Refactor Distributors to make it acutally do its jobs (#68)

Summary:
Pull Request resolved: #68

In the current Shampoo design, there are three major components: `DistributedShampoo`, which describes the high-level algorithm flow; `Distributor`, which manages the distribution of parameters for different computing paradigms (e.g., DDP, FSDP, HSDP, etc.); and `PreconditionerList`, which contains the detailed algorithm implementation.

The `Distributor` manages the distribution of parameters and then sends them to `DistributedShampoo` which invokes the corresponding algorithms implemented by each `PreconditionerList`.

Since the `Distributor` handles parameter distribution, ideally, downstream classes (i.e., `DistributedShampoo` and `PreconditionerList`) should not need to use `Distributor.distributor_selector` to compress parameters for their use. However, in the current implementation, such usages appear in both [`DistributedShampoo`](https://www.internalfb.com/code/fbsource/[20bcdb4983a16c2cd7f00995754851159da91ed3]/fbcode/hpc/optimizers/distributed_shampoo/dev/distributed_shampoo.py?lines=614-617) and [`PreconditionerList`](https://www.internalfb.com/code/fbsource/[20bcdb4983a16c2cd7f00995754851159da91ed3]/fbcode/hpc/optimizers/distributed_shampoo/dev/utils/shampoo_preconditioner_list.py?lines=145-146).

The main changes are listed as follows:
1. `Distributor` now only provides access to the local version of `params` and `block_info` to align with its responsibilities.
2. `Distributor` no longer needs to store `global_block_info_list`, the global version of `block_info`; instead, it stores `local_block_info_list`.
3. `PreconditionerList` no longer requires `distributor_selector` as an input argument because `Distributor` already performs this task.

Reviewed By: anana10c, chuanhaozhuge

Differential Revision: D64708506

fbshipit-source-id: f1f1221806b64ac9798b79d5f0fcbf2ff523c857

distributed_shampoo/distributed_shampoo.py

@@ -516,10 +516,9 @@ def _instantiate_shampoo_preconditioner_list(self) -> None:
                 )
 
             state_lists[SHAMPOO_PRECONDITIONER_LIST] = preconditioner_list_cls(
-                block_list=state_lists[DISTRIBUTOR].global_blocked_params,
+                block_list=state_lists[DISTRIBUTOR].local_blocked_params,
                 state=self.state,
-                block_info_list=state_lists[DISTRIBUTOR].global_block_info_list,
-                distributor_selector=state_lists[DISTRIBUTOR].distributor_selector,
+                block_info_list=state_lists[DISTRIBUTOR].local_block_info_list,
                 preconditioner_config=group[PRECONDITIONER_CONFIG],
                 beta2=group[BETAS][1],
                 epsilon=group[EPSILON],
@@ -537,18 +536,17 @@ def _instantiate_grafting(self) -> None:
                 state_lists[GRAFTING_PRECONDITIONER_LIST] = None
             elif type(group[GRAFTING_CONFIG]) is SGDGraftingConfig:
                 state_lists[GRAFTING_PRECONDITIONER_LIST] = SGDPreconditionerList(
-                    block_list=state_lists[DISTRIBUTOR].global_blocked_params,
+                    block_list=state_lists[DISTRIBUTOR].local_blocked_params,
                 )
             elif type(group[GRAFTING_CONFIG]) in (
                 AdaGradGraftingConfig,
                 RMSpropGraftingConfig,
                 AdamGraftingConfig,
             ):
                 state_lists[GRAFTING_PRECONDITIONER_LIST] = AdagradPreconditionerList(
-                    block_list=state_lists[DISTRIBUTOR].global_blocked_params,
+                    block_list=state_lists[DISTRIBUTOR].local_blocked_params,
                     state=self.state,
-                    block_info_list=state_lists[DISTRIBUTOR].global_block_info_list,
-                    distributor_selector=state_lists[DISTRIBUTOR].distributor_selector,
+                    block_info_list=state_lists[DISTRIBUTOR].local_block_info_list,
                     beta2=(
                         1.0
                         if type(group[GRAFTING_CONFIG]) is AdaGradGraftingConfig
@@ -565,7 +563,7 @@ def _instantiate_grafting(self) -> None:
     def _instantiate_steps(self) -> None:
         for state_lists in self._per_group_state_lists:
             assert (
-                len(state_lists[DISTRIBUTOR].global_block_info_list) > 0
+                len(state_lists[DISTRIBUTOR].local_block_info_list) > 0
             ), "There is no params in your param_group. Please check the instantiation of DistributedShampoo "
             'with param_group containing no params. For example, DistributedShampoo(params=[{"params": []}])'
             # NOTE: We instantiate a single step tensor on CPU for each group in order
@@ -575,7 +573,7 @@ def _instantiate_steps(self) -> None:
 
             # In order to ensure that the step counter is checkpointed correctly, we store it
             # as a tensor (which is replicated across all devices) under the first parameter's state.
-            block_info = state_lists[DISTRIBUTOR].global_block_info_list[0]
+            block_info = state_lists[DISTRIBUTOR].local_block_info_list[0]
             self.state[block_info.param][STEP] = state_lists[STEP]
 
     @torch.no_grad()
@@ -586,11 +584,11 @@ def _instantiate_momentum(self) -> None:
             if group[MOMENTUM] == 0.0:
                 continue
 
-            # Construct global momentum list.
-            global_momentum_list = []
+            # Construct local momentum list.
+            local_momentum_list = []
             for block, block_info in zip(
-                state_lists[DISTRIBUTOR].global_blocked_params,
-                state_lists[DISTRIBUTOR].global_block_info_list,
+                state_lists[DISTRIBUTOR].local_blocked_params,
+                state_lists[DISTRIBUTOR].local_block_info_list,
                 strict=True,
             ):
                 assert (
@@ -608,15 +606,9 @@ def _instantiate_momentum(self) -> None:
                     dtype=block.dtype,
                     device=block.device,
                 )
-                global_momentum_list.append(
-                    block_info.get_tensor(block_state[MOMENTUM])
-                )
+                local_momentum_list.append(block_state[MOMENTUM])
 
-            # We compress the momentum list to only the locally-owned parameter states.
-            state_lists[MOMENTUM_LIST] = compress_list(
-                global_momentum_list,
-                state_lists[DISTRIBUTOR].distributor_selector,
-            )
+            state_lists[MOMENTUM_LIST] = local_momentum_list
             # Here, we set masked momentum list to momentum list because we assume
             # all parameters are active.
             state_lists[MASKED_MOMENTUM_LIST] = state_lists[MOMENTUM_LIST]
@@ -629,11 +621,11 @@ def _instantiate_filtered_grads(self) -> None:
             if group[BETAS][0] == 0.0:
                 continue
 
-            # Construct global filtered gradient list.
-            global_filtered_grad_list = []
+            # Construct local filtered gradient list.
+            local_filtered_grad_list = []
             for block, block_info in zip(
-                state_lists[DISTRIBUTOR].global_blocked_params,
-                state_lists[DISTRIBUTOR].global_block_info_list,
+                state_lists[DISTRIBUTOR].local_blocked_params,
+                state_lists[DISTRIBUTOR].local_block_info_list,
                 strict=True,
             ):
                 assert (
@@ -651,15 +643,11 @@ def _instantiate_filtered_grads(self) -> None:
                     dtype=block.dtype,
                     device=block.device,
                 )
-                global_filtered_grad_list.append(
+                local_filtered_grad_list.append(
                     block_info.get_tensor(block_state[FILTERED_GRAD])
                 )
 
-            # We compress the momentum list to only the locally-owned parameter states.
-            state_lists[FILTERED_GRAD_LIST] = compress_list(
-                global_filtered_grad_list,
-                state_lists[DISTRIBUTOR].distributor_selector,
-            )
+            state_lists[FILTERED_GRAD_LIST] = local_filtered_grad_list
             # Here, we set masked filtered grad list to filtered grad list because we assume
             # all parameters are active.
             state_lists[MASKED_FILTERED_GRAD_LIST] = state_lists[FILTERED_GRAD_LIST]

distributed_shampoo/utils/shampoo_ddp_distributor.py

@@ -106,12 +106,13 @@ def __init__(
             )
         )
 
-        self._construct_global_block_info_list(buffer_size_ranks)
-
+        global_block_info_list = self._construct_global_block_info_list(
+            buffer_size_ranks
+        )
         # Initialize selectors and local blocked (masked) parameters.
         self._distributor_selector: tuple[bool, ...] = tuple(
             block_info.group_source_rank == group_rank
-            for block_info in self._global_block_info_list
+            for block_info in global_block_info_list
         )
         self._local_blocked_params: tuple[Tensor, ...] = compress_list(
             self._global_blocked_params, self._distributor_selector
@@ -122,6 +123,9 @@ def __init__(
         self._local_grad_selector: tuple[bool, ...] = (True,) * len(
             self._local_blocked_params
         )
+        self._local_block_info_list: tuple[DDPBlockInfo, ...] = compress_list(
+            global_block_info_list, self._distributor_selector
+        )
 
         self._construct_distributed_buffers(
             buffer_size_ranks=buffer_size_ranks,
@@ -257,18 +261,18 @@ def _distribute_buffer_sizes(
 
         return tuple(buffer_size_ranks)
 
+    @torch.no_grad()
     def _construct_global_block_info_list(
         self, buffer_size_ranks: tuple[tuple[int, int], ...]
-    ) -> None:
+    ) -> tuple[DDPBlockInfo, ...]:
         """Construct the global block info list.
 
         Args:
             buffer_size_ranks (tuple[tuple[int, int], ...]): A list of tuples containing the buffer size
                 and an assigned rank for each block.
 
         """
-        # Construct global block info list.
-        self._global_block_info_list: tuple[DDPBlockInfo, ...] = tuple(
+        return tuple(
             DDPBlockInfo(
                 param=param,
                 composable_block_ids=self._construct_composable_block_ids(
@@ -392,7 +396,7 @@ def _construct_distributed_buffers(
         self._global_dist_buffer = torch.zeros(
             total_buffer_size,
             dtype=torch.int8,
-            device=self._global_block_info_list[0].param.device,
+            device=self._global_blocked_params[0].device,
         )
         local_dist_buffers = torch.split(self._global_dist_buffer, max_buffer_size_sum)
         splitted_local_dist_buffers = DDPDistributor._split_local_dist_buffers(

distributed_shampoo/utils/shampoo_distributor.py

@@ -63,8 +63,8 @@ def __init__(self, param_group: dict[str, Any]) -> None:
         self._local_blocked_params: tuple[Tensor, ...]
         # Local masked blocked params are the parameters masked by the distributor selector AND the local grad selector.
         self._local_masked_blocked_params: tuple[Tensor, ...]
-        # Global block info list contains information about each global block.
-        self._global_block_info_list: tuple[BlockInfo, ...] | tuple[DDPBlockInfo, ...]
+        # Local block info list contains information about each block masked by the distributor selector.
+        self._local_block_info_list: tuple[BlockInfo, ...] | tuple[DDPBlockInfo, ...]
 
     @abstractmethod
     @torch.no_grad()
@@ -73,14 +73,6 @@ def update_params(
         masked_blocked_search_directions: tuple[Tensor, ...],
     ) -> None: ...
 
-    @property
-    def global_blocked_params(self) -> tuple[Tensor, ...]:
-        return self._global_blocked_params
-
-    @property
-    def distributor_selector(self) -> tuple[bool, ...]:
-        return self._distributor_selector
-
     @property
     def local_grad_selector(self) -> tuple[bool, ...]:
         return self._local_grad_selector
@@ -94,8 +86,8 @@ def local_masked_blocked_params(self) -> tuple[Tensor, ...]:
         return self._local_masked_blocked_params
 
     @property
-    def global_block_info_list(self) -> tuple[BlockInfo, ...]:
-        return self._global_block_info_list
+    def local_block_info_list(self) -> tuple[BlockInfo, ...]:
+        return self._local_block_info_list
 
     def _construct_composable_block_ids(
         self,
@@ -258,18 +250,18 @@ def __init__(
         param_group: dict[str, Any],
     ) -> None:
         super().__init__(param_group)
-        self._construct_global_block_info_list()
 
         # Initialize selectors and local blocked (masked) parameters.
         self._local_grad_selector: tuple[bool, ...] = (True,) * len(
             self._global_blocked_params
         )
         self._distributor_selector: tuple[bool, ...] = self._local_grad_selector
+        self._local_blocked_params: tuple[Tensor, ...] = self._global_blocked_params
         self._local_masked_blocked_params: tuple[Tensor, ...] = (
-            self._global_blocked_params
+            self._local_blocked_params
         )
-        self._local_blocked_params: tuple[Tensor, ...] = (
-            self._local_masked_blocked_params
+        self._local_block_info_list: tuple[BlockInfo, ...] = (
+            self._construct_local_block_info_list()
         )
 
     @torch.no_grad()
@@ -288,11 +280,12 @@ def update_params(
             masked_blocked_search_directions,
         )
 
-    def _construct_global_block_info_list(
+    @torch.no_grad()
+    def _construct_local_block_info_list(
         self,
-    ) -> None:
-        """Construct global block info list from param_group and num_blocks_within_param."""
-        self._global_block_info_list = tuple(
+    ) -> tuple[BlockInfo, ...]:
+        """Construct local block info list from param_group and num_blocks_within_param."""
+        return tuple(
             BlockInfo(
                 param=param,
                 composable_block_ids=self._construct_composable_block_ids(

distributed_shampoo/utils/shampoo_fsdp_distributor.py

@@ -55,7 +55,6 @@ def __init__(
         self._global_num_blocks_per_split_param: tuple[int, ...] = ()
 
         super().__init__(param_group)
-        self._construct_global_block_info_list()
 
         # Initialize selectors and local blocked (masked) parameters.
         self._local_grad_selector: tuple[bool, ...] = (True,) * len(
@@ -66,6 +65,9 @@ def __init__(
             self._global_blocked_params
         )
         self._local_blocked_params: tuple[Tensor, ...] = self._global_blocked_params
+        self._local_block_info_list: tuple[BlockInfo, ...] = (
+            self._construct_local_block_info_list()
+        )
 
     @torch.no_grad()
     def update_params(
@@ -102,12 +104,13 @@ def _construct_composable_block_ids(
         """
         return (param_index, f"rank_{rank}-block_{block_index}")
 
-    def _construct_global_block_info_list(
+    @torch.no_grad()
+    def _construct_local_block_info_list(
         self,
-    ) -> None:
-        """Construct global block info list from param_group and num_blocks_within_param."""
+    ) -> tuple[BlockInfo, ...]:
+        """Construct local block info list from param_group and num_blocks_within_param."""
         rank = dist.get_rank()
-        self._global_block_info_list = tuple(
+        return tuple(
             BlockInfo(
                 param=param,
                 composable_block_ids=self._construct_composable_block_ids(

distributed_shampoo/utils/shampoo_fully_shard_distributor.py

@@ -65,10 +65,10 @@ def _construct_composable_block_ids(
         return (param_index, f"rank_{rank}-block_{block_index}")
 
     @torch.no_grad()
-    def _construct_global_block_info_list(
+    def _construct_local_block_info_list(
         self,
-    ) -> None:
-        """Construct global block info list from param_group and num_blocks_within_param."""
+    ) -> tuple[BlockInfo, ...]:
+        """Construct local block info list from param_group and num_blocks_within_param."""
         rank = dist.get_rank()
 
         # Call `super()` instead of `self` as a performance optimization.
@@ -77,7 +77,7 @@ def _construct_global_block_info_list(
             lambda p: p.to_local().numel() > 0,  # type: ignore[arg-type]
             super()._get_params_or_grads(),
         )
-        self._global_block_info_list = tuple(
+        return tuple(
             BlockInfo(
                 param=param,
                 composable_block_ids=self._construct_composable_block_ids(

distributed_shampoo/utils/shampoo_hsdp_distributor.py

@@ -200,12 +200,13 @@ def __init__(
             )
         )
 
-        self._construct_global_block_info_list(buffer_size_ranks)
-
+        global_block_info_list = self._construct_global_block_info_list(
+            buffer_size_ranks
+        )
         # Initialize selectors and local blocked (masked) parameters.
         self._distributor_selector: tuple[bool, ...] = tuple(
             block_info.group_source_rank == comms_group_rank
-            for block_info in self._global_block_info_list
+            for block_info in global_block_info_list
         )
         self._local_blocked_params: tuple[Tensor, ...] = compress_list(
             self._global_blocked_params, self._distributor_selector
@@ -216,6 +217,9 @@ def __init__(
         self._local_grad_selector: tuple[bool, ...] = (True,) * len(
             self._local_blocked_params
         )
+        self._local_block_info_list: tuple[DDPBlockInfo, ...] = compress_list(
+            global_block_info_list, self._distributor_selector
+        )
 
         self._construct_distributed_buffers(
             buffer_size_ranks=buffer_size_ranks,
@@ -369,14 +373,15 @@ def _construct_composable_block_ids(
         """
         return (param_index, f"rank_{rank}-block_{block_index}")
 
+    @torch.no_grad()
     def _construct_global_block_info_list(
         self, buffer_size_ranks: tuple[tuple[int, int], ...]
-    ) -> None:
+    ) -> tuple[DDPBlockInfo, ...]:
         """Construct global block info list from param_group and num_blocks_within_param."""
         # Note that for HSDP, we want to get the rank within each sharded group for the block id.
         # When using a device mesh, 0 corresponds to the replicated group and 1 corresponds to the sharded group.
         sharded_group_rank = self._hsdp_device_mesh.get_local_rank(1)
-        self._global_block_info_list: tuple[DDPBlockInfo, ...] = tuple(
+        return tuple(
             DDPBlockInfo(
                 param=param,
                 composable_block_ids=self._construct_composable_block_ids(
@@ -575,7 +580,7 @@ def _construct_distributed_buffers(
         self._global_dist_buffer = torch.zeros(
             total_buffer_size,
             dtype=torch.int8,
-            device=self._global_block_info_list[0].param.device,
+            device=self._global_blocked_params[0].device,
         )
         local_dist_buffers = torch.split(self._global_dist_buffer, max_buffer_size_sum)
         splitted_local_dist_buffers = HSDPDistributor._split_local_dist_buffers(