Merge pull request #49 from xrsrke/feature/moe

[Feature] Add ExpertParallel with Top1 routing

pipegoose/nn/expert_parallel/expert_context.py

@@ -0,0 +1,32 @@
+from __future__ import annotations
+from typing import List
+
+from torchtyping import TensorType
+
+
+class ExpertContext:
+    _instance = None
+
+    def __init__(self):
+        self.aux_loss = []
+        self.z_loss = []
+
+    def push_aux_loss(self, aux_loss: TensorType):
+        self.aux_loss.append(aux_loss)
+
+    def pop_all_aux_loss(self) -> List[TensorType]:
+        aux_loss, self.aux_loss = self.aux_loss, []
+        return aux_loss
+
+    def push_z_loss(self, z_loss: TensorType):
+        self.z_loss.append(z_loss)
+
+    def pop_all_z_loss(self) -> List[TensorType]:
+        z_loss, self.z_loss = self.z_loss, []
+        return z_loss
+
+    @classmethod
+    def get_instance(cls) -> ExpertContext:
+        if not cls._instance:
+            cls._instance = ExpertContext()
+        return cls._instance

pipegoose/nn/expert_parallel/expert_parallel.py

@@ -27,7 +27,7 @@ def __init__(
         router: Union[int, Callable] = 1,
         # noise_poligy: Union[str, Callable],
         enable_tensor_parallelism: bool = False,
-        parallel_context: ParallelContext = None,
+        parallel_context: ParallelContext = None
     ):
         tensor_parallel_size = parallel_context.get_world_size(ParallelMode.TENSOR)
         assert parallel_context is not None, "parallel_context must be provided"
@@ -64,7 +64,7 @@ def parallelize(self) -> nn.Module:
                         module if self.expert is None else self.expert,
                         self.router,
                         self.enable_tensor_parallelism,
-                        self.parallel_context,
+                        self.parallel_context
                     )
                     getattr(self.module, "transformer").h[layer_idx].mlp = expert_layer
 

pipegoose/nn/expert_parallel/layers.py

@@ -5,6 +5,7 @@
 from pipegoose.nn.expert_parallel.experts import Experts
 from pipegoose.nn.expert_parallel.routers import Router
 from pipegoose.nn.expert_parallel.utils import get_num_local_experts
+from pipegoose.nn.expert_parallel.expert_context import ExpertContext
 
 
 class ExpertLayer(nn.Module):
@@ -20,7 +21,7 @@ def __init__(
         expert: nn.Module,
         router: Router,
         enable_tensor_parallel: bool,
-        parallel_context: ParallelContext,
+        parallel_context: ParallelContext
     ):
         super().__init__()
         self.router = router
@@ -39,6 +40,9 @@ def experts(self) -> nn.ModuleList:
     def forward(self, *args, **kwargs) -> TensorType["batch_size", "seq_len", "d_model"]:
         # TODO: use torch.fx to extract the inputs from args, and kwargs
         inputs = args[0]
-        dispatching_order, _, _ = self.router(inputs)
-        outputs = self._experts(inputs, dispatching_order, *args, **kwargs)
+        router_output = self.router(inputs)
+        expert_context = ExpertContext.get_instance()
+        expert_context.push_aux_loss(router_output.aux_loss)
+        expert_context.push_z_loss(router_output.z_loss)
+        outputs = self._experts(inputs, router_output.dispatching_order, *args, **kwargs)
         return outputs

pipegoose/nn/expert_parallel/loss.py

@@ -1,12 +1,19 @@
 from typing import Callable
 
-import torch
+from torchtyping import TensorType
+
+from pipegoose.nn.expert_parallel.expert_context import ExpertContext
 
 
 class ExpertLoss:
-    def __init__(self, loss: Callable, aux_weight: float):
-        self.loss = loss
+    def __init__(self, loss_func: Callable, aux_weight: float, z_weight: float):
+        self.loss_func = loss_func
         self.aux_weight = aux_weight
+        self.z_weight = z_weight
 
-    def __call__(self) -> torch.Tensor:
-        pass
+    def __call__(self, *args, **kwargs) -> TensorType:
+        loss = self.loss_func(*args, **kwargs)
+        expert_context = ExpertContext.get_instance()
+        loss += self.aux_weight * sum(expert_context.pop_all_aux_loss())
+        loss += self.z_weight * sum(expert_context.pop_all_z_loss())
+        return loss

pipegoose/nn/expert_parallel/routers.py

@@ -6,6 +6,7 @@
 import torch.nn.functional as F
 from torch import nn
 from torchtyping import TensorType
+from dataclasses import dataclass
 
 
 class RouterExplorationNoisePolicy(ABC):
@@ -32,6 +33,14 @@ def sample_like(self, input: TensorType) -> TensorType:
         return noise
 
 
+@dataclass
+class RouterOutput:
+    dispatching_order: TensorType["batch_size * seq_len", "num_experts"]
+    weight: TensorType["batch_size * seq_len", "num_experts"]
+    aux_loss: TensorType["1"]
+    z_loss: TensorType["1"]
+
+
 class Router(ABC, nn.Module):
     pass
 
@@ -93,9 +102,7 @@ def _expert_capacity(self, total_tokens: int) -> int:
 
     def forward(
         self, inputs: TensorType["batch_size", "seq_len", "d_model"]
-    ) -> Tuple[
-        TensorType["batch_size*seq_len", "num_experts"], TensorType["batch_size*seq_len", "num_experts"], TensorType["1"]
-    ]:
+    ) -> RouterOutput:
         orig_dtype = inputs.dtype
         total_tokens = inputs.shape[0] * inputs.shape[1]
 
@@ -115,15 +122,19 @@ def forward(
         topk_expert_mask = topk_expert_mask.scatter_(1, topk_idxs, True)
 
         # calculate router loss
-        loss = self.aux_loss_weight * self._aux_loss(router_prob, topk_expert_mask) + self.z_loss_weight * self._z_loss(
-            router_logits
-        )
+        aux_loss = self._aux_loss(router_prob, topk_expert_mask)
+        z_loss = self._z_loss(router_logits)
 
         if not self.expert_capacity:
             # we don't limit the capacity of the experts
             topk_weight = router_prob * topk_expert_mask
             topk_weight = topk_weight.to(orig_dtype)
-            return topk_expert_mask, topk_weight, loss
+            return RouterOutput(
+                dispatching_order=topk_expert_mask,
+                weight=topk_weight,
+                aux_loss=aux_loss,
+                z_loss=z_loss
+            )
 
         # limit the number of tokens per expert
         position_in_expert = torch.cumsum(topk_expert_mask, dim=0) * topk_expert_mask
@@ -137,7 +148,12 @@ def forward(
         topk_weight = router_prob * capacity_limited_topk_expert_mask
         topk_weight = topk_weight.to(orig_dtype)
 
-        return capacity_limited_topk_expert_mask, topk_weight, loss
+        return RouterOutput(
+            dispatching_order=capacity_limited_topk_expert_mask,
+            weight=topk_weight,
+            aux_loss=aux_loss,
+            z_loss=z_loss
+        )
 
 
 class Top1Router(_TopKRouter):

tests/nn/expert_parallel/test_expert_context.py

@@ -0,0 +1,20 @@
+from pipegoose.nn.expert_parallel.expert_context import ExpertContext
+
+
+def test_expert_context():
+    expert_context = ExpertContext.get_instance()
+
+    expert_context.push_aux_loss(1.01)
+    expert_context.push_z_loss(2.01)
+
+    expert_context.push_aux_loss(1.02)
+    expert_context.push_z_loss(2.02)
+
+    # make sure that we have a singleton!
+    expert_context = ExpertContext.get_instance()
+
+    assert expert_context.pop_all_aux_loss() == [1.01, 1.02]
+    assert expert_context.pop_all_aux_loss() == []
+
+    assert expert_context.pop_all_z_loss() == [2.01, 2.02]
+    assert expert_context.pop_all_z_loss() == []

tests/nn/expert_parallel/test_expert_loss.py

@@ -1,24 +1,35 @@
+import torch
 from torch import nn
+import torch.nn.functional as F
 
 from pipegoose.nn.expert_parallel import ExpertLoss
+from pipegoose.nn.expert_parallel.expert_context import ExpertContext
 
 
 def test_expert_loss():
-    loss_func = nn.CrossEntropyLoss()
+    torch.manual_seed(42)
+    logits = torch.randn((10, 5))
+    gt = torch.randn((10, 5))
 
-    expert_loss = ExpertLoss(loss_func, aux_weight=0.1)
+    loss_func = nn.MSELoss()
+
+    expert_loss = ExpertLoss(loss_func, aux_weight=0.1, z_weight=0.2)
+    expert_context = ExpertContext.get_instance()
 
     assert expert_loss.aux_weight == 0.1
+    assert expert_loss.z_weight == 0.2
     assert expert_loss.loss_func == loss_func
 
-    ExpertLoss.add_aux_loss(1.01)
-    ExpertLoss.add_z_loss(2.01)
+    expert_context.push_aux_loss(1.01)
+    expert_context.push_z_loss(2.01)
+
+    expert_context.push_aux_loss(1.02)
+    expert_context.push_z_loss(2.02)
 
-    assert expert_loss.get_aux_loss() == [1.01]
-    assert expert_loss.get_z_loss() == [2.01]
+    expected_loss = F.mse_loss(logits, gt) + 0.1 * (1.01 + 1.02) + 0.2 * (2.01 + 2.02)
+    loss = expert_loss(logits, gt)
 
-    ExpertLoss.add_aux_loss(1.02)
-    ExpertLoss.add_z_loss(2.02)
+    assert torch.allclose(loss, expected_loss)
 
-    assert expert_loss.get_aux_loss() == [1.01, 1.02]
-    assert expert_loss.get_z_loss() == [2.01, 2.02]
+    assert expert_context.aux_loss == []
+    assert expert_context.z_loss == []