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.
840 lines
35 KiB
840 lines
35 KiB
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
|
# This file was automatically generated from src/transformers/models/jetmoe/modular_jetmoe.py.
|
|
# Do NOT edit this file manually as any edits will be overwritten by the generation of
|
|
# the file from the modular. If any change should be done, please apply the change to the
|
|
# modular_jetmoe.py file directly. One of our CI enforces this.
|
|
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
|
# Copyright 2024 JetMoe AI and the HuggingFace Inc. 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 typing import Optional
|
|
|
|
import torch
|
|
from torch import nn
|
|
from torch.nn import functional as F
|
|
|
|
from ... import initialization as init
|
|
from ...activations import ACT2FN
|
|
from ...cache_utils import Cache, DynamicCache
|
|
from ...generation import GenerationMixin
|
|
from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub
|
|
from ...masking_utils import create_causal_mask
|
|
from ...modeling_layers import GenericForSequenceClassification, GradientCheckpointingLayer
|
|
from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
|
|
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
|
|
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
|
|
from ...processing_utils import Unpack
|
|
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
|
|
from ...utils.generic import OutputRecorder, check_model_inputs, maybe_autocast
|
|
from .configuration_jetmoe import JetMoeConfig
|
|
|
|
|
|
logger = logging.get_logger(__name__)
|
|
|
|
|
|
@use_kernel_forward_from_hub("RMSNorm")
|
|
class JetMoeRMSNorm(nn.Module):
|
|
def __init__(self, hidden_size, eps=1e-6):
|
|
"""
|
|
JetMoeRMSNorm is equivalent to T5LayerNorm
|
|
"""
|
|
super().__init__()
|
|
self.weight = nn.Parameter(torch.ones(hidden_size))
|
|
self.variance_epsilon = eps
|
|
|
|
def forward(self, hidden_states):
|
|
input_dtype = hidden_states.dtype
|
|
hidden_states = hidden_states.to(torch.float32)
|
|
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
|
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
|
return self.weight * hidden_states.to(input_dtype)
|
|
|
|
def extra_repr(self):
|
|
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
|
|
|
|
|
class JetMoeRotaryEmbedding(nn.Module):
|
|
inv_freq: torch.Tensor # fix linting for `register_buffer`
|
|
|
|
def __init__(self, config: JetMoeConfig, device=None):
|
|
super().__init__()
|
|
self.max_seq_len_cached = config.max_position_embeddings
|
|
self.original_max_seq_len = config.max_position_embeddings
|
|
|
|
self.config = config
|
|
|
|
self.rope_type = self.config.rope_parameters["rope_type"]
|
|
rope_init_fn: Callable = self.compute_default_rope_parameters
|
|
if self.rope_type != "default":
|
|
rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
|
|
inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
|
|
|
|
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
|
self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
|
|
|
|
@staticmethod
|
|
def compute_default_rope_parameters(
|
|
config: JetMoeConfig | None = None,
|
|
device: Optional["torch.device"] = None,
|
|
seq_len: int | None = None,
|
|
) -> tuple["torch.Tensor", float]:
|
|
"""
|
|
Computes the inverse frequencies according to the original RoPE implementation
|
|
Args:
|
|
config ([`~transformers.PreTrainedConfig`]):
|
|
The model configuration.
|
|
device (`torch.device`):
|
|
The device to use for initialization of the inverse frequencies.
|
|
seq_len (`int`, *optional*):
|
|
The current sequence length. Unused for this type of RoPE.
|
|
Returns:
|
|
Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
|
|
post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
|
|
"""
|
|
base = config.rope_parameters["rope_theta"]
|
|
dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
|
|
|
|
attention_factor = 1.0 # Unused in this type of RoPE
|
|
|
|
# Compute the inverse frequencies
|
|
inv_freq = 1.0 / (
|
|
base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
|
|
)
|
|
return inv_freq, attention_factor
|
|
|
|
@torch.no_grad()
|
|
@dynamic_rope_update # power user: used with advanced RoPE types (e.g. dynamic rope)
|
|
def forward(self, x, position_ids):
|
|
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
|
|
position_ids_expanded = position_ids[:, None, :].float()
|
|
|
|
device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
|
|
with maybe_autocast(device_type=device_type, enabled=False): # Force float32
|
|
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
|
|
emb = torch.cat((freqs, freqs), dim=-1)
|
|
cos = emb.cos() * self.attention_scaling
|
|
sin = emb.sin() * self.attention_scaling
|
|
|
|
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
|
|
|
|
|
class JetMoeParallelExperts(nn.Module):
|
|
def __init__(self, num_experts: int, input_size: int, output_size: int) -> None:
|
|
"""
|
|
Initialize the JetMoeParallelExperts module.
|
|
The experts weights are stored in [num_experts, output_size, input_size] format. Such that it's compatible with
|
|
many MoE libraries, such as [Megablock](https://github.com/databricks/megablocks) and
|
|
[ScatterMoE](https://github.com/shawntan/scattermoe), as well as the
|
|
[MoE kernel](https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/fused_moe/fused_moe.py)
|
|
used in vllm.
|
|
|
|
Args:
|
|
num_experts (int):
|
|
Number of experts.
|
|
input_size (int):
|
|
Size of the input.
|
|
output_size (int):
|
|
Size of the output.
|
|
"""
|
|
super().__init__()
|
|
self.weight = nn.Parameter(torch.empty(num_experts, output_size, input_size))
|
|
self.num_experts = num_experts
|
|
self.input_size = input_size
|
|
self.output_size = output_size
|
|
|
|
def forward(self, inputs, expert_size):
|
|
"""
|
|
Forward pass of the JetMoeParallelExperts module.
|
|
|
|
Args:
|
|
inputs (Tensor):
|
|
Input tensor.
|
|
expert_size:
|
|
Expert size information.
|
|
|
|
Returns:
|
|
Tensor: Output tensor.
|
|
"""
|
|
input_list = inputs.split(expert_size, dim=0)
|
|
output_list = []
|
|
for i in range(self.num_experts):
|
|
output_list.append(F.linear(input_list[i], self.weight[i]))
|
|
results = torch.cat(output_list, dim=0)
|
|
return results
|
|
|
|
|
|
class JetMoeTopKGating(nn.Module):
|
|
def __init__(self, input_size: int, num_experts: int, top_k: int):
|
|
"""
|
|
Initialize the top-k gating mechanism.
|
|
|
|
Args:
|
|
input_size (`int`):
|
|
Size of the input.
|
|
num_experts (`int`):
|
|
Number of experts.
|
|
top_k (`int`):
|
|
Number of top experts to select.
|
|
"""
|
|
super().__init__()
|
|
|
|
self.num_experts = num_experts
|
|
self.input_size = input_size
|
|
self.top_k = top_k
|
|
|
|
self.layer = nn.Linear(input_size, num_experts, bias=False)
|
|
|
|
def forward(self, hidden_states):
|
|
# compute the top_k routing decision
|
|
logits = self.layer(hidden_states).float() # [batch_size x seq_len, num_experts]
|
|
top_k_logits, top_k_indices = logits.topk(self.top_k, dim=1) # [num_tokens, top_k]
|
|
top_k_gates = torch.softmax(top_k_logits, dim=1).type_as(hidden_states) # [num_tokens, top_k]
|
|
|
|
# compute number of input given to each expert
|
|
zeros = torch.zeros(
|
|
[top_k_gates.size(0), self.num_experts], dtype=top_k_gates.dtype, device=top_k_gates.device
|
|
) # [num_tokens, num_experts]
|
|
gates = zeros.scatter(1, top_k_indices, 1) # [num_tokens, num_experts]
|
|
expert_size = gates.long().sum(0) # [num_experts,]
|
|
# (This cause torch.compile to fail with `torch._dynamo.exc.Unsupported: Backend compiler failed with a fake tensor exception at`)
|
|
# (and `DataDependentOutputException`)
|
|
expert_size = expert_size.tolist()
|
|
|
|
# sort and group input tokens according to expert assignment
|
|
top_k_experts = top_k_indices.flatten() # [num_tokens * top_k]
|
|
_, index_sorted_experts = top_k_experts.sort(0) # [num_tokens * top_k]
|
|
batch_index = index_sorted_experts.div(self.top_k, rounding_mode="trunc") # [num_tokens * top_k]
|
|
|
|
# gather the gate values for grouped input tokens
|
|
top_k_gates = top_k_gates.flatten() # [num_tokens * top_k]
|
|
batch_gates = top_k_gates[index_sorted_experts] # [num_tokens * top_k]
|
|
|
|
return index_sorted_experts, batch_index, batch_gates, expert_size, logits
|
|
|
|
|
|
class JetMoeMoE(nn.Module):
|
|
"""
|
|
A Sparsely gated mixture of experts layer with 1-layer Feed-Forward networks as experts.
|
|
|
|
Args:
|
|
config:
|
|
Configuration object with model hyperparameters.
|
|
"""
|
|
|
|
def __init__(self, config: JetMoeConfig):
|
|
super().__init__()
|
|
|
|
self.input_size = config.hidden_size
|
|
self.hidden_size = config.intermediate_size
|
|
self.activation = ACT2FN[config.activation_function]
|
|
self.bias = torch.nn.Parameter(torch.empty(self.input_size))
|
|
self.input_linear = JetMoeParallelExperts(config.num_local_experts, self.input_size, self.hidden_size * 2)
|
|
self.output_linear = JetMoeParallelExperts(config.num_local_experts, self.hidden_size, self.input_size)
|
|
|
|
self.router = JetMoeTopKGating(
|
|
input_size=self.input_size,
|
|
num_experts=config.num_local_experts,
|
|
top_k=config.num_experts_per_tok,
|
|
)
|
|
|
|
def forward(self, layer_input):
|
|
"""
|
|
Forward pass of the mixture of experts layer.
|
|
|
|
Args:
|
|
layer_input (Tensor):
|
|
Input tensor.
|
|
|
|
Returns:
|
|
Tensor:
|
|
Output tensor.
|
|
Tensor:
|
|
Router logits.
|
|
"""
|
|
bsz, length, emb_size = layer_input.size()
|
|
layer_input = layer_input.reshape(-1, emb_size)
|
|
_, batch_index, batch_gates, expert_size, router_logits = self.router(layer_input)
|
|
|
|
expert_inputs = layer_input[batch_index]
|
|
hidden_states = self.input_linear(expert_inputs, expert_size)
|
|
chunked_hidden_states = hidden_states.chunk(2, dim=-1)
|
|
hidden_states = self.activation(chunked_hidden_states[0]) * chunked_hidden_states[1]
|
|
expert_outputs = self.output_linear(hidden_states, expert_size)
|
|
|
|
expert_outputs = expert_outputs * batch_gates[:, None]
|
|
|
|
zeros = torch.zeros((bsz * length, self.input_size), dtype=expert_outputs.dtype, device=expert_outputs.device)
|
|
layer_output = zeros.index_add(0, batch_index, expert_outputs)
|
|
layer_output = layer_output.view(bsz, length, self.input_size)
|
|
layer_output = layer_output + self.bias
|
|
return layer_output
|
|
|
|
|
|
class JetMoeMoA(nn.Module):
|
|
"""
|
|
A Sparsely gated mixture of attention layer with pairs of query- and output-projections as experts.
|
|
|
|
Args:
|
|
config:
|
|
Configuration object with model hyperparameters.
|
|
"""
|
|
|
|
def __init__(self, config: JetMoeConfig):
|
|
super().__init__()
|
|
|
|
self.num_experts = config.num_local_experts
|
|
self.input_size = config.hidden_size
|
|
self.hidden_size = config.kv_channels * config.num_key_value_heads
|
|
self.top_k = config.num_experts_per_tok
|
|
self.bias = torch.nn.Parameter(torch.empty(self.input_size))
|
|
|
|
self.input_linear = JetMoeParallelExperts(self.num_experts, self.input_size, self.hidden_size)
|
|
self.output_linear = JetMoeParallelExperts(self.num_experts, self.hidden_size, self.input_size)
|
|
|
|
self.router = JetMoeTopKGating(
|
|
input_size=self.input_size,
|
|
num_experts=self.num_experts,
|
|
top_k=self.top_k,
|
|
)
|
|
|
|
def map(self, layer_input):
|
|
"""
|
|
Map inputs to attention experts according to routing decision and compute query projection inside each experts.
|
|
"""
|
|
|
|
# Compute gating topology
|
|
bsz, length, emb_size = layer_input.size()
|
|
layer_input = layer_input.reshape(-1, emb_size) # [bsz * length, emb_size]
|
|
index_sorted_experts, batch_index, batch_gates, expert_size, router_logits = self.router(layer_input)
|
|
topo_info = (index_sorted_experts, batch_index, batch_gates, expert_size)
|
|
|
|
# Group inputs according to topology and compute query projection
|
|
expert_inputs = layer_input[batch_index] # [bsz * length * top_k, emb_size]
|
|
expert_outputs = self.input_linear(expert_inputs, expert_size) # [bsz * length * top_k, hidden_size]
|
|
|
|
# Ungroup queries back to original order
|
|
zeros = torch.zeros(
|
|
(bsz * length * self.top_k, self.hidden_size), dtype=expert_outputs.dtype, device=expert_outputs.device
|
|
)
|
|
layer_output = zeros.index_add(0, index_sorted_experts, expert_outputs)
|
|
layer_output = layer_output.view(bsz, length, self.top_k, -1) # [bsz, length, top_k, hidden_size]
|
|
return layer_output, router_logits, topo_info
|
|
|
|
def reduce(self, layer_input, topo_info):
|
|
"""
|
|
Compute output projection inside each attention experts and merge the outputs of different experts.
|
|
"""
|
|
bsz, length, k, hidden_size = layer_input.size()
|
|
layer_input = layer_input.reshape(-1, hidden_size) # [bsz * length * k, hidden_size]
|
|
index_sorted_experts, batch_index, batch_gates, expert_size = topo_info
|
|
|
|
# Group inputs according to topology and compute output projection
|
|
expert_inputs = layer_input[index_sorted_experts] # [bsz * length * top_k, hidden_size]
|
|
expert_outputs = self.output_linear(expert_inputs, expert_size) # [bsz * length * top_k, emb_size]
|
|
|
|
# Apply gates to attention expert outputs
|
|
expert_outputs = expert_outputs * batch_gates[:, None]
|
|
|
|
# Ungroup and merge outputs to original order
|
|
zeros = torch.zeros((bsz * length, self.input_size), dtype=expert_outputs.dtype, device=expert_outputs.device)
|
|
layer_output = zeros.index_add(0, batch_index, expert_outputs)
|
|
layer_output = layer_output.view(bsz, length, self.input_size)
|
|
layer_output = layer_output + self.bias
|
|
return layer_output
|
|
|
|
def forward(self, layer_input):
|
|
raise NotImplementedError("This module doesn't support call and forward.")
|
|
|
|
|
|
def rotate_half(x):
|
|
"""Rotates half the hidden dims of the input."""
|
|
x1 = x[..., : x.shape[-1] // 2]
|
|
x2 = x[..., x.shape[-1] // 2 :]
|
|
return torch.cat((-x2, x1), dim=-1)
|
|
|
|
|
|
@use_kernel_func_from_hub("rotary_pos_emb")
|
|
def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
|
|
"""Applies Rotary Position Embedding to the query and key tensors.
|
|
|
|
Args:
|
|
q (`torch.Tensor`): The query tensor.
|
|
k (`torch.Tensor`): The key tensor.
|
|
cos (`torch.Tensor`): The cosine part of the rotary embedding.
|
|
sin (`torch.Tensor`): The sine part of the rotary embedding.
|
|
unsqueeze_dim (`int`, *optional*, defaults to 1):
|
|
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
|
|
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
|
|
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
|
|
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
|
|
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
|
|
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
|
|
Returns:
|
|
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
|
"""
|
|
cos = cos.unsqueeze(unsqueeze_dim)
|
|
sin = sin.unsqueeze(unsqueeze_dim)
|
|
q_embed = (q * cos) + (rotate_half(q) * sin)
|
|
k_embed = (k * cos) + (rotate_half(k) * sin)
|
|
return q_embed, k_embed
|
|
|
|
|
|
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
|
"""
|
|
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
|
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
|
"""
|
|
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
|
if n_rep == 1:
|
|
return hidden_states
|
|
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
|
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
|
|
|
|
|
def eager_attention_forward(
|
|
module: nn.Module,
|
|
query: torch.Tensor,
|
|
key: torch.Tensor,
|
|
value: torch.Tensor,
|
|
attention_mask: torch.Tensor | None,
|
|
scaling: float,
|
|
dropout: float = 0.0,
|
|
**kwargs: Unpack[TransformersKwargs],
|
|
):
|
|
key_states = repeat_kv(key, module.num_key_value_groups)
|
|
value_states = repeat_kv(value, module.num_key_value_groups)
|
|
|
|
attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
|
|
if attention_mask is not None:
|
|
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
|
attn_weights = attn_weights + causal_mask
|
|
|
|
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
|
|
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|
|
attn_output = torch.matmul(attn_weights, value_states)
|
|
attn_output = attn_output.transpose(1, 2).contiguous()
|
|
|
|
return attn_output, attn_weights
|
|
|
|
|
|
class JetMoeAttention(nn.Module):
|
|
"""
|
|
Multi-headed attention from 'Attention Is All You Need' paper.
|
|
"""
|
|
|
|
def __init__(self, config: JetMoeConfig, layer_idx: int | None = None):
|
|
"""
|
|
Initialize the JetMoeAttention module.
|
|
|
|
Args:
|
|
config:
|
|
Configuration object with model hyperparameters.
|
|
layer_idx:
|
|
Index of the layer in the model.
|
|
"""
|
|
super().__init__()
|
|
self.config = config
|
|
self.layer_idx = layer_idx
|
|
self.is_causal = True
|
|
if layer_idx is None:
|
|
logger.warning_once(
|
|
f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
|
|
"lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
|
|
"when creating this class."
|
|
)
|
|
|
|
self.num_key_value_groups = 1 # We ignore this by setting it to 1 as we have different repeat patterns
|
|
self.top_k = config.num_experts_per_tok
|
|
self.attention_dropout = config.attention_dropout
|
|
self.kv_projection_size = config.kv_channels * config.num_key_value_heads
|
|
self.num_key_value_heads = config.num_key_value_heads
|
|
self.num_heads = config.num_attention_heads
|
|
self.head_dim = config.kv_channels
|
|
self.scaling = self.head_dim**-0.5
|
|
self.experts = JetMoeMoA(config)
|
|
|
|
self.kv_proj = torch.nn.Linear(config.hidden_size, self.kv_projection_size * 2, bias=False)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
attention_mask: torch.Tensor | None = None,
|
|
position_embeddings: torch.LongTensor | None = None,
|
|
past_key_values: Cache | None = None,
|
|
cache_position: torch.LongTensor | None = None,
|
|
**kwargs,
|
|
) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
|
|
input_shape = hidden_states.shape[:-1]
|
|
hidden_shape = (*input_shape, -1, self.head_dim)
|
|
|
|
query_states, router_logits, topo_info = self.experts.map(hidden_states)
|
|
key_states, value_states = self.kv_proj(hidden_states).chunk(2, dim=-1)
|
|
|
|
query_states = query_states.view(hidden_shape).transpose(1, 2)
|
|
key_states = key_states.view(hidden_shape).transpose(1, 2)
|
|
value_states = value_states.view(hidden_shape).transpose(1, 2)
|
|
|
|
cos, sin = position_embeddings
|
|
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
|
|
|
if past_key_values is not None:
|
|
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
|
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
|
key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
|
|
|
attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
|
|
self.config._attn_implementation, eager_attention_forward
|
|
)
|
|
|
|
# This is different from other models where we repeat k/v heads
|
|
# instead of repeat interleaving them
|
|
key_states = key_states.repeat(1, self.top_k, 1, 1)
|
|
value_states = value_states.repeat(1, self.top_k, 1, 1)
|
|
|
|
attn_output, attn_weights = attention_interface(
|
|
self,
|
|
query_states,
|
|
key_states,
|
|
value_states,
|
|
attention_mask,
|
|
dropout=0.0 if not self.training else self.attention_dropout,
|
|
scaling=self.scaling,
|
|
**kwargs,
|
|
)
|
|
|
|
attn_output = attn_output.view(*input_shape, self.top_k, -1)
|
|
attn_output = self.experts.reduce(attn_output, topo_info)
|
|
attn_output = attn_output.view(*input_shape, -1)
|
|
return attn_output, attn_weights, router_logits
|
|
|
|
|
|
class JetMoeDecoderLayer(GradientCheckpointingLayer):
|
|
def __init__(self, config: JetMoeConfig, layer_idx: int | None = None):
|
|
super().__init__()
|
|
self.hidden_size = config.hidden_size
|
|
self.mlp = JetMoeMoE(config)
|
|
self.input_layernorm = JetMoeRMSNorm(config.hidden_size)
|
|
self.post_attention_layernorm = JetMoeRMSNorm(config.hidden_size)
|
|
self.self_attention = JetMoeAttention(config, layer_idx)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
attention_mask: torch.Tensor | None = None,
|
|
position_ids: torch.LongTensor | None = None,
|
|
past_key_values: Cache | None = None,
|
|
use_cache: bool | None = False,
|
|
cache_position: torch.LongTensor | None = None,
|
|
position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
|
|
**kwargs: Unpack[TransformersKwargs],
|
|
) -> torch.Tensor:
|
|
residual = hidden_states
|
|
hidden_states = self.input_layernorm(hidden_states)
|
|
# Self Attention
|
|
hidden_states, _, _ = self.self_attention(
|
|
hidden_states=hidden_states,
|
|
attention_mask=attention_mask,
|
|
position_ids=position_ids,
|
|
past_key_values=past_key_values,
|
|
use_cache=use_cache,
|
|
cache_position=cache_position,
|
|
position_embeddings=position_embeddings,
|
|
**kwargs,
|
|
)
|
|
hidden_states = residual + hidden_states
|
|
|
|
# Fully Connected
|
|
residual = hidden_states
|
|
hidden_states = self.post_attention_layernorm(hidden_states)
|
|
hidden_states = self.mlp(hidden_states)
|
|
hidden_states = residual + hidden_states
|
|
return hidden_states
|
|
|
|
|
|
@auto_docstring
|
|
class JetMoePreTrainedModel(PreTrainedModel):
|
|
config: JetMoeConfig
|
|
base_model_prefix = "model"
|
|
supports_gradient_checkpointing = False
|
|
_no_split_modules = ["JetMoeDecoderLayer"]
|
|
_skip_keys_device_placement = ["past_key_values"]
|
|
_supports_flash_attn = True
|
|
_supports_sdpa = True
|
|
_supports_flex_attn = True
|
|
_can_compile_fullgraph = False # TopK gating fails fullgraph compilation at "expert_size = expert_size.tolist()"
|
|
_supports_attention_backend = True
|
|
_can_record_outputs = {
|
|
"router_logits": OutputRecorder(nn.Linear, layer_name="gate", index=1),
|
|
"hidden_states": JetMoeDecoderLayer,
|
|
"attentions": OutputRecorder(JetMoeAttention, index=1),
|
|
}
|
|
|
|
@torch.no_grad()
|
|
def _init_weights(self, module):
|
|
"""Initialize the weights."""
|
|
super()._init_weights(module)
|
|
if isinstance(module, JetMoeParallelExperts):
|
|
init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
|
|
elif isinstance(module, JetMoeMoA | JetMoeMoE):
|
|
init.zeros_(module.bias)
|
|
|
|
|
|
@auto_docstring
|
|
class JetMoeModel(JetMoePreTrainedModel):
|
|
def __init__(self, config: JetMoeConfig):
|
|
super().__init__(config)
|
|
self.padding_idx = config.pad_token_id
|
|
self.vocab_size = config.vocab_size
|
|
|
|
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
|
self.layers = nn.ModuleList(
|
|
[JetMoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
|
)
|
|
self.norm = JetMoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
self.rotary_emb = JetMoeRotaryEmbedding(config=config)
|
|
self.gradient_checkpointing = False
|
|
self._attn_implementation = config._attn_implementation
|
|
|
|
# Initialize weights and apply final processing
|
|
self.post_init()
|
|
|
|
@check_model_inputs
|
|
@auto_docstring
|
|
def forward(
|
|
self,
|
|
input_ids: torch.LongTensor | None = None,
|
|
attention_mask: torch.Tensor | None = None,
|
|
position_ids: torch.LongTensor | None = None,
|
|
past_key_values: Cache | None = None,
|
|
inputs_embeds: torch.FloatTensor | None = None,
|
|
use_cache: bool | None = None,
|
|
cache_position: torch.LongTensor | None = None,
|
|
**kwargs: Unpack[TransformersKwargs],
|
|
) -> MoeModelOutputWithPast:
|
|
if (input_ids is None) ^ (inputs_embeds is not None):
|
|
raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
|
|
|
|
if use_cache and past_key_values is None:
|
|
past_key_values = DynamicCache(config=self.config)
|
|
|
|
if inputs_embeds is None:
|
|
inputs_embeds = self.embed_tokens(input_ids)
|
|
|
|
if cache_position is None:
|
|
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
|
cache_position = torch.arange(
|
|
past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
|
|
)
|
|
if position_ids is None:
|
|
position_ids = cache_position.unsqueeze(0)
|
|
|
|
causal_mask = create_causal_mask(
|
|
config=self.config,
|
|
input_embeds=inputs_embeds,
|
|
attention_mask=attention_mask,
|
|
cache_position=cache_position,
|
|
past_key_values=past_key_values,
|
|
position_ids=position_ids,
|
|
)
|
|
|
|
hidden_states = inputs_embeds
|
|
|
|
# create position embeddings to be shared across the decoder layers
|
|
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
|
|
|
for decoder_layer in self.layers[: self.config.num_hidden_layers]:
|
|
hidden_states = decoder_layer(
|
|
hidden_states,
|
|
position_embeddings=position_embeddings,
|
|
attention_mask=causal_mask,
|
|
past_key_values=past_key_values,
|
|
use_cache=use_cache,
|
|
cache_position=cache_position,
|
|
position_ids=position_ids,
|
|
**kwargs,
|
|
)
|
|
|
|
hidden_states = self.norm(hidden_states)
|
|
|
|
return MoeModelOutputWithPast( # only diff with Mistral is the output type, we need MoE
|
|
last_hidden_state=hidden_states,
|
|
past_key_values=past_key_values,
|
|
)
|
|
|
|
|
|
def load_balancing_loss_func(
|
|
gate_logits: torch.Tensor | tuple[torch.Tensor] | None,
|
|
num_experts: int | None = None,
|
|
top_k=2,
|
|
attention_mask: torch.Tensor | None = None,
|
|
) -> torch.Tensor | int:
|
|
r"""
|
|
Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
|
|
|
|
See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details. This function implements the loss
|
|
function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
|
|
experts is too unbalanced.
|
|
|
|
Args:
|
|
gate_logits:
|
|
Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
|
|
shape [batch_size X sequence_length, num_experts].
|
|
num_experts:
|
|
Number of experts
|
|
top_k:
|
|
The number of experts to route per-token, can be also interpreted as the `top-k` routing
|
|
parameter.
|
|
attention_mask (`torch.Tensor`, *optional*):
|
|
The attention_mask used in forward function
|
|
shape [batch_size X sequence_length] if not None.
|
|
|
|
Returns:
|
|
The auxiliary loss.
|
|
"""
|
|
if gate_logits is None or not isinstance(gate_logits, tuple):
|
|
return 0
|
|
|
|
if isinstance(gate_logits, tuple):
|
|
compute_device = gate_logits[0].device
|
|
concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)
|
|
|
|
routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)
|
|
|
|
_, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
|
|
|
|
expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)
|
|
|
|
if attention_mask is None:
|
|
# Compute the percentage of tokens routed to each experts
|
|
tokens_per_expert = torch.mean(expert_mask.float(), dim=0)
|
|
|
|
# Compute the average probability of routing to these experts
|
|
router_prob_per_expert = torch.mean(routing_weights, dim=0)
|
|
else:
|
|
batch_size, sequence_length = attention_mask.shape
|
|
num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)
|
|
|
|
# Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
|
|
expert_attention_mask = (
|
|
attention_mask[None, :, :, None, None]
|
|
.expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
|
|
.reshape(-1, top_k, num_experts)
|
|
.to(compute_device)
|
|
)
|
|
|
|
# Compute the percentage of tokens routed to each experts
|
|
tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
|
|
expert_attention_mask, dim=0
|
|
)
|
|
|
|
# Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
|
|
router_per_expert_attention_mask = (
|
|
attention_mask[None, :, :, None]
|
|
.expand((num_hidden_layers, batch_size, sequence_length, num_experts))
|
|
.reshape(-1, num_experts)
|
|
.to(compute_device)
|
|
)
|
|
|
|
# Compute the average probability of routing to these experts
|
|
router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
|
|
router_per_expert_attention_mask, dim=0
|
|
)
|
|
|
|
overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
|
|
return overall_loss * num_experts
|
|
|
|
|
|
class JetMoeForCausalLM(JetMoePreTrainedModel, GenerationMixin):
|
|
_tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
|
|
|
|
def __init__(self, config):
|
|
super().__init__(config)
|
|
self.model = JetMoeModel(config)
|
|
self.vocab_size = config.vocab_size
|
|
self.aux_loss_coef = config.aux_loss_coef
|
|
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
|
self.tie_word_embeddings = config.tie_word_embeddings
|
|
|
|
# Initialize weights and apply final processing
|
|
self.post_init()
|
|
|
|
@can_return_tuple
|
|
@auto_docstring
|
|
def forward(
|
|
self,
|
|
input_ids: torch.LongTensor | None = None,
|
|
attention_mask: torch.Tensor | None = None,
|
|
position_ids: torch.LongTensor | None = None,
|
|
past_key_values: Cache | None = None,
|
|
inputs_embeds: torch.FloatTensor | None = None,
|
|
labels: torch.LongTensor | None = None,
|
|
use_cache: bool | None = None,
|
|
cache_position: torch.LongTensor | None = None,
|
|
logits_to_keep: int | torch.Tensor = 0,
|
|
output_router_logits: bool | None = False,
|
|
**kwargs,
|
|
) -> MoeCausalLMOutputWithPast:
|
|
outputs: MoeModelOutputWithPast = self.model(
|
|
input_ids=input_ids,
|
|
attention_mask=attention_mask,
|
|
position_ids=position_ids,
|
|
past_key_values=past_key_values,
|
|
inputs_embeds=inputs_embeds,
|
|
use_cache=use_cache,
|
|
cache_position=cache_position,
|
|
**kwargs,
|
|
)
|
|
|
|
hidden_states = outputs.last_hidden_state
|
|
# Only compute necessary logits, and do not upcast them to float if we are not computing the loss
|
|
slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
|
|
logits = self.lm_head(hidden_states[:, slice_indices, :])
|
|
|
|
loss = None
|
|
if labels is not None:
|
|
loss = self.loss_function(
|
|
logits,
|
|
labels,
|
|
vocab_size=self.config.vocab_size,
|
|
**kwargs,
|
|
)
|
|
|
|
aux_loss = None
|
|
if output_router_logits:
|
|
aux_loss = load_balancing_loss_func(
|
|
outputs.router_logits,
|
|
self.num_experts,
|
|
self.num_experts_per_tok,
|
|
attention_mask,
|
|
)
|
|
if labels is not None:
|
|
loss += self.aux_loss_coef * aux_loss.to(loss.device) # make sure to reside in the same device
|
|
|
|
return MoeCausalLMOutputWithPast(
|
|
loss=loss,
|
|
aux_loss=aux_loss,
|
|
logits=logits,
|
|
past_key_values=outputs.past_key_values,
|
|
hidden_states=outputs.hidden_states,
|
|
attentions=outputs.attentions,
|
|
router_logits=outputs.router_logits,
|
|
)
|
|
|
|
|
|
class JetMoeForSequenceClassification(GenericForSequenceClassification, JetMoePreTrainedModel): ...
|
|
|
|
|
|
__all__ = ["JetMoeForCausalLM", "JetMoeModel", "JetMoePreTrainedModel", "JetMoeForSequenceClassification"]
|