dnrops
/
transformers
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

updates

This commit is contained in:
Arthur Zucker 2024-05-15 17:10:14 +02:00
parent 3dedb93c45
commit f3fe0b340a
6 changed files with 758 additions and 256 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

176
src/transformers/models/cohere/modeling_cohere.py
View File

@ -41,25 +41,29 @@ class CohereRotaryEmbedding(nn.Module):
self.base = base
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
@torch.no_grad()
def compute_cos_sin(self, x, position_ids):
# x: [bs, num_attention_heads, seq_len, head_size]
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
position_ids_expanded = position_ids[:, None, :].float()
# Force float32 since bfloat16 loses precision on long contexts
# See https://github.com/huggingface/transformers/pull/29285
device_type = x.device.type
device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
with torch.autocast(device_type=device_type, enabled=False):
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
emb = torch.cat((freqs, freqs), dim=-1)
return emb.cos(), emb.sin()
def rotate_half(self, 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)
# For BC we register cos and sin cached
self.max_seq_len_cached = max_position_embeddings
t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)
t = t / self.scaling_factor
freqs = torch.outer(t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("_cos_cached", emb.cos().to(torch.get_default_dtype()), persistent=False)
self.register_buffer("_sin_cached", emb.sin().to(torch.get_default_dtype()), persistent=False)
@property
def sin_cached(self):
logger.warning_once(
"The sin_cached attribute will be removed in 4.39. Bear in mind that its contents changed in v4.38. Use "
"the forward method of RoPE from now on instead. It is not used in the `LlamaAttention` class"
)
return self._sin_cached
@property
def cos_cached(self):
logger.warning_once(
"The cos_cached attribute will be removed in 4.39. Bear in mind that its contents changed in v4.38. Use "
"the forward method of RoPE from now on instead. It is not used in the `LlamaAttention` class"
)
return self._cos_cached
def forward(self, q, k, position_ids=None, unsqueeze_dim=1):
dtype = q.dtype
q,k = q.float(), k.float()
@ -106,7 +110,25 @@ class CohereMLP(nn.Module):
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))
if self.config.pretraining_tp > 1:
slice = self.intermediate_size // self.config.pretraining_tp
gate_proj_slices = self.gate_proj.weight.split(slice, dim=0)
up_proj_slices = self.up_proj.weight.split(slice, dim=0)
down_proj_slices = self.down_proj.weight.split(slice, dim=1)
gate_proj = torch.cat(
[F.linear(x, gate_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1
)
up_proj = torch.cat([F.linear(x, up_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1)
intermediate_states = (self.act_fn(gate_proj) * up_proj).split(slice, dim=2)
down_proj = [
F.linear(intermediate_states[i], down_proj_slices[i]) for i in range(self.config.pretraining_tp)
]
down_proj = sum(down_proj)
else:
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
return down_proj
class CohereAttention(nn.Module):
@ -181,22 +203,42 @@ class CohereAttention(nn.Module):
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
if self.config.pretraining_tp > 1:
key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
query_slices = self.q_proj.weight.split(
(self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
)
key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
query_states = torch.cat(query_states, dim=-1)
key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
key_states = torch.cat(key_states, dim=-1)
value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
value_states = torch.cat(value_states, dim=-1)
else:
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
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)
query_states, key_states = self.rotary_emb(query_states, key_states, position_ids)
cos, sin = self.rotary_emb(value_states, position_ids)
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 = {"cache_position": cache_position}
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)
@ -222,7 +264,13 @@ class CohereAttention(nn.Module):
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 self.config.pretraining_tp > 1:
attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
else:
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
@ -301,22 +349,42 @@ class CohereSdpaAttention(nn.Module):
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
if self.config.pretraining_tp > 1:
key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
query_slices = self.q_proj.weight.split(
(self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
)
key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
query_states = torch.cat(query_states, dim=-1)
key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
key_states = torch.cat(key_states, dim=-1)
value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
value_states = torch.cat(value_states, dim=-1)
else:
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
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)
query_states, key_states = self.rotary_emb(query_states, key_states, position_ids)
cos, sin = self.rotary_emb(value_states, position_ids)
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 = {"cache_position": cache_position}
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)
@ -342,7 +410,13 @@ class CohereSdpaAttention(nn.Module):
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 self.config.pretraining_tp > 1:
attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
else:
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
@ -421,22 +495,42 @@ class CohereFlashAttention2(nn.Module):
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
if self.config.pretraining_tp > 1:
key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
query_slices = self.q_proj.weight.split(
(self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
)
key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
query_states = torch.cat(query_states, dim=-1)
key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
key_states = torch.cat(key_states, dim=-1)
value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
value_states = torch.cat(value_states, dim=-1)
else:
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
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)
query_states, key_states = self.rotary_emb(query_states, key_states, position_ids)
cos, sin = self.rotary_emb(value_states, position_ids)
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 = {"cache_position": cache_position}
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)
@ -462,7 +556,13 @@ class CohereFlashAttention2(nn.Module):
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 self.config.pretraining_tp > 1:
attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
else:
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None

235
src/transformers/models/gemma/diff_gemma.py
View File

@ -5,6 +5,234 @@ from transformers.models.llama.modeling_llama import *
import torch.nn as nn
from transformers.utils import ModelConverter
# coding=utf-8
# Copyright 2024 Google Inc. 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.
""" PyTorch Gemma model."""
import math
import warnings
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
from ...modeling_attn_mask_utils import (
AttentionMaskConverter,
_prepare_4d_causal_attention_mask,
)
from ...modeling_outputs import BaseModelOutputWithPast
from ...pytorch_utils import ALL_LAYERNORM_LAYERS, is_torch_greater_or_equal_than_1_13
from ...utils import (
is_flash_attn_2_available,
logging,
)
from ...utils.import_utils import is_torch_fx_available
from .configuration_gemma import GemmaConfig
from transformers.models.llama.modeling_llama import repeat_kv, rotate_half, apply_rotary_pos_emb
logger = logging.get_logger(__name__)
class GemmaRMSNorm(nn.Module):
def __init__(self, dim: int, eps: float = 1e-6):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.zeros(dim))
def _norm(self, x):
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
def forward(self, x):
output = self._norm(x.float())
# Llama does x.to(float16) * w whilst Gemma is (x * w).to(float16)
# See https://github.com/huggingface/transformers/pull/29402
output = output * (1.0 + self.weight.float())
return output.type_as(x)
ALL_LAYERNORM_LAYERS.append(GemmaRMSNorm)
class GemmaRotaryEmbedding(nn.Module):
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
super().__init__()
self.dim = dim
self.max_position_embeddings = max_position_embeddings
self.base = base
self.register_buffer("inv_freq", None, persistent=False)
@torch.no_grad()
def forward(self, x, position_ids, seq_len=None):
# x: [bs, num_attention_heads, seq_len, head_size]
if self.inv_freq is None:
self.inv_freq = 1.0 / (
self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim)
)
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
position_ids_expanded = position_ids[:, None, :].float()
# Force float32 since bfloat16 loses precision on long contexts
# See https://github.com/huggingface/transformers/pull/29285
device_type = x.device.type
device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
with torch.autocast(device_type=device_type, enabled=False):
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
emb = torch.cat((freqs, freqs), dim=-1)
cos = emb.cos()
sin = emb.sin()
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
class GemmaMLP(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)
if config.hidden_activation is None:
logger.warning_once(
"`config.hidden_act` is ignored, you should use `config.hidden_activation` instead.\n"
"Gemma's activation function will be set to `gelu_pytorch_tanh`. Please, use\n"
"`config.hidden_activation` if you want to override this behaviour.\n"
"See https://github.com/huggingface/transformers/pull/29402 for more details."
)
config.hidden_activation = "gelu_pytorch_tanh"
hidden_activation = config.hidden_activation
self.act_fn = ACT2FN[hidden_activation]
def forward(self, x):
return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
class GemmaAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
# Ignore copy
def __init__(self, config: GemmaConfig, 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 = config.head_dim
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.hidden_size % self.num_heads != 0:
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.num_heads * self.head_dim, self.hidden_size, bias=config.attention_bias)
self.rotary_emb = GemmaRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
base=self.rope_theta,
)
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,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
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 = self.rotary_emb(value_states, position_ids, seq_len=None)
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, None)
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.view(bsz, q_len, -1)
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
return attn_output, attn_weights, past_key_value
GemmaConverter = ModelConverter(__file__)
class GemmaRMSNorm(nn.Module):
@ -23,10 +251,6 @@ class GemmaRMSNorm(nn.Module):
output = output * (1.0 + self.weight.float())
return output.type_as(x)
GemmaConverter.register("GemmaRotaryEmbedding", LlamaRotaryEmbedding)
GemmaConverter.register("GemmaMLP", LlamaMLP)
GemmaAttention = GemmaConverter.register("GemmaAttention", LlamaAttention)
GemmaFlashAttention2 = GemmaConverter.register("GemmaFlashAttention2", LlamaFlashAttention2)
GemmaSdpaAttention = GemmaConverter.register("GemmaSdpaAttention", LlamaSdpaAttention)
@ -58,4 +282,5 @@ class GemmaModel(LlamaModel):
return super().forward(None, attention_mask, position_ids, past_key_values, inputs_embeds, use_cache, output_attentions, output_hidden_states, return_dict, cache_position)
GemmaConverter.register("GemmaForCausalLM", LlamaForCausalLM)
GemmaConverter.register("GemmaForCausalLM", LlamaForCausalLM)

207
src/transformers/models/gemma/modeling_gemma.py
View File

@ -17,6 +17,56 @@ import torch.nn as nn
from transformers.utils import ModelConverter
# coding=utf-8
# Copyright 2024 Google Inc. 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.
""" PyTorch Gemma model."""
import math
import warnings
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
from ...modeling_attn_mask_utils import (
AttentionMaskConverter,
_prepare_4d_causal_attention_mask,
)
from ...modeling_outputs import BaseModelOutputWithPast
from ...pytorch_utils import ALL_LAYERNORM_LAYERS, is_torch_greater_or_equal_than_1_13
from ...utils import (
is_flash_attn_2_available,
logging,
)
from ...utils.import_utils import is_torch_fx_available
from .configuration_gemma import GemmaConfig
from transformers.models.llama.modeling_llama import repeat_kv, rotate_half, apply_rotary_pos_emb
logger = logging.get_logger(__name__)
class GemmaRMSNorm(nn.Module):
def __init__(self, dim: int, eps: float = 1e-6):
super().__init__()
@ -33,19 +83,26 @@ class GemmaRMSNorm(nn.Module):
output = output * (1.0 + self.weight.float())
return output.type_as(x)
ALL_LAYERNORM_LAYERS.append(GemmaRMSNorm)
class GemmaRotaryEmbedding(nn.Module):
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None, scaling_factor=1.0):
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
super().__init__()
self.scaling_factor = scaling_factor
self.dim = dim
self.max_position_embeddings = max_position_embeddings
self.base = base
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
self.register_buffer("inv_freq", None, persistent=False)
@torch.no_grad()
def compute_cos_sin(self, x, position_ids):
def forward(self, x, position_ids, seq_len=None):
# x: [bs, num_attention_heads, seq_len, head_size]
if self.inv_freq is None:
self.inv_freq = 1.0 / (
self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim)
)
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
position_ids_expanded = position_ids[:, None, :].float()
# Force float32 since bfloat16 loses precision on long contexts
@ -55,40 +112,11 @@ class GemmaRotaryEmbedding(nn.Module):
with torch.autocast(device_type=device_type, enabled=False):
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
emb = torch.cat((freqs, freqs), dim=-1)
return emb.cos(), emb.sin()
def rotate_half(self, 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)
cos = emb.cos()
sin = emb.sin()
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
@torch.no_grad()
@add_start_docstrings("")
def forward(self, q, k, 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.
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, sin = self.compute_cos_sin(q, position_ids)
cos = cos.unsqueeze(unsqueeze_dim).to(dtype=q.dtype)
sin = sin.unsqueeze(unsqueeze_dim).to(dtype=q.dtype)
q_embed = (q * cos) + (self.rotate_half(q) * sin)
k_embed = (k * cos) + (self.rotate_half(k) * sin)
return q_embed, k_embed
class GemmaMLP(nn.Module):
def __init__(self, config):
@ -96,19 +124,29 @@ class GemmaMLP(nn.Module):
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=config.mlp_bias)
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
self.act_fn = ACT2FN[config.hidden_act]
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)
if config.hidden_activation is None:
logger.warning_once(
"`config.hidden_act` is ignored, you should use `config.hidden_activation` instead.\n"
"Gemma's activation function will be set to `gelu_pytorch_tanh`. Please, use\n"
"`config.hidden_activation` if you want to override this behaviour.\n"
"See https://github.com/huggingface/transformers/pull/29402 for more details."
)
config.hidden_activation = "gelu_pytorch_tanh"
hidden_activation = config.hidden_activation
self.act_fn = ACT2FN[hidden_activation]
def forward(self, x):
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
return down_proj
return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
class GemmaAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
# Ignore copy
def __init__(self, config: GemmaConfig, layer_idx: Optional[int] = None):
super().__init__()
self.config = config
@ -123,14 +161,14 @@ class GemmaAttention(nn.Module):
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.head_dim = config.head_dim
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:
if self.hidden_size % self.num_heads != 0:
raise ValueError(
f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
f" and `num_heads`: {self.num_heads})."
@ -139,35 +177,12 @@ class GemmaAttention(nn.Module):
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._init_rope()
def _init_rope(self):
if self.config.rope_scaling is None:
self.rotary_emb = GemmaRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
base=self.rope_theta,
)
else:
scaling_type = self.config.rope_scaling["type"]
scaling_factor = self.config.rope_scaling["factor"]
if scaling_type == "linear":
self.rotary_emb = GemmaLinearScalingRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
scaling_factor=scaling_factor,
base=self.rope_theta,
)
elif scaling_type == "dynamic":
self.rotary_emb = GemmaDynamicNTKScalingRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
scaling_factor=scaling_factor,
base=self.rope_theta,
)
else:
raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.attention_bias)
self.rotary_emb = GemmaRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
base=self.rope_theta,
)
def forward(
self,
@ -178,6 +193,7 @@ class GemmaAttention(nn.Module):
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
@ -189,11 +205,12 @@ class GemmaAttention(nn.Module):
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)
query_states, key_states = self.rotary_emb(query_states, key_states, position_ids)
cos, sin = self.rotary_emb(value_states, position_ids, seq_len=None)
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, None)
if past_key_value is not None:
# sin and cos are specific to RoPE models; cache_position needed for the static cache
cache_kwargs = {"cache_position": cache_position}
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)
@ -218,7 +235,7 @@ class GemmaAttention(nn.Module):
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
attn_output = attn_output.view(bsz, q_len, -1)
attn_output = self.o_proj(attn_output)
if not output_attentions:
@ -226,6 +243,24 @@ class GemmaAttention(nn.Module):
return attn_output, attn_weights, past_key_value
class GemmaRMSNorm(nn.Module):
def __init__(self, dim: int, eps: float = 1e-6):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.zeros(dim))
def _norm(self, x):
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
def forward(self, x):
output = self._norm(x.float())
# Llama does x.to(float16) * w whilst Gemma is (x * w).to(float16)
# See https://github.com/huggingface/transformers/pull/29402
output = output * (1.0 + self.weight.float())
return output.type_as(x)
class GemmaFlashAttention2(GemmaAttention):
"""
Gemma flash attention module. This module inherits from `GemmaAttention` as the weights of the module stays
@ -250,6 +285,7 @@ class GemmaFlashAttention2(GemmaAttention):
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
if isinstance(past_key_value, StaticCache):
raise ValueError(
@ -272,8 +308,8 @@ class GemmaFlashAttention2(GemmaAttention):
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)
query_states, key_states = self.rotary_emb(query_states, key_states, position_ids)
cos, sin = self.rotary_emb(value_states, position_ids)
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
@ -467,12 +503,12 @@ class GemmaSdpaAttention(GemmaAttention):
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)
query_states, key_states = self.rotary_emb(query_states, key_states, position_ids)
cos, sin = self.rotary_emb(value_states, position_ids)
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 = {"cache_position": cache_position}
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)
@ -532,6 +568,7 @@ class GemmaDecoderLayer(nn.Module):
output_attentions: Optional[bool] = False,
use_cache: Optional[bool] = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
"""
Args:
@ -547,6 +584,10 @@ class GemmaDecoderLayer(nn.Module):
(see `past_key_values`).
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
"""
if "padding_mask" in kwargs:
warnings.warn(
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
)
residual = hidden_states
@ -561,6 +602,7 @@ class GemmaDecoderLayer(nn.Module):
output_attentions=output_attentions,
use_cache=use_cache,
cache_position=cache_position,
**kwargs,
)
hidden_states = residual + hidden_states
@ -1087,3 +1129,4 @@ class GemmaForCausalLM(GemmaPreTrainedModel):
)
return reordered_past

166
src/transformers/models/persimmon/modeling_persimmon.py
View File

@ -24,9 +24,34 @@ class PersimmonRotaryEmbedding(nn.Module):
self.base = base
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
# For BC we register cos and sin cached
self.max_seq_len_cached = max_position_embeddings
t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)
t = t / self.scaling_factor
freqs = torch.outer(t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("_cos_cached", emb.cos().to(torch.get_default_dtype()), persistent=False)
self.register_buffer("_sin_cached", emb.sin().to(torch.get_default_dtype()), persistent=False)
@property
def sin_cached(self):
logger.warning_once(
"The sin_cached attribute will be removed in 4.39. Bear in mind that its contents changed in v4.38. Use "
"the forward method of RoPE from now on instead. It is not used in the `PersimmonAttention` class"
)
return self._sin_cached
@property
def cos_cached(self):
logger.warning_once(
"The cos_cached attribute will be removed in 4.39. Bear in mind that its contents changed in v4.38. Use "
"the forward method of RoPE from now on instead. It is not used in the `PersimmonAttention` class"
)
return self._cos_cached
@torch.no_grad()
def compute_cos_sin(self, x, position_ids):
def forward(self, x, position_ids):
# x: [bs, num_attention_heads, seq_len, head_size]
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
position_ids_expanded = position_ids[:, None, :].float()
@ -37,40 +62,9 @@ class PersimmonRotaryEmbedding(nn.Module):
with torch.autocast(device_type=device_type, enabled=False):
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
emb = torch.cat((freqs, freqs), dim=-1)
return emb.cos(), emb.sin()
def rotate_half(self, 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)
@torch.no_grad()
@add_start_docstrings("")
def forward(self, q, k, 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.
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, sin = self.compute_cos_sin(q, position_ids)
cos = cos.unsqueeze(unsqueeze_dim).to(dtype=q.dtype)
sin = sin.unsqueeze(unsqueeze_dim).to(dtype=q.dtype)
q_embed = (q * cos) + (self.rotate_half(q) * sin)
k_embed = (k * cos) + (self.rotate_half(k) * sin)
return q_embed, k_embed
cos = emb.cos()
sin = emb.sin()
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
class PersimmonMLP(nn.Module):
def __init__(self, config):
@ -84,7 +78,25 @@ class PersimmonMLP(nn.Module):
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))
if self.config.pretraining_tp > 1:
slice = self.intermediate_size // self.config.pretraining_tp
gate_proj_slices = self.gate_proj.weight.split(slice, dim=0)
up_proj_slices = self.up_proj.weight.split(slice, dim=0)
down_proj_slices = self.down_proj.weight.split(slice, dim=1)
gate_proj = torch.cat(
[F.linear(x, gate_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1
)
up_proj = torch.cat([F.linear(x, up_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1)
intermediate_states = (self.act_fn(gate_proj) * up_proj).split(slice, dim=2)
down_proj = [
F.linear(intermediate_states[i], down_proj_slices[i]) for i in range(self.config.pretraining_tp)
]
down_proj = sum(down_proj)
else:
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
return down_proj
@ -347,22 +359,42 @@ class PersimmonSdpaAttention(nn.Module):
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
if self.config.pretraining_tp > 1:
key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
query_slices = self.q_proj.weight.split(
(self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
)
key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
query_states = torch.cat(query_states, dim=-1)
key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
key_states = torch.cat(key_states, dim=-1)
value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
value_states = torch.cat(value_states, dim=-1)
else:
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
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)
query_states, key_states = self.rotary_emb(query_states, key_states, position_ids)
cos, sin = self.rotary_emb(value_states, position_ids)
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 = {"cache_position": cache_position}
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)
@ -388,7 +420,13 @@ class PersimmonSdpaAttention(nn.Module):
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 self.config.pretraining_tp > 1:
attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
else:
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
@ -467,22 +505,42 @@ class PersimmonFlashAttention2(nn.Module):
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
if self.config.pretraining_tp > 1:
key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
query_slices = self.q_proj.weight.split(
(self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
)
key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
query_states = torch.cat(query_states, dim=-1)
key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
key_states = torch.cat(key_states, dim=-1)
value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
value_states = torch.cat(value_states, dim=-1)
else:
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
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)
query_states, key_states = self.rotary_emb(query_states, key_states, position_ids)
cos, sin = self.rotary_emb(value_states, position_ids)
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 = {"cache_position": cache_position}
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)
@ -508,7 +566,13 @@ class PersimmonFlashAttention2(nn.Module):
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 self.config.pretraining_tp > 1:
attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
else:
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
@ -538,6 +602,7 @@ class PersimmonDecoderLayer(nn.Module):
output_attentions: Optional[bool] = False,
use_cache: Optional[bool] = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
"""
Args:
@ -553,6 +618,10 @@ class PersimmonDecoderLayer(nn.Module):
(see `past_key_values`).
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
"""
if "padding_mask" in kwargs:
warnings.warn(
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
)
residual = hidden_states
@ -567,6 +636,7 @@ class PersimmonDecoderLayer(nn.Module):
output_attentions=output_attentions,
use_cache=use_cache,
cache_position=cache_position,
**kwargs,
)
hidden_states = residual + hidden_states

84
src/transformers/models/stablelm/modeling_stablelm.py
View File

@ -42,9 +42,34 @@ class StarcoderRotaryEmbedding(nn.Module):
self.base = base
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
# For BC we register cos and sin cached
self.max_seq_len_cached = max_position_embeddings
t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)
t = t / self.scaling_factor
freqs = torch.outer(t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("_cos_cached", emb.cos().to(torch.get_default_dtype()), persistent=False)
self.register_buffer("_sin_cached", emb.sin().to(torch.get_default_dtype()), persistent=False)
@property
def sin_cached(self):
logger.warning_once(
"The sin_cached attribute will be removed in 4.39. Bear in mind that its contents changed in v4.38. Use "
"the forward method of RoPE from now on instead. It is not used in the `StarcoderAttention` class"
)
return self._sin_cached
@property
def cos_cached(self):
logger.warning_once(
"The cos_cached attribute will be removed in 4.39. Bear in mind that its contents changed in v4.38. Use "
"the forward method of RoPE from now on instead. It is not used in the `StarcoderAttention` class"
)
return self._cos_cached
@torch.no_grad()
def compute_cos_sin(self, x, position_ids):
def forward(self, x, position_ids):
# x: [bs, num_attention_heads, seq_len, head_size]
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
position_ids_expanded = position_ids[:, None, :].float()
@ -55,40 +80,9 @@ class StarcoderRotaryEmbedding(nn.Module):
with torch.autocast(device_type=device_type, enabled=False):
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
emb = torch.cat((freqs, freqs), dim=-1)
return emb.cos(), emb.sin()
def rotate_half(self, 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)
@torch.no_grad()
@add_start_docstrings("")
def forward(self, q, k, 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.
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, sin = self.compute_cos_sin(q, position_ids)
cos = cos.unsqueeze(unsqueeze_dim).to(dtype=q.dtype)
sin = sin.unsqueeze(unsqueeze_dim).to(dtype=q.dtype)
q_embed = (q * cos) + (self.rotate_half(q) * sin)
k_embed = (k * cos) + (self.rotate_half(k) * sin)
return q_embed, k_embed
cos = emb.cos()
sin = emb.sin()
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
class StableLmMLP(nn.Module):
def __init__(self, config):
@ -102,7 +96,25 @@ class StableLmMLP(nn.Module):
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))
if self.config.pretraining_tp > 1:
slice = self.intermediate_size // self.config.pretraining_tp
gate_proj_slices = self.gate_proj.weight.split(slice, dim=0)
up_proj_slices = self.up_proj.weight.split(slice, dim=0)
down_proj_slices = self.down_proj.weight.split(slice, dim=1)
gate_proj = torch.cat(
[F.linear(x, gate_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1
)
up_proj = torch.cat([F.linear(x, up_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1)
intermediate_states = (self.act_fn(gate_proj) * up_proj).split(slice, dim=2)
down_proj = [
F.linear(intermediate_states[i], down_proj_slices[i]) for i in range(self.config.pretraining_tp)
]
down_proj = sum(down_proj)
else:
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
return down_proj

146
src/transformers/models/starcoder2/modeling_starcoder2.py
View File

@ -59,9 +59,34 @@ class StarcoderRotaryEmbedding(nn.Module):
self.base = base
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
# For BC we register cos and sin cached
self.max_seq_len_cached = max_position_embeddings
t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)
t = t / self.scaling_factor
freqs = torch.outer(t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("_cos_cached", emb.cos().to(torch.get_default_dtype()), persistent=False)
self.register_buffer("_sin_cached", emb.sin().to(torch.get_default_dtype()), persistent=False)
@property
def sin_cached(self):
logger.warning_once(
"The sin_cached attribute will be removed in 4.39. Bear in mind that its contents changed in v4.38. Use "
"the forward method of RoPE from now on instead. It is not used in the `StarcoderAttention` class"
)
return self._sin_cached
@property
def cos_cached(self):
logger.warning_once(
"The cos_cached attribute will be removed in 4.39. Bear in mind that its contents changed in v4.38. Use "
"the forward method of RoPE from now on instead. It is not used in the `StarcoderAttention` class"
)
return self._cos_cached
@torch.no_grad()
def compute_cos_sin(self, x, position_ids):
def forward(self, x, position_ids):
# x: [bs, num_attention_heads, seq_len, head_size]
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
position_ids_expanded = position_ids[:, None, :].float()
@ -72,40 +97,9 @@ class StarcoderRotaryEmbedding(nn.Module):
with torch.autocast(device_type=device_type, enabled=False):
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
emb = torch.cat((freqs, freqs), dim=-1)
return emb.cos(), emb.sin()
def rotate_half(self, 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)
@torch.no_grad()
@add_start_docstrings("")
def forward(self, q, k, 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.
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, sin = self.compute_cos_sin(q, position_ids)
cos = cos.unsqueeze(unsqueeze_dim).to(dtype=q.dtype)
sin = sin.unsqueeze(unsqueeze_dim).to(dtype=q.dtype)
q_embed = (q * cos) + (self.rotate_half(q) * sin)
k_embed = (k * cos) + (self.rotate_half(k) * sin)
return q_embed, k_embed
cos = emb.cos()
sin = emb.sin()
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
class Starcoder2MLP(nn.Module):
@ -438,22 +432,42 @@ class Starcoder2SdpaAttention(nn.Module):
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
if self.config.pretraining_tp > 1:
key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
query_slices = self.q_proj.weight.split(
(self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
)
key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
query_states = torch.cat(query_states, dim=-1)
key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
key_states = torch.cat(key_states, dim=-1)
value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
value_states = torch.cat(value_states, dim=-1)
else:
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
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)
query_states, key_states = self.rotary_emb(query_states, key_states, position_ids)
cos, sin = self.rotary_emb(value_states, position_ids)
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 = {"cache_position": cache_position}
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)
@ -479,7 +493,13 @@ class Starcoder2SdpaAttention(nn.Module):
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 self.config.pretraining_tp > 1:
attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
else:
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
@ -558,22 +578,42 @@ class Starcoder2FlashAttention2(nn.Module):
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
if self.config.pretraining_tp > 1:
key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
query_slices = self.q_proj.weight.split(
(self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
)
key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
query_states = torch.cat(query_states, dim=-1)
key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
key_states = torch.cat(key_states, dim=-1)
value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
value_states = torch.cat(value_states, dim=-1)
else:
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
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)
query_states, key_states = self.rotary_emb(query_states, key_states, position_ids)
cos, sin = self.rotary_emb(value_states, position_ids)
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 = {"cache_position": cache_position}
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)
@ -599,7 +639,13 @@ class Starcoder2FlashAttention2(nn.Module):
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 self.config.pretraining_tp > 1:
attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
else:
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
@ -630,6 +676,7 @@ class Starcoder2DecoderLayer(nn.Module):
output_attentions: Optional[bool] = False,
use_cache: Optional[bool] = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
"""
Args:
@ -645,6 +692,10 @@ class Starcoder2DecoderLayer(nn.Module):
(see `past_key_values`).
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
"""
if "padding_mask" in kwargs:
warnings.warn(
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
)
residual = hidden_states
@ -659,6 +710,7 @@ class Starcoder2DecoderLayer(nn.Module):
output_attentions=output_attentions,
use_cache=use_cache,
cache_position=cache_position,
**kwargs,
)
hidden_states = residual + hidden_states