class MoERunnerBase(MoERunner):
"""
Abstract base class providing common functionality for MoE runner implementations.
This class serves as the foundation for concrete MoE runner implementations by
providing shared state management and common utilities. It handles:
- Common initialization and configuration management
- Shared expert output reduction logic for tensor parallel scenarios
- Base methods for tensor model parallel reductions
- Common properties and utility functions used across different runner types
Concrete subclasses must implement the abstract methods to define their specific
execution strategies, such as standard execution, chunked processing, or other
specialized approaches. The base class provides the infrastructure while
allowing flexibility in the actual MoE computation implementation.
Key abstract methods that subclasses must implement:
- reduce_results: Determines whether results should be reduced across ranks
- _forward_impl: The core MoE computation logic specific to each runner type
"""
def __init__(
self,
layer_name: str,
moe_config: FusedMoEConfig,
router: FusedMoERouter,
routed_input_transform: torch.nn.Module | None,
gate: torch.nn.Module | None,
shared_experts: torch.nn.Module | None,
quant_method: FusedMoEMethodBase,
reduce_results: bool,
enable_dbo: bool,
):
super().__init__()
self.moe_config = moe_config
self.router = router
self.routed_input_transform = routed_input_transform
self.gate = gate
self.quant_method = quant_method
self._reduce_results = reduce_results
self.enable_dbo = enable_dbo
self._shared_experts: SharedExperts | None = None
if shared_experts is not None:
self._shared_experts = SharedExperts(
shared_experts,
moe_config=moe_config,
# Note: For now we must pass quant_method along to SharedExperts so it
# can property determine where the shared experts are supposed to be
# called, i.e. by a MK or by the MoERunner.
# Once the MK can be created upfront, we can just pass in the proper
# flags derived from the quant_method's MK.
reduce_results=reduce_results,
quant_method=quant_method,
enable_dbo=enable_dbo,
)
# Needed for string -> FusedMoE layer lookup in custom ops.
self.layer_name = layer_name
self.forward_entry = self._select_forward()
def _select_forward(self) -> Callable:
if current_platform.is_tpu() or current_platform.is_cpu():
# TODO: Once the OOM issue for the TPU backend is resolved, we
# will switch to using the moe_forward custom op.
# Note: CPU doesn't require wrapped _forward_impl.
return _moe_forward if self._shared_experts is None else _moe_forward_shared
return (
torch.ops.vllm.moe_forward
if self._shared_experts is None
else torch.ops.vllm.moe_forward_shared
)
@property
def shared_experts(self) -> SharedExperts | None:
return self._shared_experts
# TODO(bnell): temporary hack, do not call this method.
def _replace_quant_method(self, quant_method: FusedMoEMethodBase):
if self._shared_experts is not None:
self._shared_experts._quant_method = quant_method
self.quant_method = quant_method
def is_internal_router(self) -> bool:
return self.gate is not None
@property
@abstractmethod
def reduce_results(self) -> bool:
raise NotImplementedError
def must_reduce_shared_expert_outputs(self) -> bool:
"""
The shared_experts are typically computed using the RowParallelLinear
layer. The result of this function is typically used as
the reduce_results argument to the module.
When just tensor-parallel is used, it is not required to reduce
the shared_experts results immediately. Instead we reduce at the
once at the end of the MoE op. (Refer to DeepSeekV2MoE module)
With EP and all2all kernels - this is no longer viable as all
GPU ranks in DP, produce the complete set of hidden_states.
Therefore it is required that we reduce the shared_experts output
early.
"""
return (
self.quant_method.moe_kernel is not None
and self.quant_method.moe_kernel.output_is_reduced()
)
def maybe_all_reduce_tensor_model_parallel(self, final_hidden_states: torch.Tensor):
"""
Some combine kernels reduce across GPU ranks by default.
"""
if self.must_reduce_shared_expert_outputs():
return final_hidden_states
else:
return tensor_model_parallel_all_reduce(final_hidden_states)
def apply_routed_input_transform(
self, hidden_states: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor | None]:
"""Apply transform for routed experts (e.g., latent projection).
This is called by FusedMoE.forward_native. The original hidden_states
is saved separately so shared experts get [S, hidden_size] while
routed experts get the transformed [S, moe_latent_size].
TODO: For latent MoE bandwidth optimization, fc2_latent_proj could be
moved inside SharedFusedMoE to all-reduce on the smaller latent
dimension.
Returns (possibly transformed) hidden states and the input for shared
experts (or None if there are no shared experts).
"""
if self.routed_input_transform is not None:
result = self.routed_input_transform(hidden_states)
# ReplicatedLinear returns (output, extra_bias) tuple.
# We only need the output tensor; extra_bias is not used here.
if isinstance(result, tuple):
return result[0], hidden_states
return result, hidden_states
return (
hidden_states,
hidden_states if self._shared_experts is not None else None,
)
def _maybe_reduce_output(
self,
states: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
trunc_sizes: list[int],
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
def trunc(x: torch.Tensor, trunc_size: int) -> torch.Tensor:
return x[..., :trunc_size]
def reduce_and_trunc(x: torch.Tensor, trunc_size: int) -> torch.Tensor:
return trunc(self.maybe_all_reduce_tensor_model_parallel(x), trunc_size)
if (
not self.moe_config.is_sequence_parallel
and self.reduce_results
and (self.moe_config.tp_size > 1 or self.moe_config.ep_size > 1)
):
func = reduce_and_trunc
else:
func = trunc
if isinstance(states, tuple):
return tuple(
[func(s, trunc_size) for s, trunc_size in zip(states, trunc_sizes)]
)
else:
assert len(trunc_sizes) == 1
return func(states, trunc_sizes[0])
def _encode_layer_name(self) -> str | ModuleName:
if HAS_OPAQUE_TYPE:
return ModuleName(self.layer_name)
# Can be unavailable or None in unittests
if (
is_forward_context_available()
and get_forward_context().all_moe_layers is not None
):
return "from_forward_context"
return self.layer_name
def _maybe_pad_hidden_states(
self,
shared_experts_input: torch.Tensor | None,
hidden_states: torch.Tensor,
) -> tuple[torch.Tensor, list[int]]:
shared_experts_hidden_dim = (
shared_experts_input.shape[-1] if shared_experts_input is not None else 0
)
transformed_hidden_dim = hidden_states.shape[-1]
if (
not self.quant_method.skip_forward_padding
and self.moe_config.hidden_dim != transformed_hidden_dim
):
hidden_states = F.pad(
hidden_states,
(0, self.moe_config.hidden_dim - transformed_hidden_dim),
mode="constant",
value=0.0,
)
if self._shared_experts is not None:
orig_hidden_dims = [shared_experts_hidden_dim, transformed_hidden_dim]
else:
orig_hidden_dims = [transformed_hidden_dim]
return hidden_states, orig_hidden_dims
def _maybe_apply_shared_experts(
self,
shared_experts_input: torch.Tensor | None,
order: SharedExpertsOrder,
):
if self._shared_experts is not None:
assert shared_experts_input is not None
self._shared_experts.apply(shared_experts_input, order)
def _apply_quant_method(
self,
layer: torch.nn.Module,
hidden_states: torch.Tensor,
router_logits: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> tuple[torch.Tensor | None, torch.Tensor]:
# Run this before quant_method to avoid inplace issues.
# TODO(bnell): probably not needed anymore since inplace is
# disabled when shared experts are present.
self._maybe_apply_shared_experts(
shared_experts_input, SharedExpertsOrder.NO_OVERLAP
)
if self.quant_method.is_monolithic:
fused_out = self.quant_method.apply_monolithic(
layer=layer,
x=hidden_states,
router_logits=router_logits,
)
else:
topk_weights, topk_ids = self.router.select_experts(
hidden_states=hidden_states,
router_logits=router_logits,
)
# Passing shared_experts_input in case SharedExpertsOrder is
# NO_OVERLAP or MK_INTERNAL_OVERLAPPED.
fused_out = self.quant_method.apply(
layer=layer,
x=hidden_states,
topk_weights=topk_weights,
topk_ids=topk_ids,
shared_experts_input=shared_experts_input,
)
self._maybe_apply_shared_experts(
shared_experts_input,
SharedExpertsOrder.MULTI_STREAM_OVERLAPPED,
)
return (
self._shared_experts.output if self._shared_experts is not None else None,
fused_out,
)
def _sequence_parallel_context(self):
ctx = get_forward_context()
return (
ctx.dp_metadata.sp_local_sizes(self.moe_config.sp_size)
if ctx.dp_metadata
else nullcontext()
)
def _maybe_sync_shared_experts_stream(
self,
shared_experts_input: torch.Tensor | None,
):
# If router/gate provided, then apply it here.
# (Note: This code runs only when "overlapped mode" is on to allow
# parallel execution of shared experts with the FusedMoE via
# separate cuda stream)
if self._shared_experts is not None:
self._shared_experts.maybe_sync_shared_experts_stream(shared_experts_input)
def forward(
self,
hidden_states: torch.Tensor,
router_logits: torch.Tensor,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
"""Invoke the fused moe layer.
Input:
- hidden_states
- router_logits
Output:
- The new hidden_states.
or
- A tuple of (shared experts output, new hidden_states).
Calling sequence
- forward
- self.forward_entry (_moe_forward or _moe_forward_shared custom op)
- forward_dispatch
- _forward_impl
Note: The existence of _moe_forward and _moe_forward_shared custom ops are due
to the following reasons:
1. the chunking loop in ChunkingMoERunner._forward_impl cannot be compiled by
torch.compile
2. pytorch cannot handle union types in custom op signatures so _moe_forward
and _moe_forward_shared must be split.
If ChunkingMoERunner._forward_impl can be implemented via torch.scan we can
potentially get rid of _moe_forward and _moe_forward_shared and collapse the
whole sequence into the 'forward' method.
"""
# Apply transform for routed experts (e.g., latent projection for latent MoE)
hidden_states, shared_experts_input = self.apply_routed_input_transform(
hidden_states
)
hidden_states, og_hidden_dims = self._maybe_pad_hidden_states(
shared_experts_input,
hidden_states,
)
fused_output = self.forward_entry(
hidden_states,
router_logits,
shared_experts_input,
self._encode_layer_name(),
)
return self._maybe_reduce_output(fused_output, og_hidden_dims)
def forward_dispatch(
self,
layer: torch.nn.Module,
hidden_states: torch.Tensor,
router_logits: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
# TODO(bnell): this can be removed after MK migration is complete.
layer.ensure_moe_quant_config_init()
# Sync aux and main stream for shared expert multi-stream overlap.
self._maybe_sync_shared_experts_stream(shared_experts_input)
# If the Runner holds the gate, apply it after the stream sync,
# so it can run overlapped with the
# NOTE: in future PR, MoE runner will always hold the gate.
if self.gate is not None:
router_logits, _ = self.gate(hidden_states)
with self._sequence_parallel_context():
return self._forward_impl(
layer,
hidden_states,
router_logits,
shared_experts_input,
)
@abstractmethod
def _forward_impl(
self,
layer: torch.nn.Module,
hidden_states: torch.Tensor,
router_logits: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
raise NotImplementedError