ipex-llm/python/llm/src/ipex_llm/transformers/models/falcon.py

829 lines
33 KiB
Python

#
# Copyright 2016 The BigDL Authors.
#
# 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.
#
# Some parts of this file is adapted from
# https://github.com/huggingface/transformers/blob/v4.31.0/src/transformers/models/falcon/modeling_falcon.py
# which is licensed under Apache License 2.0:
#
# Copyright 2023 the Falcon authors and 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 Falcon model."""
import math
from typing import Optional, Tuple
import torch
from torch.nn import functional as F
from ipex_llm.utils.common import invalidInputError
from ipex_llm.transformers.models.utils import init_kv_cache, extend_kv_cache, append_kv_cache
import warnings
import os
KV_CACHE_ALLOC_BLOCK_LENGTH = int(os.environ.get("KV_CACHE_ALLOC_BLOCK_LENGTH", 256))
# 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, 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`):
The position indices of the tokens corresponding to the query and key tensors. For
example, this can be used to pass offsetted position ids when working with a KV-cache.
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[position_ids].unsqueeze(unsqueeze_dim)
sin = sin[position_ids].unsqueeze(unsqueeze_dim)
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
return q_embed, k_embed
def rw_attention_forward_7b(
self,
hidden_states: torch.Tensor,
alibi: torch.Tensor,
attention_mask: torch.Tensor,
layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]]=None,
head_mask: Optional[torch.Tensor]=None,
use_cache: bool=False,
output_attentions: bool=False,
):
fused_qkv = self.query_key_value(hidden_states) # [batch_size, seq_length, 3 x hidden_size]
# 3 x [batch_size, seq_length, num_heads, head_dim]
(query_layer, key_layer, value_layer) = self._split_heads(fused_qkv)
batch_size, q_length, _, _ = query_layer.shape
query_layer = query_layer.transpose(1, 2).reshape(
batch_size * self.num_heads,
q_length,
self.head_dim
)
key_layer = key_layer.transpose(1, 2).reshape(
batch_size * self.num_kv,
q_length,
self.head_dim,
)
value_layer = value_layer.transpose(1, 2).reshape(
batch_size * self.num_kv,
q_length,
self.head_dim
)
# query_layer, key_layer = self.maybe_rotary(query_layer, key_layer)
_, seq_len, _ = query_layer.shape
if layer_past is not None:
_, seq_len_past, _ = layer_past[0].shape
seq_len = seq_len + seq_len_past
query_layer, key_layer = self.maybe_rotary(query_layer, key_layer, seq_len)
_, kv_length, _ = key_layer.shape
if layer_past is not None:
kv_length += layer_past[0].shape[-2]
query_layer = query_layer.view(batch_size, self.num_heads, q_length, self.head_dim)
key_layer = key_layer.view(batch_size, self.num_kv, q_length, self.head_dim)
value_layer = value_layer.view(batch_size, self.num_kv, q_length, self.head_dim)
device = hidden_states.device
if layer_past is not None:
# reuse k, v, self_attention
cache_k = layer_past[0].view(batch_size, self.num_kv, -1, self.head_dim)
cache_v = layer_past[1].view(batch_size, self.num_kv, -1, self.head_dim)
if cache_k.stride()[1] < kv_length * cache_k.size(3):
# allocate new
new_cache_k, new_cache_v = extend_kv_cache(
batch_size,
self.num_kv,
self.head_dim,
cache_k.size(2),
kv_length + KV_CACHE_ALLOC_BLOCK_LENGTH,
dtype=cache_k.dtype,
device=device
)
new_cache_k[:] = cache_k
new_cache_v[:] = cache_v
cache_k = new_cache_k
cache_v = new_cache_v
key_layer, value_layer = append_kv_cache(cache_k, cache_v, key_layer, value_layer)
elif use_cache:
max_cache_length = kv_length + KV_CACHE_ALLOC_BLOCK_LENGTH
new_key_states, new_value_states = init_kv_cache(
batch_size,
self.num_kv,
self.head_dim,
kv_length,
max_cache_length,
dtype=key_layer.dtype,
device=device
)
new_key_states[:] = key_layer
new_value_states[:] = value_layer
key_layer = new_key_states
value_layer = new_value_states
query_layer = query_layer.view(batch_size*self.num_heads, -1, self.head_dim)
key_layer = key_layer.view(batch_size*self.num_kv, -1, self.head_dim)
value_layer = value_layer.view(batch_size*self.num_kv, -1, self.head_dim)
_, kv_length, _ = key_layer.shape
if use_cache is True:
present = (key_layer, value_layer)
else:
present = None
if alibi is None:
query_layer_ = query_layer.reshape(batch_size, self.num_heads, -1, self.head_dim)
key_layer_ = key_layer.reshape(batch_size, self.num_kv, -1, self.head_dim)
value_layer_ = value_layer.reshape(batch_size, self.num_kv, -1, self.head_dim)
# attn_output = F.scaled_dot_product_attention(
# query_layer_, key_layer_, value_layer_, None, 0.0, is_causal=True
# )
if layer_past is not None:
L = query_layer_.shape[-2]
S = key_layer_.shape[-2]
attn_mask = torch.ones(L, S, dtype=torch.bool, device=query_layer_.device)
attn_output = F.scaled_dot_product_attention(
query_layer_, key_layer_, value_layer_, attn_mask, 0.0, is_causal=False
)
else:
attn_output = F.scaled_dot_product_attention(
query_layer_, key_layer_, value_layer_, None, 0.0, is_causal=True
)
x = attn_output.view(batch_size, self.num_heads, q_length, self.head_dim)
x = x.permute(0, 2, 1, 3)
attn_output = x.reshape(batch_size, q_length, self.num_heads * self.head_dim)
output_tensor = self.dense(attn_output)
outputs = (output_tensor, present)
if output_attentions:
invalidInputError(False,
f"'output_attentions' are not supported yet")
return outputs
else:
attention_mask_float = (attention_mask * 1.0) \
.masked_fill(attention_mask, -1e9).to(torch.bfloat16)
matmul_result = query_layer @ key_layer.transpose(-1, -2)
# change view to [batch_size, num_heads, q_length, kv_length]
attention_scores = matmul_result.view(batch_size, self.num_heads, q_length, kv_length)
# cast attention scores to fp32,
# compute scaled softmax and cast back to initial dtype
# - [batch_size, num_heads, q_length, kv_length]
input_dtype = attention_scores.dtype
# `float16` has a minimum value of -65504.0,
# whereas `bfloat16` and `float32` have a minimum value of `-3.4e+38`
if input_dtype == torch.float16 or input_dtype == torch.bfloat16:
attention_scores = attention_scores.to(torch.float32)
# attn_weights = torch. \
# masked_fill(attention_scores, attention_mask, torch.finfo(attention_scores.dtype).min)
attention_probs = F.softmax(
(attention_scores + alibi) * self.inv_norm_factor + attention_mask_float,
dim=-1,
dtype=hidden_states.dtype,
)
# [batch_size, num_heads, q_length, kv_length]
attention_probs = self.attention_dropout(attention_probs)
if head_mask is not None:
attention_probs = attention_probs * head_mask
# change view [batch_size x num_heads, q_length, kv_length]
attention_probs_reshaped = attention_probs.view(
batch_size * self.num_heads,
q_length,
kv_length
)
# matmul: [batch_size * num_heads, q_length, head_dim]
context_layer = attention_probs_reshaped @ value_layer
# change view [batch_size, num_heads, q_length, head_dim]
context_layer = self._merge_heads(context_layer)
output_tensor = self.dense(context_layer)
outputs = (output_tensor, present)
if output_attentions:
outputs += (attention_probs,)
return outputs
def rw_attention_forward_40b(
self,
hidden_states: torch.Tensor,
alibi: torch.Tensor,
attention_mask: torch.Tensor,
layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]]=None,
head_mask: Optional[torch.Tensor]=None,
use_cache: bool=False,
output_attentions: bool=False,
):
# [batch_size, seq_length, 3 x hidden_size]
fused_qkv = self.query_key_value(hidden_states)
# 3 x [batch_size, seq_length, num_heads, head_dim]
(query_layer, key_layer, value_layer) = self._split_heads(fused_qkv)
batch_size, q_length, _, _ = query_layer.shape
query_layer = query_layer.transpose(1, 2).reshape(
batch_size * self.num_heads,
q_length,
self.head_dim
)
key_layer = key_layer.transpose(1, 2).reshape(
batch_size * self.num_heads,
q_length,
self.head_dim,
)
value_layer = value_layer.transpose(1, 2).reshape(
batch_size * self.num_heads,
q_length,
self.head_dim
)
# query_layer, key_layer = self.maybe_rotary(query_layer, key_layer)
_, seq_len, _ = query_layer.shape
if layer_past is not None:
_, seq_len_past, _ = layer_past[0].shape
seq_len = seq_len + seq_len_past
query_layer, key_layer = self.maybe_rotary(query_layer, key_layer, seq_len)
_, kv_length, _ = key_layer.shape
if layer_past is not None:
kv_length += layer_past[0].shape[-2]
query_layer = query_layer.view(batch_size, self.num_heads, q_length, self.head_dim)
key_layer = key_layer.view(batch_size, self.num_heads, q_length, self.head_dim)
value_layer = value_layer.view(batch_size, self.num_heads, q_length, self.head_dim)
device = hidden_states.device
if layer_past is not None:
# reuse k, v, self_attention
cache_k = layer_past[0].view(batch_size, self.num_heads, -1, self.head_dim)
cache_v = layer_past[1].view(batch_size, self.num_heads, -1, self.head_dim)
if cache_k.stride()[1] < kv_length * cache_k.size(3):
# allocate new
new_cache_k, new_cache_v = extend_kv_cache(
batch_size,
self.num_heads,
self.head_dim,
cache_k.size(2),
kv_length + KV_CACHE_ALLOC_BLOCK_LENGTH,
dtype=cache_k.dtype,
device=device
)
new_cache_k[:] = cache_k
new_cache_v[:] = cache_v
cache_k = new_cache_k
cache_v = new_cache_v
key_layer, value_layer = append_kv_cache(cache_k, cache_v, key_layer, value_layer)
elif use_cache:
max_cache_length = kv_length + KV_CACHE_ALLOC_BLOCK_LENGTH
new_key_states, new_value_states = init_kv_cache(
batch_size,
self.num_heads,
self.head_dim,
kv_length,
max_cache_length,
dtype=key_layer.dtype,
device=device
)
new_key_states[:] = key_layer
new_value_states[:] = value_layer
key_layer = new_key_states
value_layer = new_value_states
query_layer = query_layer.view(batch_size*self.num_heads, -1, self.head_dim)
key_layer = key_layer.view(batch_size*self.num_heads, -1, self.head_dim)
value_layer = value_layer.view(batch_size*self.num_heads, -1, self.head_dim)
_, kv_length, _ = key_layer.shape
if use_cache is True:
present = (key_layer, value_layer)
else:
present = None
if alibi is None:
query_layer_ = query_layer.reshape(batch_size, self.num_heads, -1, self.head_dim)
key_layer_ = key_layer.reshape(batch_size, self.num_heads, -1, self.head_dim)
value_layer_ = value_layer.reshape(batch_size, self.num_heads, -1, self.head_dim)
# attn_output = F.scaled_dot_product_attention(
# query_layer_, key_layer_, value_layer_, None, 0.0, is_causal=True
# )
if present is not None:
L = query_layer_.shape[-2]
S = key_layer_.shape[-2]
attn_mask = torch.ones(L, S, dtype=torch.bool, device=query_layer_.device)
attn_output = F.scaled_dot_product_attention(
query_layer_, key_layer_, value_layer_, attn_mask, 0.0, is_causal=False
)
else:
attn_output = F.scaled_dot_product_attention(
query_layer_, key_layer_, value_layer_, None, 0.0, is_causal=True
)
x = attn_output.view(batch_size, self.num_heads, q_length, self.head_dim)
x = x.permute(0, 2, 1, 3)
attn_output = x.reshape(batch_size, q_length, self.num_heads * self.head_dim)
output_tensor = self.dense(attn_output)
outputs = (output_tensor, present)
if output_attentions:
invalidInputError(False,
f"'output_attentions' are not supported yet")
return outputs
else:
attention_mask_float = (attention_mask * 1.0) \
.masked_fill(attention_mask, -1e9).to(torch.bfloat16)
matmul_result = query_layer @ key_layer.transpose(-1, -2)
# change view to [batch_size, num_heads, q_length, kv_length]
attention_scores = matmul_result.view(batch_size, self.num_heads, q_length, kv_length)
# cast attention scores to fp32,
# compute scaled softmax and cast back to initial dtype
# - [batch_size, num_heads, q_length, kv_length]
input_dtype = attention_scores.dtype
# `float16` has a minimum value of -65504.0,
# whereas `bfloat16` and `float32` have a minimum value of `-3.4e+38`
if input_dtype == torch.float16 or input_dtype == torch.bfloat16:
attention_scores = attention_scores.to(torch.float32)
# attn_weights = torch \
# .masked_fill(
# attention_scores, attention_mask, torch.finfo(attention_scores.dtype).min)
attention_probs = F.softmax(
(attention_scores + alibi.view(batch_size, self.num_heads, 1, -1))
* self.inv_norm_factor + attention_mask_float,
dim=-1,
dtype=hidden_states.dtype,
)
# [batch_size, num_heads, q_length, kv_length]
attention_probs = self.attention_dropout(attention_probs)
if head_mask is not None:
attention_probs = attention_probs * head_mask
# change view [batch_size x num_heads, q_length, kv_length]
attention_probs_reshaped = attention_probs.view(
batch_size * self.num_heads,
q_length,
kv_length
)
# matmul: [batch_size * num_heads, q_length, head_dim]
context_layer = attention_probs_reshaped @ value_layer
# change view [batch_size, num_heads, q_length, head_dim]
context_layer = self._merge_heads(context_layer)
output_tensor = self.dense(context_layer)
outputs = (output_tensor, present)
if output_attentions:
outputs += (attention_probs,)
return outputs
def falcon_attention_forward(
self,
hidden_states: torch.Tensor,
alibi: Optional[torch.Tensor],
attention_mask: torch.Tensor,
layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]]=None,
head_mask: Optional[torch.Tensor]=None,
use_cache: bool=False,
output_attentions: bool=False,
):
# [batch_size, seq_length, 3 x hidden_size]
fused_qkv = self.query_key_value(hidden_states)
num_kv_heads = self.num_heads if self.new_decoder_architecture else self.num_kv_heads
# 3 x [batch_size, seq_length, num_heads, head_dim]
(query_layer, key_layer, value_layer) = self._split_heads(fused_qkv)
batch_size, query_length, _, _ = query_layer.shape
query_layer = query_layer.transpose(1, 2).reshape(
batch_size * self.num_heads,
query_length,
self.head_dim
)
key_layer = key_layer.transpose(1, 2).reshape(
batch_size * num_kv_heads,
query_length,
self.head_dim,
)
value_layer = value_layer.transpose(1, 2).reshape(
batch_size * num_kv_heads,
query_length,
self.head_dim
)
past_kv_length = 0 if layer_past is None else layer_past[0].shape[1]
query_layer, key_layer = self.maybe_rotary(query_layer, key_layer, past_kv_length)
_, kv_length, _ = key_layer.shape
if layer_past is not None:
kv_length += layer_past[0].shape[-2]
query_layer = query_layer.view(batch_size, self.num_heads, query_length, self.head_dim)
key_layer = key_layer.view(batch_size, num_kv_heads, query_length, self.head_dim)
value_layer = value_layer.view(batch_size, num_kv_heads, query_length, self.head_dim)
device = hidden_states.device
if layer_past is not None:
# reuse k, v, self_attention
cache_k = layer_past[0].view(batch_size, num_kv_heads, -1, self.head_dim)
cache_v = layer_past[1].view(batch_size, num_kv_heads, -1, self.head_dim)
if cache_k.stride()[1] < kv_length * cache_k.size(3):
# allocate new
new_cache_k, new_cache_v = extend_kv_cache(
batch_size,
num_kv_heads,
self.head_dim,
cache_k.size(2),
kv_length + KV_CACHE_ALLOC_BLOCK_LENGTH,
dtype=cache_k.dtype,
device=device
)
new_cache_k[:] = cache_k
new_cache_v[:] = cache_v
cache_k = new_cache_k
cache_v = new_cache_v
key_layer, value_layer = append_kv_cache(cache_k, cache_v, key_layer, value_layer)
elif use_cache:
max_cache_length = kv_length + KV_CACHE_ALLOC_BLOCK_LENGTH
new_key_states, new_value_states = init_kv_cache(
batch_size,
num_kv_heads,
self.head_dim,
kv_length,
max_cache_length,
dtype=key_layer.dtype,
device=device
)
new_key_states[:] = key_layer
new_value_states[:] = value_layer
key_layer = new_key_states
value_layer = new_value_states
query_layer = query_layer.view(batch_size * self.num_heads, -1, self.head_dim)
key_layer = key_layer.view(batch_size * num_kv_heads, -1, self.head_dim)
value_layer = value_layer.view(batch_size * num_kv_heads, -1, self.head_dim)
_, kv_length, _ = key_layer.shape
if use_cache:
present = (key_layer, value_layer)
else:
present = None
attention_mask_float = (attention_mask * 1.0) \
.masked_fill(attention_mask, float("-1e9")).to(query_layer.dtype)
query_layer_ = query_layer.reshape(batch_size, self.num_heads, -1, self.head_dim)
key_layer_ = key_layer.reshape(batch_size, num_kv_heads, -1, self.head_dim)
value_layer_ = value_layer.reshape(batch_size, num_kv_heads, -1, self.head_dim)
if alibi is None:
if output_attentions:
# F.scaled_dot_product_attention doesn't return the attention weights, so we have
# to do it by hand if we want them
attention_scores = query_layer_ @ key_layer_.transpose(-1, -2)
attention_scores /= math.sqrt(self.head_dim)
attention_scores = F.softmax(
attention_scores + attention_mask_float, dim=-1, dtype=hidden_states.dtype
)
attn_output = attention_scores @ value_layer_
else:
attn_output = F.scaled_dot_product_attention(
query_layer_,
key_layer_,
value_layer_,
attention_mask_float,
0.0,
is_causal=False
)
attention_scores = None
attn_output = attn_output.view(batch_size, self.num_heads, query_length, self.head_dim)
attn_output = attn_output.permute(0, 2, 1, 3)
attn_output = attn_output.reshape(
batch_size,
query_length,
self.num_heads * self.head_dim
)
output_tensor = self.dense(attn_output)
if output_attentions:
return output_tensor, present, attention_scores
else:
return output_tensor, present
else:
matmul_result = query_layer_ @ key_layer_.transpose(-1, -2)
# change view to [batch_size, num_heads, q_length, kv_length]
attention_scores = matmul_result.view(
batch_size,
self.num_heads,
query_length,
kv_length
)
# cast attention scores to fp32,
# compute scaled softmax and cast back to initial dtype
# - [batch_size, num_heads, q_length, kv_length]
input_dtype = attention_scores.dtype
# `float16` has a minimum value of -65504.0,
# whereas `bfloat16` and `float32` have a minimum value of `-3.4e+38`
if input_dtype == torch.float16 or input_dtype == torch.bfloat16:
attention_scores = attention_scores.to(torch.float32)
# Matt (HF) note: We could possibly use F.scaled_dot_product_attention here too, by
# adding (alibi * self.inv_norm_factor) to attention_mask_float.
# I think this would be mathematically
# equivalent and more performant, but there might be a numerical difference.
# If you're reading this
# and you'd like to experiment and maybe file a PR, feel free!
attention_logits = attention_scores + alibi.view(batch_size, self.num_heads, 1, -1)
attention_logits *= self.inv_norm_factor
attention_probs = F.softmax(attention_logits + attention_mask_float, dim=-1,
dtype=hidden_states.dtype)
# [batch_size, num_heads, q_length, kv_length]
attention_probs = self.attention_dropout(attention_probs)
if head_mask is not None:
attention_probs = attention_probs * head_mask
# change view [batch_size, num_heads, q_length, kv_length]
attention_probs_reshaped = attention_probs.view(
batch_size,
self.num_heads,
query_length,
kv_length
)
# matmul: [batch_size * num_heads, q_length, head_dim]
context_layer = (attention_probs_reshaped @ value_layer_).flatten(0, 1)
# change view [batch_size, q_length, num_heads * head_dim]
context_layer = self._merge_heads(context_layer)
output_tensor = self.dense(context_layer)
if output_attentions:
return output_tensor, present, attention_probs
else:
return output_tensor, present
def falcon_attention_forward_4_36(
self,
hidden_states: torch.Tensor,
alibi: Optional[torch.Tensor],
attention_mask: torch.Tensor,
position_ids: Optional[torch.LongTensor]=None,
layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]]=None,
head_mask: Optional[torch.Tensor]=None,
use_cache: bool=False,
output_attentions: bool=False,
**kwargs,
):
""" based on transformers==4.36.0
https://github.com/huggingface/transformers/blob/v4.36.0/src/transformers/models/falcon/modeling_falcon.py
"""
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.`"
)
fused_qkv = self.query_key_value(hidden_states) # [batch_size, seq_length, 3 x hidden_size]
num_kv_heads = self.num_heads if self.new_decoder_architecture else self.num_kv_heads
# 3 x [batch_size, seq_length, num_heads, head_dim]
(query_layer, key_layer, value_layer) = self._split_heads(fused_qkv)
batch_size, query_length, _, _ = query_layer.shape
query_layer = query_layer.transpose(1, 2).reshape(
batch_size, self.num_heads, query_length, self.head_dim)
key_layer = key_layer.transpose(1, 2).reshape(
batch_size, num_kv_heads, query_length, self.head_dim)
value_layer = value_layer.transpose(1, 2).reshape(
batch_size, num_kv_heads, query_length, self.head_dim)
kv_seq_len = key_layer.shape[-2]
device = hidden_states.device
if layer_past is not None:
kv_seq_len += layer_past[0].shape[-2]
if alibi is None:
cos, sin = self.rotary_emb(value_layer, seq_len=kv_seq_len)
query_layer, key_layer = apply_rotary_pos_emb(
query_layer, key_layer, cos, sin, position_ids)
if layer_past is not None:
# reuse k, v, self_attention
cache_k = layer_past[0].view(batch_size, self.num_heads, -1, self.head_dim)
cache_v = layer_past[1].view(batch_size, self.num_heads, -1, self.head_dim)
if cache_k.stride()[1] <= cache_k.size(2) * cache_k.size(3):
# allocate new
new_cache_k, new_cache_v = extend_kv_cache(
batch_size,
self.num_heads,
self.head_dim,
cache_k.size(2),
kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH,
dtype=cache_k.dtype,
device=device
)
new_cache_k[:] = cache_k
new_cache_v[:] = cache_v
cache_k = new_cache_k
cache_v = new_cache_v
key_layer, value_layer = append_kv_cache(cache_k, cache_v, key_layer, value_layer)
elif use_cache:
max_cache_length = kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH
new_key_states, new_value_states = init_kv_cache(
batch_size,
self.num_heads,
self.head_dim,
kv_seq_len,
max_cache_length,
dtype=key_layer.dtype,
device=device
)
new_key_states[:] = key_layer
new_value_states[:] = value_layer
key_layer = new_key_states
value_layer = new_value_states
query_layer = query_layer.view(batch_size, self.num_heads, -1, self.head_dim)
key_layer = key_layer.view(batch_size, self.num_heads, -1, self.head_dim)
value_layer = value_layer.view(batch_size, self.num_heads, -1, self.head_dim)
kv_length = key_layer.shape[-2]
if use_cache:
present = (key_layer, value_layer)
else:
present = None
# 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_layer.device.type == "cuda" and attention_mask is not None:
query_layer = query_layer.contiguous()
key_layer = key_layer.contiguous()
value_layer = value_layer.contiguous()
if alibi is None:
if self._use_sdpa and not output_attentions:
attn_output = F.scaled_dot_product_attention(
query_layer,
key_layer,
value_layer,
attention_mask,
0.0,
# The query_length > 1 is necessary to match with
# AttentionMaskConverter.to_causal_4d that does not create a causal mask in case
# query_length == 1.
is_causal=self.is_causal and attention_mask is None and query_length > 1,
)
attention_scores = None
else:
attention_scores = query_layer @ key_layer.transpose(-1, -2)
attention_scores /= math.sqrt(self.head_dim)
attention_scores = F.softmax(
attention_scores + attention_mask, dim=-1, dtype=hidden_states.dtype)
# It is unclear why neither dropout nor head_mask is applied here
# (while it is with alibi).
attn_output = attention_scores @ value_layer
attn_output = attn_output.view(batch_size, self.num_heads, query_length, self.head_dim)
attn_output = attn_output.permute(0, 2, 1, 3)
attn_output = attn_output.reshape(batch_size, query_length, self.num_heads * self.head_dim)
attn_output = self.dense(attn_output)
if output_attentions:
return attn_output, present, attention_scores
else:
return attn_output, present
else:
if self._use_sdpa and not output_attentions and head_mask is None:
attn_output = F.scaled_dot_product_attention(
query_layer,
key_layer,
value_layer,
attn_mask=attention_mask,
dropout_p=self.attention_dropout.p if self.training else 0.0,
is_causal=self.is_causal and attention_mask is None and query_length > 1,
)
attn_output = attn_output.transpose(1, 2)
attn_output = attn_output.reshape(
batch_size, query_length, self.num_heads * self.head_dim)
attn_output = self.dense(attn_output)
else:
matmul_result = query_layer @ key_layer.transpose(-1, -2)
# change view to [batch_size, num_heads, q_length, kv_length]
attention_scores = matmul_result.view(
batch_size, self.num_heads, query_length, kv_length)
# cast attention scores to fp32, compute scaled softmax and cast back to initial dtype -
# [batch_size, num_heads, q_length, kv_length]
input_dtype = attention_scores.dtype
# `float16` has a minimum value of -65504.0, whereas `bfloat16` and `float32` have a
# minimum value of `-3.4e+38`
if input_dtype == torch.float16 or input_dtype == torch.bfloat16:
attention_scores = attention_scores.to(torch.float32)
attention_logits = attention_scores + alibi.view(batch_size, self.num_heads, 1, -1)
attention_logits *= self.inv_norm_factor
attention_probs = F.softmax(
attention_logits + attention_mask, dim=-1, dtype=hidden_states.dtype)
# [batch_size, num_heads, q_length, kv_length]
attention_probs = self.attention_dropout(attention_probs)
if head_mask is not None:
attention_probs = attention_probs * head_mask
# change view [batch_size, num_heads, q_length, kv_length]
attention_probs_reshaped = attention_probs.view(
batch_size, self.num_heads, query_length, kv_length)
# matmul: [batch_size * num_heads, q_length, head_dim]
attn_output = (attention_probs_reshaped @ value_layer).flatten(0, 1)
# change view [batch_size, q_length, num_heads * head_dim]
attn_output = self._merge_heads(attn_output)
attn_output = self.dense(attn_output)
if output_attentions:
return attn_output, present, attention_probs
else:
return attn_output, present