Megatron MoE 实现

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

class MoELayer(BaseMoELayer):

   def forward(self, hidden_states: torch.Tensor):
        """Forward pass for the MoE layer.

        The forward pass comprises four main steps:
        1. Routing & Preprocessing: Route tokens to the assigned experts and prepare for dispatch.
        2. Dispatch: Tokens are sent to the expert devices using communication collectives.
        3. Expert Computation: Experts process the dispatched tokens.
        4. Combine: The outputs from the experts are combined and returned.

        Args:
            hidden_states (torch.Tensor): The input tensor to the MoE layer.

        Returns:
            A tuple containing the output tensor and the MLP bias, if any.
        """

        hidden_states, probs, residual = self.router_and_preprocess(hidden_states)
        dispatched_input, probs = self.dispatch(hidden_states, probs)
        output, shared_expert_output, mlp_bias = self.experts_compute(
            dispatched_input, probs, residual
        )
        output = self.combine(output, shared_expert_output)

        return output, mlp_bias

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

    def router_and_preprocess(self, hidden_states: torch.Tensor):
        """Compute and preprocess token routing for dispatch.

        This method uses the router to determine which experts to send each token to,
        producing routing probabilities and a mapping. It then preprocesses the
        hidden states and probabilities for the token dispatcher. The original
        hidden states are returned as a residual connection.
        """
        residual = hidden_states
        probs, routing_map = self.router(hidden_states)
        hidden_states, probs = self.token_dispatcher.dispatch_preprocess(
            hidden_states, routing_map, probs
        )
        return hidden_states, probs, residual

1
2
3
4
5
6
7

    def dispatch(self, hidden_states: torch.Tensor, probs: torch.Tensor):
        """Dispatches tokens to assigned expert ranks via communication.
        This method performs the actual communication (e.g., All-to-All) to distribute
        tokens and their associated probabilities to the devices hosting their assigned
        experts.
        """
        return self.token_dispatcher.token_dispatch(hidden_states, probs)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

    def experts_compute(
        self, hidden_states: torch.Tensor, probs: torch.Tensor, residual: torch.Tensor
    ):
        """Computes the output of the experts on the dispatched tokens.

        This method first post-processes the dispatched input to get permuted tokens
        for each expert. It then passes the tokens through the local experts.
        If a shared expert is configured and not overlapped with communication,
        it is also applied. The output from the experts is preprocessed for the
        combine step.
        """
        shared_expert_output = None
        if self.use_shared_expert and not self.shared_expert_overlap:
            # Compute the shared expert separately when not overlapped with communication.
            shared_expert_output = self.shared_experts(residual)
        dispatched_input, tokens_per_expert, permuted_probs = (
            self.token_dispatcher.dispatch_postprocess(hidden_states, probs)
        )
        expert_output, mlp_bias = self.experts(dispatched_input, tokens_per_expert, permuted_probs)
        assert mlp_bias is None, f"mlp_bias is not supported for {type(self.token_dispatcher)}"
        output = self.token_dispatcher.combine_preprocess(expert_output)

        return output, shared_expert_output, mlp_bias

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

    def combine(self, output: torch.Tensor, shared_expert_output: Optional[torch.Tensor]):
        """Combines expert outputs via communication and adds shared expert output.

        This method uses the token dispatcher to combine the outputs from different
        experts (e.g., via an All-to-All communication). It then adds the output
        from the shared expert if it exists.
        """
        output = self.token_dispatcher.token_combine(output)
        output = self.token_dispatcher.combine_postprocess(output)
        if shared_expert_output is not None:
            output = output + shared_expert_output
        return output