[POC] Showed more memory efficient FSDP wrapping #382
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
[ghstack-poisoned]
awgu
added a commit
that referenced
this pull request
Jun 3, 2024
ghstack-source-id: 4c0382b83b7a84d9294c2dd0f15c51d527075cae Pull Request resolved: #382
facebook-github-bot
added
the
CLA Signed
[ghstack-poisoned]
awgu
added a commit
that referenced
this pull request
Jun 3, 2024
ghstack-source-id: 2133694b979250588b0067fa4b19d6c094e06797 Pull Request resolved: #382
awgu
added a commit
to pytorch/pytorch
that referenced
this pull request
Jul 11, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
awgu
added a commit
to pytorch/pytorch
that referenced
this pull request
Jul 11, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
This requires pytorch/pytorch#127786. [ghstack-poisoned]
awgu
added a commit
that referenced
this pull request
Jul 11, 2024
ghstack-source-id: 36a58a9fc9fea8b85783f9b50be96a0684603fdc Pull Request resolved: #382
This requires pytorch/pytorch#127786. **Experiment** - Llama3-8B on 8xH100, 1D FSDP, local batch size 2, selective op AC - With this PR (68.09 GiB reserved memory): ``` [rank0]:2024-07-11 10:48:25,949 - root - INFO - step: 1 loss: 12.2554 memory: 60.27GiB(63.41%) wps: 1,943 mfu: 11.38% [rank0]:2024-07-11 10:48:25,949 - root - INFO - Synchronizing and adjusting timeout for all ProcessGroups to 0:01:40 [rank0]:2024-07-11 10:48:28,506 - root - INFO - step: 2 loss: 12.0724 memory: 68.09GiB(71.65%) wps: 6,408 mfu: 37.53% [rank0]:2024-07-11 10:48:31,063 - root - INFO - step: 3 loss: 11.7467 memory: 68.09GiB(71.65%) wps: 6,410 mfu: 37.53% [rank0]:2024-07-11 10:48:33,621 - root - INFO - step: 4 loss: 11.3360 memory: 68.09GiB(71.65%) wps: 6,406 mfu: 37.51% [rank0]:2024-07-11 10:48:36,182 - root - INFO - step: 5 loss: 10.8909 memory: 68.09GiB(71.65%) wps: 6,399 mfu: 37.47% ``` - Without this PR (69.04 GiB reserved memory): ``` [rank0]:2024-07-11 10:47:04,798 - root - INFO - step: 1 loss: 12.2421 memory: 61.21GiB(64.41%) wps: 1,933 mfu: 11.32% [rank0]:2024-07-11 10:47:04,798 - root - INFO - Synchronizing and adjusting timeout for all ProcessGroups to 0:01:40 [rank0]:2024-07-11 10:47:07,357 - root - INFO - step: 2 loss: 12.0625 memory: 69.04GiB(72.65%) wps: 6,405 mfu: 37.50% [rank0]:2024-07-11 10:47:09,918 - root - INFO - step: 3 loss: 11.7242 memory: 69.04GiB(72.65%) wps: 6,397 mfu: 37.46% [rank0]:2024-07-11 10:47:12,480 - root - INFO - step: 4 loss: 11.3072 memory: 69.04GiB(72.65%) wps: 6,398 mfu: 37.47% [rank0]:2024-07-11 10:47:15,044 - root - INFO - step: 5 loss: 10.8761 memory: 69.04GiB(72.65%) wps: 6,390 mfu: 37.42% ``` [ghstack-poisoned]
awgu
added a commit
that referenced
this pull request
Jul 11, 2024
ghstack-source-id: 9c9e2fef294d9c66ede9f1b8a9ab65510c2ed230 Pull Request resolved: #382
awgu
added a commit
to pytorch/pytorch
that referenced
this pull request
Jul 15, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` To-do: add a test for shared embedding/output projection passed as list (e.g. change `multi_module` from `bool` to `None`, embedding/norm/output, or norm/output) cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
awgu
added a commit
to pytorch/pytorch
that referenced
this pull request
Jul 15, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` To-do: add a test for shared embedding/output projection passed as list (e.g. change `multi_module` from `bool` to `None`, embedding/norm/output, or norm/output) cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
awgu
added a commit
to pytorch/pytorch
that referenced
this pull request
Jul 15, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
awgu
added a commit
to pytorch/pytorch
that referenced
this pull request
Jul 15, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
awgu
added a commit
to pytorch/pytorch
that referenced
this pull request
Jul 15, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
awgu
added a commit
to pytorch/pytorch
that referenced
this pull request
Jul 15, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` cc mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 penguinwu fegin XilunWu wanchaol fduwjj wz337 tianyu-l wconstab yf225 chauhang d4l3k [ghstack-poisoned]
pytorchmergebot
pushed a commit
to pytorch/pytorch
that referenced
this pull request
Jul 15, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` Pull Request resolved: #127786 Approved by: https://github.com/yf225, https://github.com/weifengpy ghstack dependencies: #127773
awgu
added a commit
to pytorch/pytorch
that referenced
this pull request
Jul 17, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` Pull Request resolved: #127786 Approved by: https://github.com/yf225, https://github.com/weifengpy ghstack dependencies: #127773 ghstack-source-id: c70870546bed6108b163062b88415cbba3d37925
awgu
added a commit
to pytorch/pytorch
that referenced
this pull request
Jul 17, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` cc XilunWu H-Huang kwen2501 wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse penguinwu tianyu-l yf225 chauhang [ghstack-poisoned]
awgu
added a commit
to pytorch/pytorch
that referenced
this pull request
Jul 17, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` cc XilunWu H-Huang kwen2501 wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o mrshenli pritamdamania87 zhaojuanmao satgera gqchen aazzolini osalpekar jiayisuse penguinwu tianyu-l yf225 chauhang [ghstack-poisoned]
awgu
added a commit
to pytorch/pytorch
that referenced
this pull request
Jul 17, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` [ghstack-poisoned]
awgu
added a commit
to pytorch/pytorch
that referenced
this pull request
Jul 17, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` ghstack-source-id: c70870546bed6108b163062b88415cbba3d37925 Pull Request resolved: #130949
pytorchmergebot
pushed a commit
to pytorch/pytorch
that referenced
this pull request
Jul 17, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` --- **Changes for reland:** none since breakage was from PR below Pull Request resolved: #130949 Approved by: https://github.com/weifengpy ghstack dependencies: #130947
DiweiSun
pushed a commit
to DiweiSun/pytorch
that referenced
this pull request
Jul 22, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` --- **Changes for reland:** none since breakage was from PR below Pull Request resolved: pytorch#130949 Approved by: https://github.com/weifengpy ghstack dependencies: pytorch#130947
bigfootjon
pushed a commit
to pytorch/pytorch
that referenced
this pull request
Jul 24, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` --- **Changes for reland:** none since breakage was from PR below Pull Request resolved: #130949 Approved by: https://github.com/weifengpy ghstack dependencies: #130947 (cherry picked from commit 31e3330)
xuhancn
pushed a commit
to xuhancn/pytorch
that referenced
this pull request
Jul 25, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` Pull Request resolved: pytorch#127786 Approved by: https://github.com/yf225, https://github.com/weifengpy ghstack dependencies: pytorch#127773
xuhancn
pushed a commit
to xuhancn/pytorch
that referenced
this pull request
Jul 25, 2024
This PR allows `fully_shard`'s first argument to be `List[nn.Module]` instead of strictly `nn.Module`. This allows more flexible grouping of modules/parameters for communication, which can lead to memory savings and/or more efficient communication. **Approach** At a high level, we can think of a model as a tree of modules. Previously, we could only select specific module nodes in this tree as representing one FSDP parameter group. With this PR, we can select a group of module nodes, effectively becoming a single super node. To implement the runtime schedule, we define new forward hooks that run based on the following semantics: - If a module is the first to run the pre-hook, actually run the given pre-hook. Otherwise, the pre-hook is no-op. - If a module is the last to run the post-hook, actually run the given post-hook. Otherwise, the post-hook is a no-op. - First and last are determined by scoreboarding against a set of the modules. - This set must get cleared at the end of backward in the case that >=1 module in the list is never used, in which case we still want the forward hooks to run in the next forward after this backward. Beyond these new forward hooks, everything else is some simple generalization from `Module` to `List[Module]` or `Tuple[Module, ...]`. **Examples** This PR enables wrapping Llama models more efficiently by grouping the final norm and output linear together: pytorch/torchtitan#382. If at least one of the modules in the list does not run forward before backward, then there will be a warning message like: ``` 1 of the 2 modules passed to fully_shard did not run forward before backward, which is error-prone since FSDP post-forward/pre-backward logic will not run for these modules. We recommend passing only modules that run forward together. Modules that did not run forward: [FSDPLinear(in_features=1, out_features=1, bias=True)] ``` --- **Changes for reland:** none since breakage was from PR below Pull Request resolved: pytorch#130949 Approved by: https://github.com/weifengpy ghstack dependencies: pytorch#130947
tianyu-l
force-pushed
the
gh/awgu/6/base
branch
from
d31f16d
to
f336766
Compare
tianyu-l
force-pushed
the
gh/awgu/6/head
branch
from
3406987
to
baad09c
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
This requires pytorch/pytorch#127786.
Experiment
compiled_rmsnorm,torch.compileenabled per transformer block, fused AdamWcompiled_rmsnorm,torch.compileenabled per transformer block, fused AdamWFor some reason, without AC, the new wrapping actually uses more memory. This could be due to memory fragmentation or compile reasons and needs more investigation.