You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

361 lines
15 KiB

# Copyright 2025 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from collections.abc import Callable
from functools import wraps
from ..utils import logging
from ..utils.generic import GeneralInterface
from ..utils.import_utils import is_torch_available
if is_torch_available():
import torch
logger = logging.get_logger(__name__)
# Examples of experts class with its eager mm implementation
# class Experts(nn.Module):
# """Collection of expert weights stored as 3D tensors."""
# def __init__(self, config):
# super().__init__()
# self.num_experts = config.n_routed_experts
# self.hidden_dim = config.hidden_size
# self.intermediate_dim = config.moe_intermediate_size
# self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
# self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))
# self.act_fn = ACT2FN[config.hidden_act]
# def forward(
# self,
# hidden_states: torch.Tensor,
# top_k_index: torch.Tensor,
# top_k_weights: torch.Tensor,
# ) -> torch.Tensor:
# final_hidden_states = torch.zeros_like(hidden_states)
# with torch.no_grad():
# expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts)
# expert_mask = expert_mask.permute(2, 1, 0)
# expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
# for expert_idx in expert_hit:
# expert_idx = expert_idx[0]
# if expert_idx == self.num_experts:
# continue
# top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
# current_state = hidden_states[token_idx]
# gate, up = nn.functional.linear(current_state, self.gate_up_proj[expert_idx]).chunk(2, dim=-1)
# current_hidden_states = self.act_fn(gate) * up
# current_hidden_states = nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])
# current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]
# final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))
# return final_hidden_states
def _batched_linear(
input: torch.Tensor,
weight: torch.Tensor,
bias: torch.Tensor | None = None,
is_transposed: bool = False,
) -> torch.Tensor:
"""Batched linear layer supporting optional bias and transposed weights.
Args:
input (`torch.Tensor`):
Input tensor of shape (batch_size, input_dim).
weight (`torch.Tensor`):
Weight tensor of shape (batch_size, output_dim, input_dim) if transposed is `False`,
else of shape (batch_size, input_dim, output_dim).
bias (`torch.Tensor`, *optional*):
Bias tensor of shape (batch_size, output_dim). Default is `None`.
is_transposed (`bool`, *optional*, defaults to `False`):
Whether the weight tensor is transposed.
Returns:
`torch.Tensor`: Output tensor of shape (batch_size, output_dim).
"""
if is_transposed:
# (batch_size, 1, input_dim) @ (batch_size, input_dim, output_dim) -> (batch_size, 1, output_dim) -> (batch_size, output_dim)
out = torch.bmm(input.unsqueeze(1), weight).squeeze(1)
else:
# (batch_size, output_dim, input_dim) @ (batch_size, input_dim, 1) -> (batch_size, output_dim, 1) -> (batch_size, output_dim)
out = torch.bmm(weight, input.unsqueeze(-1)).squeeze(-1)
if bias is not None:
out = out + bias
return out
def batched_mm_experts_forward(
self: torch.nn.Module,
hidden_states: torch.Tensor,
top_k_index: torch.Tensor,
top_k_weights: torch.Tensor,
) -> torch.Tensor:
device = hidden_states.device
num_top_k = top_k_index.size(-1)
num_tokens = hidden_states.size(0)
hidden_dim = hidden_states.size(-1)
# Reshape for easier indexing
# S is the number of selected tokens-experts pairs (S = num_tokens * num_top_k)
token_idx = torch.arange(num_tokens, device=device).unsqueeze(1).expand(-1, num_top_k).reshape(-1) # (S,)
if top_k_weights.sum() == torch.tensor(0.0, device=top_k_weights.device):
# If all routing weights are zero local experts are not selected
return torch.zeros_like(hidden_states)
sample_weights = top_k_weights.reshape(-1) # (S,)
expert_ids = top_k_index.reshape(-1) # (S,)
# Handle invalid expert IDs from Expert Parallelism (EP)
# When EP is enabled, tokens assigned to experts on other devices are marked with sentinel value >= num_experts
valid_mask = expert_ids < self.num_experts
expert_ids_clamped = expert_ids.clamp(0, self.num_experts - 1)
# Get current hidden states for selected samples
selected_hidden_states = hidden_states[token_idx]
# Select expert weights and biases for selected samples (using clamped IDs for safe indexing)
selected_gate_up = self.gate_up_proj[expert_ids_clamped]
selected_down = self.down_proj[expert_ids_clamped]
selected_gate_up_bias = self.gate_up_proj_bias[expert_ids_clamped] if self.has_bias else None
selected_down_bias = self.down_proj_bias[expert_ids_clamped] if self.has_bias else None
# --- Up projection per expert (batched) ---
gate_up_out = _batched_linear(
selected_hidden_states, selected_gate_up, selected_gate_up_bias, is_transposed=self.is_transposed
) # (S, 2 * intermediate_dim)
# Apply gating
gated_out = self._apply_gate(gate_up_out) # (S, intermediate_dim)
# --- Down projection per expert (batched) ---
out_per_sample = _batched_linear(
gated_out, selected_down, selected_down_bias, is_transposed=self.is_transposed
) # (S, hidden_dim)
# Apply routing weights and zero out invalid expert contributions
if sample_weights.shape != expert_ids_clamped.shape:
sample_weights = sample_weights.gather(0, expert_ids_clamped)
out_per_sample = out_per_sample * sample_weights.unsqueeze(-1) # (S, hidden_dim)
out_per_sample = out_per_sample * valid_mask.unsqueeze(-1).to(out_per_sample.dtype)
# Accumulate results using deterministic reshape+sum instead of index_add_
# (index_add_ with duplicate indices is non-deterministic on CUDA due to atomicAdd)
final_hidden_states = out_per_sample.view(num_tokens, num_top_k, hidden_dim).sum(dim=1)
return final_hidden_states.to(hidden_states.dtype)
def _grouped_linear(
input: torch.Tensor,
weight: torch.Tensor,
bias: torch.Tensor | None = None,
offs: torch.Tensor | None = None,
is_transposed: bool = False,
) -> torch.Tensor:
"""Grouped linear layer supporting optional bias and transposed weights.
Args:
input (`torch.Tensor`):
Input tensor of shape (S, input_dim).
weight (`torch.Tensor`):
Weight tensor of shape (num_experts, output_dim, input_dim) if transposed is `False`,
else of shape (num_experts, input_dim, output_dim).
bias (`torch.Tensor`, *optional*):
Bias tensor of shape (num_experts, output_dim). Default is `None`.
offs (`torch.Tensor`, *optional*):
Offsets tensor indicating the boundaries of each group in the input tensor.
is_transposed (`bool`, *optional*, defaults to `False`):
Whether the weight tensor is transposed.
Returns:
`torch.Tensor`: Output tensor of shape (S, output_dim).
"""
if is_transposed:
# (S, input_dim) @ grouped (num_experts, input_dim, output_dim) -> (S, output_dim)
out = torch._grouped_mm(input, weight, offs=offs)
else:
# (S, input_dim) @ grouped (num_experts, output_dim, input_dim).T -> (S, output_dim)
out = torch._grouped_mm(input, weight.transpose(-2, -1), offs=offs)
if bias is not None:
# We should be able to pass bias to the grouped_mm call, but it's not yet supported.
out = out + bias
return out
def grouped_mm_experts_forward(
self: torch.nn.Module,
hidden_states: torch.Tensor,
top_k_index: torch.Tensor,
top_k_weights: torch.Tensor,
) -> torch.Tensor:
if not hasattr(torch, "_grouped_mm"):
raise ImportError(
"torch._grouped_mm is not available. Please make sure you are using a PyTorch version that includes it (2.9+)."
)
device = hidden_states.device
num_top_k = top_k_index.size(-1)
num_tokens = hidden_states.size(0)
hidden_dim = hidden_states.size(-1)
# Reshape for easier indexing
# S is the number of selected tokens-experts pairs (S = num_tokens * num_top_k)
token_idx = torch.arange(num_tokens, device=device).unsqueeze(1).expand(-1, num_top_k).reshape(-1) # (S,)
sample_weights = top_k_weights.reshape(-1) # (S,)
expert_ids = top_k_index.reshape(-1) # (S,)
# Get current hidden states for selected samples
selected_hidden_states = hidden_states[token_idx]
# Sort by expert for grouped processing
perm = torch.argsort(expert_ids)
inv_perm = torch.argsort(perm)
expert_ids_g = expert_ids[perm]
sample_weights_g = sample_weights[perm]
selected_hidden_states_g = selected_hidden_states[perm]
# Select expert weights and biases for selected samples
# NOTE: We keep all experts here and rely on offsets to target the active ones.
# I have already implemented a version that only passes the active experts, but
# to do so I had to use torch.unique which breaks the graph capture (data-dependent).
# Also there were no speedup gains from it in my experiments, even in eager mode.
selected_gate_up = self.gate_up_proj
selected_down = self.down_proj
selected_gate_up_bias = self.gate_up_proj_bias[expert_ids_g] if self.has_bias else None
selected_down_bias = self.down_proj_bias[expert_ids_g] if self.has_bias else None
# Compute offsets for grouped_mm
# using histc instead of bincount to avoid cuda graph issues
# With deterministic algorithms, CPU only supports float input, CUDA only supports int input.
histc_input = expert_ids_g.float() if device.type == "cpu" else expert_ids_g.int()
num_tokens_per_expert = torch.histc(histc_input, bins=self.num_experts, min=0, max=self.num_experts - 1)
offsets = torch.cumsum(num_tokens_per_expert, dim=0, dtype=torch.int32)
# --- Up projection per expert (grouped) ---
gate_up_out = _grouped_linear(
selected_hidden_states_g, selected_gate_up, selected_gate_up_bias, offsets, is_transposed=self.is_transposed
) # (S, 2 * intermediate_dim)
# Apply gating
gated_out = self._apply_gate(gate_up_out) # (S, intermediate_dim)
# --- Down projection per expert (grouped) ---
out_per_sample_g = _grouped_linear(
gated_out, selected_down, selected_down_bias, offsets, is_transposed=self.is_transposed
) # (S, hidden_dim)
# Apply routing weights
out_per_sample_g = out_per_sample_g * sample_weights_g.unsqueeze(-1) # (S, hidden_dim)
# Restore original order
out_per_sample = out_per_sample_g[inv_perm]
# Accumulate results using deterministic reshape+sum instead of index_add_
# (index_add_ with duplicate indices is non-deterministic on CUDA due to atomicAdd)
final_hidden_states = out_per_sample.view(num_tokens, num_top_k, hidden_dim).sum(dim=1)
return final_hidden_states.to(hidden_states.dtype)
class ExpertsInterface(GeneralInterface):
"""Interface for registering custom experts implementations."""
_global_mapping = {
"batched_mm": batched_mm_experts_forward,
"grouped_mm": grouped_mm_experts_forward,
}
def get_interface(self, experts_implementation: str, default: Callable) -> Callable:
"""Return the requested `experts_implementation`. Also strictly check its validity, and raise if invalid."""
if experts_implementation is None:
logger.warning_once(
"You tried to access the `ExpertsInterface` with a `config._experts_implementation` set to `None`. This "
"is expected if you use an Expert Module as a standalone Module. If this is not the case, something went "
"wrong with the dispatch of `config._experts_implementation`"
)
elif experts_implementation != "eager" and experts_implementation not in self:
raise KeyError(
f"`{experts_implementation}` is not a valid experts implementation registered in the `ExpertsInterface`"
)
return super().get(experts_implementation, default)
ALL_EXPERTS_FUNCTIONS = ExpertsInterface()
def _default_apply_gate(self, gate_up_out: torch.Tensor) -> torch.Tensor:
"""
Default gating mechanism: splits the gate_up_out into gate and up parts,
applies the activation function to the gate part, and multiplies it with the up part.
Args:
gate_up_out (`torch.Tensor`):
The output tensor from the gate and up projection of shape (S, 2 * intermediate_dim).
Returns:
`torch.Tensor`: The gated output tensor of shape (S, intermediate_dim).
"""
gate, up = gate_up_out.chunk(2, dim=-1) # (S, intermediate_dim)
return self.act_fn(gate) * up # (S, intermediate_dim)
def use_experts_implementation(
experts_class: type[torch.nn.Module] | None = None, *, is_transposed: bool = False, has_bias: bool = False
) -> type[torch.nn.Module]:
"""Decorator to modify experts class to support different experts implementations.
Args:
experts_class (`type[torch.nn.Module]`, *optional*):
The experts class to modify. If not provided, returns a decorator that can be applied to the class.
is_transposed (`bool`, *optional*, defaults to `False`):
Whether the expert weights are stored in transposed format.
has_bias (`bool`, *optional*, defaults to `False`):
Whether the expert layers include bias terms.
Returns:
`type[torch.nn.Module]`: The modified experts class.
"""
def wrapper(experts_class: type[torch.nn.Module]) -> type[torch.nn.Module]:
original_init = experts_class.__init__
original_forward = experts_class.forward
@wraps(original_init)
def __init__(self, config, *args, **kwargs):
original_init(self, config, *args, **kwargs)
self.config = config
self.has_bias = has_bias
self.is_transposed = is_transposed
@wraps(original_forward)
def forward(self, *args, **kwargs):
experts_forward = ALL_EXPERTS_FUNCTIONS.get_interface(
self.config._experts_implementation, original_forward
)
return experts_forward(self, *args, **kwargs)
if not hasattr(experts_class, "_apply_gate"):
experts_class._apply_gate = _default_apply_gate
experts_class.__init__ = __init__
experts_class.forward = forward
return experts_class
if experts_class is not None:
return wrapper(experts_class)
return wrapper