len_cached:\r\n self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)\r\n cos = F.embedding(position_ids, self.cos_cached)[0,:,:].to(dtype=x.dtype)\r\n sin = F.embedding(position_ids, self.sin_cached)[0,:,:].to(dtype=x.dtype)\r\n return cos, sin\r\n```\r\n<img width=\"896\" alt=\"image\" src=\"https://github.com/huggingface/transformers/assets/48595927/0e37d453-2712-462b-9e33-e6183e3bd983\">\r\n", "@ArthurZucker when will the next minor release (4.37.3?) of transformers come out? Will this PR be included in 4.37.3? ", "@tsengalb99 v4.37.3 would be a micro release. These are reserved for patch fixes which resolve newly introduced breaking changes in the repo or fixing regressions from the last release.\r\n\r\nThe next minor release is v4.38. We release on a roughly monthly schedule, so would be in around two weeks. \r\n\r\nIf you want to use this feature immediately, you can [install from source](https://huggingface.co./docs/transformers/installation#install-from-source). ", "Is this working in general or only for some specific hardware, etc? \r\nI tried using fresh installs with WSL2, ubuntu, Cuda 11.8 (and again with 12.1) on an RTX A5000\r\n\r\n I installed from source today and tried the test script (https://gist.github.com/ArthurZucker/2dd607c4333ac4c489af30f54a1d8a2d)\r\n\r\n but the model.generate for compiled forward throws runtimeError: attn_bias is not correctly aligned (strideH). attn_bias.stride(1) = 21, and should be a multiple of 8. \r\n\r\n", "What is the recommended way to use cuda graphs with this PR? The torch cuda graph wrapper does not appear to be working iirc due to this line `attention_mask is not None and not torch.all(attention_mask[..., 0] == 1) and q_len != 1` / the equivalent in all model files.", "This line is probably gonne be removed in #28937 \r\nI'update once this is ready.\r\nI use torch nightly so torch 2.3! ", "@ArthurZucker Great to see the PR has been merged. I tried with the latest nightlies on H100, and running into errors (full [logs](https://gist.github.com/chauhang/a0dabe50c9958af5c92a3b1a23d1e888))\r\n\r\nwith sdpa attention: \r\n`RuntimeError: Event device type CUDA does not match blocking stream's device type CPU.`\r\n\r\nwith flash_attention_2 (fails in import itself):\r\n`flash_attn_2_cuda.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZN3c104cuda9SetDeviceEi`", "Nice, could you share a full stack trace? 🤗 \r\n", "@chauhang \r\n\r\n> with flash_attention_2 (fails in import itself):\r\n\r\nThis means that the your local torch version does not match the torch version flash-attn was built against. You would need to uninstall & rebuild flash-attn." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27931/reactions", "total_count": 21, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 12, "rocket": 9, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27931", "html_url": "https://github.com/huggingface/transformers/pull/27931", "diff_url": "https://github.com/huggingface/transformers/pull/27931.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27931.patch", "merged_at": 1707389434000 }

{ "login": "kkckk1110", "id": 144304282, "node_id": "U_kgDOCJnomg", "avatar_url": "https://avatars.githubusercontent.com/u/144304282?v=4", "gravatar_id": "", "url": "https://api.github.com/users/kkckk1110", "html_url": "https://github.com/kkckk1110", "followers_url": "https://api.github.com/users/kkckk1110/followers", "following_url": "https://api.github.com/users/kkckk1110/following{/other_user}", "gists_url": "https://api.github.com/users/kkckk1110/gists{/gist_id}", "starred_url": "https://api.github.com/users/kkckk1110/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/kkckk1110/subscriptions", "organizations_url": "https://api.github.com/users/kkckk1110/orgs", "repos_url": "https://api.github.com/users/kkckk1110/repos", "events_url": "https://api.github.com/users/kkckk1110/events{/privacy}", "received_events_url": "https://api.github.com/users/kkckk1110/received_events", "type": "User", "site_admin": false }

[]

{ "login": "kashif", "id": 8100, "node_id": "MDQ6VXNlcjgxMDA=", "avatar_url": "https://avatars.githubusercontent.com/u/8100?v=4", "gravatar_id": "", "url": "https://api.github.com/users/kashif", "html_url": "https://github.com/kashif", "followers_url": "https://api.github.com/users/kashif/followers", "following_url": "https://api.github.com/users/kashif/following{/other_user}", "gists_url": "https://api.github.com/users/kashif/gists{/gist_id}", "starred_url": "https://api.github.com/users/kashif/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/kashif/subscriptions", "organizations_url": "https://api.github.com/users/kashif/orgs", "repos_url": "https://api.github.com/users/kashif/repos", "events_url": "https://api.github.com/users/kashif/events{/privacy}", "received_events_url": "https://api.github.com/users/kashif/received_events", "type": "User", "site_admin": false }

[ { "login": "kashif", "id": 8100, "node_id": "MDQ6VXNlcjgxMDA=", "avatar_url": "https://avatars.githubusercontent.com/u/8100?v=4", "gravatar_id": "", "url": "https://api.github.com/users/kashif", "html_url": "https://github.com/kashif", "followers_url": "https://api.github.com/users/kashif/followers", "following_url": "https://api.github.com/users/kashif/following{/other_user}", "gists_url": "https://api.github.com/users/kashif/gists{/gist_id}", "starred_url": "https://api.github.com/users/kashif/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/kashif/subscriptions", "organizations_url": "https://api.github.com/users/kashif/orgs", "repos_url": "https://api.github.com/users/kashif/repos", "events_url": "https://api.github.com/users/kashif/events{/privacy}", "received_events_url": "https://api.github.com/users/kashif/received_events", "type": "User", "site_admin": false } ]

[ "Hey! Could you make sure you provide a reproducer, isolating the bug? We can't really debug your code in your stead 🤗 ", "Thanks to your quick reply! Here are my detailed codes related to the bug.\r\n\r\n```python\r\nfrom transformers import TimeSeriesTransformerConfig, TimeSeriesTransformerForPrediction\r\n\r\nconfig = TimeSeriesTransformerConfig(\r\n prediction_length = prediction_length,\r\n context_length = prediction_length,\r\n lags_sequence = lags_sequence,\r\n num_time_features = len(time_features) + 1,\r\n num_static_categorical_features = 1,\r\n num_dynamic_real_features = 48,\r\n \r\n cardinality = [len(train_dataset)],\r\n embedding_dimension = [2],\r\n\r\n # transformer params:\r\n encoder_layers = 4,\r\n decoder_layers = 4,\r\n d_model = 32,\r\n)\r\n\r\nmodel = TimeSeriesTransformerForPrediction(config)\r\n\r\ndef create_test_dataloader(\r\n config: PretrainedConfig,\r\n freq,\r\n data,\r\n batch_size: int,\r\n **kwargs,\r\n):\r\n PREDICTION_INPUT_NAMES = [\r\n \"past_time_features\",\r\n \"past_values\", \r\n \"past_observed_mask\",\r\n \"future_time_features\",\r\n ]\r\n if config.num_static_categorical_features > 0:\r\n PREDICTION_INPUT_NAMES.append(\"static_categorical_features\")\r\n\r\n if config.num_static_real_features > 0:\r\n PREDICTION_INPUT_NAMES.append(\"static_real_features\")\r\n \r\n transformation = create_transformation(freq, config)\r\n transformed_data = transformation.apply(data, is_train=False) #!!!\r\n\r\n # we create a Test Instance splitter which will sample the very last\r\n # context window seen during training only for the encoder.\r\n instance_sampler = create_instance_splitter(config, \"test\")\r\n\r\n # we apply the transformations in test mode\r\n testing_instances = instance_sampler.apply(transformed_data, is_train=False)\r\n\r\n return as_stacked_batches(\r\n testing_instances,\r\n batch_size=batch_size,\r\n output_type=torch.tensor,\r\n field_names=PREDICTION_INPUT_NAMES,\r\n )\r\n```", "Actually, most of the codes are identical with those in https://huggingface.co./blog/time-series-transformers, except that I set num_dynamic_real_features = 48 in the Config", "Yep, but this changes the value of `self._number_of_features`, which affect `self.feature_size` in return. The `past_time_features` given to the model need to be updated as well. I'm betting the issue is with the dataset rather than anything. The `feat_dynamic_real`'s shape is expected to be `2,75` apparently. I'm not super familiar with this model but if you can share the dataset you are using as well it would help 🤗 ", "Thank you very much! I will provide more details about my dataset. I am working on my own dataset, using the following codes.\r\n```python\r\nfrom gluonts.dataset.pandas import PandasDataset\r\n\r\nprediction_length = 8\r\ntrain = PandasDataset.from_long_dataframe(train, target=\"sales\", item_id=\"city\",\r\n feat_dynamic_real = [col for col in train.columns if ((col != 'sales') & (col != 'city'))])\r\ntest = PandasDataset.from_long_dataframe(test, target=\"sales\", item_id=\"city\",\r\n feat_dynamic_real = [col for col in test.columns if ((col != 'sales') & (col != 'city'))])\r\n\r\nclass ProcessStartField():\r\n ts_id = 0\r\n def __call__(self, data):\r\n data[\"start\"] = data[\"start\"].to_timestamp()\r\n data[\"feat_static_cat\"] = [self.ts_id]\r\n dynamic_feature_values = np.array(data[\"feat_dynamic_real\"])\r\n data[\"feat_dynamic_real\"] = dynamic_feature_values\r\n self.ts_id += 1\r\n return data\r\n\r\nfrom gluonts.itertools import Map\r\nprocess_start = ProcessStartField()\r\nlist_train = list(Map(process_start, train))\r\nprocess_start = ProcessStartField()\r\nlist_test = list(Map(process_start, test))\r\n\r\ntrainset = Dataset.from_list(list_train) #, features=features\r\ntestset = Dataset.from_list(list_test)\r\n```\r\nAnd my dataset looks like:\r\n```\r\ntrain_example: {'start': datetime.datetime(2017, 1, 1, 0, 0),\r\n 'target': [806, 806...] (a list with length of **59**)\r\n'item_id': cityname\r\n'feat_dynamic_real': [[215,219...],[217,192...]...] (a list comprising of _48_ sublists, each sublist has length of **59**)}\r\n'feat_static_cat': [0]}\r\n```\r\nThe corresponding test_example is :\r\n```\r\n {'start': datetime.datetime(2017, 1, 1, 0, 0),\r\n 'target': [806, 806...] (a list with length of **67**)\r\n'item_id': cityname\r\n'feat_dynamic_real': [[215,219...],[217,192...]...] (a list comprising of _48_ sublists, each sublist has length of **67**)}\r\n'feat_static_cat': [0]}\r\n```\r\nI also tried other number of context length, and I found that, the codes always transformed the testdata to the length of 67+context_length, which in turn incurred an error. \r\nI am really confused about the bug for a long time. I would really appreciate it if you can help! Thanks!", "cc @NielsRogge if you have any idea?!", "cc @kashif, the master of time series", "@kkckk1110 so it could be that the arrays `feat_dynamic_real` in your dataset have to be transposed... let me double-check but you can quickly try it if that helps.", "Thank you very much for your attention! I have already tried but came across errors though.", "can you paste a bigger error trace? is it from the `VStack` transformation?\r\n", "Exactly! the complete error information is as follows.\r\n\r\n---------------------------------------------------------------------------\r\nValueError Traceback (most recent call last)\r\nCell In[51], line 1\r\n----> 1 batch = next(iter(test_dataloader))\r\n 2 for k, v in batch.items():\r\n 3 print(k, v.shape, v.type())\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/itertools.py:415, in IterableSlice.__iter__(self)\r\n 414 def __iter__(self):\r\n--> 415 yield from itertools.islice(self.iterable, self.length)\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/transform/_base.py:111, in TransformedDataset.__iter__(self)\r\n 110 def __iter__(self) -> Iterator[DataEntry]:\r\n--> 111 yield from self.transformation(\r\n 112 self.base_dataset, is_train=self.is_train\r\n 113 )\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/transform/_base.py:132, in MapTransformation.__call__(self, data_it, is_train)\r\n 129 def __call__(\r\n 130 self, data_it: Iterable[DataEntry], is_train: bool\r\n 131 ) -> Iterator:\r\n--> 132 for data_entry in data_it:\r\n 133 try:\r\n 134 yield self.map_transform(data_entry.copy(), is_train)\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/dataset/loader.py:55, in Stack.__call__(self, data, is_train)\r\n 54 def __call__(self, data, is_train):\r\n---> 55 for batch in data:\r\n 56 yield rows_to_columns(batch, np.array)\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/dataset/loader.py:50, in Batch.__call__(self, data, is_train)\r\n 49 def __call__(self, data, is_train):\r\n---> 50 yield from batcher(data, self.batch_size)\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/itertools.py:131, in batcher.<locals>.get_batch()\r\n 130 def get_batch():\r\n--> 131 return list(itertools.islice(it, batch_size))\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/transform/_base.py:132, in MapTransformation.__call__(self, data_it, is_train)\r\n 129 def __call__(\r\n 130 self, data_it: Iterable[DataEntry], is_train: bool\r\n 131 ) -> Iterator:\r\n--> 132 for data_entry in data_it:\r\n 133 try:\r\n 134 yield self.map_transform(data_entry.copy(), is_train)\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/transform/_base.py:111, in TransformedDataset.__iter__(self)\r\n 110 def __iter__(self) -> Iterator[DataEntry]:\r\n--> 111 yield from self.transformation(\r\n 112 self.base_dataset, is_train=self.is_train\r\n 113 )\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/transform/_base.py:186, in FlatMapTransformation.__call__(self, data_it, is_train)\r\n 182 def __call__(\r\n 183 self, data_it: Iterable[DataEntry], is_train: bool\r\n 184 ) -> Iterator:\r\n 185 num_idle_transforms = 0\r\n--> 186 for data_entry in data_it:\r\n 187 num_idle_transforms += 1\r\n 188 for result in self.flatmap_transform(data_entry.copy(), is_train):\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/transform/_base.py:111, in TransformedDataset.__iter__(self)\r\n 110 def __iter__(self) -> Iterator[DataEntry]:\r\n--> 111 yield from self.transformation(\r\n 112 self.base_dataset, is_train=self.is_train\r\n 113 )\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/transform/_base.py:132, in MapTransformation.__call__(self, data_it, is_train)\r\n 129 def __call__(\r\n 130 self, data_it: Iterable[DataEntry], is_train: bool\r\n 131 ) -> Iterator:\r\n--> 132 for data_entry in data_it:\r\n 133 try:\r\n 134 yield self.map_transform(data_entry.copy(), is_train)\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/transform/_base.py:136, in MapTransformation.__call__(self, data_it, is_train)\r\n 134 yield self.map_transform(data_entry.copy(), is_train)\r\n 135 except Exception as e:\r\n--> 136 raise e\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/transform/_base.py:134, in MapTransformation.__call__(self, data_it, is_train)\r\n 132 for data_entry in data_it:\r\n 133 try:\r\n--> 134 yield self.map_transform(data_entry.copy(), is_train)\r\n 135 except Exception as e:\r\n 136 raise e\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/transform/_base.py:149, in SimpleTransformation.map_transform(self, data, is_train)\r\n 148 def map_transform(self, data: DataEntry, is_train: bool) -> DataEntry:\r\n--> 149 return self.transform(data)\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/gluonts/transform/convert.py:219, in VstackFeatures.transform(self, data)\r\n 213 def transform(self, data: DataEntry) -> DataEntry:\r\n 214 r = [\r\n 215 data[fname]\r\n 216 for fname in self.input_fields\r\n 217 if data[fname] is not None\r\n 218 ]\r\n--> 219 output = np.vstack(r) if not self.h_stack else np.hstack(r)\r\n 220 data[self.output_field] = output\r\n 221 for fname in self.cols_to_drop:\r\n\r\nFile ~/miniforge3/envs/tensorflow/lib/python3.9/site-packages/numpy/core/shape_base.py:289, in vstack(tup, dtype, casting)\r\n 287 if not isinstance(arrs, list):\r\n 288 arrs = [arrs]\r\n--> 289 return _nx.concatenate(arrs, 0, dtype=dtype, casting=casting)\r\n\r\nValueError: all the input array dimensions except for the concatenation axis must match exactly, but along dimension 1, the array at index 0 has size 75 and the array at index 2 has size 67", "so yes the shape of the `data[\"feat_dynamic_real\"].shape` has to be (`num_dynamic_real_features`, `time`), can you confirm that is the case in your dataset?\r\n", "Sure! \r\nnp.array(train_example['feat_dynamic_real']).shape : (48, 59)\r\nnp.array(validation_example['feat_dynamic_real']).shape : (48, 67)\r\nwhich are expected and consistent with (num_dynamic_real_features, time)", "@kkckk1110 ok I figured it out, my bad, i'll update the blog post: In the mean time use this for the back-testing:\r\n\r\n```python\r\n def create_backtest_dataloader(\r\n config: PretrainedConfig,\r\n freq,\r\n data,\r\n batch_size: int,\r\n **kwargs,\r\n):\r\n PREDICTION_INPUT_NAMES = [\r\n \"past_time_features\",\r\n \"past_values\",\r\n \"past_observed_mask\",\r\n \"future_time_features\",\r\n ]\r\n if config.num_static_categorical_features > 0:\r\n PREDICTION_INPUT_NAMES.append(\"static_categorical_features\")\r\n\r\n if config.num_static_real_features > 0:\r\n PREDICTION_INPUT_NAMES.append(\"static_real_features\")\r\n\r\n transformation = create_transformation(freq, config)\r\n transformed_data = transformation.apply(data, is_train=True)\r\n\r\n # we create a Test Instance splitter which will sample the very last\r\n # context window seen during training only for the encoder.\r\n instance_sampler = create_instance_splitter(config, \"validation\")\r\n\r\n # we apply the transformations in train mode\r\n backtesting_instances = instance_sampler.apply(transformed_data, is_train=True)\r\n \r\n return as_stacked_batches(\r\n backtesting_instances,\r\n batch_size=batch_size,\r\n output_type=torch.tensor,\r\n field_names=PREDICTION_INPUT_NAMES,\r\n )\r\n``` ", "It works! Great! Can you elaborate on why the modified codes solved the problems?", "yes the splitter for back-testing still needs to be applied in \"Training\"\r\n\r\n```\r\n backtesting_instances = instance_sampler.apply(transformed_data, is_train=True)\r\n ```\r\n \r\n else the dynamic features are not split while the target is... I'll check with the gluonts folks ", "I really appreciate your efforts! I hate to bother you but I have another problem. I wonder how can I interpret time series models, such as analyzing feature importance? Are there any tools like SHAP, for time series forecasting?", "there is none at the moment for transformer-based models as far as i know... one does have the attention weights so can figure out which time points were used to make a particular prediction" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27930/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "thundergolfer", "id": 12058921, "node_id": "MDQ6VXNlcjEyMDU4OTIx", "avatar_url": "https://avatars.githubusercontent.com/u/12058921?v=4", "gravatar_id": "", "url": "https://api.github.com/users/thundergolfer", "html_url": "https://github.com/thundergolfer", "followers_url": "https://api.github.com/users/thundergolfer/followers", "following_url": "https://api.github.com/users/thundergolfer/following{/other_user}", "gists_url": "https://api.github.com/users/thundergolfer/gists{/gist_id}", "starred_url": "https://api.github.com/users/thundergolfer/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/thundergolfer/subscriptions", "organizations_url": "https://api.github.com/users/thundergolfer/orgs", "repos_url": "https://api.github.com/users/thundergolfer/repos", "events_url": "https://api.github.com/users/thundergolfer/events{/privacy}", "received_events_url": "https://api.github.com/users/thundergolfer/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Test failures are for \"documentation\" and \"transformers metadata\", same as last time (https://github.com/huggingface/transformers/pull/27820#issuecomment-1841186551)", "> have you confirmed this works on a multi-GPU system?\r\n\r\nYes, that's detailed in the PR description, starting with the sentence: \"I didn't setup a multi-GPU VM to run the test, ...\"\r\n\r\nAlso if you agree with the TODO, I'm happy to make a follow-up PR addressing it 🙂 ", "Sorry for missing it! I'll run it locally here and get back to you on if the solution indeed works. If so, yes a follow up PR for that would be great :) ", "Any Update here? Thanks!", "any update here.? waiting for the PR merge", "I tried the changes from this PR, but I got other issue as follows:\r\n```\r\n[E ProcessGroupNCCL.cpp:475] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=387992, OpType=_ALLGATHER_BASE, NumelIn=6291456, NumelOut=25165824, Timeout(ms)=1800000) ran for 1800865 milliseconds before timing out.\r\n[E ProcessGroupNCCL.cpp:489] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data.\r\n[E ProcessGroupNCCL.cpp:495] To avoid data inconsistency, we are taking the entire process down.\r\n[E ProcessGroupNCCL.cpp:916] [Rank 0] NCCL watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=387891, OpType=_ALLGATHER_BASE, NumelIn=1024, NumelOut=4096, Timeout(ms)=1800000) ran for 1801725 milliseconds before timing out.\r\nterminate called after throwing an instance of 'std::runtime_error'\r\n what(): [Rank 0] NCCL watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=387891, OpType=_ALLGATHER_BASE, NumelIn=1024, NumelOut=4096, Timeout(ms)=1800000) ran for 1801725 milliseconds before timing out.\r\n```\r\n\r\n\r\nI realized that why we don't just check the existence of the folder like this:\r\n```python\r\n ...\r\n if os.path.exists(staging_output_dir):\r\n if self.args.should_save:\r\n self.state.save_to_json(os.path.join(staging_output_dir, TRAINER_STATE_NAME))\r\n \r\n if self.args.push_to_hub:\r\n self._push_from_checkpoint(staging_output_dir)\r\n \r\n # Place checkpoint in final location after all saving is finished.\r\n if staging_output_dir != output_dir:\r\n os.rename(staging_output_dir, output_dir)\r\n ...\r\n```\r\nThis works smoothly.", "@thundergolfer I have a different fix coming in that works, the issue is you were not checking that the rename of the staging folder was happening just on the main process: https://github.com/huggingface/transformers/pull/28009", "Closing in favor of https://github.com/huggingface/transformers/pull/28009 as this change still doesn't handle all multi-GPU scenarios. " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27929/reactions", "total_count": 1, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 1, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27929", "html_url": "https://github.com/huggingface/transformers/pull/27929", "diff_url": "https://github.com/huggingface/transformers/pull/27929.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27929.patch", "merged_at": null }

{ "login": "daehuikim", "id": 40377750, "node_id": "MDQ6VXNlcjQwMzc3NzUw", "avatar_url": "https://avatars.githubusercontent.com/u/40377750?v=4", "gravatar_id": "", "url": "https://api.github.com/users/daehuikim", "html_url": "https://github.com/daehuikim", "followers_url": "https://api.github.com/users/daehuikim/followers", "following_url": "https://api.github.com/users/daehuikim/following{/other_user}", "gists_url": "https://api.github.com/users/daehuikim/gists{/gist_id}", "starred_url": "https://api.github.com/users/daehuikim/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/daehuikim/subscriptions", "organizations_url": "https://api.github.com/users/daehuikim/orgs", "repos_url": "https://api.github.com/users/daehuikim/repos", "events_url": "https://api.github.com/users/daehuikim/events{/privacy}", "received_events_url": "https://api.github.com/users/daehuikim/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @younesbelkada ", "Hmm looks interesting. I think, when you print it out, first thing is considering original+ adapter and last thing is considering only structure of it’s backbone. \r\nDid you considered a random state, before inference?\r\n\r\nSometimes, same input’s result can be different when you inference it. ", "> Hmm looks interesting. I think, when you print it out, first thing is considering original+ adapter and last thing is considering only structure of it’s backbone. Did you considered a random state, before inference?\r\n> \r\n> Sometimes, same input’s result can be different when you inference it.\r\n\r\nThe reason why the structure of the last one feels exactly the same as the structure of the backbone model is because the lora adapter of this one has already been merged, so there is no numerical difference.\r\nThe only difference is the Object they are carried (```PeftModelForCausalLM```, ```AutoModelForCasualLM```).\r\n\r\nAnd 2 models are initialized from different sources. one from ```PeftModelForCausalLM.from_pretrained``` and one for from ```AutoModelForCasualLM.from_pretrained```. However, every params, configuration files are the same except the object.", "Hi @daehuikim \r\nThanks for your issue, recall the formulation of the LoRA adapters in the figure below:\r\n\r\n\r\nThe fundamental difference between `AutoModelForCausalLM` and `PeftModelForCausalLM`, in your case is that the `AutoModelForCausalLM` contains the merged adapter whereas the second model contains the model with LoRA attached on them. Technically the inference results should be the same as the LoRA operations can be simply rewritten as a refactorization of simple matrix multiplication:\r\n\r\n\r\n\r\nNow regarding your issue, it seems you have further trained your `lm_head` as well. I think `merge_and_unload` does not properly take care of that, but I am not sure. Can you confirm which PEFT version are you using now?", "> Hi @daehuikim Thanks for your issue, recall the formulation of the LoRA adapters in the figure below:  [  ](https://camo.githubusercontent.com/0c4168380b2d5a2d93f963838ccd920bc6384e56ebea50f4904031289db70673/68747470733a2f2f63646e2d75706c6f6164732e68756767696e67666163652e636f2f70726f64756374696f6e2f75706c6f6164732f3632343431643164396664656662353561306237643132632f454d7473444f6634576837617665583077792d4f5a2e676966) [ ](https://camo.githubusercontent.com/0c4168380b2d5a2d93f963838ccd920bc6384e56ebea50f4904031289db70673/68747470733a2f2f63646e2d75706c6f6164732e68756767696e67666163652e636f2f70726f64756374696f6e2f75706c6f6164732f3632343431643164396664656662353561306237643132632f454d7473444f6634576837617665583077792d4f5a2e676966)\r\n> \r\n> The fundamental difference between `AutoModelForCausalLM` and `PeftModelForCausalLM`, in your case is that the `AutoModelForCausalLM` contains the merged adapter whereas the second model contains the model with LoRA attached on them. Technically the inference results should be the same as the LoRA operations can be simply rewritten as a refactorization of simple matrix multiplication:\r\n> \r\n> \r\n> \r\n> Now regarding your issue, it seems you have further trained your `lm_head` as well. I think `merge_and_unload` does not properly take care of that, but I am not sure. Can you confirm which PEFT version are you using now?\r\n\r\nThanks for describing the issue correctly.\r\nMy peftconfig for training is like below: (```lm_head``` is not trained by trainer) \r\n```\r\npeft_config = LoraConfig(\r\n lora_alpha=lora_alpha,\r\n lora_dropout=lora_dropout,\r\n r=lora_r,\r\n bias=\"none\",\r\n task_type=\"CAUSAL_LM\",\r\n target_modules= [\r\n \"q_proj\",\r\n \"k_proj\",\r\n \"v_proj\",\r\n \"o_proj\",\r\n \"gate_proj\",\r\n \"up_proj\",\r\n \"down_proj\",\r\n ],\r\n modules_to_save=[\r\n \"embed_tokens\",\r\n \"lm_head\"\r\n ]\r\n )\r\n```\r\nAnd this is my peft version\r\n```\r\n$ pip freeze | grep peft\r\npeft==0.6.2\r\n```\r\n\r\nThanks for answering! @younesbelkada ", "Thanks @daehuikim \r\nYou are adding \r\n\r\n```py\r\n modules_to_save=[\r\n \"embed_tokens\",\r\n \"lm_head\"\r\n ]\r\n```\r\nin the lora config, this will make the lm head and the input embedding layer trainable, hence the strange behaviour you are facing (I am not sure we support merge and unload for models that have `modules_to_save`). Can you try without `modules_to_save`?", "> Thanks @daehuikim You are adding\r\n> \r\n> ```python\r\n> modules_to_save=[\r\n> \"embed_tokens\",\r\n> \"lm_head\"\r\n> ]\r\n> ```\r\n> \r\n> in the lora config, this will make the lm head and the input embedding layer trainable, hence the strange behaviour you are facing (I am not sure we support merge and unload for models that have `modules_to_save`). Can you try without `modules_to_save`?\r\n\r\nThanks again @younesbelkada \r\n\r\n```modules_to_save``` wrapper save the modules not to be trained from trainer.\r\nAlso, ```embed_tokens, lm_head``` must to be saved when there exist added_tokens.\r\nTherefore, ```modules_to_save``` block is necessary for my task.\r\nI will see what happens in ```merge_and_unload``` for finding this out.\r\nThank you again!", "https://github.com/huggingface/peft/blob/21c304f6f6ea62c7dcbef8e201d178a7575b471d/src/peft/tuners/lora/model.py#L419\r\n\r\n```\r\nelif isinstance(target, ModulesToSaveWrapper):\r\n # save any additional trainable modules part of `modules_to_save`\r\n setattr(parent, target_name, target.modules_to_save[target.active_adapter])\r\n```\r\n@younesbelkada \r\n```modules_to_save``` preserve modules from parent model.\r\nTheoretically both model should reproduce same outputs.\r\nBut why this still happen? I am curious.", "Hi, I'm having the same problem as you, in my scenario I also need to add some tokens for training and I've also set the\r\n```\r\nmodules_to_save=[\r\n \"embed_tokens\",\r\n \"lm_head\"\r\n ]\r\n```\r\nMay I ask if the problem is solved now? How did you end up solving it, looking forward to your reply.", "> Hi, I'm having the same problem as you, in my scenario I also need to add some tokens for training and I've also set the\r\n> \r\n> ```\r\n> modules_to_save=[\r\n> \"embed_tokens\",\r\n> \"lm_head\"\r\n> ]\r\n> ```\r\n> \r\n> May I ask if the problem is solved now? How did you end up solving it, looking forward to your reply.\r\n\r\nI have no idea yet. whole codes seems same in peft and transformers. I guess it could be some weight initialization problem", "> > Hi, I'm having the same problem as you, in my scenario I also need to add some tokens for training and I've also set the\r\n> > ```\r\n> > modules_to_save=[\r\n> > \"embed_tokens\",\r\n> > \"lm_head\"\r\n> > ]\r\n> > ```\r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > May I ask if the problem is solved now? How did you end up solving it, looking forward to your reply.\r\n> \r\n> I have no idea yet. whole codes seems same in peft and transformers. I guess it could be some weight initialization problem\r\n\r\nRoger that. Thank you for your reply.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored.", "Hi there! Is the issue solved? Let me know if you need more clarification", "@younesbelkada I think we can close this issue by now. Thanks :)", "Thanks @daehuikim !" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27928/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "fakerybakery", "id": 76186054, "node_id": "MDQ6VXNlcjc2MTg2MDU0", "avatar_url": "https://avatars.githubusercontent.com/u/76186054?v=4", "gravatar_id": "", "url": "https://api.github.com/users/fakerybakery", "html_url": "https://github.com/fakerybakery", "followers_url": "https://api.github.com/users/fakerybakery/followers", "following_url": "https://api.github.com/users/fakerybakery/following{/other_user}", "gists_url": "https://api.github.com/users/fakerybakery/gists{/gist_id}", "starred_url": "https://api.github.com/users/fakerybakery/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/fakerybakery/subscriptions", "organizations_url": "https://api.github.com/users/fakerybakery/orgs", "repos_url": "https://api.github.com/users/fakerybakery/repos", "events_url": "https://api.github.com/users/fakerybakery/events{/privacy}", "received_events_url": "https://api.github.com/users/fakerybakery/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey! I don't think this is possible through the TextIteratorStreamer but should be handled within gradio 🤗 ", "Ok, thanks for the suggestion! I'll close this issue then" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27927/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "Cospui", "id": 36847795, "node_id": "MDQ6VXNlcjM2ODQ3Nzk1", "avatar_url": "https://avatars.githubusercontent.com/u/36847795?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Cospui", "html_url": "https://github.com/Cospui", "followers_url": "https://api.github.com/users/Cospui/followers", "following_url": "https://api.github.com/users/Cospui/following{/other_user}", "gists_url": "https://api.github.com/users/Cospui/gists{/gist_id}", "starred_url": "https://api.github.com/users/Cospui/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Cospui/subscriptions", "organizations_url": "https://api.github.com/users/Cospui/orgs", "repos_url": "https://api.github.com/users/Cospui/repos", "events_url": "https://api.github.com/users/Cospui/events{/privacy}", "received_events_url": "https://api.github.com/users/Cospui/received_events", "type": "User", "site_admin": false }

[]

{ "login": "muellerzr", "id": 7831895, "node_id": "MDQ6VXNlcjc4MzE4OTU=", "avatar_url": "https://avatars.githubusercontent.com/u/7831895?v=4", "gravatar_id": "", "url": "https://api.github.com/users/muellerzr", "html_url": "https://github.com/muellerzr", "followers_url": "https://api.github.com/users/muellerzr/followers", "following_url": "https://api.github.com/users/muellerzr/following{/other_user}", "gists_url": "https://api.github.com/users/muellerzr/gists{/gist_id}", "starred_url": "https://api.github.com/users/muellerzr/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/muellerzr/subscriptions", "organizations_url": "https://api.github.com/users/muellerzr/orgs", "repos_url": "https://api.github.com/users/muellerzr/repos", "events_url": "https://api.github.com/users/muellerzr/events{/privacy}", "received_events_url": "https://api.github.com/users/muellerzr/received_events", "type": "User", "site_admin": false }

[ { "login": "muellerzr", "id": 7831895, "node_id": "MDQ6VXNlcjc4MzE4OTU=", "avatar_url": "https://avatars.githubusercontent.com/u/7831895?v=4", "gravatar_id": "", "url": "https://api.github.com/users/muellerzr", "html_url": "https://github.com/muellerzr", "followers_url": "https://api.github.com/users/muellerzr/followers", "following_url": "https://api.github.com/users/muellerzr/following{/other_user}", "gists_url": "https://api.github.com/users/muellerzr/gists{/gist_id}", "starred_url": "https://api.github.com/users/muellerzr/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/muellerzr/subscriptions", "organizations_url": "https://api.github.com/users/muellerzr/orgs", "repos_url": "https://api.github.com/users/muellerzr/repos", "events_url": "https://api.github.com/users/muellerzr/events{/privacy}", "received_events_url": "https://api.github.com/users/muellerzr/received_events", "type": "User", "site_admin": false } ]

[ "I had this same issue, I temporarily fixed it by neutering the different staging directory:\r\n\r\n```python\r\nif os.path.exists(output_dir) and len(os.listdir(output_dir)) > 0:\r\n logger.warning(\r\n f\"Checkpoint destination directory {output_dir} already exists and is non-empty.\"\r\n \"Saving will proceed but saved results may be invalid.\"\r\n )\r\n staging_output_dir = output_dir\r\nelse:\r\n # staging_output_dir = os.path.join(run_dir, f\"tmp-{checkpoint_folder}\")\r\n staging_output_dir = output_dir\r\n```", "> I had this same issue, I temporarily fixed it by neutering the different staging directory:\r\n> \r\n> ```python\r\n> if os.path.exists(output_dir) and len(os.listdir(output_dir)) > 0:\r\n> logger.warning(\r\n> f\"Checkpoint destination directory {output_dir} already exists and is non-empty.\"\r\n> \"Saving will proceed but saved results may be invalid.\"\r\n> )\r\n> staging_output_dir = output_dir\r\n> else:\r\n> # staging_output_dir = os.path.join(run_dir, f\"tmp-{checkpoint_folder}\")\r\n> staging_output_dir = output_dir\r\n> ```\r\n\r\nWhere did you insert this?", "Facing same issue in multi-node training:\r\n`File \"/home/user/.local/lib/python3.8/site-packages/transformers/trainer.py\", line 2353, in _save_checkpoint\r\n self.save_model(staging_output_dir, _internal_call=True)\r\nRuntimeError: Parent directory tmp-checkpoint-200 does\r\nnot exist.`\r\nIt added annoying tmp- in front of the checkpoint", "This is a showstopper for training on multi-GPU nodes. The culprit seems to be the following merged PR #27820.", "There is an open PR #27929, which seems to fix the issue.\r\n@ArthurZucker @sgugger @younesbelkada ", "Hi all, can you please do `pip install git+https://github.com/huggingface/transformers` and rerun your code? This should fix your issue now.\r\n\r\nThank you very much for your patience and flagging this!", "@muellerzr @thundergolfer I still get the same issue of saving checkpoint using the latest version of transformers `4.36` and even with `‘4.37.0.dev0’`\r\n\r\nI used three workers each one has two GPUs, I tried fine-tuning to be saved on a shared storage and a non-shared storage, and for both cases I still got the same error!\r\n\r\n **FileNotFoundError: [Errno 2] No such file or directory: 'model/tmp-checkpoint-49' -> 'model/checkpoint-49'**\r\n\r\n ```\r\nFile \"/opt/conda/lib/python3.10/site-packages/transformers/trainer.py\", line 1537, in train\r\n return inner_training_loop(\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/trainer.py\", line 1929, in _inner_training_loop\r\n self._maybe_log_save_evaluate(tr_loss, model, trial, epoch, ignore_keys_for_eval)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/trainer.py\", line 2279, in _maybe_log_save_evaluate\r\n self._save_checkpoint(model, trial, metrics=metrics)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/trainer.py\", line 2395, in _save_checkpoint\r\n os.rename(staging_output_dir, output_dir)\r\nFileNotFoundError: [Errno 2] No such file or directory: 'model/tmp-checkpoint-49' -> 'model/checkpoint-49'\r\n```\r\n\r\nalthough the `model/checkpoint-49 `is already created!", "@hahmad2008 can you try doing either `pip install transformers -U` or reinstall from git? From the line numbers it's not adding up that you're using a version that includes the fix", "I encountered this issue with the trainer with the following command-line. This was after recently updating transformers with pip install transformers --upgrade\r\n\r\n```--save_strategy epoch --save_total_limit 1```\r\n\r\ntransformers==4.36.2\r\n\r\nEdit: \r\nOne thing to note this was with 2 nodes with 8x A100s per node.\r\nLooking at the code around the error, I have a feeling this was because I may have used local=True when using with main_process_first. Going to try disabling save_on_each_node.\r\n```\r\n if staging_output_dir != output_dir:\r\n with self.args.main_process_first(\r\n desc=\"Renaming model checkpoint folder to true location\", local=self.args.save_on_each_node\r\n ):\r\n if os.path.exists(staging_output_dir):\r\n os.rename(staging_output_dir, output_dir)\r\n```\r\n\r\nedit edit: \r\nLooks like its still not working even when specifying save_on_each_node to false.\r\n\r\nHere is the full command, launched from a slurm sbatch job:\r\n```\r\nsrun --kill-on-bad-exit=1 --jobid $SLURM_JOB_ID bash -c \"accelerate launch --use_deepspeed --zero_stage 1 --deepspeed_hostfile hostfile --deepspeed_multinode_launcher openmpi --gradient_accumulation_steps 1 --num_processes $(( $NUM_GPUS * $COUNT_NODE )) --num_machines $COUNT_NODE --num_cpu_threads_per_process $CPU_COUNT --mixed_precision bf16 --machine_rank \\$SLURM_PROCID --main_process_ip $MASTER_ADDR --main_process_port $MASTER_PORT main.py --source_datasets_filepath source_data/clm --output_dir testing_output_cluster --model_number 2 --overwrite_output_dir --dataloader_num_workers 10 --bf16 --data_fraction 0.1 --save_strategy steps --save_total_limit 1 --save_on_each_node false --dataloader_num_workers 2 --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --max_token_length 1024 --num_train_epochs 1\"\r\n```\r\n\r\n", "I encountered a similar error when using the trainer from DeepSpeed.\r\nThe error occurs at the exact moment after `if os.path.exists(staging_output_dir):` is evaluated and another process finishes renaming.\r\n\r\nI had no other choice, so I resorted to using a try block to get around it.\r\n\r\n```python\r\nif staging_output_dir != output_dir:\r\n with self.args.main_process_first(\r\n desc=\"Renaming model checkpoint folder to true location\", local=self.args.save_on_each_node\r\n ):\r\n if os.path.exists(staging_output_dir):\r\n try:\r\n os.rename(staging_output_dir, output_dir)\r\n except Exception as e:\r\n logger.info(f\"Could not rename checkpoint directory from {staging_output_dir} to {output_dir}. Reason: {e}\")\r\n \r\n```\r\n\r\ntransformers-4.37.0.dev0", "Hi, @snowyday , @tblattner , and @muellerzr . I think `main_process_first` may be broken. \r\n\r\nI run the trainer with 2 nodes X 8 V100 GPUs and deepspeed. When I turned on `log_level=debug`, I found that only one process entered the waiting mode, while all other processes tried to save the checkpoint.\r\n\r\nThe log from process that waited: \r\n\r\n```\r\n[DEBUG|training_args.py:2119] 2023-12-27 15:11:30,917 >> 4: waiting for the main process to perform Renaming model checkpoint folder to true location\r\n```", "I also encounter this with 4.36.2 and HEAD in a multi-node multi-GPU setup. Looks like an obvious race condition, as it happens indeterminately (sometimes 2nd save, sometimes 7th save etc).", "Hi Any update or final conclusion here? :>", "any solutions? facing the same issue on multinode training using deepspeed", "same here, any solutions?", "I've been using a try-except approach for bypassing the issue, and it's been working well for me. However, as xk-huang mentioned, it seems that the root cause is that self.args.main_process_first is not handling multi-node setups properly.", "Curious if there is any reason why we must do ```os.path.exists``` and ```os.rename``` for each process, why not just the main process(es)?\r\n\r\n\r\nHaven't tested this code yet as my compute resources are currently filled and I have a long-running experiment set to finish in a couple days, but wanted to get some thoughts on this potential solution.\r\n```\r\n # Only rename from main process to avoid race condition from other processes especially for distributed filesystems\r\n if staging_output_dir != output_dir:\r\n if self.args.distributed_state.is_local_main_process if self.args.save_on_each_node else self.args.distributed_state.is_main_process:\r\n if os.path.exists(staging_output_dir):\r\n os.rename(staging_output_dir, output_dir)\r\n\r\n self.args.distributed_state.wait_for_everyone()\r\n```", "I'm using transformers's Trainer, is there any work around for this?", "For work around with Trainer, I just subclassed it and replace the _save_checkpoint method that added try exception.\r\n\r\n```\r\nclass CustomTrainer(Trainer):\r\n def _save_checkpoint(self, model, trial, metrics=None):\r\n # In all cases, including ddp/dp/deepspeed, self.model is always a reference to the model we\r\n # want to save except FullyShardedDDP.\r\n # assert unwrap_model(model) is self.model, \"internal model should be a reference to self.model\"\r\n\r\n # Save model checkpoint\r\n checkpoint_folder = f\"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}\"\r\n\r\n if self.hp_search_backend is None and trial is None:\r\n self.store_flos()\r\n\r\n run_dir = self._get_output_dir(trial=trial)\r\n output_dir = os.path.join(run_dir, checkpoint_folder)\r\n if os.path.exists(output_dir) and len(os.listdir(output_dir)) > 0:\r\n logger.warning(\r\n f\"Checkpoint destination directory {output_dir} already exists and is non-empty.\"\r\n \"Saving will proceed but saved results may be invalid.\"\r\n )\r\n staging_output_dir = output_dir\r\n else:\r\n staging_output_dir = os.path.join(\r\n run_dir, f\"tmp-{checkpoint_folder}\")\r\n self.save_model(staging_output_dir, _internal_call=True)\r\n\r\n if not self.args.save_only_model:\r\n # Save optimizer and scheduler\r\n self._save_optimizer_and_scheduler(staging_output_dir)\r\n # Save RNG state\r\n self._save_rng_state(staging_output_dir)\r\n\r\n # Determine the new best metric / best model checkpoint\r\n if metrics is not None and self.args.metric_for_best_model is not None:\r\n metric_to_check = self.args.metric_for_best_model\r\n if not metric_to_check.startswith(\"eval_\"):\r\n metric_to_check = f\"eval_{metric_to_check}\"\r\n metric_value = metrics[metric_to_check]\r\n\r\n operator = np.greater if self.args.greater_is_better else np.less\r\n if (\r\n self.state.best_metric is None\r\n or self.state.best_model_checkpoint is None\r\n or operator(metric_value, self.state.best_metric)\r\n ):\r\n self.state.best_metric = metric_value\r\n self.state.best_model_checkpoint = output_dir\r\n\r\n # Save the Trainer state\r\n if self.args.should_save:\r\n self.state.save_to_json(os.path.join(\r\n staging_output_dir, TRAINER_STATE_NAME))\r\n\r\n if self.args.push_to_hub:\r\n self._push_from_checkpoint(staging_output_dir)\r\n\r\n # Place checkpoint in final location after all saving is finished.\r\n # First wait for everyone to finish writing\r\n self.args.distributed_state.wait_for_everyone()\r\n # Then go through the rewriting process starting on process 0\r\n try:\r\n if staging_output_dir != output_dir:\r\n with self.args.main_process_first(\r\n desc=\"Renaming model checkpoint folder to true location\", local=self.args.save_on_each_node\r\n ):\r\n if os.path.exists(staging_output_dir):\r\n os.rename(staging_output_dir, output_dir)\r\n\r\n # Maybe delete some older checkpoints.\r\n if self.args.should_save:\r\n self._rotate_checkpoints(use_mtime=True, output_dir=run_dir)\r\n except Exception:\r\n print(\"Error rotating checkpoints skipping\")\r\n pass\r\n```", "I've checked the `main_process_first` using the code snippet below:\r\nNumber of nodes: 3\r\nProcesses per node (GPUs): 4\r\nTotal: 12 processes\r\n\r\n```python\r\nimport logging\r\n\r\nimport deepspeed\r\nimport transformers\r\nimport torch\r\n\r\n\r\nlogging.basicConfig(level=logging.INFO)\r\nlogger = logging.getLogger()\r\n\r\nif __name__ == \"__main__\":\r\n deepspeed.init_distributed()\r\n node_rank = torch.distributed.get_rank() \r\n training_args = transformers.TrainingArguments(per_device_train_batch_size=8,\r\n gradient_accumulation_steps=2,\r\n num_train_epochs=3,\r\n deepspeed=\"ds_config/ds_config_zero3.json\",\r\n output_dir=\"logs\")\r\n\r\n with training_args.main_process_first():\r\n logger.info(f\"Check `main_process_first`. Node rank {node_rank}\")\r\n```\r\n\r\n```\r\nAddress family not supported by protocol).\r\n[INFO:root:Check `main_process_first`. Node rank 8\r\nINFO:root:Check `main_process_first`. Node rank 0\r\nINFO:root:Check `main_process_first`. Node rank 4\r\nINFO:root:Check `main_process_first`. Node rank 6\r\nINFO:root:Check `main_process_first`. Node rank 10\r\nINFO:root:Check `main_process_first`. Node rank 5\r\nINFO:root:Check `main_process_first`. Node rank 9\r\nINFO:root:Check `main_process_first`. Node rank 1\r\nINFO:root:Check `main_process_first`. Node rank 2\r\nINFO:root:Check `main_process_first`. Node rank 3\r\nINFO:root:Check `main_process_first`. Node rank 7\r\nINFO:root:Check `main_process_first`. Node rank 11\r\n```\r\n\r\nThe node rankings appear to be correctly allocated, with Node rank 0 going to node 1, Node rank 4 to node 2, and Node rank 8 to node 3; however, there are inaccuracies with the global rankings. In the context of a shared filesystem, if we proceed without waiting for the result from global rank 0, it could cause conflicts during the os.rename operation.\r\n\r\n```python\r\nif staging_output_dir != output_dir:\r\n with self.args.main_process_first(\r\n desc=\"Renaming model checkpoint folder to true location\", local=self.args.save_on_each_node\r\n ):\r\n if os.path.exists(staging_output_dir):\r\n os.rename(staging_output_dir, output_dir)\r\n \r\n ```", "> however, there are inaccuracies with the global rankings.\r\n\r\n@snowyday as indicated by the fact that `rank 8` is printed first?", "@thundergolfer \r\n`Rank 0` should pop up first, and the others should hang tight until the renaming wraps up.\r\nI should set `args.save_on_each_node=False`:\r\n\r\n```python\r\nwith self.args.main_process_first(\r\n desc=\"Renaming model checkpoint folder to true location\", local=self.args.save_on_each_node\r\n ):\r\n```", "Without having tested, this looks like the right direction.", "In the end, simply setting `save_on_each_node=False` worked out for everything.\r\n\r\n```python\r\ntraining_args = transformers.TrainingArguments(..., save_on_each_node=False, ...)\r\n```\r\n\r\nBy setting `save_on_each_node=False` in `TrainingArguments`, it ensures that in the `Trainer`’s `def _save_checkpoint method`, `main_process_first`'s `local` will be set to `False`. Consequently, following the explanation provided, it works correctly.\r\n\r\n_**if `False` first means process of rank 0 of node rank 0 In multi-node environment with a shared filesystem you most likely will want to use `local=False` so that only the main process of the first node will do the processing.**_\r\n\r\n```python\r\n @contextlib.contextmanager\r\n def main_process_first(self, local=True, desc=\"work\"):\r\n \"\"\"\r\n A context manager for torch distributed environment where on needs to do something on the main process, while\r\n blocking replicas, and when it's finished releasing the replicas.\r\n\r\n One such use is for `datasets`'s `map` feature which to be efficient should be run once on the main process,\r\n which upon completion saves a cached version of results and which then automatically gets loaded by the\r\n replicas.\r\n\r\n Args:\r\n local (`bool`, *optional*, defaults to `True`):\r\n if `True` first means process of rank 0 of each node if `False` first means process of rank 0 of node\r\n rank 0 In multi-node environment with a shared filesystem you most likely will want to use\r\n `local=False` so that only the main process of the first node will do the processing. If however, the\r\n filesystem is not shared, then the main process of each node will need to do the processing, which is\r\n the default behavior.\r\n desc (`str`, *optional*, defaults to `\"work\"`):\r\n a work description to be used in debug logs\r\n\r\n \"\"\"\r\n```", "Would this work in a setting without a shared file system?", "I'v checked it on a GPU cluster with a shared file system. For multi-node setups with independent file systems, the default `save_on_each_node=True` is fine; `main_process_first` make sure to serialize the execution for each node. If that still doesn't work, then I think there might still be an issue with `main_process_first`.", "I don't think there is an issue with main_process_first as I've been using it across a lot of dataset processing steps. \r\n\r\nI believe that on network/shared file systems os.rename is not atomic. So its possible that the file system in this case might not be reflected after os.rename returns, causing other processes to observe the wrong state. I haven't found a good way to ensure the rename is completed. Catching the exception would handle it though, but not my ideal way to deal with the race condition.", "In the case of processes sharing a filesystem, it seems prudent for only one process to wait for a rename operation to complete. However, why `main_process_first` is being used? On a shared filesystem, if the `rename()` fails, options are limited. Is this why multiple processes are making repeated attempts?", "I'm not sure if it fails or not. From what I understand, the network attached storage node might not actually complete the operation before the next process comes to check if the path exists. It will complete, just not in the timeframe allowed (sometimes). But that outlines the core issue here.\r\n\r\nMy suggestion is to use something like this:\r\n```if self.args.distributed_state.is_local_main_process if self.args.save_on_each_node else self.args.distributed_state.is_main_process:```\r\n\r\nThen ```self.args.distributed_state.wait_for_everyone()``` to synchronize everyone afterwards. \r\n\r\nThis would only use the main process if save_on_each_node is false, otherwise only the local main processes. Which I think is the intended behavior. The part I'm not sure of is if the renamed file is used later downstream, then that could introduce a race condition there...\r\n\r\nIt would be nice if we could have an fsync for the shared filesystem to ensure the rename actually completed.\r\n", "It is, so we could have a race condition. An `fsync` could be done certainly and your logic makes sense. @tblattner would you like to open a PR on this by chance? " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27925/reactions", "total_count": 1, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 1 }

{ "login": "staghado", "id": 84044788, "node_id": "MDQ6VXNlcjg0MDQ0Nzg4", "avatar_url": "https://avatars.githubusercontent.com/u/84044788?v=4", "gravatar_id": "", "url": "https://api.github.com/users/staghado", "html_url": "https://github.com/staghado", "followers_url": "https://api.github.com/users/staghado/followers", "following_url": "https://api.github.com/users/staghado/following{/other_user}", "gists_url": "https://api.github.com/users/staghado/gists{/gist_id}", "starred_url": "https://api.github.com/users/staghado/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/staghado/subscriptions", "organizations_url": "https://api.github.com/users/staghado/orgs", "repos_url": "https://api.github.com/users/staghado/repos", "events_url": "https://api.github.com/users/staghado/events{/privacy}", "received_events_url": "https://api.github.com/users/staghado/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey @staghado, thanks for taking care of this, let us know when it's ready to be reviewed!", "I have conducted some tests on an A10 GPU : \r\n - The code seems to work without errors when `_supports_flash_attn_2` is set to `True` for `MusicgenForConditionalGeneration` but does not load the model with FA2 if not specified by hand. Maybe it needs to be added at the class level in MusicgenForConditionalGeneration?\r\n - There is no difference in generation speed between eager attention and FA2 : \r\n \r\n\r\n\r\n", "cc @ylacombe could you possibly circle back here when you get the chance!", "Hi @ylacombe,\r\n\r\nI confirm that the model was instantiated as described [here](https://huggingface.co./docs/transformers/v4.36.1/en/perf_infer_gpu_one#flashattention-2) with the exception of `torch_dtype=torch.float16` instead of `torch_dtype=torch.bfloat16` because some operations did not seem to implement bfloat16.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27924/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27924", "html_url": "https://github.com/huggingface/transformers/pull/27924", "diff_url": "https://github.com/huggingface/transformers/pull/27924.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27924.patch", "merged_at": null }

{ "login": "odellus", "id": 4686956, "node_id": "MDQ6VXNlcjQ2ODY5NTY=", "avatar_url": "https://avatars.githubusercontent.com/u/4686956?v=4", "gravatar_id": "", "url": "https://api.github.com/users/odellus", "html_url": "https://github.com/odellus", "followers_url": "https://api.github.com/users/odellus/followers", "following_url": "https://api.github.com/users/odellus/following{/other_user}", "gists_url": "https://api.github.com/users/odellus/gists{/gist_id}", "starred_url": "https://api.github.com/users/odellus/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/odellus/subscriptions", "organizations_url": "https://api.github.com/users/odellus/orgs", "repos_url": "https://api.github.com/users/odellus/repos", "events_url": "https://api.github.com/users/odellus/events{/privacy}", "received_events_url": "https://api.github.com/users/odellus/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @Rocketknight1 if you have time to look at this! ", "I'm using `transformers==4.34.0`", "Upgrading to 4.35.2 fixed this issue for me. Closing." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27926/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "KyrieCui", "id": 37808472, "node_id": "MDQ6VXNlcjM3ODA4NDcy", "avatar_url": "https://avatars.githubusercontent.com/u/37808472?v=4", "gravatar_id": "", "url": "https://api.github.com/users/KyrieCui", "html_url": "https://github.com/KyrieCui", "followers_url": "https://api.github.com/users/KyrieCui/followers", "following_url": "https://api.github.com/users/KyrieCui/following{/other_user}", "gists_url": "https://api.github.com/users/KyrieCui/gists{/gist_id}", "starred_url": "https://api.github.com/users/KyrieCui/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/KyrieCui/subscriptions", "organizations_url": "https://api.github.com/users/KyrieCui/orgs", "repos_url": "https://api.github.com/users/KyrieCui/repos", "events_url": "https://api.github.com/users/KyrieCui/events{/privacy}", "received_events_url": "https://api.github.com/users/KyrieCui/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey, could you try to use the latest release of safetensors? \r\nOtherwise not sure I can help you with a checkpoint that is private ! ", "Hello, Sir. \r\nI met this error when loading the 6th checkpoint, so I re-downloaded the sixth safetensors file. But still the same error. Should I try to re- download all of the safetensors files?", "If you download them from the hub anyway, you should just use from_pretrained, they will automatically be cache in your .cache and re-used accordingly. ", "@KyrieCui this might be a storage issue as well, can you make sure you have enough Disk space on your device?", "thanks for all of your support! I fixed this issue by re-downloading all of the safetensors files. " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27923/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "ONE-THING-9", "id": 123763769, "node_id": "U_kgDOB2B8OQ", "avatar_url": "https://avatars.githubusercontent.com/u/123763769?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ONE-THING-9", "html_url": "https://github.com/ONE-THING-9", "followers_url": "https://api.github.com/users/ONE-THING-9/followers", "following_url": "https://api.github.com/users/ONE-THING-9/following{/other_user}", "gists_url": "https://api.github.com/users/ONE-THING-9/gists{/gist_id}", "starred_url": "https://api.github.com/users/ONE-THING-9/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ONE-THING-9/subscriptions", "organizations_url": "https://api.github.com/users/ONE-THING-9/orgs", "repos_url": "https://api.github.com/users/ONE-THING-9/repos", "events_url": "https://api.github.com/users/ONE-THING-9/events{/privacy}", "received_events_url": "https://api.github.com/users/ONE-THING-9/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey, I think it already supports this! see [here](https://huggingface.co./docs/transformers/chat_templating) (you just need to customize the roles) ", "Got it, i missed that.", "There is a big problem here. Some of these models reject the system prompt in their chat template. It's wildly inconsistent and there doesn't seem to be a way to tell if a model accepts a system prompt or not. You end up getting this error if you pass in a system prompt and format it using the chat template function:\r\n\r\njinja2.exceptions.TemplateError: Conversation roles must alternate user/assistant/user/assistant/\r\n\r\nI'm really surprised I haven't seen anybody raise this. Am I missing something? How do you tell if a model takes a system prompt or not? ", "Thanks for the feedback! cc @Rocketknight1 if that's something you noticed", "Hi @unoriginalscreenname - unfortunately, this is an unavoidable consequence of how these models were trained. Some models were trained with system prompts as part of the training data, and other models were not. When a model was not trained with a system prompt, it will not have any tokens that it can use to represent a system prompt, and trying to insert a prompt will confuse the model and probably significantly reduce the output quality.\r\n\r\nIn the cases when a model was trained without a system prompt, the model's chat template can be configured to raise an error if a `system` message is included in the input, and this is indeed what happens with some models (e.g. LLaMA/Mistral/Mixtral). This is correct and intended behaviour, and there isn't really any way to \"fix\" it without retraining the models!\r\n\r\nThe only solution I can suggest is that there is usually a different fine-tune of most models that supports a system prompt. For example, instead of Mistral-instruct you can use [Zephyr-beta](https://huggingface.co./HuggingFaceH4/zephyr-7b-beta), and instead of Mixtral-instruct you can use [Dolphin](https://huggingface.co./cognitivecomputations/dolphin-2.7-mixtral-8x7b). Both of these models were trained with system prompts, and will understand them correctly (and apply them in their chat template)." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27922/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "gau-nernst", "id": 26946864, "node_id": "MDQ6VXNlcjI2OTQ2ODY0", "avatar_url": "https://avatars.githubusercontent.com/u/26946864?v=4", "gravatar_id": "", "url": "https://api.github.com/users/gau-nernst", "html_url": "https://github.com/gau-nernst", "followers_url": "https://api.github.com/users/gau-nernst/followers", "following_url": "https://api.github.com/users/gau-nernst/following{/other_user}", "gists_url": "https://api.github.com/users/gau-nernst/gists{/gist_id}", "starred_url": "https://api.github.com/users/gau-nernst/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/gau-nernst/subscriptions", "organizations_url": "https://api.github.com/users/gau-nernst/orgs", "repos_url": "https://api.github.com/users/gau-nernst/repos", "events_url": "https://api.github.com/users/gau-nernst/events{/privacy}", "received_events_url": "https://api.github.com/users/gau-nernst/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The tests are failing because when I copy LayoutLMv2Tokenizer to LayoutLMTokenizer, `bbox` input is required. To preserve backward-compatibility, we need to check if `bbox` is None, and apply the old logic in that case. I'm not familiar with the tokenizer code, so help would be appreciated 🙏.", "(feel free to ping me again for another review 😉 )", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27921/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27921", "html_url": "https://github.com/huggingface/transformers/pull/27921", "diff_url": "https://github.com/huggingface/transformers/pull/27921.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27921.patch", "merged_at": null }

{ "login": "asusevski", "id": 77211520, "node_id": "MDQ6VXNlcjc3MjExNTIw", "avatar_url": "https://avatars.githubusercontent.com/u/77211520?v=4", "gravatar_id": "", "url": "https://api.github.com/users/asusevski", "html_url": "https://github.com/asusevski", "followers_url": "https://api.github.com/users/asusevski/followers", "following_url": "https://api.github.com/users/asusevski/following{/other_user}", "gists_url": "https://api.github.com/users/asusevski/gists{/gist_id}", "starred_url": "https://api.github.com/users/asusevski/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/asusevski/subscriptions", "organizations_url": "https://api.github.com/users/asusevski/orgs", "repos_url": "https://api.github.com/users/asusevski/repos", "events_url": "https://api.github.com/users/asusevski/events{/privacy}", "received_events_url": "https://api.github.com/users/asusevski/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27920/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27920", "html_url": "https://github.com/huggingface/transformers/pull/27920", "diff_url": "https://github.com/huggingface/transformers/pull/27920.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27920.patch", "merged_at": 1702338263000 }

{ "login": "asusevski", "id": 77211520, "node_id": "MDQ6VXNlcjc3MjExNTIw", "avatar_url": "https://avatars.githubusercontent.com/u/77211520?v=4", "gravatar_id": "", "url": "https://api.github.com/users/asusevski", "html_url": "https://github.com/asusevski", "followers_url": "https://api.github.com/users/asusevski/followers", "following_url": "https://api.github.com/users/asusevski/following{/other_user}", "gists_url": "https://api.github.com/users/asusevski/gists{/gist_id}", "starred_url": "https://api.github.com/users/asusevski/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/asusevski/subscriptions", "organizations_url": "https://api.github.com/users/asusevski/orgs", "repos_url": "https://api.github.com/users/asusevski/repos", "events_url": "https://api.github.com/users/asusevski/events{/privacy}", "received_events_url": "https://api.github.com/users/asusevski/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27919/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "f4hy", "id": 36440, "node_id": "MDQ6VXNlcjM2NDQw", "avatar_url": "https://avatars.githubusercontent.com/u/36440?v=4", "gravatar_id": "", "url": "https://api.github.com/users/f4hy", "html_url": "https://github.com/f4hy", "followers_url": "https://api.github.com/users/f4hy/followers", "following_url": "https://api.github.com/users/f4hy/following{/other_user}", "gists_url": "https://api.github.com/users/f4hy/gists{/gist_id}", "starred_url": "https://api.github.com/users/f4hy/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/f4hy/subscriptions", "organizations_url": "https://api.github.com/users/f4hy/orgs", "repos_url": "https://api.github.com/users/f4hy/repos", "events_url": "https://api.github.com/users/f4hy/events{/privacy}", "received_events_url": "https://api.github.com/users/f4hy/received_events", "type": "User", "site_admin": false }

[]

[]

[ "thanks" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27918/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27918", "html_url": "https://github.com/huggingface/transformers/pull/27918", "diff_url": "https://github.com/huggingface/transformers/pull/27918.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27918.patch", "merged_at": 1702119564000 }

{ "login": "py4", "id": 747819, "node_id": "MDQ6VXNlcjc0NzgxOQ==", "avatar_url": "https://avatars.githubusercontent.com/u/747819?v=4", "gravatar_id": "", "url": "https://api.github.com/users/py4", "html_url": "https://github.com/py4", "followers_url": "https://api.github.com/users/py4/followers", "following_url": "https://api.github.com/users/py4/following{/other_user}", "gists_url": "https://api.github.com/users/py4/gists{/gist_id}", "starred_url": "https://api.github.com/users/py4/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/py4/subscriptions", "organizations_url": "https://api.github.com/users/py4/orgs", "repos_url": "https://api.github.com/users/py4/repos", "events_url": "https://api.github.com/users/py4/events{/privacy}", "received_events_url": "https://api.github.com/users/py4/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @younesbelkada I think this worked / works on main no? \r\n\r\nFYI @gante and @tomaarsen we'll work on a fix with @younesbelkada " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27917/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "Pulsar110", "id": 125087940, "node_id": "U_kgDOB3SwxA", "avatar_url": "https://avatars.githubusercontent.com/u/125087940?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Pulsar110", "html_url": "https://github.com/Pulsar110", "followers_url": "https://api.github.com/users/Pulsar110/followers", "following_url": "https://api.github.com/users/Pulsar110/following{/other_user}", "gists_url": "https://api.github.com/users/Pulsar110/gists{/gist_id}", "starred_url": "https://api.github.com/users/Pulsar110/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Pulsar110/subscriptions", "organizations_url": "https://api.github.com/users/Pulsar110/orgs", "repos_url": "https://api.github.com/users/Pulsar110/repos", "events_url": "https://api.github.com/users/Pulsar110/events{/privacy}", "received_events_url": "https://api.github.com/users/Pulsar110/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey 🤗 thanks for opening an issue! We try to keep the github issues for bugs/feature requests.\r\nCould you ask your question on the [forum](https://discuss.huggingface.co/) instead? I'm sure the community will be of help!\r\n\r\nThanks!", "Thank you. I have created a post here: https://discuss.huggingface.co/t/question-about-the-output-of-the-decision-transformer/65384\r\nSo far no one has commented there. I'm not sure if there is a bug in the code, or maybe I do not understand it correctly, that's also why I wanted to post here. ", "I don't know this model at all so pinging @edbeeching the author of the PR! ", "Hi @Pulsar110 , thanks for your question. It would probably be best to reach out to the authors with this question as our implementation aims to match the author's codebase: https://github.com/kzl/decision-transformer/blob/e2d82e68f330c00f763507b3b01d774740bee53f/gym/decision_transformer/models/decision_transformer.py#L97\r\n\r\nIf I were to hazard a guess I would think that there is a mistake in their implementation and we should be indexing entry 0 at some point. \r\n\r\nLet us know what they say and perhaps we can update our implementation with any changes they suggest. I will close the issue for now but feel free to reopen it with more questions or if you hear back from them.\r\n" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27916/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "AlpinDale", "id": 52078762, "node_id": "MDQ6VXNlcjUyMDc4NzYy", "avatar_url": "https://avatars.githubusercontent.com/u/52078762?v=4", "gravatar_id": "", "url": "https://api.github.com/users/AlpinDale", "html_url": "https://github.com/AlpinDale", "followers_url": "https://api.github.com/users/AlpinDale/followers", "following_url": "https://api.github.com/users/AlpinDale/following{/other_user}", "gists_url": "https://api.github.com/users/AlpinDale/gists{/gist_id}", "starred_url": "https://api.github.com/users/AlpinDale/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/AlpinDale/subscriptions", "organizations_url": "https://api.github.com/users/AlpinDale/orgs", "repos_url": "https://api.github.com/users/AlpinDale/repos", "events_url": "https://api.github.com/users/AlpinDale/events{/privacy}", "received_events_url": "https://api.github.com/users/AlpinDale/received_events", "type": "User", "site_admin": false }

[ { "id": 1843244711, "node_id": "MDU6TGFiZWwxODQzMjQ0NzEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/New%20model", "name": "New model", "color": "fbca04", "default": false, "description": "" } ]

[]

[ "I've been using a quick drop-in replacement lifted from the dmoe.py implementation from megablocks.\r\n\r\nRunning `torch.cat` for the expert weights on each forward pass adds a ~5% overhead, since I didn't want to deal with managing the state dicts. Overall training is 2-3x faster.\r\n\r\n```python\r\nclass MixtralSparseMoeBlock(torch.nn.Module):\r\n def __init__(self, config: MixtralConfig):\r\n super(MixtralSparseMoeBlock, self).__init__()\r\n\r\n self.config = config\r\n\r\n self.hidden_dim = config.hidden_size\r\n self.ffn_dim = config.intermediate_size\r\n self.num_experts = config.num_local_experts\r\n self.top_k = config.num_experts_per_tok\r\n\r\n self.gate = nn.Linear(config.hidden_size, config.num_local_experts, bias=False)\r\n self.experts = nn.ModuleList(\r\n [MixtralBLockSparseTop2MLP(config) for _ in range(self.num_experts)]\r\n )\r\n\r\n self.sort_end_bit = max(int(np.ceil(np.log2(self.num_experts))), 1)\r\n self.blocking = 128\r\n self.quantize_scatter_num_bits = -1\r\n max_column_index = (self.ffn_dim * self.num_experts) // self.blocking\r\n self.transpose_sort_end_bit = max(int(np.ceil(np.log2(max_column_index))), 1)\r\n\r\n # From https://github.com/stanford-futuredata/megablocks/blob/7c25169ce87c32c31e8845ef34785d3095b1a2cb/megablocks/layers/dmoe.py#L31\r\n def sparse_transpose(self, size, row_indices, column_indices):\r\n block_columns = size[1] // self.blocking\r\n\r\n # Sort row indices by column indices to get the transposed matrix's\r\n # column indices.\r\n #\r\n # NOTE: Our sort operation uses the same width indices as the input values.\r\n # To avoid overflow when we have large activation matrices we cast to\r\n # 32-bit before sorting.\r\n _, gather_indices = ops.sort(column_indices.int(), self.transpose_sort_end_bit)\r\n\r\n # There are a constant number of blocks in every row of the sparse matrix.\r\n # A blocks offset is:\r\n #\r\n # row_index * blocks_per_row + column_index % blocks_per_row\r\n #\r\n # Once we have the block offsets ordered for transposition we can divide\r\n # by blocks_per_row to get the transposed column indices.\r\n column_indices_t = row_indices.gather(0, gather_indices.long())\r\n block_offsets_t = gather_indices.int()\r\n\r\n zero = torch.zeros((1,), dtype=torch.int32, device=row_indices.device)\r\n nnz_per_column = ops.histogram(column_indices, block_columns)\r\n nnz_per_column = ops.inclusive_cumsum(nnz_per_column, 0)\r\n offsets_t = torch.cat([zero, nnz_per_column])\r\n return column_indices_t, offsets_t, block_offsets_t\r\n\r\n # From https://github.com/stanford-futuredata/megablocks/blob/7c25169ce87c32c31e8845ef34785d3095b1a2cb/megablocks/layers/dmoe.py#L59\r\n def topology(self, x: torch.Tensor, padded_bins: torch.Tensor):\r\n padded_tokens, _ = x.size()\r\n assert padded_tokens % self.blocking == 0\r\n assert self.ffn_dim % self.blocking == 0\r\n\r\n # Offsets for the sparse matrix. All rows have the\r\n # same number of nonzero blocks dictated by the\r\n # dimensionality of a single expert.\r\n block_rows = padded_tokens // self.blocking\r\n blocks_per_row = self.ffn_dim // self.blocking\r\n offsets = torch.arange(\r\n 0,\r\n block_rows * blocks_per_row + 1,\r\n blocks_per_row,\r\n dtype=torch.int32,\r\n device=x.device,\r\n )\r\n\r\n # Indices for the sparse matrix. The indices for\r\n # the intermediate matrix are dynamic depending\r\n # on the mapping of tokens to experts.\r\n column_indices = ops.topology(\r\n padded_bins, self.blocking, block_rows, blocks_per_row\r\n )\r\n\r\n # TODO(tgale): This is unused. Remove the need for this in stk.\r\n # For now, use meta init to save the device memory.\r\n data = torch.empty(\r\n column_indices.numel(),\r\n self.blocking,\r\n self.blocking,\r\n dtype=x.dtype,\r\n device=\"meta\",\r\n )\r\n shape = (padded_tokens, self.ffn_dim * self.num_experts)\r\n row_indices = stk.ops.row_indices(shape, data, offsets, column_indices)\r\n column_indices_t, offsets_t, block_offsets_t = self.sparse_transpose(\r\n shape, row_indices, column_indices\r\n )\r\n return stk.Matrix(\r\n shape,\r\n data,\r\n row_indices,\r\n column_indices,\r\n offsets,\r\n column_indices_t,\r\n offsets_t,\r\n block_offsets_t,\r\n )\r\n\r\n # From https://github.com/stanford-futuredata/megablocks/blob/7c25169ce87c32c31e8845ef34785d3095b1a2cb/megablocks/layers/dmoe.py#L103\r\n def indices_and_padded_bins(self, top_experts: torch.Tensor):\r\n # Sort the expert ids to produce the scatter/gather\r\n # indices for the permutation.\r\n top_experts = top_experts.int()\r\n bin_ids, indices = ops.sort(top_experts, self.sort_end_bit)\r\n\r\n # Histogram the expert ids to identify the number of\r\n # tokens routed to each expert.\r\n tokens_per_expert = ops.histogram(top_experts, self.num_experts)\r\n\r\n # Round the token counts up to the block size used in\r\n # the matrix muliplications. Caculate the starting\r\n # position of each bin.\r\n padded_tokens_per_expert = ops.round_up(tokens_per_expert, self.blocking)\r\n padded_bins = ops.inclusive_cumsum(padded_tokens_per_expert, 0)\r\n padded_bins = promote_scalar(padded_bins)\r\n\r\n # Calculate the bin bounds for the sorted tokens.\r\n bins = ops.inclusive_cumsum(tokens_per_expert, 0)\r\n bins = promote_scalar(bins)\r\n return indices, bin_ids, bins, padded_bins, tokens_per_expert\r\n\r\n # From https://github.com/stanford-futuredata/megablocks/blob/7c25169ce87c32c31e8845ef34785d3095b1a2cb/megablocks/layers/dmoe.py#L126\r\n def sparse_forward(\r\n self,\r\n hidden_states: torch.Tensor,\r\n expert_weights: torch.Tensor,\r\n top_experts: torch.Tensor,\r\n ):\r\n # x: [sl, bs, hs]\r\n # expert_weights: [sl * bs, top-k]\r\n # top_experts: [sl * bs, top-k]\r\n expert_weights = expert_weights.flatten().to(hidden_states.dtype)\r\n top_experts = top_experts.flatten()\r\n\r\n with torch.no_grad():\r\n (\r\n indices,\r\n bin_ids,\r\n bins,\r\n padded_bins,\r\n _,\r\n ) = self.indices_and_padded_bins(top_experts)\r\n\r\n # Permute tokens and pad to prepare expert computation\r\n # (top_k * sequence_length padding, model_dim)\r\n # Route the tokens for MoE computation.\r\n hidden_states = ops.padded_gather(\r\n hidden_states, indices, bin_ids, bins, padded_bins, self.top_k\r\n )\r\n\r\n # Create the sparse matrix topology\r\n with torch.no_grad():\r\n topo = self.topology(hidden_states, padded_bins)\r\n\r\n w1 = torch.cat([expert.w1.weight.T for expert in self.experts], dim=1)\r\n w2 = torch.cat([expert.w2.weight for expert in self.experts], dim=1).T\r\n w3 = torch.cat([expert.w3.weight.T for expert in self.experts], dim=1)\r\n\r\n # Perform the expert computation\r\n hidden_states = stk.Matrix( # type: ignore\r\n topo.size(),\r\n F.silu(stk.ops.sdd(hidden_states, w1, topo).data)\r\n * stk.ops.sdd(hidden_states, w3, topo).data,\r\n topo.row_indices,\r\n topo.column_indices,\r\n topo.offsets,\r\n topo.column_indices_t,\r\n topo.offsets_t,\r\n topo.block_offsets_t,\r\n )\r\n hidden_states = stk.ops.dsd(hidden_states, w2)\r\n\r\n # Permute back and remove padding\r\n # (top_k * sequence_length, model_dim)\r\n hidden_states: torch.Tensor = ops.padded_scatter( # type: ignore\r\n hidden_states,\r\n indices,\r\n bin_ids,\r\n expert_weights,\r\n bins,\r\n padded_bins,\r\n self.top_k,\r\n self.quantize_scatter_num_bits,\r\n )\r\n return hidden_states\r\n\r\n def forward(self, hidden_states: torch.Tensor):\r\n orig_shape = hidden_states.shape\r\n batch_size, sequence_length, hidden_dim = orig_shape\r\n\r\n hidden_states = hidden_states.view(-1, hidden_dim)\r\n\r\n router_logits = self.gate(hidden_states)\r\n\r\n routing_weights = router_logits.softmax(dim=-1).to(hidden_states.dtype)\r\n routing_weights, expert_indices = torch.topk(\r\n routing_weights, self.top_k, dim=-1\r\n )\r\n routing_weights /= routing_weights.sum(dim=-1, keepdim=True)\r\n\r\n hidden_states = self.sparse_forward(\r\n hidden_states, routing_weights, expert_indices\r\n )\r\n\r\n return hidden_states.view(*orig_shape), router_logits\r\n```", "@kevinhu what do you think would be the best way to modify the state dict to avoid the `.cat`s? Merging the individual w1, w2, and w3 tensors that are currently in a list of `MixtralBlockSparseTop2MLP` into `w1, w3=nn.Linear(self.hidden_dim, self.ffn_dim * self.num_experts)` and `w2=nn.Linear(self.ffn_dim * self.num_experts, self.hidden_dim)` under the `MixtralSparseMoeBlock` and then using `.view`s of it for each expert? Since you're effectively just using the `MixtralBLockSparseTop2MLP` class as a dataclass for storing the expert weights and not actually using its `forward()` method." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27915/reactions", "total_count": 1, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "nicholasneo78", "id": 45549785, "node_id": "MDQ6VXNlcjQ1NTQ5Nzg1", "avatar_url": "https://avatars.githubusercontent.com/u/45549785?v=4", "gravatar_id": "", "url": "https://api.github.com/users/nicholasneo78", "html_url": "https://github.com/nicholasneo78", "followers_url": "https://api.github.com/users/nicholasneo78/followers", "following_url": "https://api.github.com/users/nicholasneo78/following{/other_user}", "gists_url": "https://api.github.com/users/nicholasneo78/gists{/gist_id}", "starred_url": "https://api.github.com/users/nicholasneo78/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/nicholasneo78/subscriptions", "organizations_url": "https://api.github.com/users/nicholasneo78/orgs", "repos_url": "https://api.github.com/users/nicholasneo78/repos", "events_url": "https://api.github.com/users/nicholasneo78/events{/privacy}", "received_events_url": "https://api.github.com/users/nicholasneo78/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Awesome @nicholasneo78!", "cc @ylacombe ", "Hi @ylacombe,\r\n\r\nYes this works for me. Have updated the code `feature_extraction_seamless_m4t.py` and added test in `test_feature_extraction_seamless_m4t.py`\r\n\r\nThanks for the suggestion!", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27914). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27914/reactions", "total_count": 1, "+1": 0, "-1": 0, "laugh": 0, "hooray": 1, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27914", "html_url": "https://github.com/huggingface/transformers/pull/27914", "diff_url": "https://github.com/huggingface/transformers/pull/27914.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27914.patch", "merged_at": 1703242051000 }

{ "login": "khyatikhandelwal", "id": 65815098, "node_id": "MDQ6VXNlcjY1ODE1MDk4", "avatar_url": "https://avatars.githubusercontent.com/u/65815098?v=4", "gravatar_id": "", "url": "https://api.github.com/users/khyatikhandelwal", "html_url": "https://github.com/khyatikhandelwal", "followers_url": "https://api.github.com/users/khyatikhandelwal/followers", "following_url": "https://api.github.com/users/khyatikhandelwal/following{/other_user}", "gists_url": "https://api.github.com/users/khyatikhandelwal/gists{/gist_id}", "starred_url": "https://api.github.com/users/khyatikhandelwal/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/khyatikhandelwal/subscriptions", "organizations_url": "https://api.github.com/users/khyatikhandelwal/orgs", "repos_url": "https://api.github.com/users/khyatikhandelwal/repos", "events_url": "https://api.github.com/users/khyatikhandelwal/events{/privacy}", "received_events_url": "https://api.github.com/users/khyatikhandelwal/received_events", "type": "User", "site_admin": false }

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27913/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27913", "html_url": "https://github.com/huggingface/transformers/pull/27913", "diff_url": "https://github.com/huggingface/transformers/pull/27913.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27913.patch", "merged_at": null }

{ "login": "ydshieh", "id": 2521628, "node_id": "MDQ6VXNlcjI1MjE2Mjg=", "avatar_url": "https://avatars.githubusercontent.com/u/2521628?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ydshieh", "html_url": "https://github.com/ydshieh", "followers_url": "https://api.github.com/users/ydshieh/followers", "following_url": "https://api.github.com/users/ydshieh/following{/other_user}", "gists_url": "https://api.github.com/users/ydshieh/gists{/gist_id}", "starred_url": "https://api.github.com/users/ydshieh/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ydshieh/subscriptions", "organizations_url": "https://api.github.com/users/ydshieh/orgs", "repos_url": "https://api.github.com/users/ydshieh/repos", "events_url": "https://api.github.com/users/ydshieh/events{/privacy}", "received_events_url": "https://api.github.com/users/ydshieh/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27912/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27912", "html_url": "https://github.com/huggingface/transformers/pull/27912", "diff_url": "https://github.com/huggingface/transformers/pull/27912.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27912.patch", "merged_at": 1702282658000 }

{ "login": "ylacombe", "id": 52246514, "node_id": "MDQ6VXNlcjUyMjQ2NTE0", "avatar_url": "https://avatars.githubusercontent.com/u/52246514?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ylacombe", "html_url": "https://github.com/ylacombe", "followers_url": "https://api.github.com/users/ylacombe/followers", "following_url": "https://api.github.com/users/ylacombe/following{/other_user}", "gists_url": "https://api.github.com/users/ylacombe/gists{/gist_id}", "starred_url": "https://api.github.com/users/ylacombe/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ylacombe/subscriptions", "organizations_url": "https://api.github.com/users/ylacombe/orgs", "repos_url": "https://api.github.com/users/ylacombe/repos", "events_url": "https://api.github.com/users/ylacombe/events{/privacy}", "received_events_url": "https://api.github.com/users/ylacombe/received_events", "type": "User", "site_admin": false }

[]

[]

[ "OC, I've already verified that the tests pass (on a 24GB TITAN RTX)! I've also monitored by hand GPU utilization (around 10GB) so I believe it should pass on a 16GB GPU.\r\n\r\nFYI, the tests that I've changed verify that every task-specific model has a 1-to-1 correspondence with the task-agnostic model" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27911/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27911", "html_url": "https://github.com/huggingface/transformers/pull/27911", "diff_url": "https://github.com/huggingface/transformers/pull/27911.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27911.patch", "merged_at": 1702282722000 }

{ "login": "pcuenca", "id": 1177582, "node_id": "MDQ6VXNlcjExNzc1ODI=", "avatar_url": "https://avatars.githubusercontent.com/u/1177582?v=4", "gravatar_id": "", "url": "https://api.github.com/users/pcuenca", "html_url": "https://github.com/pcuenca", "followers_url": "https://api.github.com/users/pcuenca/followers", "following_url": "https://api.github.com/users/pcuenca/following{/other_user}", "gists_url": "https://api.github.com/users/pcuenca/gists{/gist_id}", "starred_url": "https://api.github.com/users/pcuenca/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/pcuenca/subscriptions", "organizations_url": "https://api.github.com/users/pcuenca/orgs", "repos_url": "https://api.github.com/users/pcuenca/repos", "events_url": "https://api.github.com/users/pcuenca/events{/privacy}", "received_events_url": "https://api.github.com/users/pcuenca/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27910/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27910", "html_url": "https://github.com/huggingface/transformers/pull/27910", "diff_url": "https://github.com/huggingface/transformers/pull/27910.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27910.patch", "merged_at": 1702047770000 }

{ "login": "ArthurZucker", "id": 48595927, "node_id": "MDQ6VXNlcjQ4NTk1OTI3", "avatar_url": "https://avatars.githubusercontent.com/u/48595927?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ArthurZucker", "html_url": "https://github.com/ArthurZucker", "followers_url": "https://api.github.com/users/ArthurZucker/followers", "following_url": "https://api.github.com/users/ArthurZucker/following{/other_user}", "gists_url": "https://api.github.com/users/ArthurZucker/gists{/gist_id}", "starred_url": "https://api.github.com/users/ArthurZucker/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ArthurZucker/subscriptions", "organizations_url": "https://api.github.com/users/ArthurZucker/orgs", "repos_url": "https://api.github.com/users/ArthurZucker/repos", "events_url": "https://api.github.com/users/ArthurZucker/events{/privacy}", "received_events_url": "https://api.github.com/users/ArthurZucker/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27909). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27909/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27909", "html_url": "https://github.com/huggingface/transformers/pull/27909", "diff_url": "https://github.com/huggingface/transformers/pull/27909.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27909.patch", "merged_at": 1702053154000 }

{ "login": "plroit", "id": 1734563, "node_id": "MDQ6VXNlcjE3MzQ1NjM=", "avatar_url": "https://avatars.githubusercontent.com/u/1734563?v=4", "gravatar_id": "", "url": "https://api.github.com/users/plroit", "html_url": "https://github.com/plroit", "followers_url": "https://api.github.com/users/plroit/followers", "following_url": "https://api.github.com/users/plroit/following{/other_user}", "gists_url": "https://api.github.com/users/plroit/gists{/gist_id}", "starred_url": "https://api.github.com/users/plroit/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/plroit/subscriptions", "organizations_url": "https://api.github.com/users/plroit/orgs", "repos_url": "https://api.github.com/users/plroit/repos", "events_url": "https://api.github.com/users/plroit/events{/privacy}", "received_events_url": "https://api.github.com/users/plroit/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey, you added a new token, this you need to resize the token embedding layer of the model with `model.resize_token_embeddings(len(tokenizer))`", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27908/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "gante", "id": 12240844, "node_id": "MDQ6VXNlcjEyMjQwODQ0", "avatar_url": "https://avatars.githubusercontent.com/u/12240844?v=4", "gravatar_id": "", "url": "https://api.github.com/users/gante", "html_url": "https://github.com/gante", "followers_url": "https://api.github.com/users/gante/followers", "following_url": "https://api.github.com/users/gante/following{/other_user}", "gists_url": "https://api.github.com/users/gante/gists{/gist_id}", "starred_url": "https://api.github.com/users/gante/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/gante/subscriptions", "organizations_url": "https://api.github.com/users/gante/orgs", "repos_url": "https://api.github.com/users/gante/repos", "events_url": "https://api.github.com/users/gante/events{/privacy}", "received_events_url": "https://api.github.com/users/gante/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Some additional results @gante \r\n\r\n\r\n\r\n* `transformers_old` refers to before #27700 (thanks @kevinhu for the change!)\r\n* `transformers_56be5e80` refers to commit 56be5e80, i.e. near main\r\n* `transformers_attn_sink_1024_4_pr-27907` refers to this PR, with `SinkCache(1024, 4)`\r\n\r\nThis experiment is with calling the models with individual tokens exclusively. Using `SinkCache` makes the memory usage linear at a very low cost in perplexity.\r\n\r\n---\r\n\r\nAdditionally, I'm doing a test with calling Mistral-7B in this case with this PR using multiple tokens at once. I took indices 0 to 63k of a book from pg19, only kept 20% of all indices, and then fed the model with the tokens between subsequent indices. The running cache is also included. The NLL loss is then converted to perplexity.\r\n**Note:** We can't compare this with the perplexities from the previous graph: we should only try and observe whether the model eventually increases in perplexity.\r\n\r\n\r\n\r\nThe same script crashes on `main`. In this test, the perplexity stays constant, which is good. \r\nEdit: I have now continued with more tests:\r\n\r\n\r\n\r\n* `transformers_multi_attn_sink_1024_4_pr-27907_left_cache`: This PR, with `SinkCache(1024, 4)`.\r\n* `transformers_multi_attn_sink_1024_4_pr-27907_right_cache`: This PR, with `SinkCache(1024, 4)` & the change @ArthurZucker proposed regarding slicing the cache from the right.\r\n* `transformers_multi_b31905d1`: `main` without a special SinkCache.\r\n\r\nThe perplexity diverges quite heavily between the SinkCache and non, which is not ideal. Perhaps this is indicative of some error/bug, or perhaps not. It's a bit hard to tell. Beyond that, the left and right cache implementations behave identically (unless I made some measuring mistake), which is a bit odd. I don't have 100% confidence in this fix anymore I'm afraid.\r\n\r\n- Tom Aarsen", "@ArthurZucker as per your suggestion, I've reworked the PR to avoid post hoc `attention_mask` slicing -- there is a new function to get the usable cache length, and that function is used to obtain `kv_seq_len`\r\n\r\n@tomaarsen the rework seems to have resulted in a qualitative result upgrade (e.g. see the test case), so I suspect that I've inadvertently fixed a bug 👀 Would you be able to rerun your benchmarks for `SinkCache`? ", "@gante I get `ValueError: Attention weights should be of size (1, 32, 5, 1027), but is torch.Size([1, 32, 5, 1024])` upon running my modified script. Do you get an error like this with the multi-step generation script from your PR?", "@tomaarsen no, the script in the PR header runs endlessly without issues 🤔 LMK if you can find a reproducer", "I have the same, that script works fine. Hmmm", "Got a reproducer: change `max_new_tokens` in the script above to 512 👀 having a look!", "@tomaarsen should be fixed", "@gante Works great now!\r\n\r\n\r\nRed is the baseline, I can only run it to about ~15k seq length until my PC completely freezes.", "🙌 " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27907/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27907", "html_url": "https://github.com/huggingface/transformers/pull/27907", "diff_url": "https://github.com/huggingface/transformers/pull/27907.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27907.patch", "merged_at": 1702065740000 }

{ "login": "ydshieh", "id": 2521628, "node_id": "MDQ6VXNlcjI1MjE2Mjg=", "avatar_url": "https://avatars.githubusercontent.com/u/2521628?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ydshieh", "html_url": "https://github.com/ydshieh", "followers_url": "https://api.github.com/users/ydshieh/followers", "following_url": "https://api.github.com/users/ydshieh/following{/other_user}", "gists_url": "https://api.github.com/users/ydshieh/gists{/gist_id}", "starred_url": "https://api.github.com/users/ydshieh/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ydshieh/subscriptions", "organizations_url": "https://api.github.com/users/ydshieh/orgs", "repos_url": "https://api.github.com/users/ydshieh/repos", "events_url": "https://api.github.com/users/ydshieh/events{/privacy}", "received_events_url": "https://api.github.com/users/ydshieh/received_events", "type": "User", "site_admin": false }

[]

[]

[ "For the record: below code shows the issue\r\n\r\n```python\r\nimport torch\r\n\r\nDIM = 32\r\nSTD = 1e-10\r\nB = 2\r\nN_ITER = 100\r\n\r\n\r\nfor _ in range(100):\r\n for pow in range(4, 15):\r\n m = 0\r\n seq_len = 2 ** pow\r\n for i in range(N_ITER):\r\n t = torch.zeros(size=(seq_len, DIM))\r\n o = torch.abs(torch.nn.init.trunc_normal_(t.to(torch.float32), mean=0.0, std=STD, a=-B, b=B))\r\n n = torch.sum(o >= B)\r\n m += n\r\n\r\n print(f\"seq_len = {seq_len}: {m / seq_len / DIM / N_ITER}\")\r\n\r\n print(\"-\" * 80)\r\n```", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27906). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27906/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27906", "html_url": "https://github.com/huggingface/transformers/pull/27906", "diff_url": "https://github.com/huggingface/transformers/pull/27906.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27906.patch", "merged_at": 1702043672000 }

{ "login": "sanchit-gandhi", "id": 93869735, "node_id": "U_kgDOBZhWpw", "avatar_url": "https://avatars.githubusercontent.com/u/93869735?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sanchit-gandhi", "html_url": "https://github.com/sanchit-gandhi", "followers_url": "https://api.github.com/users/sanchit-gandhi/followers", "following_url": "https://api.github.com/users/sanchit-gandhi/following{/other_user}", "gists_url": "https://api.github.com/users/sanchit-gandhi/gists{/gist_id}", "starred_url": "https://api.github.com/users/sanchit-gandhi/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sanchit-gandhi/subscriptions", "organizations_url": "https://api.github.com/users/sanchit-gandhi/orgs", "repos_url": "https://api.github.com/users/sanchit-gandhi/repos", "events_url": "https://api.github.com/users/sanchit-gandhi/events{/privacy}", "received_events_url": "https://api.github.com/users/sanchit-gandhi/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27905). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27905/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27905", "html_url": "https://github.com/huggingface/transformers/pull/27905", "diff_url": "https://github.com/huggingface/transformers/pull/27905.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27905.patch", "merged_at": 1702566853000 }

{ "login": "zhaosheng-thu", "id": 144892591, "node_id": "U_kgDOCKLirw", "avatar_url": "https://avatars.githubusercontent.com/u/144892591?v=4", "gravatar_id": "", "url": "https://api.github.com/users/zhaosheng-thu", "html_url": "https://github.com/zhaosheng-thu", "followers_url": "https://api.github.com/users/zhaosheng-thu/followers", "following_url": "https://api.github.com/users/zhaosheng-thu/following{/other_user}", "gists_url": "https://api.github.com/users/zhaosheng-thu/gists{/gist_id}", "starred_url": "https://api.github.com/users/zhaosheng-thu/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/zhaosheng-thu/subscriptions", "organizations_url": "https://api.github.com/users/zhaosheng-thu/orgs", "repos_url": "https://api.github.com/users/zhaosheng-thu/repos", "events_url": "https://api.github.com/users/zhaosheng-thu/events{/privacy}", "received_events_url": "https://api.github.com/users/zhaosheng-thu/received_events", "type": "User", "site_admin": false }

[]

[]

[ "hey which version of python are you working with? And what hardware are you using? ", "> hey which version of python are you working with? And what hardware are you using?\r\n\r\nthanks a lot! I have just solved this problem, and wish you a good day!", "How did you solve this?\r\n", "> hey which version of python are you working with? And what hardware are you using?\r\n\r\ni am getting same issue. i am using python 3.12" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27904/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "ydshieh", "id": 2521628, "node_id": "MDQ6VXNlcjI1MjE2Mjg=", "avatar_url": "https://avatars.githubusercontent.com/u/2521628?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ydshieh", "html_url": "https://github.com/ydshieh", "followers_url": "https://api.github.com/users/ydshieh/followers", "following_url": "https://api.github.com/users/ydshieh/following{/other_user}", "gists_url": "https://api.github.com/users/ydshieh/gists{/gist_id}", "starred_url": "https://api.github.com/users/ydshieh/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ydshieh/subscriptions", "organizations_url": "https://api.github.com/users/ydshieh/orgs", "repos_url": "https://api.github.com/users/ydshieh/repos", "events_url": "https://api.github.com/users/ydshieh/events{/privacy}", "received_events_url": "https://api.github.com/users/ydshieh/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27903/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27903", "html_url": "https://github.com/huggingface/transformers/pull/27903", "diff_url": "https://github.com/huggingface/transformers/pull/27903.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27903.patch", "merged_at": 1702043702000 }

{ "login": "RmZeta2718", "id": 42400165, "node_id": "MDQ6VXNlcjQyNDAwMTY1", "avatar_url": "https://avatars.githubusercontent.com/u/42400165?v=4", "gravatar_id": "", "url": "https://api.github.com/users/RmZeta2718", "html_url": "https://github.com/RmZeta2718", "followers_url": "https://api.github.com/users/RmZeta2718/followers", "following_url": "https://api.github.com/users/RmZeta2718/following{/other_user}", "gists_url": "https://api.github.com/users/RmZeta2718/gists{/gist_id}", "starred_url": "https://api.github.com/users/RmZeta2718/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/RmZeta2718/subscriptions", "organizations_url": "https://api.github.com/users/RmZeta2718/orgs", "repos_url": "https://api.github.com/users/RmZeta2718/repos", "events_url": "https://api.github.com/users/RmZeta2718/events{/privacy}", "received_events_url": "https://api.github.com/users/RmZeta2718/received_events", "type": "User", "site_admin": false }

[ { "id": 1990918270, "node_id": "MDU6TGFiZWwxOTkwOTE4Mjcw", "url": "https://api.github.com/repos/huggingface/transformers/labels/Good%20First%20Issue", "name": "Good First Issue", "color": "bbf794", "default": false, "description": "" }, { "id": 3551105283, "node_id": "LA_kwDOCUB6oc7TqZED", "url": "https://api.github.com/repos/huggingface/transformers/labels/Good%20First%20Documentation%20Issue", "name": "Good First Documentation Issue", "color": "AB0BA8", "default": false, "description": "" } ]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored.", "Gently pinging @muellerzr and @pacman100 ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored.", "Marking as good first issue, doc can be updated 😉 ", "Hey, @pacman100. Can I work on this issue?", "Sure feel free to open a PR to update the doc 😉 " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27902/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "okotaku", "id": 24734142, "node_id": "MDQ6VXNlcjI0NzM0MTQy", "avatar_url": "https://avatars.githubusercontent.com/u/24734142?v=4", "gravatar_id": "", "url": "https://api.github.com/users/okotaku", "html_url": "https://github.com/okotaku", "followers_url": "https://api.github.com/users/okotaku/followers", "following_url": "https://api.github.com/users/okotaku/following{/other_user}", "gists_url": "https://api.github.com/users/okotaku/gists{/gist_id}", "starred_url": "https://api.github.com/users/okotaku/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/okotaku/subscriptions", "organizations_url": "https://api.github.com/users/okotaku/orgs", "repos_url": "https://api.github.com/users/okotaku/repos", "events_url": "https://api.github.com/users/okotaku/events{/privacy}", "received_events_url": "https://api.github.com/users/okotaku/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@ArthurZucker Here is the minimum code.\r\n\r\n```\r\nimport torch\r\nfrom transformers.models.llava.modeling_llava import LlavaForConditionalGeneration\r\n\r\n\r\nclass TestLlavaForConditionalGeneration(LlavaForConditionalGeneration):\r\n def forward(\r\n self,\r\n input_ids= None,\r\n pixel_values= None,\r\n attention_mask = None,\r\n position_ids= None,\r\n past_key_values= None,\r\n inputs_embeds= None,\r\n vision_feature_layer= None,\r\n vision_feature_select_strategy= None,\r\n labels= None,\r\n use_cache= None,\r\n output_attentions= None,\r\n output_hidden_states= None,\r\n return_dict= None,\r\n ):\r\n output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions\r\n output_hidden_states = (\r\n output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states\r\n )\r\n return_dict = return_dict if return_dict is not None else self.config.use_return_dict\r\n vision_feature_layer = (\r\n vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer\r\n )\r\n vision_feature_select_strategy = (\r\n vision_feature_select_strategy\r\n if vision_feature_select_strategy is not None\r\n else self.config.vision_feature_select_strategy\r\n )\r\n\r\n if inputs_embeds is None:\r\n # 1. Extra the input embeddings\r\n inputs_embeds = self.get_input_embeddings()(input_ids)\r\n\r\n # 2. Merge text and images\r\n if pixel_values is not None and input_ids.shape[1] != 1:\r\n image_outputs = self.vision_tower(pixel_values, output_hidden_states=True)\r\n # this is not memory efficient at all (output_hidden_states=True) will save all the hidden stated.\r\n selected_image_feature = image_outputs.hidden_states[vision_feature_layer]\r\n\r\n if vision_feature_select_strategy == \"default\":\r\n selected_image_feature = selected_image_feature[:, 1:]\r\n elif vision_feature_select_strategy == \"full\":\r\n selected_image_feature = selected_image_feature\r\n else:\r\n raise ValueError(\r\n f\"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}\"\r\n )\r\n\r\n image_features = self.multi_modal_projector(selected_image_feature)\r\n inputs_embeds, attention_mask, position_ids = self._merge_input_ids_with_image_features(\r\n image_features, inputs_embeds, input_ids, attention_mask, position_ids\r\n )\r\n if labels is None:\r\n labels = torch.full_like(attention_mask, self.config.ignore_index).to(torch.long)\r\n else:\r\n # In case input_ids.shape[1] == 1 & pixel_values==None & past_key_values != None, we are in the case of\r\n # generation with cache\r\n if past_key_values is not None and pixel_values is not None and input_ids.shape[1] == 1:\r\n # Retrieve the first layer to inspect the logits and mask out the hidden states\r\n # that are set to 0\r\n first_layer_past_key_value = past_key_values[0][0][:, 0, :, 0]\r\n batch_index, non_attended_tokens = torch.where(first_layer_past_key_value == 0)\r\n\r\n ############################\r\n # Add here\r\n # non_attended_tokens = non_attended_tokens - attention_mask.shape[1]\r\n ############################\r\n\r\n # Get the target length\r\n target_seqlen = first_layer_past_key_value.shape[-1] + 1\r\n\r\n extended_attention_mask = torch.ones(\r\n (attention_mask.shape[0], target_seqlen - attention_mask.shape[1]),\r\n dtype=attention_mask.dtype,\r\n device=attention_mask.device,\r\n )\r\n\r\n # Zero-out the places where we don't need to attend\r\n extended_attention_mask[batch_index, non_attended_tokens] = 0\r\n\r\n attention_mask = torch.cat((attention_mask, extended_attention_mask), dim=1)\r\n position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1\r\n print(position_ids)\r\n\r\nllava = TestLlavaForConditionalGeneration.from_pretrained(\r\n \"llava-hf/llava-1.5-13b-hf\", torch_dtype=torch.float16)\r\nllava.to(\"cuda\")\r\ninput_ids = torch.tensor([[319]]).cuda().long()\r\npixel_values = torch.zeros((1, 3, 224, 224)).cuda().half()\r\npast_key_values = torch.rand((1, 1, 1, 40, 642, 128)).cuda().half()\r\npast_key_values[:, :, :, 0, 600, 0] = 0.\r\nattention_mask = torch.ones((1, 68)).cuda().half()\r\nllava(input_ids=input_ids,\r\n pixel_values=pixel_values,\r\n past_key_values=past_key_values,\r\n attention_mask=attention_mask)\r\n```\r\n\r\nOutput is:\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"/workspace/a.py\", line 99, in <module>\r\n llava(input_ids=input_ids,\r\n File \"/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py\", line 1501, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/workspace/a.py\", line 87, in forward\r\n attention_mask = torch.cat((attention_mask, extended_attention_mask), dim=1)\r\nRuntimeError: CUDA error: device-side assert triggered\r\nCUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.\r\nFor debugging consider passing CUDA_LAUNCH_BLOCKING=1.\r\nCompile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.\r\n```", "cc @younesbelkada ", "@younesbelkada \r\nThank you!\r\n\r\nI feel there is an indexing error between `extended_attention_mask` and `non_attended_tokens`.\r\n\r\n```\r\nbatch_index, non_attended_tokens = torch.where(first_layer_past_key_value == 0)\r\n# Get the target length\r\ntarget_seqlen = first_layer_past_key_value.shape[-1] + 1\r\n\r\n# Now the index of `non_attended_tokens` corresponds to the index of `target_seqlen` == first_layer_past_key_value.axis(-1).\r\n\r\n# However, the index of extended_attention_mask is target_seqlen - attention_mask.shape[1].\r\n# Are the indices of `non_attended_tokens` and `extended_attention_mask` different?\r\nextended_attention_mask = torch.ones(\r\n (attention_mask.shape[0], target_seqlen - attention_mask.shape[1]),\r\n dtype=attention_mask.dtype,\r\n device=attention_mask.device,\r\n)\r\n\r\n# Zero-out the places where we don't need to attend\r\nextended_attention_mask[batch_index, non_attended_tokens] = 0\r\nattention_mask = torch.cat((attention_mask, extended_attention_mask), dim=1)\r\n```", "Hi @okotaku \r\nI think https://github.com/huggingface/transformers/pull/28032 fixes the same issue, can you try out on transformers main ? 🙏 " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27901/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27901", "html_url": "https://github.com/huggingface/transformers/pull/27901", "diff_url": "https://github.com/huggingface/transformers/pull/27901.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27901.patch", "merged_at": null }

{ "login": "phoongkhangzhie", "id": 25717121, "node_id": "MDQ6VXNlcjI1NzE3MTIx", "avatar_url": "https://avatars.githubusercontent.com/u/25717121?v=4", "gravatar_id": "", "url": "https://api.github.com/users/phoongkhangzhie", "html_url": "https://github.com/phoongkhangzhie", "followers_url": "https://api.github.com/users/phoongkhangzhie/followers", "following_url": "https://api.github.com/users/phoongkhangzhie/following{/other_user}", "gists_url": "https://api.github.com/users/phoongkhangzhie/gists{/gist_id}", "starred_url": "https://api.github.com/users/phoongkhangzhie/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/phoongkhangzhie/subscriptions", "organizations_url": "https://api.github.com/users/phoongkhangzhie/orgs", "repos_url": "https://api.github.com/users/phoongkhangzhie/repos", "events_url": "https://api.github.com/users/phoongkhangzhie/events{/privacy}", "received_events_url": "https://api.github.com/users/phoongkhangzhie/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Ahhh I'll have a look that looks a bit nasty indeed\r\n", "Hi @ArthurZucker , any updates on this? Thank you!", "Hey, I can't reproduce this yet. I don't have your local dataset, and I don't have the loading script so \r\n```python\r\ndef python_generator():\r\n # Load local files for code_search_net/python\r\n # https://huggingface.co./datasets/code_search_net\r\n dataset = load_dataset(\"code_search_net/python.py\", \"python\")\r\n dataset = dataset[\"train\"]\r\n for start_idx in range(0, len(dataset), 1000):\r\n samples = dataset[start_idx: start_idx + 1000]\r\n yield samples[\"whole_func_string\"]\r\n```\r\n\r\nfails with \r\n`FileNotFoundError: Couldn't find a dataset script at /Users/arthurzucker/Work/transformers/deci-7b/code_search_net/python.py`", "I cannot help you without a proper reproducer\r\n", "One thing that is certain is that Bytefallback does not seem to be activated (properly) because the bytes should be part of the vocab, the trainer should have a logic to handle that which it does not at the moment ", "> I cannot help you without a proper reproducer\r\n\r\nI've updated the script above. Hopefully it works now!", "Same here! There are tokens in the vocabulary that consist of some joined words, like `this▁is▁a▁test`", "What did you train your tokenizer on? ", "@phoongkhangzhie I had to update your script it does not work out of the box, ", "@ArthurZucker on batches of strings. It seems it's not splitting words", "I think a quick fix would be to disable the normalizer and use a metaspace pre-tokenizer instead. \r\n```python3\r\nfrom tokenizers import pre_tokenizers, normalizers\r\nfrom transformers import AutoTokenizer\r\nold_tokenizer = AutoTokenizer.from_pretrained(\"meta-llama/Llama-2-7b-hf\")\r\nold_tokenizer._tokenizer.normalizer = normalizers.Sequence([])\r\nold_tokenizer._tokenizer.pre_tokenizer = pre_tokenizers.Metaspace(\"▁\", True, prepend_scheme = \"first\")\r\n```\r\n", "It works, the vocabulary is correctly generated now. However, it does not pretokenize punctuation:\r\n\r\n```bash\r\n(Pdb) old_tokenizer.convert_ids_to_tokens(old_tokenizer(\"This is a test.\")[\"input_ids\"])\r\n['<s>', '▁This', '▁is', '▁a', '▁test', '.']\r\n(Pdb) new_tokenizer.convert_ids_to_tokens(new_tokenizer(\"This is a test.\")[\"input_ids\"])\r\n['<s>', '▁Th', 'is', '▁is', '▁a', '▁tes', 't.']\r\n```", "That's because it is probably missing a replace normalizer. so something like this: \r\n```python \r\nfrom tokenizers import pre_tokenizers, normalizers\r\nfrom transformers import AutoTokenizer\r\nold_tokenizer = AutoTokenizer.from_pretrained(\"meta-llama/Llama-2-7b-hf\")\r\nold_tokenizer._tokenizer.normalizer = normalizers.Sequence([normalizers.Strip(left=False, right=True), normalizers.Replace(Regex(\" {2,}\"), \"▁\")])\r\nold_tokenizer._tokenizer.pre_tokenizer = pre_tokenizers.Metaspace(\"▁\", True, prepend_scheme = \"first\")\r\n```\r\n(make sure you don't use \"_\" but \"▁\"", "#26678 should provide the fix. \r\ncc @xenova as this seems to give us a headache hahaa ", "I've added the noramlizer as you said. I solves the final dot issue. However, inner punctuation is not tokenized. There are tokens like `▁(house)` in the final vocabulary: I think we need to add `pre_tokenizers.Punctuation()` in the `pre_tokenizers`: \r\n\r\n```python\r\nold_tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)\r\nold_tokenizer._tokenizer.normalizer = normalizers.Sequence([normalizers.Strip(left=False, right=True), normalizers.Replace(tokenizers.Regex(\" {2,}\"), \"▁\")])\r\nold_tokenizer._tokenizer.pre_tokenizer = pre_tokenizers.Sequence([pre_tokenizers.Punctuation(), pre_tokenizers.Metaspace(prepend_scheme=\"first\")])\r\n```", "Thank you @ArthurZucker and @anderleich for your inputs.\r\n\r\nI think there are still issues with the tokenizer even after the various fixes.\r\n\r\n> I think a quick fix would be to disable the normalizer and use a metaspace pre-tokenizer instead.\r\n> \r\n> ```python\r\n> from tokenizers import pre_tokenizers, normalizers\r\n> from transformers import AutoTokenizer\r\n> old_tokenizer = AutoTokenizer.from_pretrained(\"meta-llama/Llama-2-7b-hf\")\r\n> old_tokenizer._tokenizer.normalizer = normalizers.Sequence([])\r\n> old_tokenizer._tokenizer.pre_tokenizer = pre_tokenizers.Metaspace(\"▁\", True, prepend_scheme = \"first\")\r\n> ```\r\n\r\nWith the above fix, the outputs are:\r\n```\r\nExample 1:\r\ndef add_numbers(a, b):\r\n \"\"\"Add the two numbers `a` and `b`.\"\"\"\r\n return a + b\r\n\r\nold: ['▁', '<0x0A>', '▁▁▁▁▁▁▁', '▁def', '▁add', '_', 'numbers', '(', 'a', ',', '▁b', '):', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁\"\"\"', 'Add', '▁the', '▁two', '▁numbers', '▁`', 'a', '`', '▁and', '▁`', 'b', '`', '.\"', '\"\"', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁return', '▁a', '▁+', '▁b', '<0x0A>', '▁▁▁▁▁▁▁▁']\r\nnew: ['▁\\n', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁def', '▁add_', 'number', 's(', 'a,', '▁b):\\n', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁\"\"\"Add', '▁the', '▁two', '▁numbers', '▁`a`', '▁and', '▁`b', '`.\"\"\"\\n', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁return', '▁a', '▁+', '▁b\\n', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁']\r\n\r\nExample 2:\r\nclass LinearLayer():\r\n def __init__(self, input_size, output_size):\r\n self.weight = torch.randn(input_size, output_size)\r\n self.bias = torch.zeros(output_size)\r\n\r\n def __call__(self, x):\r\n return x @ self.weights + self.bias\r\n\r\nold: ['▁', '<0x0A>', '▁▁▁▁▁▁▁', '▁class', '▁Linear', 'Layer', '():', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁def', '▁__', 'init', '__(', 'self', ',', '▁input', '_', 'size', ',', '▁output', '_', 'size', '):', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁', '▁self', '.', 'weight', '▁=', '▁tor', 'ch', '.', 'rand', 'n', '(', 'input', '_', 'size', ',', '▁output', '_', 'size', ')', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁', '▁self', '.', 'b', 'ias', '▁=', '▁tor', 'ch', '.', 'zer', 'os', '(', 'output', '_', 'size', ')', '<0x0A>', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁def', '▁__', 'call', '__(', 'self', ',', '▁x', '):', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁', '▁return', '▁x', '▁@', '▁self', '.', 'we', 'ights', '▁+', '▁self', '.', 'b', 'ias', '<0x0A>', '▁▁▁▁▁▁▁▁']\r\nnew: ['▁\\n', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁class', '▁Linear', 'Layer', '():\\n', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁def', '▁__init__(self,', '▁input', '_size,', '▁output', '_size):\\n', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁self.weight', '▁=', '▁torch', '.randn', '(input', '_size,', '▁output', '_size)\\n', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁self.b', 'ias', '▁=', '▁torch.', 'zeros(', 'output', '_size)\\n\\n', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁def', '▁__call', '__(self,', '▁x):\\n', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁return', '▁x', '▁@', '▁self.', 'weights', '▁+', '▁self.b', 'ias', '\\n', '▁', '▁', '▁', '▁', '▁', '▁', '▁', '▁']\r\n```\r\nThis fix prepends all whitespace characters with `'▁'`, but all of them are separate tokens in the final output where instead some of them should be merged instead to represent indentations or double indentation in code. Also, the newline character `\\n` is not treated as a whitespace character.\r\n\r\n> That's because it is probably missing a replace normalizer. so something like this:\r\n> \r\n> ```python\r\n> from tokenizers import pre_tokenizers, normalizers\r\n> from transformers import AutoTokenizer\r\n> old_tokenizer = AutoTokenizer.from_pretrained(\"meta-llama/Llama-2-7b-hf\")\r\n> old_tokenizer._tokenizer.normalizer = normalizers.Sequence([normalizers.Strip(left=False, right=True), normalizers.Replace(Regex(\" {2,}\"), \"▁\")])\r\n> old_tokenizer._tokenizer.pre_tokenizer = pre_tokenizers.Metaspace(\"▁\", True, prepend_scheme = \"first\")\r\n> ```\r\n> \r\n> (make sure you don't use \"_\" but \"▁\"\r\n\r\nWith the above fix, the outputs are:\r\n```\r\nExample 1:\r\ndef add_numbers(a, b):\r\n \"\"\"Add the two numbers `a` and `b`.\"\"\"\r\n return a + b\r\n\r\nold: ['▁', '<0x0A>', '▁▁▁▁▁▁▁', '▁def', '▁add', '_', 'numbers', '(', 'a', ',', '▁b', '):', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁\"\"\"', 'Add', '▁the', '▁two', '▁numbers', '▁`', 'a', '`', '▁and', '▁`', 'b', '`', '.\"', '\"\"', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁return', '▁a', '▁+', '▁b', '<0x0A>', '▁▁▁▁▁▁▁▁']\r\nnew: ['▁\\n', '▁def', '▁add_', 'number', 's(', 'a,', '▁b):\\n', '▁\"\"\"Add', '▁the', '▁two', '▁numbers', '▁`a`', '▁and', '▁`b', '`.\"\"\"\\n', '▁return', '▁a', '▁+', '▁b']\r\n\r\nExample 2:\r\nclass LinearLayer():\r\n def __init__(self, input_size, output_size):\r\n self.weight = torch.randn(input_size, output_size)\r\n self.bias = torch.zeros(output_size)\r\n\r\n def __call__(self, x):\r\n return x @ self.weights + self.bias\r\n\r\nold: ['▁', '<0x0A>', '▁▁▁▁▁▁▁', '▁class', '▁Linear', 'Layer', '():', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁def', '▁__', 'init', '__(', 'self', ',', '▁input', '_', 'size', ',', '▁output', '_', 'size', '):', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁', '▁self', '.', 'weight', '▁=', '▁tor', 'ch', '.', 'rand', 'n', '(', 'input', '_', 'size', ',', '▁output', '_', 'size', ')', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁', '▁self', '.', 'b', 'ias', '▁=', '▁tor', 'ch', '.', 'zer', 'os', '(', 'output', '_', 'size', ')', '<0x0A>', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁def', '▁__', 'call', '__(', 'self', ',', '▁x', '):', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁', '▁return', '▁x', '▁@', '▁self', '.', 'we', 'ights', '▁+', '▁self', '.', 'b', 'ias', '<0x0A>', '▁▁▁▁▁▁▁▁']\r\nnew: ['▁\\n', '▁class', '▁Linear', 'Layer', '():\\n', '▁def', '▁__init__(self,', '▁input', '_size,', '▁output', '_size):\\n', '▁self.weight', '▁=', '▁torch', '.randn', '(input', '_size,', '▁output', '_size)\\n', '▁self.b', 'ias', '▁=', '▁torch.', 'zeros(', 'output', '_size)\\n\\n', '▁def', '▁__call', '__(self,', '▁x):\\n', '▁return', '▁x', '▁@', '▁self.', 'weights', '▁+', '▁self.b', 'ias']\r\n```\r\nThis fix collapses all the whitespace characters into a single `'▁'` character. However, this removes the importance of whitespace in code such as the different indentation levels. Again, the newline character `\\n` is not treated as a whitespace character.\r\n\r\n> I've added the noramlizer as you said. I solves the final dot issue. However, inner punctuation is not tokenized. There are tokens like `▁(house)` in the final vocabulary: I think we need to add `pre_tokenizers.Punctuation()` in the `pre_tokenizers`:\r\n> \r\n> ```python\r\n> old_tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)\r\n> old_tokenizer._tokenizer.normalizer = normalizers.Sequence([normalizers.Strip(left=False, right=True), normalizers.Replace(tokenizers.Regex(\" {2,}\"), \"▁\")])\r\n> old_tokenizer._tokenizer.pre_tokenizer = pre_tokenizers.Sequence([pre_tokenizers.Punctuation(), pre_tokenizers.Metaspace(prepend_scheme=\"first\")])\r\n> ```\r\n\r\nAnd with this fix, the outputs are:\r\n```\r\nExample 1:\r\ndef add_numbers(a, b):\r\n \"\"\"Add the two numbers `a` and `b`.\"\"\"\r\n return a + b\r\n\r\nold: ['▁', '<0x0A>', '▁▁▁▁▁▁▁', '▁def', '▁add', '_', 'numbers', '(', 'a', ',', '▁b', '):', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁\"\"\"', 'Add', '▁the', '▁two', '▁numbers', '▁`', 'a', '`', '▁and', '▁`', 'b', '`', '.\"', '\"\"', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁return', '▁a', '▁+', '▁b', '<0x0A>', '▁▁▁▁▁▁▁▁']\r\nnew: ['▁\\n', '▁def', '▁add', '_', 'numbers', '(', 'a', ',', '▁b', ')', ':', '\\n', '▁', '\"', '\"', '\"', 'Add', '▁the', '▁two', '▁numbers', '▁', '`', 'a', '`', '▁and', '▁', '`', 'b', '`', '.', '\"', '\"', '\"', '\\n', '▁return', '▁a', '▁', '+', '▁b']\r\n\r\nExample 2:\r\nclass LinearLayer():\r\n def __init__(self, input_size, output_size):\r\n self.weight = torch.randn(input_size, output_size)\r\n self.bias = torch.zeros(output_size)\r\n\r\n def __call__(self, x):\r\n return x @ self.weights + self.bias\r\n\r\nold: ['▁', '<0x0A>', '▁▁▁▁▁▁▁', '▁class', '▁Linear', 'Layer', '():', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁def', '▁__', 'init', '__(', 'self', ',', '▁input', '_', 'size', ',', '▁output', '_', 'size', '):', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁', '▁self', '.', 'weight', '▁=', '▁tor', 'ch', '.', 'rand', 'n', '(', 'input', '_', 'size', ',', '▁output', '_', 'size', ')', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁', '▁self', '.', 'b', 'ias', '▁=', '▁tor', 'ch', '.', 'zer', 'os', '(', 'output', '_', 'size', ')', '<0x0A>', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁def', '▁__', 'call', '__(', 'self', ',', '▁x', '):', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁', '▁return', '▁x', '▁@', '▁self', '.', 'we', 'ights', '▁+', '▁self', '.', 'b', 'ias', '<0x0A>', '▁▁▁▁▁▁▁▁']\r\nnew: ['▁\\n', '▁class', '▁Linear', 'Layer', '(', ')', ':', '\\n', '▁def', '▁', '_', '_', 'init', '_', '_', '(', 'self', ',', '▁input', '_', 'size', ',', '▁output', '_', 'size', ')', ':', '\\n', '▁self', '.', 'weight', '▁', '=', '▁torch', '.', 'randn', '(', 'input', '_', 'size', ',', '▁output', '_', 'size', ')', '\\n', '▁self', '.', 'bias', '▁', '=', '▁torch', '.', 'zeros', '(', 'output', '_', 'size', ')', '\\n\\n', '▁def', '▁', '_', '_', 'call', '_', '_', '(', 'self', ',', '▁x', ')', ':', '\\n', '▁return', '▁x', '▁', '@', '▁self', '.', 'weights', '▁', '+', '▁self', '.', 'bias']\r\n```\r\nWhile this tokenization might be better than the above one, I think it is too aggressive with the splitting of the punctuation. Like the above fixes, the newline character `\\n` is not treated as a whitespace character.\r\n\r\nIdeally, the outputs should be like this (similar to the GPT2 tokenization):\r\n```\r\nExample 1:\r\ndef add_numbers(a, b):\r\n \"\"\"Add the two numbers `a` and `b`.\"\"\"\r\n return a + b\r\n\r\nold: ['▁', '<0x0A>', '▁▁▁▁▁▁▁', '▁def', '▁add', '_', 'numbers', '(', 'a', ',', '▁b', '):', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁\"\"\"', 'Add', '▁the', '▁two', '▁numbers', '▁`', 'a', '`', '▁and', '▁`', 'b', '`', '.\"', '\"\"', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁return', '▁a', '▁+', '▁b', '<0x0A>', '▁▁▁▁▁▁▁▁']\r\nnew: ['▁\\n', '▁def', '▁add', '_', 'numbers', '(', 'a', ',', '▁b', ')', ':', '▁\\n', '▁\"\"\"', 'Add', '▁the', '▁two', '▁numbers', '▁', '`', 'a', '`', '▁and', '▁', '`', 'b', '`', '.\"\"\"', '▁\\n', '▁return', '▁a', '▁+', '▁b']\r\n\r\nExample 2:\r\nclass LinearLayer():\r\n def __init__(self, input_size, output_size):\r\n self.weight = torch.randn(input_size, output_size)\r\n self.bias = torch.zeros(output_size)\r\n\r\n def __call__(self, x):\r\n return x @ self.weights + self.bias\r\n\r\nold: ['▁', '<0x0A>', '▁▁▁▁▁▁▁', '▁class', '▁Linear', 'Layer', '():', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁def', '▁__', 'init', '__(', 'self', ',', '▁input', '_', 'size', ',', '▁output', '_', 'size', '):', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁', '▁self', '.', 'weight', '▁=', '▁tor', 'ch', '.', 'rand', 'n', '(', 'input', '_', 'size', ',', '▁output', '_', 'size', ')', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁', '▁self', '.', 'b', 'ias', '▁=', '▁tor', 'ch', '.', 'zer', 'os', '(', 'output', '_', 'size', ')', '<0x0A>', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁', '▁def', '▁__', 'call', '__(', 'self', ',', '▁x', '):', '<0x0A>', '▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁', '▁return', '▁x', '▁@', '▁self', '.', 'we', 'ights', '▁+', '▁self', '.', 'b', 'ias', '<0x0A>', '▁▁▁▁▁▁▁▁']\r\nnew: ['▁\\n', '▁class', '▁Linear', 'Layer', '():', '▁\\n', '▁def', '▁__', 'init', '__(', 'self', ',', '▁input', '_', 'size', ',', '▁output', '_', 'size', '):', '▁\\n', '▁self', '.', 'weight', '▁=', '▁torch', '.', 'randn', '(', 'input', '_', 'size', ',', '▁output', '_', 'size', ')', '▁\\n', '▁self', '.', 'bias', '▁=', '▁torch', '.', 'zeros', '(', 'output', '_', 'size', ')', '▁\\n\\n', '▁def', '▁__', 'call', '__(', 'self', ',', '▁x', '):', '▁\\n', '▁return', '▁x', '▁@', '▁self', '.', 'weights', '▁+', '▁self', '.', 'bias']\r\n```\r\nWill there be any other fixes for this?", "If you want to keep the white space, `normalizers.Replace(Regex(\" {2,}\"), \"▁\")` should not be used indeed. \r\nLeftStripping can be kept, but you would need to also add the bytefallback tokens to the vocab ('<0x0A>' is a new line via bytefallback) if you want it to have the same behaviour! \r\n\r\nRegarding the merges, it might be the frequency of the `▁▁▁▁▁▁▁` token that prevents the model from learning it but should not be related to the pre-processing. \r\n\r\nSo the last issue is probably the bytefallback. \r\n\r\n", "@ArthurZucker are there any plans to add all those fixes to the `train_new_from_iterator` function for Llama2 models?", "There are plans to add these fixes to the LlamaTokenizer as a whole (specifically the pretokenizer vs normalizer) here #26678. The bytefallback thing needs to be adde to `tokenizers` and there is a plan but I don't have bandwidth just yet! 🤗 " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27900/reactions", "total_count": 1, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "zhc7", "id": 53651354, "node_id": "MDQ6VXNlcjUzNjUxMzU0", "avatar_url": "https://avatars.githubusercontent.com/u/53651354?v=4", "gravatar_id": "", "url": "https://api.github.com/users/zhc7", "html_url": "https://github.com/zhc7", "followers_url": "https://api.github.com/users/zhc7/followers", "following_url": "https://api.github.com/users/zhc7/following{/other_user}", "gists_url": "https://api.github.com/users/zhc7/gists{/gist_id}", "starred_url": "https://api.github.com/users/zhc7/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/zhc7/subscriptions", "organizations_url": "https://api.github.com/users/zhc7/orgs", "repos_url": "https://api.github.com/users/zhc7/repos", "events_url": "https://api.github.com/users/zhc7/events{/privacy}", "received_events_url": "https://api.github.com/users/zhc7/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27899/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27899", "html_url": "https://github.com/huggingface/transformers/pull/27899", "diff_url": "https://github.com/huggingface/transformers/pull/27899.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27899.patch", "merged_at": 1702060368000 }

{ "login": "wandouguo", "id": 9603390, "node_id": "MDQ6VXNlcjk2MDMzOTA=", "avatar_url": "https://avatars.githubusercontent.com/u/9603390?v=4", "gravatar_id": "", "url": "https://api.github.com/users/wandouguo", "html_url": "https://github.com/wandouguo", "followers_url": "https://api.github.com/users/wandouguo/followers", "following_url": "https://api.github.com/users/wandouguo/following{/other_user}", "gists_url": "https://api.github.com/users/wandouguo/gists{/gist_id}", "starred_url": "https://api.github.com/users/wandouguo/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/wandouguo/subscriptions", "organizations_url": "https://api.github.com/users/wandouguo/orgs", "repos_url": "https://api.github.com/users/wandouguo/repos", "events_url": "https://api.github.com/users/wandouguo/events{/privacy}", "received_events_url": "https://api.github.com/users/wandouguo/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hello, if you want our help we are going to need a full reproducer (as the contribution guidelines point out)", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27898/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "xuchenhao001", "id": 20737194, "node_id": "MDQ6VXNlcjIwNzM3MTk0", "avatar_url": "https://avatars.githubusercontent.com/u/20737194?v=4", "gravatar_id": "", "url": "https://api.github.com/users/xuchenhao001", "html_url": "https://github.com/xuchenhao001", "followers_url": "https://api.github.com/users/xuchenhao001/followers", "following_url": "https://api.github.com/users/xuchenhao001/following{/other_user}", "gists_url": "https://api.github.com/users/xuchenhao001/gists{/gist_id}", "starred_url": "https://api.github.com/users/xuchenhao001/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/xuchenhao001/subscriptions", "organizations_url": "https://api.github.com/users/xuchenhao001/orgs", "repos_url": "https://api.github.com/users/xuchenhao001/repos", "events_url": "https://api.github.com/users/xuchenhao001/events{/privacy}", "received_events_url": "https://api.github.com/users/xuchenhao001/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27897/reactions", "total_count": 1, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 1, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27897", "html_url": "https://github.com/huggingface/transformers/pull/27897", "diff_url": "https://github.com/huggingface/transformers/pull/27897.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27897.patch", "merged_at": 1702311556000 }

{ "login": "stevhliu", "id": 59462357, "node_id": "MDQ6VXNlcjU5NDYyMzU3", "avatar_url": "https://avatars.githubusercontent.com/u/59462357?v=4", "gravatar_id": "", "url": "https://api.github.com/users/stevhliu", "html_url": "https://github.com/stevhliu", "followers_url": "https://api.github.com/users/stevhliu/followers", "following_url": "https://api.github.com/users/stevhliu/following{/other_user}", "gists_url": "https://api.github.com/users/stevhliu/gists{/gist_id}", "starred_url": "https://api.github.com/users/stevhliu/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/stevhliu/subscriptions", "organizations_url": "https://api.github.com/users/stevhliu/orgs", "repos_url": "https://api.github.com/users/stevhliu/repos", "events_url": "https://api.github.com/users/stevhliu/events{/privacy}", "received_events_url": "https://api.github.com/users/stevhliu/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27896/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27896", "html_url": "https://github.com/huggingface/transformers/pull/27896", "diff_url": "https://github.com/huggingface/transformers/pull/27896.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27896.patch", "merged_at": 1702320094000 }

{ "login": "NielsRogge", "id": 48327001, "node_id": "MDQ6VXNlcjQ4MzI3MDAx", "avatar_url": "https://avatars.githubusercontent.com/u/48327001?v=4", "gravatar_id": "", "url": "https://api.github.com/users/NielsRogge", "html_url": "https://github.com/NielsRogge", "followers_url": "https://api.github.com/users/NielsRogge/followers", "following_url": "https://api.github.com/users/NielsRogge/following{/other_user}", "gists_url": "https://api.github.com/users/NielsRogge/gists{/gist_id}", "starred_url": "https://api.github.com/users/NielsRogge/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/NielsRogge/subscriptions", "organizations_url": "https://api.github.com/users/NielsRogge/orgs", "repos_url": "https://api.github.com/users/NielsRogge/repos", "events_url": "https://api.github.com/users/NielsRogge/events{/privacy}", "received_events_url": "https://api.github.com/users/NielsRogge/received_events", "type": "User", "site_admin": false }

[]

[]

[ "They are failing on my setup:\r\n```\r\nFAILED tests/models/llava/test_modeling_llava.py::LlavaForConditionalGenerationIntegrationTest::test_small_model_integration_test_batch - AssertionError: Lists differ: ['USER: \\nWhat are the things I should be [267 chars]ock'] != ['\\nUSER: What are the things I should be c[306 chars] R.']\r\nFAILED tests/models/llava/test_modeling_llava.py::LlavaForConditionalGenerationIntegrationTest::test_small_model_integration_test_llama - AssertionError: 'USER[116 chars]hich appears to be a dock or pier extending ov[562 chars]ier.' != 'USER[116 chars]hich is a pier or dock extending over a body o[572 chars]ies.'\r\nFAILED tests/models/llava/test_modeling_llava.py::LlavaForConditionalGenerationIntegrationTest::test_small_model_integration_test_llama_batched - AssertionError: Lists differ: ['USE[114 chars]ANT: When visiting this serene location, one s[177 chars]ed.'] != ['USE[114 chars]ANT: the water is calm and clear\\n\\nThe image [154 chars]ed.']\r\n```\r\nCould possibly also be the case on main, will check", "Yeah it still fails for me both on main and my branch even after #27909. So @ArthurZucker could you perhaps run the slow tests from my branch before merging? But I'm pretty sure it doesn't affect integration tests. " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27895/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27895", "html_url": "https://github.com/huggingface/transformers/pull/27895", "diff_url": "https://github.com/huggingface/transformers/pull/27895.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27895.patch", "merged_at": 1702286546000 }

{ "login": "Wovchena", "id": 10669582, "node_id": "MDQ6VXNlcjEwNjY5NTgy", "avatar_url": "https://avatars.githubusercontent.com/u/10669582?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Wovchena", "html_url": "https://github.com/Wovchena", "followers_url": "https://api.github.com/users/Wovchena/followers", "following_url": "https://api.github.com/users/Wovchena/following{/other_user}", "gists_url": "https://api.github.com/users/Wovchena/gists{/gist_id}", "starred_url": "https://api.github.com/users/Wovchena/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Wovchena/subscriptions", "organizations_url": "https://api.github.com/users/Wovchena/orgs", "repos_url": "https://api.github.com/users/Wovchena/repos", "events_url": "https://api.github.com/users/Wovchena/events{/privacy}", "received_events_url": "https://api.github.com/users/Wovchena/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey, both padding tokens are not `non-printable` you just decided to skip them using `skip_special_tokens = True`. \r\nYou should try to set the padding to the left as it is recommended for generation and the padding token on the right will always impact the model. more details [here](https://github.com/huggingface/transformers/issues/25420#issuecomment-1775317535) ", "Hi. `skip_special_tokens=True` is intentional. How do I modify the reproducer to set the padding to the left?", "`tokenizer.padding_side = \"left\"`", "The results are still different:\r\n```py\r\nimport transformers\r\n\r\ntokenizer = transformers.LlamaTokenizer.from_pretrained('TinyLlama/TinyLlama-1.1B-Chat-v0.6')\r\ntokenizer.padding_side = \"left\"\r\ninput_ids = tokenizer('Hi', return_tensors='pt')['input_ids']\r\nmodel = transformers.LlamaForCausalLM.from_pretrained('TinyLlama/TinyLlama-1.1B-Chat-v0.6')\r\nprint(model.generation_config.eos_token_id)\r\npad_token_id = 0\r\nassert pad_token_id != model.generation_config.eos_token_id\r\nzero = [tokenizer.decode(beam, skip_special_tokens=True) for beam in model.generate(input_ids, max_new_tokens=25, num_beam_groups=9, num_beams=99, num_return_sequences=99, diversity_penalty=1.0, no_repeat_ngram_size=3, do_sample=False, pad_token_id=pad_token_id)]\r\nprint(zero)\r\npad_token_id = 1\r\nassert pad_token_id != model.generation_config.eos_token_id\r\none = [tokenizer.decode(beam, skip_special_tokens=True) for beam in model.generate(input_ids, max_new_tokens=25, num_beam_groups=9, num_beams=99, num_return_sequences=99, diversity_penalty=1.0, no_repeat_ngram_size=3, do_sample=False, pad_token_id=pad_token_id)]\r\nprint(one)\r\nassert zero == one\r\n```\r\nMy explanation why your suggestion didn't help is that the problem I'm referring to lies inside of `model.generate()`, which doesn't involve `tokenizer`.", "Hi @Wovchena 👋 \r\n\r\nThe issue with your script was that you were not passing the attention mask to `generate` :) \r\n\r\nThroughout transformers, we do our best effort to infer the attention mask when it is not passed: if the token is equal to the pad token the attention mask is 0 (and 1 otherwise). In your particular example, you were setting the pad token to the bos token (both `1`), the token that the tokenizer uses to signal the beginning of the sequence. As such, the inferred attention mask was different in the two inputs, leading to a different output.\r\n\r\nWe always recommend passing the attention mask 🤗 \r\n\r\nWorking example (passing the attention mask):\r\n```py\r\nimport transformers\r\n\r\ntokenizer = transformers.LlamaTokenizer.from_pretrained('TinyLlama/TinyLlama-1.1B-Chat-v0.6')\r\nmodel = transformers.LlamaForCausalLM.from_pretrained('TinyLlama/TinyLlama-1.1B-Chat-v0.6')\r\n\r\ntokenizer.padding_side = \"left\"\r\ninput_ids = tokenizer('Hi', return_tensors='pt')\r\nprint(model.generation_config.eos_token_id)\r\n\r\npad_token_id = 0\r\nassert pad_token_id != model.generation_config.eos_token_id\r\ngen_zero = model.generate(**input_ids, max_new_tokens=25, num_beam_groups=9, num_beams=99, num_return_sequences=99, diversity_penalty=1.0, no_repeat_ngram_size=3, do_sample=False, pad_token_id=pad_token_id)\r\nzero = [tokenizer.decode(beam, skip_special_tokens=True) for beam in gen_zero]\r\n\r\npad_token_id = 1\r\nassert pad_token_id != model.generation_config.eos_token_id\r\ngen_one = model.generate(**input_ids, max_new_tokens=25, num_beam_groups=9, num_beams=99, num_return_sequences=99, diversity_penalty=1.0, no_repeat_ngram_size=3, do_sample=False, pad_token_id=pad_token_id)\r\none = [tokenizer.decode(beam, skip_special_tokens=True) for beam in gen_one]\r\n\r\nassert zero == one\r\n```", "Thank you!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27894/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "asusdisciple", "id": 138434950, "node_id": "U_kgDOCEBZhg", "avatar_url": "https://avatars.githubusercontent.com/u/138434950?v=4", "gravatar_id": "", "url": "https://api.github.com/users/asusdisciple", "html_url": "https://github.com/asusdisciple", "followers_url": "https://api.github.com/users/asusdisciple/followers", "following_url": "https://api.github.com/users/asusdisciple/following{/other_user}", "gists_url": "https://api.github.com/users/asusdisciple/gists{/gist_id}", "starred_url": "https://api.github.com/users/asusdisciple/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/asusdisciple/subscriptions", "organizations_url": "https://api.github.com/users/asusdisciple/orgs", "repos_url": "https://api.github.com/users/asusdisciple/repos", "events_url": "https://api.github.com/users/asusdisciple/events{/privacy}", "received_events_url": "https://api.github.com/users/asusdisciple/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey could you share a full reproducer, with a traceback and the output of `transformers-cli env`", "```\r\nfrom transformers import pipeline\r\nimport torch\r\n\r\nmodel = \"openai/whisper-large\"\r\n\r\nobj = pipeline(model=model,\r\n torch_dtype=torch.float16,\r\n device=\"cuda:0\", # or mps for Mac devices\r\n chunk_length_s=30,\r\n batch_size=8,\r\n return_timestamps=False,\r\n model_kwargs={\"use_flash_attention_2\": False},\r\n\r\n )\r\nobj.model = obj.model.to_bettertransformer()\r\n\r\naudio = obj(\"/.../csmd024.wav\", generate_kwargs={\"language\": \"yue\"})\r\n\r\nprint(audio)\r\n```\r\n\r\nError log:\r\n```\r\nSpecial tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.\r\nThe BetterTransformer implementation does not support padding during training, as the fused kernels do not support attention masks. Beware that passing padded batched data during training may result in unexpected outputs. Please refer to https://huggingface.co./docs/optimum/bettertransformer/overview for more details.\r\nTraceback (most recent call last):\r\n File \"/usr/lib/python3.10/contextlib.py\", line 153, in __exit__\r\n self.gen.throw(typ, value, traceback)\r\n File \"/.../Whisper/venv/lib/python3.10/site-packages/transformers/pipelines/base.py\", line 924, in device_placement\r\n yield\r\n File \"/.../Whisper/venv/lib/python3.10/site-packages/transformers/pipelines/base.py\", line 1046, in forward\r\n model_outputs = self._forward(model_inputs, **forward_params)\r\n File \"/.../Whisper/venv/lib/python3.10/site-packages/transformers/pipelines/automatic_speech_recognition.py\", line 569, in _forward\r\n tokens = self.model.generate(\r\n File \"/.../Whisper/venv/lib/python3.10/site-packages/transformers/models/whisper/modeling_whisper.py\", line 2050, in generate\r\n forced_decoder_ids.append((1, generation_config.lang_to_id[language_token]))\r\nKeyError: '<|yue|>'\r\n```\r\n", "That is expected you are not using `whisper-large-v3`. Other version do not support this language", "I think the behaviour is still strange because usually if the language is not supported it gives you an error with the languages that are supported.", "That's because it's supported in the actual code but not in the lang to id. But for sure the error message should be improved! ", "The issue is with this:\r\n```python \r\n if generation_config.language in generation_config.lang_to_id.keys():\r\n language_token = generation_config.language\r\n elif generation_config.language in TO_LANGUAGE_CODE.keys():\r\n language_token = f\"<|{TO_LANGUAGE_CODE[generation_config.language]}|>\"\r\n elif generation_config.language in TO_LANGUAGE_CODE.values():\r\n language_token = f\"<|{generation_config.language}|>\"\r\n```\r\nthe language token is in the supported language but not in the lang to id. " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27893/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "ferjorosa", "id": 24965845, "node_id": "MDQ6VXNlcjI0OTY1ODQ1", "avatar_url": "https://avatars.githubusercontent.com/u/24965845?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ferjorosa", "html_url": "https://github.com/ferjorosa", "followers_url": "https://api.github.com/users/ferjorosa/followers", "following_url": "https://api.github.com/users/ferjorosa/following{/other_user}", "gists_url": "https://api.github.com/users/ferjorosa/gists{/gist_id}", "starred_url": "https://api.github.com/users/ferjorosa/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ferjorosa/subscriptions", "organizations_url": "https://api.github.com/users/ferjorosa/orgs", "repos_url": "https://api.github.com/users/ferjorosa/repos", "events_url": "https://api.github.com/users/ferjorosa/events{/privacy}", "received_events_url": "https://api.github.com/users/ferjorosa/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@ArthurZucker do you know which steps are left in the process?" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27892/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27892", "html_url": "https://github.com/huggingface/transformers/pull/27892", "diff_url": "https://github.com/huggingface/transformers/pull/27892.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27892.patch", "merged_at": 1704454570000 }

{ "login": "pacman100", "id": 13534540, "node_id": "MDQ6VXNlcjEzNTM0NTQw", "avatar_url": "https://avatars.githubusercontent.com/u/13534540?v=4", "gravatar_id": "", "url": "https://api.github.com/users/pacman100", "html_url": "https://github.com/pacman100", "followers_url": "https://api.github.com/users/pacman100/followers", "following_url": "https://api.github.com/users/pacman100/following{/other_user}", "gists_url": "https://api.github.com/users/pacman100/gists{/gist_id}", "starred_url": "https://api.github.com/users/pacman100/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/pacman100/subscriptions", "organizations_url": "https://api.github.com/users/pacman100/orgs", "repos_url": "https://api.github.com/users/pacman100/repos", "events_url": "https://api.github.com/users/pacman100/events{/privacy}", "received_events_url": "https://api.github.com/users/pacman100/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27891). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27891/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27891", "html_url": "https://github.com/huggingface/transformers/pull/27891", "diff_url": "https://github.com/huggingface/transformers/pull/27891.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27891.patch", "merged_at": 1702735903000 }

{ "login": "aaronjimv", "id": 67152883, "node_id": "MDQ6VXNlcjY3MTUyODgz", "avatar_url": "https://avatars.githubusercontent.com/u/67152883?v=4", "gravatar_id": "", "url": "https://api.github.com/users/aaronjimv", "html_url": "https://github.com/aaronjimv", "followers_url": "https://api.github.com/users/aaronjimv/followers", "following_url": "https://api.github.com/users/aaronjimv/following{/other_user}", "gists_url": "https://api.github.com/users/aaronjimv/gists{/gist_id}", "starred_url": "https://api.github.com/users/aaronjimv/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/aaronjimv/subscriptions", "organizations_url": "https://api.github.com/users/aaronjimv/orgs", "repos_url": "https://api.github.com/users/aaronjimv/repos", "events_url": "https://api.github.com/users/aaronjimv/events{/privacy}", "received_events_url": "https://api.github.com/users/aaronjimv/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hello. I am open to any feedback.", "Hi @stevhliu, thanks!", "Hi @osanseviero, thanks for the help. Please let me know anything else." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27890/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27890", "html_url": "https://github.com/huggingface/transformers/pull/27890", "diff_url": "https://github.com/huggingface/transformers/pull/27890.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27890.patch", "merged_at": 1702060338000 }

{ "login": "ydshieh", "id": 2521628, "node_id": "MDQ6VXNlcjI1MjE2Mjg=", "avatar_url": "https://avatars.githubusercontent.com/u/2521628?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ydshieh", "html_url": "https://github.com/ydshieh", "followers_url": "https://api.github.com/users/ydshieh/followers", "following_url": "https://api.github.com/users/ydshieh/following{/other_user}", "gists_url": "https://api.github.com/users/ydshieh/gists{/gist_id}", "starred_url": "https://api.github.com/users/ydshieh/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ydshieh/subscriptions", "organizations_url": "https://api.github.com/users/ydshieh/orgs", "repos_url": "https://api.github.com/users/ydshieh/repos", "events_url": "https://api.github.com/users/ydshieh/events{/privacy}", "received_events_url": "https://api.github.com/users/ydshieh/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27889/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27889", "html_url": "https://github.com/huggingface/transformers/pull/27889", "diff_url": "https://github.com/huggingface/transformers/pull/27889.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27889.patch", "merged_at": 1702043729000 }

{ "login": "dblakely", "id": 20539855, "node_id": "MDQ6VXNlcjIwNTM5ODU1", "avatar_url": "https://avatars.githubusercontent.com/u/20539855?v=4", "gravatar_id": "", "url": "https://api.github.com/users/dblakely", "html_url": "https://github.com/dblakely", "followers_url": "https://api.github.com/users/dblakely/followers", "following_url": "https://api.github.com/users/dblakely/following{/other_user}", "gists_url": "https://api.github.com/users/dblakely/gists{/gist_id}", "starred_url": "https://api.github.com/users/dblakely/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/dblakely/subscriptions", "organizations_url": "https://api.github.com/users/dblakely/orgs", "repos_url": "https://api.github.com/users/dblakely/repos", "events_url": "https://api.github.com/users/dblakely/events{/privacy}", "received_events_url": "https://api.github.com/users/dblakely/received_events", "type": "User", "site_admin": false }

[]

[]

[ "On going through the source code for `add_special_tokens` in the link [here](https://github.com/huggingface/transformers/blob/main/src/transformers/tokenization_utils_base.py#L873), it seems like this function has a param `replace_additional_special_tokens` which defaults to `True`.\r\n\r\nThis is an excerpt from the docstring for this param:\r\n\r\nIf `True`, the existing list of additional special tokens will be replaced by the list provided in `special_tokens_dict`. Otherwise, `self._additional_special_tokens` is just extended. In the former case, the tokens will NOT be removed from the tokenizer's full vocabulary - they are only being flagged as non-special tokens.\r\n\r\nSo, what happened is after you added the same token again, as you expected, it wasn't added, which is why an empty list was outputted.\r\n\r\nSo, it seems like what you wanna do is explicitly set this param to `False`.\r\n\r\nAlso, you can execute `print(tokenizer_2)` to verify that the token you added prior to saving is retained. This is what I see using your code sample:\r\n\r\n```\r\nLlamaTokenizer(name_or_path='tokenizer_1', vocab_size=32000, model_max_length=1000000000000000019884624838656, is_fast=False, padding_side='right', truncation_side='right', special_tokens={'bos_token': '<s>', 'eos_token': '</s>', 'unk_token': '<unk>'}, clean_up_tokenization_spaces=False), added_tokens_decoder={\r\n\t0: AddedToken(\"<unk>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\r\n\t1: AddedToken(\"<s>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\r\n\t2: AddedToken(\"</s>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\r\n\t32000: AddedToken(\"<tok>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\r\n}\r\n```\r\nClearly your special token is retained after loading.", "I actually discovered that as well and that is how I solved the problem on my end. But I still felt like the situation seemed like a bug--they initialize an empty set of additional special tokens but then don't end up adding a special token that already exists. And so the set remains empty, which in effect removes the special token from the set. My guess is that this wasn't intended.", "Yep, I can reproduce and that's indeed a good catch. I have no idea why this is happening yet, opening a pr now 🤗 ", "Sorry! Christmas delays 🤗 making sure this is fixed for next release " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27888/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "fxmarty", "id": 9808326, "node_id": "MDQ6VXNlcjk4MDgzMjY=", "avatar_url": "https://avatars.githubusercontent.com/u/9808326?v=4", "gravatar_id": "", "url": "https://api.github.com/users/fxmarty", "html_url": "https://github.com/fxmarty", "followers_url": "https://api.github.com/users/fxmarty/followers", "following_url": "https://api.github.com/users/fxmarty/following{/other_user}", "gists_url": "https://api.github.com/users/fxmarty/gists{/gist_id}", "starred_url": "https://api.github.com/users/fxmarty/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/fxmarty/subscriptions", "organizations_url": "https://api.github.com/users/fxmarty/orgs", "repos_url": "https://api.github.com/users/fxmarty/repos", "events_url": "https://api.github.com/users/fxmarty/events{/privacy}", "received_events_url": "https://api.github.com/users/fxmarty/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Strange that we don't see the issue on our CI. Do you know why @fxmarty?", "@ydshieh Can you try `RUN_SLOW=1 CUDA_VISIBLE_DEVICES=0 pytest tests/models/llama -s -vvvvv -k \"test_resize_tokens_embeddings\"` on main?", "I am not sure it's weird. The `push` tests should run on GPU so those should have failed...\r\nhttps://github.com/huggingface/transformers/blob/fc71e815f6ff5fc8b743786e2146cff9644ec598/.github/workflows/self-push.yml#L133", "As discussed offline it did actually fail (https://github.com/huggingface/transformers/actions/runs/7124821720) so all good!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27887/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27887", "html_url": "https://github.com/huggingface/transformers/pull/27887", "diff_url": "https://github.com/huggingface/transformers/pull/27887.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27887.patch", "merged_at": 1701952484000 }

{ "login": "ArthurZucker", "id": 48595927, "node_id": "MDQ6VXNlcjQ4NTk1OTI3", "avatar_url": "https://avatars.githubusercontent.com/u/48595927?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ArthurZucker", "html_url": "https://github.com/ArthurZucker", "followers_url": "https://api.github.com/users/ArthurZucker/followers", "following_url": "https://api.github.com/users/ArthurZucker/following{/other_user}", "gists_url": "https://api.github.com/users/ArthurZucker/gists{/gist_id}", "starred_url": "https://api.github.com/users/ArthurZucker/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ArthurZucker/subscriptions", "organizations_url": "https://api.github.com/users/ArthurZucker/orgs", "repos_url": "https://api.github.com/users/ArthurZucker/repos", "events_url": "https://api.github.com/users/ArthurZucker/events{/privacy}", "received_events_url": "https://api.github.com/users/ArthurZucker/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27886/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27886", "html_url": "https://github.com/huggingface/transformers/pull/27886", "diff_url": "https://github.com/huggingface/transformers/pull/27886.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27886.patch", "merged_at": null }

{ "login": "rajveer43", "id": 64583161, "node_id": "MDQ6VXNlcjY0NTgzMTYx", "avatar_url": "https://avatars.githubusercontent.com/u/64583161?v=4", "gravatar_id": "", "url": "https://api.github.com/users/rajveer43", "html_url": "https://github.com/rajveer43", "followers_url": "https://api.github.com/users/rajveer43/followers", "following_url": "https://api.github.com/users/rajveer43/following{/other_user}", "gists_url": "https://api.github.com/users/rajveer43/gists{/gist_id}", "starred_url": "https://api.github.com/users/rajveer43/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/rajveer43/subscriptions", "organizations_url": "https://api.github.com/users/rajveer43/orgs", "repos_url": "https://api.github.com/users/rajveer43/repos", "events_url": "https://api.github.com/users/rajveer43/events{/privacy}", "received_events_url": "https://api.github.com/users/rajveer43/received_events", "type": "User", "site_admin": false }

[ { "id": 2796628563, "node_id": "MDU6TGFiZWwyNzk2NjI4NTYz", "url": "https://api.github.com/repos/huggingface/transformers/labels/WIP", "name": "WIP", "color": "234C99", "default": false, "description": "Label your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progress" } ]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27885/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "rajveer43", "id": 64583161, "node_id": "MDQ6VXNlcjY0NTgzMTYx", "avatar_url": "https://avatars.githubusercontent.com/u/64583161?v=4", "gravatar_id": "", "url": "https://api.github.com/users/rajveer43", "html_url": "https://github.com/rajveer43", "followers_url": "https://api.github.com/users/rajveer43/followers", "following_url": "https://api.github.com/users/rajveer43/following{/other_user}", "gists_url": "https://api.github.com/users/rajveer43/gists{/gist_id}", "starred_url": "https://api.github.com/users/rajveer43/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/rajveer43/subscriptions", "organizations_url": "https://api.github.com/users/rajveer43/orgs", "repos_url": "https://api.github.com/users/rajveer43/repos", "events_url": "https://api.github.com/users/rajveer43/events{/privacy}", "received_events_url": "https://api.github.com/users/rajveer43/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@stevhliu, I built docs locally, it should pass now.", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27884). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "Did the final change as well, as we discussed on the mail, can you open a separate issue for a model doc files with listing of the `model_docs` files that are yet to be translated? " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27884/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27884", "html_url": "https://github.com/huggingface/transformers/pull/27884", "diff_url": "https://github.com/huggingface/transformers/pull/27884.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27884.patch", "merged_at": 1702490550000 }

{ "login": "ArthurZucker", "id": 48595927, "node_id": "MDQ6VXNlcjQ4NTk1OTI3", "avatar_url": "https://avatars.githubusercontent.com/u/48595927?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ArthurZucker", "html_url": "https://github.com/ArthurZucker", "followers_url": "https://api.github.com/users/ArthurZucker/followers", "following_url": "https://api.github.com/users/ArthurZucker/following{/other_user}", "gists_url": "https://api.github.com/users/ArthurZucker/gists{/gist_id}", "starred_url": "https://api.github.com/users/ArthurZucker/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ArthurZucker/subscriptions", "organizations_url": "https://api.github.com/users/ArthurZucker/orgs", "repos_url": "https://api.github.com/users/ArthurZucker/repos", "events_url": "https://api.github.com/users/ArthurZucker/events{/privacy}", "received_events_url": "https://api.github.com/users/ArthurZucker/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27883). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "Hi~ o(*￣▽￣*)ブ Could you tell me what your plans? May I know when the PR can be merged? I have PRs depended on this feature. ", "@younesbelkada and I just came back from holidays, we are hoping end of the week maybe later! ", "cc @ArthurZucker could you give a first pass ? 🙏 " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27883/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27883", "html_url": "https://github.com/huggingface/transformers/pull/27883", "diff_url": "https://github.com/huggingface/transformers/pull/27883.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27883.patch", "merged_at": null }

{ "login": "rajveer43", "id": 64583161, "node_id": "MDQ6VXNlcjY0NTgzMTYx", "avatar_url": "https://avatars.githubusercontent.com/u/64583161?v=4", "gravatar_id": "", "url": "https://api.github.com/users/rajveer43", "html_url": "https://github.com/rajveer43", "followers_url": "https://api.github.com/users/rajveer43/followers", "following_url": "https://api.github.com/users/rajveer43/following{/other_user}", "gists_url": "https://api.github.com/users/rajveer43/gists{/gist_id}", "starred_url": "https://api.github.com/users/rajveer43/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/rajveer43/subscriptions", "organizations_url": "https://api.github.com/users/rajveer43/orgs", "repos_url": "https://api.github.com/users/rajveer43/repos", "events_url": "https://api.github.com/users/rajveer43/events{/privacy}", "received_events_url": "https://api.github.com/users/rajveer43/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27882/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27882", "html_url": "https://github.com/huggingface/transformers/pull/27882", "diff_url": "https://github.com/huggingface/transformers/pull/27882.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27882.patch", "merged_at": null }

{ "login": "ydshieh", "id": 2521628, "node_id": "MDQ6VXNlcjI1MjE2Mjg=", "avatar_url": "https://avatars.githubusercontent.com/u/2521628?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ydshieh", "html_url": "https://github.com/ydshieh", "followers_url": "https://api.github.com/users/ydshieh/followers", "following_url": "https://api.github.com/users/ydshieh/following{/other_user}", "gists_url": "https://api.github.com/users/ydshieh/gists{/gist_id}", "starred_url": "https://api.github.com/users/ydshieh/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ydshieh/subscriptions", "organizations_url": "https://api.github.com/users/ydshieh/orgs", "repos_url": "https://api.github.com/users/ydshieh/repos", "events_url": "https://api.github.com/users/ydshieh/events{/privacy}", "received_events_url": "https://api.github.com/users/ydshieh/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27881). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27881/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27881", "html_url": "https://github.com/huggingface/transformers/pull/27881", "diff_url": "https://github.com/huggingface/transformers/pull/27881.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27881.patch", "merged_at": 1701958171000 }

{ "login": "hengjiUSTC", "id": 25971665, "node_id": "MDQ6VXNlcjI1OTcxNjY1", "avatar_url": "https://avatars.githubusercontent.com/u/25971665?v=4", "gravatar_id": "", "url": "https://api.github.com/users/hengjiUSTC", "html_url": "https://github.com/hengjiUSTC", "followers_url": "https://api.github.com/users/hengjiUSTC/followers", "following_url": "https://api.github.com/users/hengjiUSTC/following{/other_user}", "gists_url": "https://api.github.com/users/hengjiUSTC/gists{/gist_id}", "starred_url": "https://api.github.com/users/hengjiUSTC/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/hengjiUSTC/subscriptions", "organizations_url": "https://api.github.com/users/hengjiUSTC/orgs", "repos_url": "https://api.github.com/users/hengjiUSTC/repos", "events_url": "https://api.github.com/users/hengjiUSTC/events{/privacy}", "received_events_url": "https://api.github.com/users/hengjiUSTC/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey! This is a duplicate of a lot of issue, the eos token is set to `AddedToken(content = \"</s>\", normalized = True,...)`. This will work:\r\n```python \r\nfrom transformers import AutoTokenizer, AddedToken \r\ntokenizer = AutoTokenizer.from_pretrained(\"NousResearch/Llama-2-7b-chat-hf\", eos_token = AddedToken(\"</s>\", normalized = False, special=True))\r\n```" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27880/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "grahamannett", "id": 7343667, "node_id": "MDQ6VXNlcjczNDM2Njc=", "avatar_url": "https://avatars.githubusercontent.com/u/7343667?v=4", "gravatar_id": "", "url": "https://api.github.com/users/grahamannett", "html_url": "https://github.com/grahamannett", "followers_url": "https://api.github.com/users/grahamannett/followers", "following_url": "https://api.github.com/users/grahamannett/following{/other_user}", "gists_url": "https://api.github.com/users/grahamannett/gists{/gist_id}", "starred_url": "https://api.github.com/users/grahamannett/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/grahamannett/subscriptions", "organizations_url": "https://api.github.com/users/grahamannett/orgs", "repos_url": "https://api.github.com/users/grahamannett/repos", "events_url": "https://api.github.com/users/grahamannett/events{/privacy}", "received_events_url": "https://api.github.com/users/grahamannett/received_events", "type": "User", "site_admin": false }

[ { "id": 3081136536, "node_id": "MDU6TGFiZWwzMDgxMTM2NTM2", "url": "https://api.github.com/repos/huggingface/transformers/labels/Good%20Difficult%20Issue", "name": "Good Difficult Issue", "color": "684CC7", "default": false, "description": "" } ]

[]

[ "if anyone runs into this, its because of the `if end != start` then `continue` will just keep going. both these functions should be rewritten ", "cc @molbap for visibility, and @grahamannett if you can share a full reproducer would be helpful! ", "@ArthurZucker its hard to reproduce since the encoding throws an error if you try to create a similar input but here:\r\n\r\n```python\r\nprocessor = AutoProcessor.from_pretrained(\"adept/fuyu-8b\")\r\nx = torch.tensor([[71011, 71019, 1, 138997, 75694, 71374, 71118, 70004, 70050, 70013, 70004, 70013, 71119, 71122]])\r\noutput = processor.post_process_box_coordinates(x)\r\n```\r\n\r\nthe tokens are supposed to resemble something like the model outputting 'here is the example <box>1, 2, 11, 12, 13</box>' (or something similar think x in my example is 2's + 21).\r\n\r\nI closed the issue tho so think we are all good 👍", "Yep thanks I was able to reproduce. It's a bit of an edge case, but we could / should have a check for this. Do you want to open a PR for that? 🤗 ", "@ArthurZucker Its not really an edge case as the generated output from the model can actually generate this sort of output and anecdotally seems like it becomes much more common if you further finetune the model on OCR related tasks/instructions. \r\n\r\nI do not think the PR I make will be merged so not sure if I should PR (hence closed this issue).", "If this should be fix, we'll merge the PR that fixed it! We'll iterate together on the PR if you want! \r\nOpening this again and marking as good difficult issue ", "I think a lot of the current FuyuProcessor and FuyuImageProcessor (the image processor preprocess/preprocess_with_tokenizer_info I believe has some big issues and unless I am told it is how they trained it, it will make fine-tuning with images not really work) are broken/not working correctly. \r\n\r\nI am just going to close this as it seems like the best way to fix this would be to use a different implementation.", "It is how it was trained / the code that was shared by the authors 🤗 ", "@ArthurZucker interesting. There are a lot of parts in it that seem to to not work/be confusing to me e.g. how `get_sample_encoding` https://github.com/huggingface/transformers/blob/7e0ddf89f483f53107870cddabb2e1cc93069705/src/transformers/models/fuyu/processing_fuyu.py#L385 is done makes me think that because of https://github.com/huggingface/transformers/blob/7e0ddf89f483f53107870cddabb2e1cc93069705/src/transformers/models/fuyu/processing_fuyu.py#L246 the processor is not really able to be used for training? Or why `_tokenize_prompts_with_image_and_batch` adds `\"|ENDOFTEXT|\"` token to a text prompt even though it then is removed by just using the original prompt size?\r\n" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27879/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "Aria-K-Alethia", "id": 22995879, "node_id": "MDQ6VXNlcjIyOTk1ODc5", "avatar_url": "https://avatars.githubusercontent.com/u/22995879?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Aria-K-Alethia", "html_url": "https://github.com/Aria-K-Alethia", "followers_url": "https://api.github.com/users/Aria-K-Alethia/followers", "following_url": "https://api.github.com/users/Aria-K-Alethia/following{/other_user}", "gists_url": "https://api.github.com/users/Aria-K-Alethia/gists{/gist_id}", "starred_url": "https://api.github.com/users/Aria-K-Alethia/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Aria-K-Alethia/subscriptions", "organizations_url": "https://api.github.com/users/Aria-K-Alethia/orgs", "repos_url": "https://api.github.com/users/Aria-K-Alethia/repos", "events_url": "https://api.github.com/users/Aria-K-Alethia/events{/privacy}", "received_events_url": "https://api.github.com/users/Aria-K-Alethia/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hello, please let us know if the above PR resolves the issue.", "Thank you for your follow up.\r\nI tried the new branch, but it gave another error:\r\n```python\r\nTraceback (most recent call last):\r\n File \"/home/user/lm_finetune/train_tts.py\", line 183, in <module>\r\n train()\r\n File \"/home/user/lm_finetune/train_tts.py\", line 175, in train\r\n trainer.train(resume_from_checkpoint=ckpt_dir)\r\n File \"/opt/conda/envs/alpaca/lib/python3.10/site-packages/transformers/trainer.py\", line 1533, in train\r\n return inner_training_loop(\r\n File \"/opt/conda/envs/alpaca/lib/python3.10/site-packages/transformers/trainer.py\", line 1691, in _inner_training_loop\r\n self._load_optimizer_and_scheduler(resume_from_checkpoint)\r\n File \"/opt/conda/envs/alpaca/lib/python3.10/site-packages/transformers/trainer.py\", line 2518, in _load_optimizer_and_scheduler\r\n load_fsdp_optimizer(\r\n File \"/opt/conda/envs/alpaca/lib/python3.10/site-packages/accelerate/utils/fsdp_utils.py\", line 190, in load_fsdp_optimizer\r\n optimizer.load_state_dict(flattened_osd)\r\n File \"/opt/conda/envs/alpaca/lib/python3.10/site-packages/accelerate/optimizer.py\", line 107, in load_state_dict\r\n self.optimizer.load_state_dict(state_dict)\r\n File \"/opt/conda/envs/alpaca/lib/python3.10/site-packages/torch/_compile.py\", line 24, in inner\r\n return torch._dynamo.disable(fn, recursive)(*args, **kwargs)\r\n File \"/opt/conda/envs/alpaca/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py\", line 328, in _fn\r\n return fn(*args, **kwargs)\r\n File \"/opt/conda/envs/alpaca/lib/python3.10/site-packages/torch/optim/optimizer.py\", line 735, in load_state_dict\r\n raise ValueError(\"loaded state dict contains a parameter group \"\r\nValueError: loaded state dict contains a parameter group that doesn't match the size of optimizer's group\r\n```" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27878/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "cyyever", "id": 17618148, "node_id": "MDQ6VXNlcjE3NjE4MTQ4", "avatar_url": "https://avatars.githubusercontent.com/u/17618148?v=4", "gravatar_id": "", "url": "https://api.github.com/users/cyyever", "html_url": "https://github.com/cyyever", "followers_url": "https://api.github.com/users/cyyever/followers", "following_url": "https://api.github.com/users/cyyever/following{/other_user}", "gists_url": "https://api.github.com/users/cyyever/gists{/gist_id}", "starred_url": "https://api.github.com/users/cyyever/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/cyyever/subscriptions", "organizations_url": "https://api.github.com/users/cyyever/orgs", "repos_url": "https://api.github.com/users/cyyever/repos", "events_url": "https://api.github.com/users/cyyever/events{/privacy}", "received_events_url": "https://api.github.com/users/cyyever/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey, which version of `ruff` are you using? The main CI usually make sure that we have the correct format and it does not seem to be failing. ", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27877). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "> Hey, which version of `ruff` are you using? The main CI usually make sure that we have the correct format and it does not seem to be failing.\r\n\r\nruff 0.1.7", "you need to have 0.1.5 we pinned it for now", ":joy: Any plan to update?", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27877/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27877", "html_url": "https://github.com/huggingface/transformers/pull/27877", "diff_url": "https://github.com/huggingface/transformers/pull/27877.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27877.patch", "merged_at": null }

{ "login": "antoine-lizee", "id": 2957716, "node_id": "MDQ6VXNlcjI5NTc3MTY=", "avatar_url": "https://avatars.githubusercontent.com/u/2957716?v=4", "gravatar_id": "", "url": "https://api.github.com/users/antoine-lizee", "html_url": "https://github.com/antoine-lizee", "followers_url": "https://api.github.com/users/antoine-lizee/followers", "following_url": "https://api.github.com/users/antoine-lizee/following{/other_user}", "gists_url": "https://api.github.com/users/antoine-lizee/gists{/gist_id}", "starred_url": "https://api.github.com/users/antoine-lizee/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/antoine-lizee/subscriptions", "organizations_url": "https://api.github.com/users/antoine-lizee/orgs", "repos_url": "https://api.github.com/users/antoine-lizee/repos", "events_url": "https://api.github.com/users/antoine-lizee/events{/privacy}", "received_events_url": "https://api.github.com/users/antoine-lizee/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hello Antoine,\r\n\r\nThanks for reporting this issue. I don't have the time to look at it in details. I think this is due to the tokenizer not supporting \"real words\" hence falling into the heuristic [here](https://github.com/huggingface/transformers/blob/75336c17945c6b1b5552dbf0236d25f869168aab/src/transformers/pipelines/token_classification.py#L387).\r\n\r\nFor now, I can suggest you to change your model to use a compatible tokenizer. If you need it absolutely, you can get the entity from the text with the offset mapping like `word = text[entity[\"start\"]:entity[\"end\"]]`. For \"Dermatologie\" fused with \".\", I don't think we can do something immediately with this type of tokenizer.\r\n\r\nHave a good day,\r\nLuc\r\n", "Hello Luc,\r\n\r\nI do remember seeing the warning, so you're likely right.\r\nI'm a bit surprised that the heuristic can't do better but I won't have to look in depth either - tbh this was more of a bug report as `simple` is good enough in general and I'm moving to more suitable tokenizers.\r\n\r\nThanks you for your answer, closing as won't fix." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27876/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "atawari", "id": 26859544, "node_id": "MDQ6VXNlcjI2ODU5NTQ0", "avatar_url": "https://avatars.githubusercontent.com/u/26859544?v=4", "gravatar_id": "", "url": "https://api.github.com/users/atawari", "html_url": "https://github.com/atawari", "followers_url": "https://api.github.com/users/atawari/followers", "following_url": "https://api.github.com/users/atawari/following{/other_user}", "gists_url": "https://api.github.com/users/atawari/gists{/gist_id}", "starred_url": "https://api.github.com/users/atawari/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/atawari/subscriptions", "organizations_url": "https://api.github.com/users/atawari/orgs", "repos_url": "https://api.github.com/users/atawari/repos", "events_url": "https://api.github.com/users/atawari/events{/privacy}", "received_events_url": "https://api.github.com/users/atawari/received_events", "type": "User", "site_admin": false }

[]

[]

[ "tagging @amyeroberts as the change is in the vision module.", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27875). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27875/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27875", "html_url": "https://github.com/huggingface/transformers/pull/27875", "diff_url": "https://github.com/huggingface/transformers/pull/27875.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27875.patch", "merged_at": 1702311631000 }

{ "login": "yundai424", "id": 43726198, "node_id": "MDQ6VXNlcjQzNzI2MTk4", "avatar_url": "https://avatars.githubusercontent.com/u/43726198?v=4", "gravatar_id": "", "url": "https://api.github.com/users/yundai424", "html_url": "https://github.com/yundai424", "followers_url": "https://api.github.com/users/yundai424/followers", "following_url": "https://api.github.com/users/yundai424/following{/other_user}", "gists_url": "https://api.github.com/users/yundai424/gists{/gist_id}", "starred_url": "https://api.github.com/users/yundai424/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/yundai424/subscriptions", "organizations_url": "https://api.github.com/users/yundai424/orgs", "repos_url": "https://api.github.com/users/yundai424/repos", "events_url": "https://api.github.com/users/yundai424/events{/privacy}", "received_events_url": "https://api.github.com/users/yundai424/received_events", "type": "User", "site_admin": false }

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored.", "Gently pinging @pacman100 and @younesbelkada ", "i will let @pacman100 answer on this as he is more familiar than I am on DS, @yundai424 can you try on latest transformers and peft ? `pip install -U transformers peft`", "@younesbelkada that gives the same result. Though I found https://huggingface.co./docs/accelerate/usage_guides/deepspeed#saving-and-loading gives a more official solution (i.e. using accelerate `unwrap_model` so if this is the ultimate solution i can close this issue :D" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27874/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "NielsRogge", "id": 48327001, "node_id": "MDQ6VXNlcjQ4MzI3MDAx", "avatar_url": "https://avatars.githubusercontent.com/u/48327001?v=4", "gravatar_id": "", "url": "https://api.github.com/users/NielsRogge", "html_url": "https://github.com/NielsRogge", "followers_url": "https://api.github.com/users/NielsRogge/followers", "following_url": "https://api.github.com/users/NielsRogge/following{/other_user}", "gists_url": "https://api.github.com/users/NielsRogge/gists{/gist_id}", "starred_url": "https://api.github.com/users/NielsRogge/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/NielsRogge/subscriptions", "organizations_url": "https://api.github.com/users/NielsRogge/orgs", "repos_url": "https://api.github.com/users/NielsRogge/repos", "events_url": "https://api.github.com/users/NielsRogge/events{/privacy}", "received_events_url": "https://api.github.com/users/NielsRogge/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Yep I like this optimisation, non breaking overall ", "@ArthurZucker it is a breaking change in its current state, since `out_indices` currently defaults to the last stage index if the user doesn't specify them (think @amyeroberts added that [here](https://github.com/huggingface/transformers/blob/0410a29a2d5c798b2c0c1ca28398e0ddcf3384f2/src/transformers/utils/backbone_utils.py#L77-L79)). So if we were to add this with backwards compatibility, we would have to update the default `out_indices` to all stages in case they are not specified.", "We can set it to `-1` to return everything maybe but I mean we can make it BC! ", "I'd like to have @amyeroberts's opinion on this one", "> it is a breaking change in its current state, since out_indices currently defaults to the last stage index if the user doesn't specify them (think @amyeroberts added that [here](https://github.com/huggingface/transformers/blob/0410a29a2d5c798b2c0c1ca28398e0ddcf3384f2/src/transformers/utils/backbone_utils.py#L77-L79)).\r\n\r\nThis was just matching the logic that was originally implemented for the `out_features` (selecting the last layer). As you added this @NielsRogge you'll know the motivation for this better than me :) \r\n\r\nAs it stands this, I'm not in favour of this as this requires adding in backbone API / logic into standard model APIs. This is essentially making things leaky: why do I need to know about `out_indices` to get my hidden states if I'm not loading a backbone?\r\n\r\nMoreover, this is going to break a tonne of stuff, as users who have created checkpoints which are not backbones will still have `out_indices` set in the model config. This isn't easy to rectify: how would we know if the values in the config are what the user wanted e.g. just the last hidden state, or it just happened to be the default when the config was created? \r\n\r\nIt introduces inconsistencies in our models forward passes, which makes the code harder to understand and is tying non-backbone logic to an API which still isn't 100% stable at the moment. \r\n\r\nAn alternative approach would be to have a different argument in the config which defaults to all the layers but then can be overridden by the config's `out_indices` when loading a backbone. ", "Actually, not another config parameter - because then the source of truth isn't clear and behaviour for the user can be unexpected.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27873/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27873", "html_url": "https://github.com/huggingface/transformers/pull/27873", "diff_url": "https://github.com/huggingface/transformers/pull/27873.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27873.patch", "merged_at": null }

{ "login": "bofenghuang", "id": 38185248, "node_id": "MDQ6VXNlcjM4MTg1MjQ4", "avatar_url": "https://avatars.githubusercontent.com/u/38185248?v=4", "gravatar_id": "", "url": "https://api.github.com/users/bofenghuang", "html_url": "https://github.com/bofenghuang", "followers_url": "https://api.github.com/users/bofenghuang/followers", "following_url": "https://api.github.com/users/bofenghuang/following{/other_user}", "gists_url": "https://api.github.com/users/bofenghuang/gists{/gist_id}", "starred_url": "https://api.github.com/users/bofenghuang/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/bofenghuang/subscriptions", "organizations_url": "https://api.github.com/users/bofenghuang/orgs", "repos_url": "https://api.github.com/users/bofenghuang/repos", "events_url": "https://api.github.com/users/bofenghuang/events{/privacy}", "received_events_url": "https://api.github.com/users/bofenghuang/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Kindly ping @muellerzr ", "Hi @muellerzr, thanks for the review! Just updated other no_trainer scripts which handle gradient accumulation by `with accelerator.accumulate(model)`", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27872/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27872", "html_url": "https://github.com/huggingface/transformers/pull/27872", "diff_url": "https://github.com/huggingface/transformers/pull/27872.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27872.patch", "merged_at": 1705933338000 }