github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-08 20:40:34 +00:00
Code Issues Packages Projects Releases Wiki Activity

[shardformer] support whisper (#4212)

Browse Source 
* support whisper

* fix bug in vocabembedding

* support downstream model of whisper

* update readme
This commit is contained in:
FoolPlayer
2023-07-17 14:25:32 +08:00
committed by Hongxin Liu
parent dd2bf02679
commit 9ee4ebea83
7 changed files with 443 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
colossalai/shardformer/README.md
View File

@@ -102,7 +102,7 @@ We will follow this roadmap to develop Shardformer:
- [ ] SwinTransformer
- [ ] SwinTransformer V2
- [ ] Audio
- [ ] Whisper
- [x] Whisper
- [ ] Multi-modal
- [ ] To be added

10
colossalai/shardformer/layer/embedding.py
View File

@@ -202,7 +202,6 @@ class VocabParallelEmbedding1D(ParallelModule):
super().__init__()
self.num_embeddings = num_embeddings
self.embedding_dim = embedding_dim
self.padding_idx = padding_idx
self.embed_args = args
self.embed_kwargs = kwargs
self.process_group = process_group
@@ -276,6 +275,15 @@ class VocabParallelEmbedding1D(ParallelModule):
with torch.no_grad():
self.weight[self.padding_idx - self.vocab_start_index].fill_(0)
def _select_padding_idx(self, padding_idx: int):
# select padding index according to the rank
if padding_idx is None:
return None
elif padding_idx < self.vocab_end_index and padding_idx >= self.vocab_start_index:
return padding_idx - self.vocab_start_index
else:
return None
def forward(self, input_: Tensor) -> Tensor:
# Build the mask.
input_mask = (input_ < self.vocab_start_index) | (input_ >= self.vocab_end_index)

8
colossalai/shardformer/policies/auto_policy.py
View File

@@ -105,6 +105,14 @@ _POLICY_LIST = {
"transformers.models.bloom.modeling_bloom.BloomForQuestionAnswering":
PolicyLocation(file_name="bloom", class_name="BloomForQuestionAnsweringPolicy"),
# Whisper
"transformers.models.whisper.modeling_whisper.WhisperModel":
PolicyLocation(file_name="whisper", class_name="WhisperModelPolicy"),
"transformers.models.whisper.modeling_whisper.WhisperForConditionalGeneration":
PolicyLocation(file_name="whisper", class_name="WhisperForConditionalGenerationPolicy"),
"transformers.models.whisper.modeling_whisper.WhisperForAudioClassification":
PolicyLocation(file_name="whisper", class_name="WhisperForAudioClassificationPolicy"),
# Sam
"transformers.models.sam.modeling_sam.SamModel":
PolicyLocation(file_name="sam", class_name="SamModelPolicy"),

232
colossalai/shardformer/policies/whisper.py Normal file
View File

@@ -0,0 +1,232 @@
import torch.nn as nn
import colossalai.shardformer.layer as col_nn
from .._utils import getattr_, setattr_
from .basepolicy import ModulePolicyDescription, Policy, SubModuleReplacementDescription
__all__ = [
'WhisperPolicy', 'WhisperModelPolicy', 'WhisperForConditionalGenerationPolicy', 'WhisperForAudioClassification'
]
class WhisperPolicy(Policy):
def config_sanity_check(self):
pass
def preprocess(self):
# reshape the embedding layer
r"""
Reshape the Embedding layer to make the embedding dimension divisible by world_size
"""
# TODO:
vocab_size = self.model.config.vocab_size
world_size = self.shard_config.tensor_parallel_size
if vocab_size % world_size != 0:
new_vocab_size = vocab_size + world_size - vocab_size % world_size
self.model.resize_token_embeddings(new_vocab_size)
return self.model
def module_policy(self):
from transformers.models.whisper.modeling_whisper import (
WhisperDecoder,
WhisperDecoderLayer,
WhisperEncoder,
WhisperEncoderLayer,
)
policy = {}
if self.shard_config.enable_tensor_parallelism:
policy[WhisperEncoderLayer] = ModulePolicyDescription(attribute_replacement={
"self_attn.embed_dim":
self.model.config.d_model // self.shard_config.tensor_parallel_size,
"self_attn.num_heads":
self.model.config.encoder_attention_heads // self.shard_config.tensor_parallel_size,
},
sub_module_replacement=[
SubModuleReplacementDescription(
suffix="self_attn.q_proj",
target_module=col_nn.Linear1D_Col,
),
SubModuleReplacementDescription(
suffix="self_attn.k_proj",
target_module=col_nn.Linear1D_Col,
),
SubModuleReplacementDescription(
suffix="self_attn.v_proj",
target_module=col_nn.Linear1D_Col,
),
SubModuleReplacementDescription(
suffix="self_attn.out_proj",
target_module=col_nn.Linear1D_Row,
),
SubModuleReplacementDescription(
suffix="fc1",
target_module=col_nn.Linear1D_Col,
),
SubModuleReplacementDescription(
suffix="fc2",
target_module=col_nn.Linear1D_Row,
),
])
policy[WhisperDecoderLayer] = ModulePolicyDescription(attribute_replacement={
"self_attn.embed_dim":
self.model.config.d_model // self.shard_config.tensor_parallel_size,
"self_attn.num_heads":
self.model.config.decoder_attention_heads // self.shard_config.tensor_parallel_size,
"encoder_attn.embed_dim":
self.model.config.d_model // self.shard_config.tensor_parallel_size,
"encoder_attn.num_heads":
self.model.config.encoder_attention_heads // self.shard_config.tensor_parallel_size,
},
sub_module_replacement=[
SubModuleReplacementDescription(
suffix="self_attn.q_proj",
target_module=col_nn.Linear1D_Col,
),
SubModuleReplacementDescription(
suffix="self_attn.k_proj",
target_module=col_nn.Linear1D_Col,
),
SubModuleReplacementDescription(
suffix="self_attn.v_proj",
target_module=col_nn.Linear1D_Col,
),
SubModuleReplacementDescription(
suffix="self_attn.out_proj",
target_module=col_nn.Linear1D_Row,
),
SubModuleReplacementDescription(
suffix="encoder_attn.q_proj",
target_module=col_nn.Linear1D_Col,
),
SubModuleReplacementDescription(
suffix="encoder_attn.k_proj",
target_module=col_nn.Linear1D_Col,
),
SubModuleReplacementDescription(
suffix="encoder_attn.v_proj",
target_module=col_nn.Linear1D_Col,
),
SubModuleReplacementDescription(
suffix="encoder_attn.out_proj",
target_module=col_nn.Linear1D_Row,
),
SubModuleReplacementDescription(
suffix="fc1",
target_module=col_nn.Linear1D_Col,
),
SubModuleReplacementDescription(
suffix="fc2",
target_module=col_nn.Linear1D_Row,
),
])
policy[WhisperDecoder] = ModulePolicyDescription(sub_module_replacement=[
SubModuleReplacementDescription(
suffix="embed_tokens",
target_module=col_nn.VocabParallelEmbedding1D,
),
])
# optimization configuration
if self.shard_config.enable_fused_normalization:
# Handle encoder layer
self.append_or_create_submodule_replacement(description=[
SubModuleReplacementDescription(
suffix="self_attn_layer_norm",
target_module=col_nn.FusedLayerNorm,
),
SubModuleReplacementDescription(
suffix="final_layer_norm",
target_module=col_nn.FusedLayerNorm,
)
],
policy=policy,
target_key=WhisperEncoderLayer)
# Handle decoder layer
self.append_or_create_submodule_replacement(description=[
SubModuleReplacementDescription(
suffix="self_attn_layer_norm",
target_module=col_nn.FusedLayerNorm,
),
SubModuleReplacementDescription(
suffix="final_layer_norm",
target_module=col_nn.FusedLayerNorm,
)
],
policy=policy,
target_key=WhisperDecoderLayer)
# handle encoder layer
self.append_or_create_submodule_replacement(description=[
SubModuleReplacementDescription(
suffix="layer_norm",
target_module=col_nn.FusedLayerNorm,
)
],
policy=policy,
target_key=WhisperEncoder)
# handle decoder layer
self.append_or_create_submodule_replacement(description=[
SubModuleReplacementDescription(
suffix="layer_norm",
target_module=col_nn.FusedLayerNorm,
)
],
policy=policy,
target_key=WhisperDecoder)
return policy
def add_lm_head_policy(self, base_policy):
from transformers.models.whisper.modeling_whisper import WhisperForConditionalGeneration
# optimize for tensor parallelism
if self.shard_config.enable_tensor_parallelism:
self.append_or_create_submodule_replacement(description=SubModuleReplacementDescription(
suffix="proj_out", target_module=col_nn.Linear1D_Col, kwargs={"gather_output": True}),
policy=base_policy,
target_key=WhisperForConditionalGeneration)
return base_policy
def postprocess(self):
return self.model
# WhisperModel
class WhisperModelPolicy(WhisperPolicy):
def __init__(self) -> None:
super().__init__()
# WhisperForConditionalGeneration
class WhisperForConditionalGenerationPolicy(WhisperPolicy):
def __init__(self) -> None:
super().__init__()
def module_policy(self):
module_policy = super().module_policy()
module_policy = self.add_lm_head_policy(module_policy)
return module_policy
def postprocess(self):
binding_map = {"model.decoder.embed_tokens.weight": "proj_out.weight"}
for k, v in binding_map.items():
param = getattr_(self.model, k)
setattr_(self.model, v, param)
return self.model
# WhisperForAudioClassification
class WhisperForAudioClassificationPolicy(WhisperPolicy):
def __init__(self) -> None:
super().__init__()