github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-26 20:23:26 +00:00
Code Issues Packages Projects Releases Wiki Activity

[Sharderformer] Support zbv in Sharderformer Policy (#6150)

Browse Source 
* [feat] Sharderformer support zbv

* [feat] support chatglm2, command, deepseek for zbv

* [feat] support zbv in shardformer policy:
falcon,gptj,mistral,opt,qwen2,t5, vit, whisper

* [feat] support GPT2FusedLinearConv1D

* [feat] support GPT2FusedLinear (without tp)

* [fix] debug FusedConvLinear

* [shardfromer] support gpt2 policy for zbv, support GPT2FusedLinearConv
Col and Row.

* [Shardformer] support FusedLinear1D base for zbv

* [shardformer] support zbv in FusedLinear1D base, Col, Row

* [shardformer] support zbv in blip2 and sam policy

* [shardformer] fix bug incorrect number of gradients; add fusedLinear
base testcase;

* [fix] fix incorrect number of gradients ;

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [Shardformer] add en doc for zbv;

* [fix] fix typo in Model compatibility table

* [fix] fix API Reference typo

* [Shardformer] add zh-Han doc for zbv

* [fix] fix Linear name; update en & zh doc

* [fix] fix shardformer doc import err

* [fix] fix shardconfig import in doc

* [fix] fix shardformer doc

* [fix] fix shardconfig doc

* [fix] fix config

* [fix] remove shardconfig

* [fix] fix doc

* [feat] add zbv doc string

* [fix] rm doc

* [fix] fix doc

* [fix] empty zbv doc

* [fix] ifx torch version

* [fix] fix torch version

* [fix] fix torch versions

* [fix] fix torch versions

* [fix] fix pyramid versions

* [fix] fix pyramid, zope version

* [fix] try fix workflow

* [fix] try import ShardConfig in yml

* [fix] fix workflow

* [fix] fix workflow

* [fix] fix workflow

* [fix] fix workflow

* [fix] fix ci

* [fix] fix zbv doc

* [fix] fix param for qkv linear, gpt2fused linear; fix requirments;

* [fix] fix policy use fused_linear

* [fix] fix weight grad none, err caused by  weight ptr change

* [fix] fix comm in WeightGradStore

* [fix] fix WeightGradStore pop param

* [fix] remove useless param in doc; fix gpt2 qkv test;

* [shardformer] simplify execute_w_pass_grad_accum;

* [fix] rm useless comments

* [shardformer] simplify execute_w_pass_grad_accum & execute_w_pass

* [shardformer] Run meaningful doc test

* [shadformer] fix doc test cmd;

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
This commit is contained in:
duanjunwen
2025-01-02 10:22:26 +08:00
committed by GitHub
parent af06d162cf
commit a9bedc7a43
27 changed files with 3511 additions and 316 deletions
Download Patch File Download Diff File Expand all files Collapse all files

178
colossalai/shardformer/policies/vit.py
View File

@@ -43,6 +43,8 @@ class ViTPolicy(Policy):
self.shard_config.enable_sequence_parallelism = False
warnings.warn("Vit doesn't support sequence parallelism now, will ignore the sequence parallelism flag.")
use_zbv = self.pipeline_stage_manager is not None and self.pipeline_stage_manager.use_zbv
if self.shard_config.enable_tensor_parallelism:
assert (
self.model.config.num_attention_heads % self.shard_config.tensor_parallel_size == 0
@@ -72,6 +74,7 @@ class ViTPolicy(Policy):
target_module=col_nn.Linear1D_Col,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@@ -79,6 +82,7 @@ class ViTPolicy(Policy):
target_module=col_nn.Linear1D_Col,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@@ -86,6 +90,7 @@ class ViTPolicy(Policy):
target_module=col_nn.Linear1D_Col,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@@ -97,6 +102,7 @@ class ViTPolicy(Policy):
target_module=col_nn.Linear1D_Row,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@@ -109,6 +115,7 @@ class ViTPolicy(Policy):
kwargs={
"skip_bias_add": self.enable_bias_gelu_fused,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@@ -116,6 +123,7 @@ class ViTPolicy(Policy):
target_module=col_nn.Linear1D_Row,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@@ -132,7 +140,92 @@ class ViTPolicy(Policy):
policy=policy,
target_key=ViTIntermediate,
)
elif use_zbv:
policy[ViTEmbeddings] = ModulePolicyDescription(
attribute_replacement={},
param_replacement=[],
sub_module_replacement=[
SubModuleReplacementDescription(
suffix="dropout",
target_module=DropoutForReplicatedInput,
)
],
)
policy[ViTLayer] = ModulePolicyDescription(
param_replacement=[],
sub_module_replacement=[
SubModuleReplacementDescription(
suffix="attention.attention.query",
target_module=col_nn.LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="attention.attention.key",
target_module=col_nn.LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="attention.attention.value",
target_module=col_nn.LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="attention.attention.dropout",
target_module=col_nn.DropoutForParallelInput,
),
SubModuleReplacementDescription(
suffix="attention.output.dense",
target_module=col_nn.LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="attention.output.dropout",
target_module=col_nn.DropoutForReplicatedInput,
),
SubModuleReplacementDescription(
suffix="intermediate.dense",
target_module=col_nn.LinearWithGradAccum,
kwargs={
"skip_bias_add": self.enable_bias_gelu_fused,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="output.dense",
target_module=col_nn.LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="output.dropout",
target_module=col_nn.DropoutForReplicatedInput,
),
],
)
if self.enable_bias_gelu_fused:
self.append_or_create_method_replacement(
description={
"forward": get_jit_fused_vit_intermediate_forward(),
},
policy=policy,
target_key=ViTIntermediate,
)
# use flash attention
if self.shard_config.enable_flash_attention:
self.append_or_create_method_replacement(
@@ -173,11 +266,20 @@ class ViTPolicy(Policy):
stage_manager = self.pipeline_stage_manager
held_layers = []
layers_per_stage = stage_manager.distribute_layers(len(module.encoder.layer))
if stage_manager.is_first_stage():
held_layers.append(module.embeddings)
start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage)
held_layers.extend(module.encoder.layer[start_idx:end_idx])
if stage_manager.is_interleave:
assert stage_manager.num_model_chunks is not None
layers_per_stage = stage_manager.distribute_layers(len(module.encoder.layer))
stage_indices = stage_manager.get_stage_index(layers_per_stage)
if stage_manager.is_first_stage(ignore_chunk=True):
held_layers.append(module.embeddings)
for start_idx, end_idx in stage_indices:
held_layers.extend(module.encoder.layer[start_idx:end_idx])
else:
layers_per_stage = stage_manager.distribute_layers(len(module.encoder.layer))
if stage_manager.is_first_stage():
held_layers.append(module.embeddings)
start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage)
held_layers.extend(module.encoder.layer[start_idx:end_idx])
return held_layers
def set_pipeline_forward(self, model_cls: nn.Module, pipeline_forward: Callable, policy: Dict):
@@ -213,9 +315,16 @@ class ViTModelPolicy(ViTPolicy):
module = self.model
stage_manager = self.pipeline_stage_manager
if stage_manager.is_last_stage():
held_layers.append(module.layernorm)
held_layers.append(module.pooler)
if stage_manager.is_interleave:
if (stage_manager.use_zbv and stage_manager.is_first_stage(ignore_chunk=True)) or (
not stage_manager.use_zbv and stage_manager.is_last_stage(ignore_chunk=True)
):
held_layers.append(module.layernorm)
held_layers.append(module.pooler)
else:
if stage_manager.is_last_stage():
held_layers.append(module.layernorm)
held_layers.append(module.pooler)
return held_layers
@@ -226,6 +335,9 @@ class ViTForImageClassificationPolicy(ViTPolicy):
from transformers.models.vit.modeling_vit import ViTForImageClassification, ViTModel
policy = super().module_policy()
use_zbv = self.pipeline_stage_manager is not None and self.pipeline_stage_manager.use_zbv
if self.shard_config.enable_tensor_parallelism:
new_item = {
ViTForImageClassification: ModulePolicyDescription(
@@ -233,13 +345,33 @@ class ViTForImageClassificationPolicy(ViTPolicy):
SubModuleReplacementDescription(
suffix="classifier",
target_module=Linear1D_Col,
kwargs=dict(gather_output=True, fp8_communication=self.shard_config.fp8_communication),
kwargs=dict(
gather_output=True,
fp8_communication=self.shard_config.fp8_communication,
use_zbv=use_zbv,
),
)
]
)
}
policy.update(new_item)
elif use_zbv:
new_item = {
ViTForImageClassification: ModulePolicyDescription(
sub_module_replacement=[
SubModuleReplacementDescription(
suffix="classifier",
target_module=col_nn.LinearWithGradAccum,
kwargs=dict(
gather_output=True,
fp8_communication=self.shard_config.fp8_communication,
use_zbv=use_zbv,
),
)
]
)
}
policy.update(new_item)
if self.shard_config.pipeline_stage_manager is not None:
self.set_pipeline_forward(model_cls=ViTModel, pipeline_forward=ViTModel_pipeline_forward, policy=policy)
self.set_pipeline_forward(
@@ -256,9 +388,16 @@ class ViTForImageClassificationPolicy(ViTPolicy):
module = self.model.vit
stage_manager = self.pipeline_stage_manager
if stage_manager.is_last_stage():
held_layers.append(module.layernorm)
held_layers.append(self.model.classifier)
if stage_manager.is_interleave:
if (stage_manager.use_zbv and stage_manager.is_first_stage(ignore_chunk=True)) or (
not stage_manager.use_zbv and stage_manager.is_last_stage(ignore_chunk=True)
):
held_layers.append(module.layernorm)
held_layers.append(self.model.classifier)
else:
if stage_manager.is_last_stage():
held_layers.append(module.layernorm)
held_layers.append(self.model.classifier)
return held_layers
@@ -285,8 +424,15 @@ class ViTForMaskedImageModelingPolicy(ViTPolicy):
module = self.model.vit
stage_manager = self.pipeline_stage_manager
if stage_manager.is_last_stage():
held_layers.append(module.layernorm)
held_layers.append(self.model.decoder)
if stage_manager.is_interleave:
if (stage_manager.use_zbv and stage_manager.is_first_stage(ignore_chunk=True)) or (
not stage_manager.use_zbv and stage_manager.is_last_stage(ignore_chunk=True)
):
held_layers.append(module.layernorm)
held_layers.append(self.model.decoder)
else:
if stage_manager.is_last_stage():
held_layers.append(module.layernorm)
held_layers.append(self.model.decoder)
return held_layers