# 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.
"""PyTorch OLMoE model."""
import math
from typing import List, Optional, Tuple, Union
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache, StaticCache
from ...generation import GenerationMixin
from ...modeling_attn_mask_utils import AttentionMaskConverter
from ...modeling_flash_attention_utils import flash_attn_supports_top_left_mask, is_flash_attn_available
from ...modeling_outputs import (
MoeCausalLMOutputWithPast,
MoeModelOutputWithPast,
)
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from ...modeling_utils import PreTrainedModel
from ...pytorch_utils import ALL_LAYERNORM_LAYERS
from ...utils import (
add_start_docstrings,
add_start_docstrings_to_model_forward,
logging,
replace_return_docstrings,
)
from ...utils.deprecation import deprecate_kwarg
from .configuration_olmoe import OlmoeConfig
if is_flash_attn_available():
from ...modeling_flash_attention_utils import _flash_attention_forward
logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "OlmoeConfig"
# Copied from transformers.models.mixtral.modeling_mixtral.load_balancing_loss_func
def load_balancing_loss_func(
gate_logits: Union[torch.Tensor, Tuple[torch.Tensor], None],
num_experts: Optional[int] = None,
top_k=2,
attention_mask: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, int]:
r"""
Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
See Switch Transformer (https://arxiv.org/abs/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 OlmoeRMSNorm(nn.Module):
def __init__(self, hidden_size, eps=1e-5):
"""
OlmoeRMSNorm 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}"
ALL_LAYERNORM_LAYERS.append(OlmoeRMSNorm)
# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Olmoe
class OlmoeRotaryEmbedding(nn.Module):
def __init__(self, config: OlmoeConfig, device=None):
super().__init__()
# BC: "rope_type" was originally "type"
if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
else:
self.rope_type = "default"
self.max_seq_len_cached = config.max_position_embeddings
self.original_max_seq_len = config.max_position_embeddings
self.config = config
self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
self.register_buffer("inv_freq", inv_freq, persistent=False)
self.original_inv_freq = self.inv_freq
@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 torch.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)
# Copied from transformers.models.llama.modeling_llama.rotate_half
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)
# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, 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.
position_ids (`torch.Tensor`, *optional*):
Deprecated and unused.
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
# Copied from transformers.models.olmo.modeling_olmo.OlmoMLP with Olmo->Olmoe
class OlmoeMLP(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.hidden_size = config.hidden_size
self.intermediate_size = config.intermediate_size
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
self.act_fn = ACT2FN[config.hidden_act]
def forward(self, x):
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
return down_proj
# Copied from transformers.models.llama.modeling_llama.repeat_kv
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)
class OlmoeAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
def __init__(self, config: OlmoeConfig, layer_idx: Optional[int] = None):
super().__init__()
self.config = config
self.layer_idx = layer_idx
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.attention_dropout = config.attention_dropout
self.hidden_size = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.hidden_size // self.num_heads
self.num_key_value_heads = config.num_key_value_heads
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
self.max_position_embeddings = config.max_position_embeddings
self.rope_theta = config.rope_theta
self.is_causal = True
if (self.head_dim * self.num_heads) != self.hidden_size:
raise ValueError(
f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
f" and `num_heads`: {self.num_heads})."
)
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=config.attention_bias)
self.q_norm = OlmoeRMSNorm(self.hidden_size, eps=config.rms_norm_eps)
self.k_norm = OlmoeRMSNorm(
(self.hidden_size // self.num_heads) * self.num_key_value_heads, eps=config.rms_norm_eps
)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value: Optional[Cache] = None,
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
query_states = self.q_norm(self.q_proj(hidden_states))
key_states = self.k_norm(self.k_proj(hidden_states))
value_states = self.v_proj(hidden_states)
if self.config.clip_qkv is not None:
query_states.clamp_(min=-self.config.clip_qkv, max=self.config.clip_qkv)
key_states.clamp_(min=-self.config.clip_qkv, max=self.config.clip_qkv)
value_states.clamp_(min=-self.config.clip_qkv, max=self.config.clip_qkv)
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
cos, sin = position_embeddings
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
if past_key_value 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_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
key_states = repeat_kv(key_states, self.num_key_value_groups)
value_states = repeat_kv(value_states, self.num_key_value_groups)
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
if attention_mask is not None: # no matter the length, we just slice it
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
attn_weights = attn_weights + causal_mask
# upcast attention to fp32
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
attn_output = torch.matmul(attn_weights, value_states)
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
return attn_output, attn_weights, past_key_value
class OlmoeFlashAttention2(OlmoeAttention):
"""
OLMoE flash attention module. This module inherits from `OlmoeAttention` as the weights of the module stays
untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
flash attention and deal with padding tokens in case the input contains any of them.
"""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
# flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignment, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
# Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
self._flash_attn_uses_top_left_mask = flash_attn_supports_top_left_mask()
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value: Optional[Cache] = None,
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
output_attentions = False
bsz, q_len, _ = hidden_states.size()
query_states = self.q_norm(self.q_proj(hidden_states))
key_states = self.k_norm(self.k_proj(hidden_states))
value_states = self.v_proj(hidden_states)
if self.config.clip_qkv is not None:
query_states.clamp_(min=-self.config.clip_qkv, max=self.config.clip_qkv)
key_states.clamp_(min=-self.config.clip_qkv, max=self.config.clip_qkv)
value_states.clamp_(min=-self.config.clip_qkv, max=self.config.clip_qkv)
# Flash attention requires the input to have the shape
# batch_size x seq_length x head_dim x hidden_dim
# therefore we just need to keep the original shape
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
cos, sin = position_embeddings
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
if past_key_value 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_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
# TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
# to be able to avoid many of these transpose/reshape/view.
query_states = query_states.transpose(1, 2)
key_states = key_states.transpose(1, 2)
value_states = value_states.transpose(1, 2)
dropout_rate = self.attention_dropout if self.training else 0.0
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
# therefore the input hidden states gets silently casted in float32. Hence, we need
# cast them back in the correct dtype just to be sure everything works as expected.
# This might slowdown training & inference so it is recommended to not cast the LayerNorms
# in fp32. (OlmoeRMSNorm handles it correctly)
input_dtype = query_states.dtype
if input_dtype == torch.float32:
if torch.is_autocast_enabled():
target_dtype = torch.get_autocast_gpu_dtype()
# Handle the case where the model is quantized
elif hasattr(self.config, "_pre_quantization_dtype"):
target_dtype = self.config._pre_quantization_dtype
else:
target_dtype = self.q_proj.weight.dtype
logger.warning_once(
f"The input hidden states seems to be silently casted in float32, this might be related to"
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
f" {target_dtype}."
)
query_states = query_states.to(target_dtype)
key_states = key_states.to(target_dtype)
value_states = value_states.to(target_dtype)
attn_output = _flash_attention_forward(
query_states,
key_states,
value_states,
attention_mask,
q_len,
dropout=dropout_rate,
use_top_left_mask=self._flash_attn_uses_top_left_mask,
is_causal=self.is_causal,
)
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
return attn_output, attn_weights, past_key_value
class OlmoeSdpaAttention(OlmoeAttention):
"""
OLMoE attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
`OlmoeAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
SDPA API.
"""
# Adapted from OlmoeAttention.forward
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value: Optional[Cache] = None,
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
if output_attentions:
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
logger.warning_once(
"OlmoeModel is using OlmoeSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
)
return super().forward(
hidden_states=hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_value=past_key_value,
output_attentions=output_attentions,
use_cache=use_cache,
cache_position=cache_position,
position_embeddings=position_embeddings,
)
bsz, q_len, _ = hidden_states.size()
query_states = self.q_norm(self.q_proj(hidden_states))
key_states = self.k_norm(self.k_proj(hidden_states))
value_states = self.v_proj(hidden_states)
if self.config.clip_qkv is not None:
query_states.clamp_(min=-self.config.clip_qkv, max=self.config.clip_qkv)
key_states.clamp_(min=-self.config.clip_qkv, max=self.config.clip_qkv)
value_states.clamp_(min=-self.config.clip_qkv, max=self.config.clip_qkv)
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
cos, sin = position_embeddings
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
if past_key_value 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_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
key_states = repeat_kv(key_states, self.num_key_value_groups)
value_states = repeat_kv(value_states, self.num_key_value_groups)
causal_mask = attention_mask
# if attention_mask is not None and cache_position is not None:
if attention_mask is not None:
causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
# Reference: https://github.com/pytorch/pytorch/issues/112577.
if query_states.device.type == "cuda" and causal_mask is not None:
query_states = query_states.contiguous()
key_states = key_states.contiguous()
value_states = value_states.contiguous()
# We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
# in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
is_causal = True if causal_mask is None and q_len > 1 else False
attn_output = torch.nn.functional.scaled_dot_product_attention(
query_states,
key_states,
value_states,
attn_mask=causal_mask,
dropout_p=self.attention_dropout if self.training else 0.0,
is_causal=is_causal,
)
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.view(bsz, q_len, self.hidden_size)
attn_output = self.o_proj(attn_output)
return attn_output, None, past_key_value
OLMOE_ATTENTION_CLASSES = {
"eager": OlmoeAttention,
"flash_attention_2": OlmoeFlashAttention2,
"sdpa": OlmoeSdpaAttention,
}
class OlmoeSparseMoeBlock(nn.Module):
def __init__(self, config):
super().__init__()
self.num_experts = config.num_experts
self.top_k = config.num_experts_per_tok
self.norm_topk_prob = config.norm_topk_prob
self.gate = nn.Linear(config.hidden_size, self.num_experts, bias=False)
self.experts = nn.ModuleList([OlmoeMLP(config) for _ in range(self.num_experts)])
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
batch_size, sequence_length, hidden_dim = hidden_states.shape
hidden_states = hidden_states.view(-1, hidden_dim)
# router_logits: (batch * sequence_length, n_experts)
router_logits = self.gate(hidden_states)
routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
if self.norm_topk_prob:
routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
# we cast back to the input dtype
routing_weights = routing_weights.to(hidden_states.dtype)
final_hidden_states = torch.zeros(
(batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
)
# One hot encode the selected experts to create an expert mask
# this will be used to easily index which expert is going to be selected
expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
# Loop over all available experts in the model and perform the computation on each expert
for expert_idx in range(self.num_experts):
expert_layer = self.experts[expert_idx]
idx, top_x = torch.where(expert_mask[expert_idx])
# Index the correct hidden states and compute the expert hidden state for
# the current expert. We need to make sure to multiply the output hidden
# states by `routing_weights` on the corresponding tokens (top-1 and top-2)
current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
# However `index_add_` only support torch tensors for indexing so we'll use
# the `top_x` tensor here.
final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
return final_hidden_states, router_logits
class OlmoeDecoderLayer(nn.Module):
def __init__(self, config: OlmoeConfig, layer_idx: int):
super().__init__()
self.hidden_size = config.hidden_size
self.self_attn = OLMOE_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
self.mlp = OlmoeSparseMoeBlock(config)
self.input_layernorm = OlmoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.post_attention_layernorm = OlmoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value: Optional[Cache] = None,
output_attentions: Optional[bool] = False,
output_router_logits: Optional[bool] = False,
use_cache: Optional[bool] = False,
cache_position: Optional[torch.LongTensor] = None,
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
**kwargs,
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
"""
Args:
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
attention_mask (`torch.FloatTensor`, *optional*):
attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
query_sequence_length, key_sequence_length)` if default attention is used.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_router_logits (`bool`, *optional*):
Whether or not to return the logits of all the routers. They are useful for computing the router loss,
and should not be returned during inference.
use_cache (`bool`, *optional*):
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
(see `past_key_values`).
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
Indices depicting the position of the input sequence tokens in the sequence
position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
with `head_dim` being the embedding dimension of each attention head.
kwargs (`dict`, *optional*):
Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
into the model
"""
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
# Self Attention
hidden_states, self_attn_weights, present_key_value = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_value=past_key_value,
output_attentions=output_attentions,
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, router_logits = self.mlp(hidden_states)
hidden_states = residual + hidden_states
outputs = (hidden_states,)
if output_attentions:
outputs += (self_attn_weights,)
if use_cache:
outputs += (present_key_value,)
if output_router_logits:
outputs += (router_logits,)
return outputs
OLMOE_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`OlmoeConfig`]):
Model configuration class with all the parameters of the model. Initializing with a config file does not
load the weights associated with the model, only the configuration. Check out the
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
@add_start_docstrings(
"The bare Olmoe Model outputting raw hidden-states without any specific head on top.",
OLMOE_START_DOCSTRING,
)
class OlmoePreTrainedModel(PreTrainedModel):
config_class = OlmoeConfig
base_model_prefix = "model"
supports_gradient_checkpointing = True
_no_split_modules = ["OlmoeDecoderLayer"]
_skip_keys_device_placement = ["past_key_values"]
_supports_flash_attn_2 = True
_supports_sdpa = True
_supports_cache_class = True
_supports_quantized_cache = True
_supports_static_cache = False # MoE models don't work with torch.compile (`torch.where(condition)` not supported)
def _init_weights(self, module):
std = self.config.initializer_range
if isinstance(module, nn.Linear):
module.weight.data.normal_(mean=0.0, std=std)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=std)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
OLMOE_INPUTS_DOCSTRING = r"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
it.
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
`past_key_values`).
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
information on the default strategy.
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
config.n_positions - 1]`.
[What are position IDs?](../glossary#position-ids)
past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
Two formats are allowed:
- a [`~cache_utils.Cache`] instance;
- Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
cache format.
The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
legacy cache format will be returned.
If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
of shape `(batch_size, sequence_length)`.
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
model's internal embedding lookup matrix.
use_cache (`bool`, *optional*):
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
`past_key_values`).
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
output_router_logits (`bool`, *optional*):
Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
should not be returned during inference.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
the complete sequence length.
"""
@add_start_docstrings(
"The bare Olmoe Model outputting raw hidden-states without any specific head on top.",
OLMOE_START_DOCSTRING,
)
# TODO: re-enable check: Copied from transformers.models.llama.modeling_llama.LlamaModel with Llama->Olmoe
class OlmoeModel(OlmoePreTrainedModel):
"""
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`OlmoeDecoderLayer`]
Args:
config: OlmoeConfig
"""
def __init__(self, config: OlmoeConfig):
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(
[OlmoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
)
self.norm = OlmoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.rotary_emb = OlmoeRotaryEmbedding(config=config)
self.gradient_checkpointing = False
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.embed_tokens
def set_input_embeddings(self, value):
self.embed_tokens = value
@add_start_docstrings_to_model_forward(OLMOE_INPUTS_DOCSTRING)
# Ignore copy
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
output_router_logits: Optional[bool] = None,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.LongTensor] = None,
) -> Union[Tuple, MoeModelOutputWithPast]:
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_router_logits = (
output_router_logits if output_router_logits is not None else self.config.output_router_logits
)
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if (input_ids is None) ^ (inputs_embeds is not None):
raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
if self.gradient_checkpointing and self.training and use_cache:
logger.warning_once(
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
)
use_cache = False
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids)
# kept for BC (non `Cache` `past_key_values` inputs)
return_legacy_cache = False
if use_cache and not isinstance(past_key_values, Cache):
return_legacy_cache = True
if past_key_values is None:
past_key_values = DynamicCache()
else:
past_key_values = DynamicCache.from_legacy_cache(past_key_values)
logger.warning_once(
"We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and "
"will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class "
"(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)"
)
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 = self._update_causal_mask(
attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
)
# embed positions
hidden_states = inputs_embeds
# create position embeddings to be shared across the decoder layers
position_embeddings = self.rotary_emb(hidden_states, position_ids)
# decoder layers
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
all_router_logits = () if output_router_logits else None
next_decoder_cache = None
for decoder_layer in self.layers[: self.config.num_hidden_layers]:
if output_hidden_states:
all_hidden_states += (hidden_states,)
if self.gradient_checkpointing and self.training:
layer_outputs = self._gradient_checkpointing_func(
decoder_layer.__call__,
hidden_states,
causal_mask,
position_ids,
past_key_values,
output_attentions,
output_router_logits,
use_cache,
cache_position,
position_embeddings,
)
else:
layer_outputs = decoder_layer(
hidden_states,
attention_mask=causal_mask,
position_ids=position_ids,
past_key_value=past_key_values,
output_attentions=output_attentions,
output_router_logits=output_router_logits,
use_cache=use_cache,
cache_position=cache_position,
position_embeddings=position_embeddings,
)
hidden_states = layer_outputs[0]
if use_cache:
next_decoder_cache = layer_outputs[2 if output_attentions else 1]
if output_attentions:
all_self_attns += (layer_outputs[1],)
if output_router_logits and layer_outputs[-1] is not None:
all_router_logits += (layer_outputs[-1],)
hidden_states = self.norm(hidden_states)
# add hidden states from the last decoder layer
if output_hidden_states:
all_hidden_states += (hidden_states,)
next_cache = next_decoder_cache if use_cache else None
if return_legacy_cache:
next_cache = next_cache.to_legacy_cache()
if not return_dict:
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
return MoeModelOutputWithPast(
last_hidden_state=hidden_states,
past_key_values=next_cache,
hidden_states=all_hidden_states,
attentions=all_self_attns,
router_logits=all_router_logits,
)
def _update_causal_mask(
self,
attention_mask: torch.Tensor,
input_tensor: torch.Tensor,
cache_position: torch.Tensor,
past_key_values: Cache,
output_attentions: bool,
):
if self.config._attn_implementation == "flash_attention_2":
if attention_mask is not None and 0.0 in attention_mask:
return attention_mask
return None
# For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
# order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
# to infer the attention mask.
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
using_static_cache = isinstance(past_key_values, StaticCache)
# When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
if AttentionMaskConverter._ignore_causal_mask_sdpa(
attention_mask,
inputs_embeds=input_tensor,
past_key_values_length=past_seen_tokens,
is_training=self.training,
):
return None
dtype, device = input_tensor.dtype, input_tensor.device
sequence_length = input_tensor.shape[1]
if using_static_cache:
target_length = past_key_values.get_max_cache_shape()
else:
target_length = (
attention_mask.shape[-1]
if isinstance(attention_mask, torch.Tensor)
else past_seen_tokens + sequence_length + 1
)
# In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
attention_mask,
sequence_length=sequence_length,
target_length=target_length,
dtype=dtype,
device=device,
cache_position=cache_position,
batch_size=input_tensor.shape[0],
)
if (
self.config._attn_implementation == "sdpa"
and attention_mask is not None
and attention_mask.device.type in ["cuda", "xpu"]
and not output_attentions
):
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
# Details: https://github.com/pytorch/pytorch/issues/110213
min_dtype = torch.finfo(dtype).min
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
return causal_mask
@staticmethod
def _prepare_4d_causal_attention_mask_with_cache_position(
attention_mask: torch.Tensor,
sequence_length: int,
target_length: int,
dtype: torch.dtype,
device: torch.device,
cache_position: torch.Tensor,
batch_size: int,
**kwargs,
):
"""
Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
`(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
Args:
attention_mask (`torch.Tensor`):
A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
`(batch_size, 1, query_length, key_value_length)`.
sequence_length (`int`):
The sequence length being processed.
target_length (`int`):
The target length: when generating with static cache, the mask should be as long as the static cache,
to account for the 0 padding, the part of the cache that is not filled yet.
dtype (`torch.dtype`):
The dtype to use for the 4D attention mask.
device (`torch.device`):
The device to place the 4D attention mask on.
cache_position (`torch.Tensor`):
Indices depicting the position of the input sequence tokens in the sequence.
batch_size (`torch.Tensor`):
Batch size.
"""
if attention_mask is not None and attention_mask.dim() == 4:
# In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
causal_mask = attention_mask
else:
min_dtype = torch.finfo(dtype).min
causal_mask = torch.full(
(sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
)
if sequence_length != 1:
causal_mask = torch.triu(causal_mask, diagonal=1)
causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
if attention_mask is not None:
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
mask_length = attention_mask.shape[-1]
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
padding_mask = padding_mask == 0
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
padding_mask, min_dtype
)
return causal_mask
class OlmoeForCausalLM(OlmoePreTrainedModel, GenerationMixin):
_tied_weights_keys = ["lm_head.weight"]
def __init__(self, config):
super().__init__(config)
self.model = OlmoeModel(config)
self.vocab_size = config.vocab_size
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
self.router_aux_loss_coef = config.router_aux_loss_coef
self.num_experts = config.num_experts
self.num_experts_per_tok = config.num_experts_per_tok
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.model.embed_tokens
def set_input_embeddings(self, value):
self.model.embed_tokens = value
def get_output_embeddings(self):
return self.lm_head
def set_output_embeddings(self, new_embeddings):
self.lm_head = new_embeddings
def set_decoder(self, decoder):
self.model = decoder
def get_decoder(self):
return self.model
@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
@add_start_docstrings_to_model_forward(OLMOE_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[List[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
output_router_logits: Optional[bool] = None,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.LongTensor] = None,
logits_to_keep: Union[int, torch.Tensor] = 0,
**loss_kwargs,
) -> Union[Tuple, MoeCausalLMOutputWithPast]:
r"""
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
logits_to_keep (`int` or `torch.Tensor`, *optional*):
If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
`input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
This is useful when using packed tensor format (single dimension for batch and sequence length).
Returns:
Example:
```python
>>> from transformers import AutoTokenizer, OlmoeForCausalLM
>>> model = OlmoeForCausalLM.from_pretrained("allenai/OLMoE-1B-7B-0924")
>>> tokenizer = AutoTokenizer.from_pretrained("allenai/OLMoE-1B-7B-0924")
>>> prompt = "Hey, are you conscious? Can you talk to me?"
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> # Generate
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
'Hey, are you conscious? Can you talk to me?\nI’m not sure if you’re conscious of this, but I’m'
```
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_router_logits = (
output_router_logits if output_router_logits is not None else self.config.output_router_logits
)
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
outputs = 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,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
output_router_logits=output_router_logits,
return_dict=return_dict,
cache_position=cache_position,
)
hidden_states = outputs[0]
# 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, self.vocab_size, **loss_kwargs)
aux_loss = None
if output_router_logits:
aux_loss = load_balancing_loss_func(
outputs.router_logits if return_dict else outputs[-1],
self.num_experts,
self.num_experts_per_tok,
attention_mask,
)
if labels is not None:
loss += self.router_aux_loss_coef * aux_loss.to(loss.device) # make sure to reside in the same device
if not return_dict:
output = (logits,) + outputs[1:]
if output_router_logits:
output = (aux_loss,) + output
return (loss,) + output if loss is not None else output
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,
)
__all__ = ["OlmoeForCausalLM", "OlmoeModel", "OlmoePreTrainedModel"]