114 lines
4.3 KiB
Python
114 lines
4.3 KiB
Python
import warnings
|
|
from typing import Dict, Union
|
|
|
|
import torch.nn as nn
|
|
from colossalai.shardformer.layer import Linear1D_Col, Linear1D_Row, VocabParallelEmbedding1D
|
|
from colossalai.shardformer.policies.base_policy import ModulePolicyDescription, Policy, SubModuleReplacementDescription
|
|
|
|
__all__ = ["LlavaMistralPolicy", "LlavaMistralForCausalLMPolicy"]
|
|
|
|
|
|
class LlavaMistralPolicy(Policy):
|
|
def config_sanity_check(self):
|
|
pass
|
|
|
|
def preprocess(self):
|
|
if self.shard_config.enable_tensor_parallelism:
|
|
# Resize embedding
|
|
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) -> Dict[Union[str, nn.Module], ModulePolicyDescription]:
|
|
from transformers.models.mistral.modeling_mistral import MistralDecoderLayer, MistralModel
|
|
|
|
policy = {}
|
|
|
|
if self.shard_config.enable_sequence_parallelism:
|
|
self.shard_config.enable_sequence_parallelism = False
|
|
warnings.warn(
|
|
"Mistral doesn't support sequence parallelism now, will ignore the sequence parallelism flag."
|
|
)
|
|
|
|
if self.shard_config.enable_tensor_parallelism:
|
|
decoder_attribute_replacement = {
|
|
"self_attn.hidden_size": self.model.config.hidden_size // self.shard_config.tensor_parallel_size,
|
|
"self_attn.num_heads": self.model.config.num_attention_heads // self.shard_config.tensor_parallel_size,
|
|
"self_attn.num_key_value_heads": self.model.config.num_key_value_heads
|
|
// self.shard_config.tensor_parallel_size,
|
|
}
|
|
|
|
policy[MistralDecoderLayer] = ModulePolicyDescription(
|
|
attribute_replacement=decoder_attribute_replacement,
|
|
sub_module_replacement=[
|
|
SubModuleReplacementDescription(
|
|
suffix="self_attn.q_proj",
|
|
target_module=Linear1D_Col,
|
|
),
|
|
SubModuleReplacementDescription(
|
|
suffix="self_attn.k_proj",
|
|
target_module=Linear1D_Col,
|
|
),
|
|
SubModuleReplacementDescription(
|
|
suffix="self_attn.v_proj",
|
|
target_module=Linear1D_Col,
|
|
),
|
|
SubModuleReplacementDescription(
|
|
suffix="self_attn.o_proj",
|
|
target_module=Linear1D_Row,
|
|
),
|
|
SubModuleReplacementDescription(
|
|
suffix="mlp.gate_proj",
|
|
target_module=Linear1D_Col,
|
|
),
|
|
SubModuleReplacementDescription(
|
|
suffix="mlp.up_proj",
|
|
target_module=Linear1D_Col,
|
|
),
|
|
SubModuleReplacementDescription(
|
|
suffix="mlp.down_proj",
|
|
target_module=Linear1D_Row,
|
|
),
|
|
],
|
|
)
|
|
|
|
self.append_or_create_submodule_replacement(
|
|
description=SubModuleReplacementDescription(
|
|
suffix="embed_tokens",
|
|
target_module=VocabParallelEmbedding1D,
|
|
),
|
|
policy=policy,
|
|
target_key=MistralModel,
|
|
)
|
|
|
|
return policy
|
|
|
|
def postprocess(self):
|
|
return self.model
|
|
|
|
|
|
class LlavaMistralForCausalLMPolicy(LlavaMistralPolicy):
|
|
def module_policy(self):
|
|
from transformers import MistralForCausalLM
|
|
|
|
policy = super().module_policy()
|
|
|
|
if self.shard_config.enable_tensor_parallelism:
|
|
# add a new item for casual lm
|
|
new_item = {
|
|
MistralForCausalLM: ModulePolicyDescription(
|
|
sub_module_replacement=[
|
|
SubModuleReplacementDescription(
|
|
suffix="lm_head", target_module=Linear1D_Col, kwargs=dict(gather_output=True)
|
|
)
|
|
]
|
|
)
|
|
}
|
|
policy.update(new_item)
|
|
return policy
|