Skip to content

Dcsl #19

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
nnkhanhduy wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Dcsl #19

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
0dff9fd
Add support for multiple Layer 2 devices
nnkhanhduy Mar 13, 2026
7d98b9e
Update Bert and kwt
nnkhanhduy Mar 16, 2026
03b44f9
feat: Add Speech Commands and AG News datasets with their respective …
nnkhanhduy Mar 21, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
60 changes: 56 additions & 4 deletions other/DCSL/src/RpcClient.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,8 @@
import src.Log
from src.model import *

from peft import LoraConfig, get_peft_model


class RpcClient:
def __init__(self, client_id, layer_id, channel, train_func, device):
Expand All @@ -23,6 +25,7 @@ def __init__(self, client_id, layer_id, channel, train_func, device):
self.response = None
self.model = None
self.label_count = None
self.peft_config = None

self.train_set = None
self.label_to_indices = None
Expand Down Expand Up @@ -75,13 +78,41 @@ def response_message(self, body):
])
self.train_set = torchvision.datasets.CIFAR10(root='./data', train=True, download=True,
transform=transform_train)
elif data_name == "SPEECHCOMMANDS":
from src.dataset.SPEECHCOMMANDS import SpeechCommandsDataset
self.train_set = SpeechCommandsDataset(root='./data', subset='training')
elif data_name == "AGNEWS":
from datasets import load_dataset
from transformers import BertTokenizer
from src.dataset.AGNEWS import AGNEWS_DATASET

dataset = load_dataset('ag_news', download_mode='reuse_dataset_if_exists', cache_dir='./hf_cache')
tokenizer = BertTokenizer.from_pretrained('bert-base-cased')

train_data = dataset['train']
texts = train_data['text']
labels = train_data['label']

self.train_set = AGNEWS_DATASET(texts, labels, tokenizer, max_length=128)
else:
self.train_set = None
raise ValueError(f"Data name '{data_name}' is not valid.")

self.label_to_indices = defaultdict(list)
for idx, (_, label) in tqdm(enumerate(self.train_set)):
self.label_to_indices[int(label)].append(idx)
if data_name == "AGNEWS":
for idx, label in enumerate(self.train_set.labels):
self.label_to_indices[int(label)].append(idx)
elif data_name == "SPEECHCOMMANDS":
from src.dataset.SPEECHCOMMANDS import CLASSES
for idx, (audio_path, label_name) in enumerate(self.train_set.samples):
if label_name in CLASSES:
label_idx = CLASSES.index(label_name)
else:
label_idx = CLASSES.index('unknown')
self.label_to_indices[label_idx].append(idx)
else:
for idx, (_, label) in tqdm(enumerate(self.train_set)):
self.label_to_indices[int(label)].append(idx)

# Load model
if self.model is None:
Expand All @@ -99,11 +130,28 @@ def response_message(self, body):
lr = self.response["lr"]
momentum = self.response["momentum"]
sda_size = self.response.get("sda_size", 1)
layer2_devices = self.response.get("layer2_devices", [])

# Read parameters and load to model
if state_dict:
self.model.load_state_dict(state_dict)

# Apply LoRA for BERT model
if model_name == 'BERT':
if self.peft_config is None:
self.peft_config = LoraConfig(
task_type="SEQ_CLS",
r=8, lora_alpha=16, lora_dropout=0.1,
bias="none",
target_modules=["query", "key", "value", "dense"]
)
self.model = get_peft_model(self.model, self.peft_config)
if self.layer_id == 2:
for param in self.model.layer15.parameters():
param.requires_grad = True

self.model.to(self.device)

# Start training
if self.layer_id == 1:
selected_indices = []
Expand All @@ -113,10 +161,14 @@ def response_message(self, body):
subset = torch.utils.data.Subset(self.train_set, selected_indices)
train_loader = torch.utils.data.DataLoader(subset, batch_size=batch_size, shuffle=True)

result, size = self.train_func(self.model, lr, momentum, train_loader, local_round=local_round)
result, size = self.train_func(self.model, lr, momentum, train_loader, local_round=local_round, layer2_devices=layer2_devices, model_name=model_name)

else:
result, size = self.train_func(self.model, lr, momentum, None, local_round=local_round, sda_size=sda_size)
result, size = self.train_func(self.model, lr, momentum, None, local_round=local_round, sda_size=sda_size, model_name=model_name)

# Merge LoRA weights back for BERT
if model_name == 'BERT':
self.model = self.model.merge_and_unload()

# Stop training, then send parameters to server
model_state_dict = copy.deepcopy(self.model.state_dict())
Expand Down
95 changes: 55 additions & 40 deletions other/DCSL/src/Scheduler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,9 +18,12 @@ def __init__(self, client_id, layer_id, channel, device):
self.device = device
self.data_count = 0

def send_intermediate_output(self, output, labels, trace, data_id=None):
def send_intermediate_output(self, output, labels, trace, data_id=None, target_device_id=None):

forward_queue_name = f'intermediate_queue_{self.layer_id}'
if target_device_id is not None:
forward_queue_name = f'intermediate_queue_{target_device_id}'
else:
forward_queue_name = f'intermediate_queue_{self.layer_id}'

self.channel.queue_declare(forward_queue_name, durable=False)

Expand Down Expand Up @@ -63,49 +66,67 @@ def send_to_server(self, message):
routing_key='rpc_queue',
body=pickle.dumps(message))

def train_on_first_layer(self, model, lr, momentum, train_loader=None, local_round=3):
"""
Synchronous training: forward 1 batch → wait for gradient → backward → next batch.
Edge device does NOT send multiple batches before receiving gradient.
"""
optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum)
def train_on_first_layer(self, model, lr, momentum, train_loader=None, local_round=3, layer2_devices=None, model_name=None):
if model_name == 'BERT':
optimizer = optim.AdamW(model.parameters(), lr=lr, weight_decay=0.01)
else:
optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum)

backward_queue_name = f'gradient_queue_{self.layer_id}_{self.client_id}'
self.channel.queue_declare(queue=backward_queue_name, durable=False)
self.channel.basic_qos(prefetch_count=1)

model.to(self.device)

batch_counter = 0

for i in range(local_round):
src.Log.print_with_color(f'Epoch {i}', 'green')

with tqdm(total=len(train_loader), desc="Processing", unit="step") as pbar:
for training_data, labels in train_loader:
training_data = training_data.to(self.device)
for batch in train_loader:
if isinstance(batch, dict) and 'input_ids' in batch:
training_data = batch['input_ids'].to(self.device)
attention_mask = batch['attention_mask'].to(self.device)
labels = batch['labels'].to(self.device)
kwargs = {'input_ids': training_data, 'attention_mask': attention_mask}
else:
training_data, labels = batch
training_data = training_data.to(self.device)
labels = labels.to(self.device)
kwargs = {}

# Step 1: Forward
data_id = str(uuid.uuid4())
intermediate_output = model(training_data)
with torch.no_grad():
if 'input_ids' in kwargs:
intermediate_output = model(**kwargs)
else:
intermediate_output = model(training_data, **kwargs)
intermediate_output = intermediate_output.detach().requires_grad_(True)

self.data_count += 1
pbar.update(1)

# Step 2: Send smashed data to server
self.send_intermediate_output(intermediate_output, labels, trace=None, data_id=data_id)
target_device_id = None
if layer2_devices:
target_device_id = layer2_devices[batch_counter % len(layer2_devices)]
batch_counter += 1

self.send_intermediate_output(intermediate_output, labels, trace=None, data_id=data_id, target_device_id=target_device_id)

# Step 3: Wait for gradient (blocking)
while True:
method_frame, header_frame, body = self.channel.basic_get(
queue=backward_queue_name, auto_ack=True)
if method_frame and body:
received_data = pickle.loads(body)
gradient = torch.tensor(received_data["data"]).to(self.device)

# Step 4: Backward
model.train()
optimizer.zero_grad()
output = model(training_data)
if 'input_ids' in kwargs:
output = model(**kwargs)
else:
output = model(training_data, **kwargs)
output.backward(gradient=gradient)
optimizer.step()
break
Expand All @@ -128,17 +149,11 @@ def train_on_first_layer(self, model, lr, momentum, train_loader=None, local_rou
return True
time.sleep(0.5)

def _process_sda_batch(self, model, optimizer, criterion, collected):
"""
SDA (Smashed Data Aggregation) — Eq. 4-5 from paper.
Concatenate smashed data from all clients, forward once,
split gradient back to each client.
"""
def _process_sda_batch(self, model, optimizer, criterion, collected, model_name=None):
batch_sizes = [item["data"].shape[0] for item in collected]
traces = [item["trace"] for item in collected]
data_ids = [item["data_id"] for item in collected]

# Eq. 4: S_c = concat(σ_1, σ_2, ..., σ_|D_c|)
all_data = np.concatenate([item["data"] for item in collected], axis=0)
all_labels = np.concatenate([item["label"] for item in collected], axis=0)

Expand All @@ -149,8 +164,10 @@ def _process_sda_batch(self, model, optimizer, criterion, collected):
optimizer.zero_grad()
concat_intermediate.retain_grad()

# Eq. 5: ŷ = f(S_c | W)
output = model(concat_intermediate)
if model_name == 'BERT':
output = model(input_ids=concat_intermediate)
else:
output = model(concat_intermediate)
loss = criterion(output, concat_labels.long())
print(f"Loss (SDA, {len(collected)} clients, {sum(batch_sizes)} samples): {loss.item():.4f}")

Expand All @@ -173,20 +190,18 @@ def _process_sda_batch(self, model, optimizer, criterion, collected):

return result

def train_on_last_layer(self, model, lr, momentum, sda_size=1):
"""
SDA: collect exactly 1 batch from each client,
concat and forward once, split gradient back.
Since edge devices are synchronous, no overflow needed.
"""
optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum)
def train_on_last_layer(self, model, lr, momentum, sda_size=1, model_name=None):
if model_name == 'BERT':
optimizer = optim.AdamW(model.parameters(), lr=lr, weight_decay=0.01)
else:
optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum)
result = True
criterion = nn.CrossEntropyLoss()

forward_queue_name = f'intermediate_queue_{self.layer_id - 1}'
forward_queue_name = f'intermediate_queue_{self.client_id}'
self.channel.queue_declare(queue=forward_queue_name, durable=False)
self.channel.basic_qos(prefetch_count=1)
print(f'Waiting for intermediate output (SDA size={sda_size}). To exit press CTRL+C')
print(f'Waiting for intermediate output on queue {forward_queue_name} (SDA size={sda_size}). To exit press CTRL+C')
model.to(self.device)

sda_batch = {} # {client_id: data} — exactly 1 batch per client
Expand All @@ -200,7 +215,7 @@ def train_on_last_layer(self, model, lr, momentum, sda_size=1):

# When we have 1 batch from each client → SDA forward
if len(sda_batch) >= sda_size:
batch_result = self._process_sda_batch(model, optimizer, criterion, list(sda_batch.values()))
batch_result = self._process_sda_batch(model, optimizer, criterion, list(sda_batch.values()), model_name=model_name)
if not batch_result:
result = False
sda_batch = {}
Expand All @@ -214,16 +229,16 @@ def train_on_last_layer(self, model, lr, momentum, sda_size=1):
if received_data["action"] == "PAUSE":
# Process remaining
if sda_batch:
batch_result = self._process_sda_batch(model, optimizer, criterion, list(sda_batch.values()))
batch_result = self._process_sda_batch(model, optimizer, criterion, list(sda_batch.values()), model_name=model_name)
if not batch_result:
result = False
return result

def train_on_device(self, model, lr, momentum, train_loader=None, local_round=None, sda_size=1):
def train_on_device(self, model, lr, momentum, train_loader=None, local_round=None, sda_size=1, layer2_devices=None, model_name=None):
self.data_count = 0
if self.layer_id == 1:
result = self.train_on_first_layer(model, lr, momentum, train_loader, local_round)
result = self.train_on_first_layer(model, lr, momentum, train_loader, local_round, layer2_devices=layer2_devices, model_name=model_name)
else:
result = self.train_on_last_layer(model, lr, momentum, sda_size)
result = self.train_on_last_layer(model, lr, momentum, sda_size, model_name=model_name)

return result, self.data_count
Loading