FastAPI部署中国最强通用语言大模型Baichuan2
首先，需要安装 FastAPI 和 Uvicorn（用于运行 FastAPI 应用程序）：
pip install fastapi uvicorn
使用 FastAPI 框架：
from fastapi import FastAPI, HTTPException, Request
from pydantic import BaseModel
import json
import mindspore as ms
from mindformers import AutoConfig, AutoModel, AutoTokenizer
app = FastAPI()
ms.set_context(mode=ms.GRAPH_MODE, device_target="Ascend", device_id=0)
config = AutoConfig.from_pretrained("baichuan2_6b")
config.checkpoint_name_or_path = "../mindformers/checkpoint_download/baichuan2baichuan2_6b.ckpt"
model = AutoModel.from_config(config)
tokenizer = AutoTokenizer.from_pretrained("baichuan2_6b")
class RequestData(BaseModel):
text: str
@app.post('/baichuan2_bot')
async def say_hello_func(data: RequestData):
print("----------- in hello func ----------")
text = data.text
inputs = tokenizer(tokenizer.build_prompt(text))["input_ids"]
print(tokenizer.decode(inputs))
outputs = model.generate(inputs, max_length=14096)
outputs_text = tokenizer.decode(outputs)
return {"response": outputs_text}
@app.get('/goodbye')
async def say_goodbye_func():
print("----------- in goodbye func ----------")
return {'message': 'Goodbye!'}
@app.post('/')
async def default_func(request: Request):
print("----------- in default func ----------")
data = await request.json()
return {'message': 'Called default func!', 'data': data}
if __name__ == '__main__':
import uvicorn
uvicorn.run(app, host="0.0.0.0", port=8080)
百川大模型
"""Baichuan2_13b models' APIs."""
from typing import Optional
import math
import numpy as np
import mindspore.common.dtype as mstype
try:
from mindspore._checkparam import Validator
except ImportError:
import mindspore._checkparam as Validator
from mindspore import Tensor, nn, ops
from mindspore.common.parameter import Parameter
from mindspore.ops import operations as P
from mindspore.common.initializer import initializer, HeUniform
try:
from mindspore.nn.layer.flash_attention import FlashAttention
FLASHATTENTION_VALID = True
except ImportError:
FLASHATTENTION_VALID = False
from mindformers.core.loss.loss import CrossEntropyLoss
from mindformers.mindformer_book import MindFormerBook
from mindformers.models.base_model import BaseModel
from mindformers.models.utils import cell_reuse
from mindformers.modules.transformer.op_parallel_config import _check_config
from mindformers.modules.transformer import AttentionMask, TransformerOpParallelConfig
from mindformers.modules.layers import Linear, _check_input_dtype, _check_past_none_input_none, AlibiTensorV2
from mindformers.tools.register.register import MindFormerModuleType, MindFormerRegister
from mindformers.models.llama.llama import layer_compute_dtype
from mindformers.models.llama.llama_config import LlamaConfig
from mindformers.models.llama.llama_layer import LlamaEmbedding, LlamaFeedForward, LlamaRMSNorm
from mindformers.tools.logger import logger
__all__ = ['Baichuan13BV2ForCausalLM', 'Baichuan13BV2Model']
@MindFormerRegister.register(MindFormerModuleType.MODELS)
class Baichuan13BV2ForCausalLM(BaseModel):
r"""
Provide baichuan2_13B training loss or logits through network.
Args:
config (LlamaConfig): The config of baichuan2_13B model.
Inputs:
input_ids(Tensor): the tokenized inputs with datatype int32, Tensor of shape :math:`(batch, seq\_length)`.
labels(Tensor): the tokenized labels with datatype int32, Tensor of shape :math:`(batch, seq\_length)`.
input_position(Tensor): current position, used by model.predict.
position_ids(Tensor): Reserved param, not used.
attention_mask(Tensor): Reserved param, not used.
input_embeds(Tensor): Reserved param, not used.
init_reset(bool, optional): A bool tensor with shape [1], used to clear the past key parameter and
past value parameter used in the incremental prediction. Default True.
batch_valid_length(Tensor): the past calculated the index with datatype int32, used for incremental
prediction. Tensor of shape :math:`(batch_size,)`. Default None.
Returns:
Tensor, the loss or logits of the network.
Examples:
>>> from mindformers.models.llama import LlamaConfig
>>> from research.baichuan2.baichuan2_13b import Baichuan13BV2ForCausalLM
>>> config = LlamaConfig(batch_size=2)
>>> network = Baichuan13BV2ForCausalLM(config=config)
"""
_support_list = MindFormerBook.get_model_support_list()['llama']
@cell_reuse()
def __init__(self, config: LlamaConfig = None):
super(Baichuan13BV2ForCausalLM, self).__init__(config, auto_prefix=True)
_check_config(config.parallel_config)
self.ignore_token_id = config.ignore_token_id
self.pad_token_id = config.pad_token_id
self.dtype = config.compute_dtype
self.reshape = P.Reshape()
self.cast = P.Cast()
self.slice = P.StridedSlice()
self.not_equal = P.NotEqual()
self.mul = P.Mul()
self.add = P.Add()
self.model = Baichuan13BV2Model(config=config)
self.lm_head = NormHead(hidden_size=config.hidden_size,
vocab_size=config.vocab_size,
compute_dtype=config.compute_dtype)
self.loss = CrossEntropyLoss(parallel_config=config.parallel_config)
dp = config.parallel_config.data_parallel
self.slice.shard(((dp, 1),))
self.not_equal.shard(((dp, 1), ()))
self.mul.shard(((dp, 1), (dp, 1)))
self.add.shard(((dp, 1), ()))
self.lm_head.shard(config.parallel_config)
self.load_checkpoint(config)
# pylint: disable=W0613
def prepare_inputs_for_generation(self, input_ids, **kwargs):
return {
"input_ids": Tensor(input_ids, mstype.int32)
}
# pylint: disable=W0613
def construct(self, input_ids, labels=None, input_position=None, position_ids=None, attention_mask=None,
input_embeds=None, init_reset=True, batch_valid_length=None):
"""Baichuan13BV2ForCausalLM forward."""
bsz, seqlen = input_ids.shape
if self.training:
tokens = self.slice(input_ids, (0, 0), (bsz, seqlen - 1), (1, 1))
else:
tokens = input_ids
output = self.model(tokens, input_position, init_reset, batch_valid_length)
logits = self.lm_head(output)
input_mask = self.cast(self.not_equal(tokens, self.pad_token_id), self.dtype)
if labels is None:
labels = self.slice(input_ids, (0, 1), (bsz, seqlen), (1, 1))
else:
if labels.ndim > 1:
if self.training:
labels = self.slice(labels, (0, 1), (bsz, seqlen), (1, 1))
label_mask = self.cast(self.not_equal(labels, self.ignore_token_id), self.dtype)
input_mask = self.mul(input_mask, label_mask)
logits = self.cast(logits, mstype.float32)
if not self.training:
logits = self.reshape(logits, (bsz, seqlen, -1))
# makes cast effective to avoid allgather issue in Mindspore1.10
input_mask = self.add(input_mask, 1)
return logits, tokens, input_mask
if logits.ndim > 2:
logits = self.reshape(logits, (-1, logits.shape[-1]))
labels = self.reshape(labels, (-1,))
input_mask = self.reshape(input_mask, (-1,))
loss = self.loss(logits, labels, input_mask)
return loss
class Baichuan13BV2Model(BaseModel):
r"""
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Baichuan13BV2DecoderLayer`]
Args:
config(LlamaConfig): the config of network
Inputs:
input_ids: the tokenized inputs with datatype int32
Returns:
output: Tensor, the output of baichuan2_13b decoderlayer
"""
_support_list = MindFormerBook.get_model_support_list()['llama']
def __init__(self,
config: LlamaConfig = None):
super().__init__(config, auto_prefix=True)
_check_config(config.parallel_config)
if config.batch_size or config.use_past:
Validator.check_positive_int(config.batch_size)
self.dtype = config.compute_dtype
self.num_layers = config.num_layers
self.pad_token_id = config.pad_token_id
self.is_first_iteration = True
self.use_past = config.use_past
self.use_flash_attention = config.use_flash_attention and FLASHATTENTION_VALID
if self.use_flash_attention:
logger.info("Enable flash attention.")
elif config.use_flash_attention:
logger.info("Current MindSpore do not support flash attention.")
self.get_attention_mask = AttentionMask(
config.seq_length, parallel_config=config.parallel_config.dp_mp_config).to_float(config.compute_dtype)
self.multiply_data = Tensor([-10000.0], dtype=config.compute_dtype)
self.one = Tensor([1.0], dtype=config.compute_dtype)
self.all_ones_attention_mask_alibi = P.Ones()((1, config.seq_length), self.dtype)
self.reshape = P.Reshape()
self.cast = P.Cast()
self.tile = P.Tile()
self.mul_mask = P.Mul()
self.mul_alibi = P.Mul()
self.sub = P.Sub()
self.expand_dims = P.ExpandDims()
self.not_equal = P.NotEqual()
self.gather = P.Gather()
self.transpose = P.Transpose()
self.tok_embeddings = LlamaEmbedding(
config.vocab_size, config.hidden_size, param_init_type=config.param_init_type)
self.layers = nn.CellList()
for layer_id in range(config.num_layers):
layer = Baichuan13BDecodeLayer(config.batch_size,
config.seq_length,
layer_id,
dim=config.hidden_size,
n_heads=config.num_heads,
multiple_of=config.multiple_of,
n_kv_heads=config.n_kv_heads,
ffn_dim_multiplier=config.ffn_dim_multiplier,
norm_eps=config.rms_norm_eps,
compute_dtype=config.compute_dtype,
layernorm_compute_dtype=config.layernorm_compute_type,
softmax_compute_dtype=config.softmax_compute_type,
param_init_type=config.param_init_type,
use_past=config.use_past,
use_flash_attention=config.use_flash_attention,
compute_in_2d=config.compute_in_2d,
use_past_shard=config.use_past_shard,
parallel_config=config.parallel_config)
layer_compute_dtype(layer, layer_id, config.offset, config.parallel_config,
config.num_layers, select_recompute=config.parallel_config.recompute.select_recompute)
self.layers.append(layer)
self.norm_out = LlamaRMSNorm(config.hidden_size, config.rms_norm_eps,
compute_type=config.layernorm_compute_type)
self.build_alibi_tensor = AlibiTensorV2(seq_length=config.seq_length,
num_heads=config.num_heads)
dp = config.parallel_config.data_parallel
mp = config.parallel_config.model_parallel
self.tok_embeddings.pipeline_stage = 0
if config.parallel_config.pipeline_stage > 1:
self.norm_out.pipeline_stage = config.parallel_config.pipeline_stage - 1
self.tok_embeddings.set_comm_fusion(2)
self.norm_out.set_comm_fusion(2)
else:
self.tok_embeddings.set_comm_fusion(config.parallel_config.gradient_aggregation_group)
self.norm_out.set_comm_fusion(config.parallel_config.gradient_aggregation_group)
self.tok_embeddings.shard(config.parallel_config)
self.build_alibi_tensor.shard(config.parallel_config)
self.tile.shard(((1, 1, 1, 1), ()))
self.sub.shard(((1,), (dp, 1, 1)))
self.mul_mask.shard(((dp, 1, 1, 1), (1,)))
self.mul_alibi.shard(((dp, mp, 1, 1), (dp, 1, 1, 1))) # (dp, mp, 1, 1)
self.expand_dims.shard(((dp, 1, 1),))
self.not_equal.shard(((dp, 1), ()))
self.gather.shard(((dp, mp, 1, 1), (1,)))
self.transpose.shard(((1, mp, dp, 1),))
if config.compute_in_2d:
self.norm_out.shard((dp, 1))
else:
self.norm_out.shard((dp, 1, 1))
if self.use_past:
seq_range = np.arange(config.seq_length).reshape(1, 1, -1)
self.ones = P.Ones()
self.range = Tensor(np.tile(seq_range, (config.batch_size, 1, 1)), mstype.int32)
self.gather_past = P.Gather()
self.expand_dims = P.ExpandDims()
self.le_past = P.LessEqual()
# pylint: disable=W0613
def construct(self, tokens: Tensor, input_position=None, init_reset=True, batch_valid_length=None):
"""Forward of baichuan2_13b model."""
# preprocess
bs, _ = tokens.shape
if self.use_past:
if not isinstance(init_reset, Tensor):
init_reset = Tensor([init_reset], mstype.bool_)
if not isinstance(batch_valid_length, Tensor):
batch_valid_length = self.ones((bs, 1), mstype.int32)
if self.is_first_iteration:
input_mask = self.cast(self.not_equal(tokens, self.pad_token_id), self.dtype)
mask = self.get_attention_mask(input_mask)
alibi_tensor = self.build_alibi_tensor(input_mask, self.dtype)
# mask: [bs, seq, seq]
else:
cur_pos = batch_valid_length - 1
valid_length = self.reshape(cur_pos, (-1, 1, 1))
mask = self.cast(self.le_past(self.range, valid_length), self.dtype)
alibi_tensor = self.build_alibi_tensor(self.all_ones_attention_mask_alibi, self.dtype)
alibi_tensor = self.tile(alibi_tensor, (bs, 1, 1, 1))
alibi_tensor = self.gather(alibi_tensor, cur_pos[0], 2)
alibi_tensor = self.transpose(alibi_tensor, (2, 1, 0, 3))
# mask: [bs, 1, 1]
mask = self.sub(self.one, self.cast(mask, self.dtype))
if not self.use_flash_attention:
mask = self.expand_dims(mask, 1)
mask = self.mul_mask(mask, self.multiply_data)
# tokens: [bs, seq/1]
h = self.tok_embeddings(tokens)
# h: [bs, seq/1, hidden_dim]
for i in range(self.num_layers):
h, _ = self.layers[i](h, alibi_tensor, mask,
init_reset=init_reset, batch_valid_length=batch_valid_length)
output = self.norm_out(h)
return output
class Baichuan13BDecodeLayer(nn.Cell):
r"""
Transformer Layer. This is an implementation of the single layer of the transformer
encoder layer, including multihead attention and feedward layer.
Args:
batch_size(int): The batch size of the input tensor when do increnmental prediction. Should be a positive
value. When do training or prediction, the argument will not work and the user can just pass None to
the argument.
seq_length(int): The input sequence length.
layer_id(int): The layer id of current transformer block layer.
dim(int): The hidden size of the input.
num_heads(int): The number of the heads.
multiple_of(int): The SwiGLU hidden layer size multiple of large power of 2.
norm_eps (float): The epsilon value of the denominator. Default 1e-5.
compute_dtype(dtype.Number): The computation type of the layer.
Should be mstype.float32 or mstype.float16. Default mstype.float32.
layernorm_compute_type(dtype.Number): The computation type of the norm.
Should be mstype.float32 or mstype.float16. Default mstype.float32.
softmax_compute_type(dtype.Number): The computation type of the softmax in the attention.
Should be mstype.float32 or mstype.float16. Default mstype.float32.
param_init_type(dtype.Number): The parameter initialization type of the module.
Should be mstype.float32 or mstype.float16. Default mstype.float32.
use_past(bool): Use the past state to compute, used for incremental prediction. For example, if we have two
words and want to generate the ten more words. We just need to compute the two words' state only once,
and generate the next word one by one. When use_past is True, there are two steps to run the prediction.
In the first step, set the is_first_iteration to be True by
`model.add_flags_recursive(is_first_iteration=True)`, and pass the full inputs. Then, set the
is_first_iteration to be False by `model.add_flags_recursive(is_first_iteration=False)`.
At this moment, pass the single step's input tensor, and loop it. Default False.
parallel_config(OpParallelConfig, MoEParallelConfig): The parallel configure. When MoE is applied,
MoEParallelConfig is effective, otherwise OpParallelConfig is effective. Default `default_dpmp_config`,
an instance of `OpParallelConfig` with default args.
Inputs:
- **x** (Tensor) - Float Tensor, shape should be [batch_size, seq_length, hidden_size] or
[batch_size * seq_length, hidden_size], if the use_past is False or is_first_iteration=True. Otherwise,
should be [batch_size, 1, hidden_size]
- **alibi_tensor** (Tensor) - Alibi Tensor for position embedding used in attention.
- **mask** (Tensor) - Float Tensor, If the use_past is False or is_first_iteration=True,
the attention mask matrix should ba [batch_size, seq_length, seq_length], or None. None means there will
be no mask in softmax computation. Otherwise, should be [batch_size, 1, hidden_size]
- **init_reset** (Tensor) - A bool tensor with shape [1], used to clear the past key parameter and
past value parameter used in the incremental prediction. Only valid when use_past is True. Default True.
- **batch_valid_length** (Tensor) - Int32 tensor with shape [batch_size] the past calculated the index.
Used for incremental prediction when the use_past is True. Default None.
Outputs:
Tuple, a tuple contains(`output`, `layer_present`).
- **output** (Tensor) - The float tensor of the output of the layer with
shape (batch_size, seq_length, hidden_size) or (batch_size * seq_length, hidden_size), if the use_past is
False or is_first_iteration=True. Otherwise, it will be (batch_size, 1, hidden_size)
- **layer_present** (Tuple) - A tuple of the Tensor of the projected key and value vector with
((batch_size, num_heads, head_dim, seq_length),
(batch_size, num_heads, seq_length, head_dim)).
"""
def __init__(self,
batch_size,
seq_length,
layer_id,
dim: int = 512,
n_heads: int = 8,
multiple_of: int = 256,
n_kv_heads: Optional[int] = None,
ffn_dim_multiplier: Optional[int] = None,
norm_eps: float = 1e-5,
compute_dtype=mstype.float16,
layernorm_compute_dtype=mstype.float32,
softmax_compute_dtype=mstype.float32,
param_init_type=mstype.float32,
use_past=False,
use_flash_attention=False,
compute_in_2d=False,
use_past_shard=False,
parallel_config=TransformerOpParallelConfig()):
super().__init__()
if batch_size or use_past:
Validator.check_positive_int(batch_size)
self.batch_size = batch_size
self.compute_in_2d = compute_in_2d
self.seq_length = seq_length
self.layer_id = layer_id
self.hidden_size = dim
self.n_head = n_heads
self.head_dim = self.hidden_size // self.n_head
self.n_kv_head = n_heads if n_kv_heads is None else n_kv_heads
self.dtype = compute_dtype
self.is_first_iteration = True
self.use_past = use_past
self.compute_in_2d = compute_in_2d
self.key_past = None
self.value_past = None
self.reshape = P.Reshape()
self.add = P.Add()
self.attention_norm = LlamaRMSNorm(self.hidden_size, norm_eps, compute_type=layernorm_compute_dtype)
self.ffn_norm = LlamaRMSNorm(self.hidden_size, norm_eps, compute_type=layernorm_compute_dtype)
self.attention = Baichuan13BAttention(batch_size=batch_size,
seq_length=seq_length,
dim=dim,
n_heads=n_heads,
n_kv_heads=n_kv_heads,
compute_dtype=compute_dtype,
softmax_compute_dtype=softmax_compute_dtype,
param_init_type=param_init_type,
use_past=use_past,
use_flash_attention=use_flash_attention,
compute_in_2d=compute_in_2d,
use_past_shard=use_past_shard,
parallel_config=parallel_config)
self.feed_forward = LlamaFeedForward(dim=self.hidden_size,
hidden_dim=4 * self.hidden_size,
multiple_of=multiple_of,
ffn_dim_multiplier=ffn_dim_multiplier,
compute_dtype=compute_dtype,
param_init_type=param_init_type)
dp = parallel_config.data_parallel
mp = parallel_config.model_parallel
self.feed_forward.shard(parallel_config)
if self.compute_in_2d:
self.add.shard(((dp, 1), (dp, 1)))
self.attention_norm.shard((dp, 1))
self.ffn_norm.shard((dp, 1))
else:
self.add.shard(((dp, 1, 1), (dp, 1, 1)))
self.attention_norm.shard((dp, 1, 1))
self.ffn_norm.shard((dp, 1, 1))
self.feed_forward.mul.shard(((dp, 1, mp), (dp, 1, mp)))
if parallel_config.use_seq_parallel and self.is_first_iteration:
if self.compute_in_2d:
self.add.shard(((dp * mp, 1), (dp * mp, 1)))
self.attention_norm.shard((dp * mp, 1))
self.ffn_norm.shard((dp * mp, 1))
else:
self.add.shard(((dp, mp, 1), (dp, mp, 1)))
self.attention_norm.shard((dp, mp, 1))
self.ffn_norm.shard((dp, mp, 1))
self.feed_forward.w2.shard(((dp, mp), (1, mp)), out_strategy_matmul=((dp * mp, 1),))
if self.use_past:
kv_shape = (batch_size, self.n_kv_head, seq_length, self.head_dim)
self.key_past = Parameter(Tensor(np.zeros(kv_shape), self.dtype), name="key_past")
self.value_past = Parameter(Tensor(np.zeros(kv_shape), self.dtype), name="value_past")
self.mul_past = P.Mul().shard(((dp, 1, 1, 1), (1,)))
self.assign_past = P.Assign().shard(((dp, 1, 1, 1), (dp, 1, 1, 1)))
if use_past_shard:
self.mul_past.shard(((dp, mp, 1, 1), (1,)))
self.assign_past.shard(((dp, mp, 1, 1), (dp, mp, 1, 1)))
def construct(self, x, alibi_tensor, mask=None, init_reset=True, batch_valid_length=None):
""" Forward of transformer block. """
bs = x.shape[0]
if self.use_past:
if not isinstance(init_reset, Tensor):
init_reset = Tensor([init_reset], mstype.bool_)
if not isinstance(batch_valid_length, Tensor):
batch_valid_length = self.ones((bs, 1), mstype.int32)
self._check_input(x, alibi_tensor, mask, init_reset, batch_valid_length)
# [bs, seq/1, hidden_dim] (first) [bs * seq/1, hidden_dim] (others)
if self.compute_in_2d and x.ndim != 2:
x = self.reshape(x, (-1, x.shape[-1]))
# [bs, seq/1, hidden_dim] or [bs * seq/1, hidden_dim]
input_x = self.attention_norm(x)
key_reset = None
value_reset = None
if self.use_past and self.is_first_iteration:
# reset states, init_reset True for reuse and False for reset
self.assign_past(self.key_past, self.mul_past(self.key_past, self.cast(init_reset, self.dtype)))
self.assign_past(self.value_past, self.mul_past(self.value_past, self.cast(init_reset, self.dtype)))
key_reset = self.key_past
value_reset = self.value_past
# add dependency for desired execution order
input_x = ops.depend(input_x, key_reset)
input_x = ops.depend(input_x, value_reset)
# [bs, seq/1, hidden_dim] or [bs * seq/1, hidden_dim]
h, layer_present = self.attention(input_x, alibi_tensor, mask,
self.key_past, self.value_past, batch_valid_length)
h = self.add(x, h)
ffn_norm = self.ffn_norm(h)
# [bs, seq/1, hidden_dim] or [bs * seq/1, hidden_dim]
ffn_out = self.feed_forward(ffn_norm)
value_update = None
key_update = None
if self.use_past:
# current key and value
key_present, value_present = layer_present
# update key and value calculated this step
self.assign_past(self.key_past, key_present)
self.assign_past(self.value_past, value_present)
key_update = self.key_past
value_update = self.value_past
# add dependency for desired execution order
key_update = ops.depend(key_update, key_reset)
value_update = ops.depend(value_update, value_reset)
# add dependency for desired execution order
ffn_out = ops.depend(ffn_out, value_update)
ffn_out = ops.depend(ffn_out, key_update)
# [bs, seq/1, hidden_dim] or [bs * seq/1, hidden_dim]
out = self.add(h, ffn_out)
return out, layer_present
def _check_input(self, x, alibi_tensor, mask, init_reset, batch_valid_length):
r"""Check inputs"""
_check_input_dtype(
x.dtype, "x", [mstype.float32, mstype.float16], self.cls_name)
_check_input_dtype(alibi_tensor.dtype, "alibi_tensor",
[mstype.float32, mstype.float16], self.cls_name)
if mask is not None:
_check_input_dtype(mask.dtype, "input_mask", [mstype.float32, mstype.float16], self.cls_name)
init_reset_is_tensor = isinstance(init_reset, Tensor)
init_reset_is_default = init_reset is True
batch_valid_length_is_tensor = isinstance(batch_valid_length, Tensor)
batch_is_default = batch_valid_length is None
_check_past_none_input_none(self.use_past, "init_reset", self.cls_name, True, init_reset_is_tensor,
init_reset_is_default)
_check_past_none_input_none(self.use_past, "batch_valid_length", self.cls_name, None,
batch_valid_length_is_tensor, batch_is_default)
if self.use_past:
_check_input_dtype(init_reset.dtype, "init_reset", [mstype.bool_], self.cls_name)
_check_input_dtype(batch_valid_length.dtype, "batch_valid_length", [mstype.int32], self.cls_name)
return True
class Baichuan13BAttention(nn.Cell):
r"""
This is an implementation of multihead attention in Baichuan.
Args:
- **batch_size** (int): The batch size of the input tensor when do increnmental prediction. Should be a
positive value.
When do training or prediction, the argument will not work and the user can just pass None to the
argument.
- **src_seq_length** (int): The sequence length of the query vector.
- **tgt_seq_length** (int): The sequence length of the key and value vector.
- **dim** (int): The hidden size of the input.
- **head_dim** (int): The dim of head.
- **n_heads** (int): The number of the heads.
- **compute_dtype** (dtype.Number): The computation type of dense. Default mstype.float16.
Should be mstype.float32 or mstype.float16.
- **softmax_compute_type** (dtype.Number): The type of softmax computation module. Default mstype.float32.
Should be mstype.float32 or mstype.float16.
- **param_init_type** (dtype.Number): The parameter initialization type of the module. Default mstype.
float32. Should be mstype.float32 or mstype.float16.
- **use_past** (bool): Use the past state to compute, used for incremental prediction.
For example, if we have two words and want to generate the ten more words.
We just need to compute the two words' state only once, and generate the next word one by one.
When use_past is True, there are two steps to run the prediction.
In the first step, set the is_first_iteration to be True by
`model.add_flags_recursive(is_first_iteration=True)`, and pass the full inputs. Then, set the
is_first_iteration to be False by `model.add_flags_recursive(is_first_iteration=False)`. At this moment,
pass the single step's input tensor, and loop it. Default False.
- **parallel_config** (OpParallelConfig): The parallel configure. Default `default_dpmp_config`,
an instance of `OpParallelConfig` with default args.
Inputs:
- **x** (Tensor) - The input tokens with shape (batch_size, src_seq_length, hidden_size) or
(batch_size * src_seq_length, hidden_size), if the use_past is False or is_first_iteration=True.
Otherwise, must be (batch_size, 1, hidden_size)
- **alibi_tensor** (Tensor) - Alibi Tensor for position embedding used in attention.
- **mask** (Tensor) - If the use_past is False or is_first_iteration=True, the attention mask
matrix should ba (batch_size, src_seq_length, tgt_seq_length), or None. None means there will be no mask
in softmax computation. Otherwise, the mask must be (batch_size, 1, tgt_seq_length)
- **key_past** (Tensor) - Float16 tensor with shape (batch_size, num_heads, head_dim, tgt_seq_length).
The past calculated key vector. Used for incremental prediction when the use_past is True.
Default None.
- **value_past** (Tensor) - Float16 tensor with shape (batch_size, num_heads, tgt_seq_length,
head_dim).
The past calculated value vector. Used for incremental prediction when the use_past is True.
Default None.
- **batch_valid_length** (Tensor) - Int32 tensor with shape (batch_size,) the past calculated the index.
Used for incremental prediction when the use_past is True. Default None.
Outputs:
Tuple, a tuple contains(`output`, `layer_present`)
- **output** (Tensor) - Tensor, the float tensor of the output of the layer with
shape (batch_size, src_seq_length, hidden_size) or (batch_size * src_seq_length, hidden_size),
if the use_past is False or is_first_iteration=True. Otherwise, it will be (batch_size, 1, hidden_size).
- **layer_present** (Tuple) - A tuple of the Tensor of the projected key and value vector with
((batch_size, num_heads, head_dim, tgt_seq_length),
(batch_size, num_heads, tgt_seq_length, head_dim)).
"""
def __init__(self,
batch_size,
seq_length,
dim: int = 512,
n_heads: int = 8,
n_kv_heads: Optional[int] = None,
compute_dtype=mstype.float16,
softmax_compute_dtype=mstype.float32,
param_init_type=mstype.float32,
use_past=False,
use_flash_attention=False,
compute_in_2d=False,
use_past_shard=False,
parallel_config=TransformerOpParallelConfig()):
super().__init__()
self.seq_length = seq_length
self.hidden_size = dim
self.n_head = n_heads
self.head_dim = dim // n_heads
self.n_kv_head = n_heads if n_kv_heads is None else n_kv_heads
self.n_rep = self.n_head // self.n_kv_head
self.dtype = compute_dtype
self.softmax_dtype = softmax_compute_dtype
self.is_first_iteration = True
self.use_past = use_past
self.compute_in_2d = compute_in_2d
self.use_flash_attention = use_flash_attention and FLASHATTENTION_VALID
if self.hidden_size % self.n_head != 0:
raise ValueError("For 'MultiHeadAttention', the class variable 'hidden_size' must be a multiple "
"of 'n_head', but got the hidden_size is {} and the n_head is {}."
.format(self.hidden_size, self.n_head))
if self.n_kv_head % parallel_config.model_parallel != 0:
raise ValueError("For 'MultiHeadAttention', the class variable 'n_kv_head' must be a multiple of "
"'parallel_config.model_parallel', but got the n_kv_head is {} "
"and the parallel_config.model_parallel is {}."
.format(self.n_kv_head, parallel_config.model_parallel))
self.inv_norm_factor = Tensor(1.0 / math.sqrt(self.head_dim), dtype=compute_dtype)
self.reshape = P.Reshape()
self.transpose = P.Transpose()
self.merger_head_transpose = P.Transpose()
self.batch_matmul = P.BatchMatMul()
self.batch_matmul_q_k = P.BatchMatMul(transpose_b=True)
self.mul = P.Mul()
self.add = P.Add()
self.add_alibi = P.Add()
self.softmax = nn.Softmax().to_float(softmax_compute_dtype)
self.cast = P.Cast()
self.cast_attn = P.Cast()
self.tile_kv = P.Tile()
self.wo = Linear(in_channels=self.hidden_size,
out_channels=self.hidden_size,
has_bias=False,
compute_dtype=compute_dtype,
param_init_type=param_init_type)
self.wq = Linear(self.hidden_size,
self.hidden_size,
has_bias=False,
compute_dtype=compute_dtype,
param_init_type=param_init_type)
self.wk = Linear(self.hidden_size,
self.n_kv_head * self.head_dim,
has_bias=False,
compute_dtype=compute_dtype,
param_init_type=param_init_type)
self.wv = Linear(self.hidden_size,
self.n_kv_head * self.head_dim,
has_bias=False,
compute_dtype=compute_dtype,
param_init_type=param_init_type)
dp = parallel_config.data_parallel
mp = parallel_config.model_parallel
self.transpose.shard(((dp, 1, mp, 1),))
self.merger_head_transpose.shard(((dp, mp, 1, 1),))
self.batch_matmul_q_k.shard(((dp, mp, 1, 1), (dp, mp, 1, 1)))
self.batch_matmul.shard(((dp, mp, 1, 1), (dp, mp, 1, 1)))
self.mul.shard(((dp, mp, 1, 1), ()))
self.add.shard(((dp, 1, 1, 1), (dp, mp, 1, 1)))
self.add_alibi.shard(((dp, mp, 1, 1), (dp, mp, 1, 1)))
self.softmax.softmax.shard(((dp, mp, 1, 1),))
self.tile_kv.shard(((dp * mp, 1, 1, 1),))
self.wq.shard(((dp, 1), (mp, 1)))
self.wk.shard(((dp, 1), (mp, 1)))
self.wv.shard(((dp, 1), (mp, 1)))
self.wo.shard(((dp, mp), (1, mp)))
if parallel_config.use_seq_parallel and self.is_first_iteration:
self.wo.shard(((dp, mp), (1, mp)), out_strategy_matmul=((dp * mp, 1),))
if parallel_config.recompute.select_recompute:
self.tile_kv.recompute()
self.batch_matmul_q_k.recompute()
self.mul.recompute()
self.add.recompute()
self.cast_attn.recompute()
self.softmax.softmax.recompute()
self.batch_matmul.recompute()
if self.use_flash_attention:
self.flash_attention = FlashAttention(self.head_dim, dp=dp, mp=mp, next_block_num=0)
self.flash_attention.shard(((dp, mp, 1, 1), (dp, mp, 1, 1), (dp, mp, 1, 1), (dp, 1, 1), ()))
if parallel_config.recompute.select_recompute:
self.flash_attention.recompute()
if self.use_past:
# operators used for state reuse
seq_range = np.arange(seq_length).reshape(1, 1, -1)
self.range = Tensor(np.tile(seq_range, (batch_size, 1, 1)), mstype.int32)
self.expand_dims = P.ExpandDims().shard(((dp, 1, 1),))
self.add_past = P.Add().shard(((dp, 1, 1, 1), (dp, 1, 1, 1)))
self.equal = P.Equal().shard(((dp, 1, 1), (dp, 1, 1)))
self.less = P.Less().shard(((dp, 1, 1), (dp, 1, 1)))
self.mul_past = P.Mul().shard(((dp, 1, 1, 1), (dp, 1, 1, 1)))
if use_past_shard:
self.add_past.shard(((dp, mp, 1, 1), (dp, mp, 1, 1)))
self.mul_past.shard(((dp, mp, 1, 1), (dp, 1, 1, 1)))
def construct(self, x: Tensor, alibi_tensor: Tensor, mask=None,
key_past=None, value_past=None, batch_valid_length=None):
"""Forward process of the MultiHeadAttention"""
ori_dtype = x.dtype
# [bs, seq/1, hidden_dim] or [bs * seq/1, hidden_dim]
x = self.reshape(x, (-1, x.shape[-1]))
# [bs * seq/1, hidden_dim]
query = self.cast(self.wq(x), self.dtype) # dp, 1 -> dp, mp
key = self.cast(self.wk(x), self.dtype) # dp, 1 -> dp, mp
value = self.cast(self.wv(x), self.dtype) # dp, 1 -> dp, mp
query = self.reshape(query, (-1, self._get_seq_length_under_incremental(self.seq_length),
self.n_head, self.head_dim))
key = self.reshape(key, (-1, self._get_seq_length_under_incremental(self.seq_length),
self.n_kv_head, self.head_dim))
value = self.reshape(value, (-1, self._get_seq_length_under_incremental(self.seq_length),
self.n_kv_head, self.head_dim))
# [bs, seq/1, n_head/n_kv_head, head_dim]
query = self.transpose(query, (0, 2, 1, 3))
key = self.transpose(key, (0, 2, 1, 3))
value = self.transpose(value, (0, 2, 1, 3))
# kv cache: [bs, n_kv_head, 1, head_dim] -> [bs, n_kv_head, seq, head_dim]
key_present = key
value_present = value
if self.use_past:
# The first graph with the input size of (bs, seq_length)
if self.is_first_iteration:
# Get the valid input length without padding
valid_length_vector = (
self.less(self.range, batch_valid_length.view(-1, 1, 1))).astype(self.dtype)
# Cover the key and value numbers corresponding to the padding position
key_present = self.mul_past(key, self.expand_dims(valid_length_vector, 3))
value_present = self.mul_past(value, self.expand_dims(valid_length_vector, 3))
# The second graph with the inpus size of (bs, 1)
else:
# Get the current token position index
valid_length = batch_valid_length - 1
valid_length = self.reshape(valid_length, (-1, 1, 1))
valid_length_vector = (self.equal(self.range, valid_length)).astype(self.dtype)
# Pad the key and value to seq_length with only the position index not zero
current_key = self.mul_past(key, self.expand_dims(valid_length_vector, 3))
current_value = self.mul_past(value, self.expand_dims(valid_length_vector, 3))
# Concat the previous saved state and current state
key = self.add_past(key_past, current_key)
value = self.add_past(value_past, current_value)
# Update key_present and value_present for state update
key_present = key
value_present = value
layer_present = (key_present, value_present)
# kv share: [bs, n_kv_head, seq, head_dim] -> [bs, n_head, seq, head_dim]
key = self._repeat_kv(key, self.n_rep)
value = self._repeat_kv(value, self.n_rep)
# q, k, v: [bs, n_head, seq/1, head_dim], [bs, n_head, seq, head_dim], [bs, n_head, seq, head_dim]
if self.use_flash_attention:
attention = self.flash_attention(query, key, value, mask)
attention = self._merge_heads(attention)
else:
attention = self._attn(query, key, value, alibi_tensor, mask)
# [bs, seq/1, hidden_dim] or [bs * seq/1, hidden_dim]
output = self.wo(attention) # dp, mp -> dp, 1 / dp * mp, 1
output = self.cast(output, ori_dtype)
return output, layer_present
def _repeat_kv(self, x, rep):
if rep == 1:
return x
bs, n_kv_head, seqlen, head_dim = x.shape
x = self.reshape(x, (bs * n_kv_head, 1, seqlen, head_dim))
x = self.tile_kv(x, (1, rep, 1, 1))
x = self.reshape(x, (bs, n_kv_head * rep, seqlen, head_dim))
return x
def _get_seq_length_under_incremental(self, length):
r"""Return the length of the tensor.
For the incremental prediction, the seq length for the input is 1.
"""
if self.use_past and not self.is_first_iteration:
return 1
return length
def _merge_heads(self, x):
"""
convert a 4d input to a 2d or 3d output
Inputs:
x: input tensor
Output:
x_merge: the 2d output
"""
# [bs, n_head, seq/1, head_dim]
x = self.merger_head_transpose(x, (0, 2, 1, 3)) # dp,mp,1,1 -> dp,1,mp,1
# [bs, seq/1, n_head, head_dim]
x_shape = x.shape
if self.compute_in_2d:
# [bs * seq/1, hidden_dim]
new_shape = (-1, x_shape[-2] * x_shape[-1])
else:
# [bs, seq/1, hidden_dim]
new_shape = (x_shape[0], x_shape[1], -1)
x_merge = self.reshape(x, new_shape)
return x_merge
def _attn(self, query, key, value, alibi_tensor, mask):
"""
Get the weighted score along the seq_length
Inputs:
query: the query matrix
key: the key matrix
value: the value matrix
mask: the attention mask adder matrix with shape (batch_size,
1, seq_length, seq_length)
Outputs:
weighted_values: Tensor, the weighted sum scores
"""
# q, k: [bs, n_head, seq/1, head_dim], [bs, n_head, seq, head_dim]
score = self.batch_matmul_q_k(query, key)
# score: [bs, n_head, seq/1, seq]
score = self.mul(score, self.inv_norm_factor)
score = self.add_alibi(score, alibi_tensor)
score = self.add(mask, score)
attention_probs = self.softmax(self.cast_attn(score, self.softmax_dtype))
# score, v: [bs, n_head, seq/1, seq], [bs, n_head, seq, head_dim]
weighted_values = self.batch_matmul(self.cast(attention_probs, self.dtype), value)
# [bs, n_head, seq/1, head_dim]
attention_merge = self._merge_heads(weighted_values)
# [bs, seq/1, hidden_dim] or [bs * seq/1, hidden_dim]
return attention_merge
class NormHead(nn.Cell):
"""
NormHead Layer.
Args:
hidden_size (int): The hidden size of the input.
vocab_size (int): Size of the dictionary of embeddings.
compute_type (dtype.Number): The compute type.
eps (number): A small positive value prevents division by zero.
Inputs:
- hidden_states (Tensor) - Tensor of shape :math:`(batch, seq_length, hidden_size)`.
Outputs:
Tensor of shape :math:`(batch, seq_length, vocab_size)`.
"""
def __init__(self,
hidden_size,
vocab_size,
compute_dtype=mstype.float32,
eps=1e-5):
super().__init__()
self.weight = Parameter(
initializer(HeUniform(negative_slope=math.sqrt(5)),
[vocab_size, hidden_size],
mstype.float32),
name='weight',
parallel_optimizer=False)
self.square = P.Square()
self.sqrt = P.Sqrt()
self.add = P.Add()
self.real_div = P.RealDiv()
self.eps = Tensor([eps], mstype.float32)
self.matmul = P.MatMul(transpose_b=True)
self.cast = P.Cast()
self.compute_dtype = compute_dtype
self.hidden_size = hidden_size
self.vocab_size = vocab_size
def construct(self, hidden_states):
"""Forward process of the NormHead"""
out_shape = P.Shape()(hidden_states)[:-1] + (self.vocab_size,)
hidden_states = P.Reshape()(hidden_states, (-1, self.hidden_size))
variance = self.square(self.weight).sum(axis=1).reshape(-1, 1)
variance_eps = self.sqrt(self.add(variance, self.eps))
norm_weight = self.real_div(self.weight, variance_eps)
ori_type = hidden_states.dtype
out = self.matmul(hidden_states.astype(self.compute_dtype),
norm_weight.astype(self.compute_dtype))
out = P.Reshape()(out, out_shape)
return self.cast(out, ori_type)
def shard(self, parallel_config):
"""sharding for norm head"""
if parallel_config.vocab_emb_dp:
self.square.shard(((1, 1),))
self.sqrt.shard(((1, 1),))
self.add.shard(((1, 1), (1,)))
self.real_div.shard(((1, 1), (1, 1)))
self.matmul.shard(((parallel_config.data_parallel, 1), (1, 1)))
else:
self.square.shard(((parallel_config.model_parallel, 1),))
self.sqrt.shard(((parallel_config.model_parallel, 1),))
self.add.shard(((parallel_config.model_parallel, 1), (1,)))
self.real_div.shard(((parallel_config.model_parallel, 1),
(parallel_config.model_parallel, 1)))
self.matmul.shard(((parallel_config.data_parallel, 1),
(parallel_config.model_parallel, 1)))
if parallel_config.pipeline_stage > 1:
self.matmul.pipeline_stage = parallel_config.pipeline_stage - 1