chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Support DP MLA (#1970)

Browse Source
This commit is contained in:
Ke Bao
2024-11-16 09:01:43 +00:00
committed by GitHub
parent 2f2e07439c
commit 976bc302e5
12 changed files with 395 additions and 63 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/data_parallel_controller.py
+43 -8
View File
@@ -81,20 +81,34 @@ class DataParallelController:
# Start data parallel workers
base_gpu_id = 0
self.workers = []
scheduler_pipe_readers = []
for dp_rank in range(server_args.dp_size):
tmp_port_args = PortArgs.init_new(server_args)
tmp_port_args.tokenizer_ipc_name = port_args.tokenizer_ipc_name
tmp_port_args.detokenizer_ipc_name = port_args.detokenizer_ipc_name
send_to = self.launch_tensor_parallel_group(
server_args,
tmp_port_args,
base_gpu_id,
dp_rank,
)
if server_args.enable_dp_attention:
# Share workers for DP and TP
send_to, reader = self.launch_tensor_parallel_process(
server_args,
tmp_port_args,
base_gpu_id,
dp_rank,
)
base_gpu_id += 1
scheduler_pipe_readers.append(reader)
else:
send_to = self.launch_tensor_parallel_group(
server_args,
tmp_port_args,
base_gpu_id,
dp_rank,
)
base_gpu_id += server_args.tp_size
self.workers.append(send_to)
base_gpu_id += server_args.tp_size
for reader in scheduler_pipe_readers:
reader.recv()
def launch_tensor_parallel_group(
self,
@@ -132,6 +146,27 @@ class DataParallelController:
return send_to
def launch_tensor_parallel_process(
self,
server_args: ServerArgs,
port_args: PortArgs,
base_gpu_id: int,
dp_rank: int,
):
reader, writer = mp.Pipe(duplex=False)
gpu_id = base_gpu_id
tp_rank = dp_rank
proc = mp.Process(
target=run_scheduler_process,
args=(server_args, port_args, gpu_id, tp_rank, dp_rank, writer),
)
proc.start()
send_to = get_zmq_socket(
self.context, zmq.PUSH, port_args.scheduler_input_ipc_name
)
return send_to, reader
def round_robin_scheduler(self, req):
self.workers[self.round_robin_counter].send_pyobj(req)
self.round_robin_counter = (self.round_robin_counter + 1) % len(self.workers)
python/sglang/srt/managers/schedule_batch.py
+24 -5
View File
@@ -56,6 +56,7 @@ global_server_args_dict = {
"disable_mla": ServerArgs.disable_mla,
"torchao_config": ServerArgs.torchao_config,
"disable_nan_detection": ServerArgs.disable_nan_detection,
"enable_dp_attention": ServerArgs.enable_dp_attention,
}
@@ -450,6 +451,9 @@ class ScheduleBatch:
# The sum of all sequence lengths
seq_lens_sum: int = None
# For DP attention
global_num_tokens: Optional[List[int]] = None
# For processing logprobs
return_logprob: bool = False
top_logprobs_nums: Optional[List[int]] = None
@@ -858,6 +862,16 @@ class ScheduleBatch:
# Reset the encoder cached status
self.encoder_cached = [True] * len(self.reqs)
def prepare_for_idle(self):
self.forward_mode = ForwardMode.IDLE
self.input_ids = torch.empty(0, dtype=torch.int32).to(
self.device, non_blocking=True
)
self.seq_lens = torch.empty(0, dtype=torch.int32).to(
self.device, non_blocking=True
)
self.extend_num_tokens = 0
def prepare_for_decode(self, enable_overlap: bool = False):
self.forward_mode = ForwardMode.DECODE
@@ -969,17 +983,18 @@ class ScheduleBatch:
self.has_grammar = self.has_grammar or other.has_grammar
def get_model_worker_batch(self):
if self.forward_mode.is_decode():
if self.forward_mode.is_decode() or self.forward_mode.is_idle():
extend_seq_lens = extend_prefix_lens = extend_logprob_start_lens = None
else:
extend_seq_lens = self.extend_lens
extend_prefix_lens = self.prefix_lens
extend_logprob_start_lens = self.extend_logprob_start_lens
if self.has_grammar:
self.sampling_info.grammars = [req.grammar for req in self.reqs]
else:
self.sampling_info.grammars = None
if self.sampling_info is not None:
if self.has_grammar:
self.sampling_info.grammars = [req.grammar for req in self.reqs]
else:
self.sampling_info.grammars = None
global bid
bid += 1
@@ -995,6 +1010,7 @@ class ScheduleBatch:
req_to_token_pool_records=self.req_to_token_pool.get_write_records(),
return_logprob=self.return_logprob,
top_logprobs_nums=self.top_logprobs_nums,
global_num_tokens=self.global_num_tokens,
extend_num_tokens=self.extend_num_tokens,
extend_seq_lens=extend_seq_lens,
extend_prefix_lens=extend_prefix_lens,
@@ -1051,6 +1067,9 @@ class ModelWorkerBatch:
return_logprob: bool
top_logprobs_nums: Optional[List[int]]
# For DP attention
global_num_tokens: Optional[List[int]]
# For extend
extend_num_tokens: Optional[int]
extend_seq_lens: Optional[List[int]]
python/sglang/srt/managers/scheduler.py
+55 -3
View File
@@ -110,7 +110,7 @@ class Scheduler:
# Init inter-process communication
context = zmq.Context(2)
if self.tp_rank == 0:
if self.tp_rank == 0 or self.server_args.enable_dp_attention:
self.recv_from_tokenizer = get_zmq_socket(
context, zmq.PULL, port_args.scheduler_input_ipc_name
)
@@ -347,6 +347,10 @@ class Scheduler:
self.process_input_requests(recv_reqs)
batch = self.get_next_batch_to_run()
if self.server_args.enable_dp_attention:
batch = self.prepare_dp_attn_batch(batch)
self.cur_batch = batch
if batch:
@@ -361,6 +365,8 @@ class Scheduler:
self.update_running_batch()
if not self.running_batch:
break
if self.server_args.enable_dp_attention:
batch = self.prepare_dp_attn_batch(batch)
result = self.run_batch(batch)
self.process_batch_result(batch, result)
else:
@@ -396,8 +402,48 @@ class Scheduler:
self.last_batch = batch
def prepare_dp_attn_batch(self, local_batch: ScheduleBatch):
# Check if other DP workers have running batches
if local_batch is None:
num_tokens = 0
elif local_batch.forward_mode.is_decode():
num_tokens = local_batch.batch_size()
else:
num_tokens = local_batch.extend_num_tokens
local_num_tokens = torch.tensor(
num_tokens, dtype=torch.int64, device=self.device
)
global_num_tokens = torch.empty(
self.tp_size, dtype=torch.int64, device=self.device
)
torch.distributed.all_gather_into_tensor(
global_num_tokens,
local_num_tokens,
group=self.tp_worker.get_tp_device_group(),
)
if local_batch is None and global_num_tokens.max().item() > 0:
local_batch = self.get_idle_batch()
if local_batch is not None:
local_batch.global_num_tokens = global_num_tokens.tolist()
return local_batch
def get_idle_batch(self):
idle_batch = ScheduleBatch.init_new(
[],
self.req_to_token_pool,
self.token_to_kv_pool,
self.tree_cache,
self.model_config,
)
idle_batch.prepare_for_idle()
return idle_batch
def recv_requests(self):
if self.tp_rank == 0:
if self.tp_rank == 0 or self.server_args.enable_dp_attention:
recv_reqs = []
while True:
@@ -409,7 +455,7 @@ class Scheduler:
else:
recv_reqs = None
if self.tp_size != 1:
if self.tp_size != 1 and not self.server_args.enable_dp_attention:
recv_reqs = broadcast_pyobj(recv_reqs, self.tp_rank, self.tp_cpu_group)
return recv_reqs
@@ -812,6 +858,10 @@ class Scheduler:
logits_output, next_token_ids = self.tp_worker.forward_batch_generation(
model_worker_batch
)
elif batch.forward_mode.is_idle():
model_worker_batch = batch.get_model_worker_batch()
self.tp_worker.forward_batch_idle(model_worker_batch)
return
else:
logits_output = None
if self.skip_tokenizer_init:
@@ -830,6 +880,8 @@ class Scheduler:
return ret
def process_batch_result(self, batch: ScheduleBatch, result):
if batch.forward_mode.is_idle():
return
if batch.forward_mode.is_decode():
self.process_batch_result_decode(batch, result)
if batch.is_empty():
python/sglang/srt/managers/tp_worker.py
+7
View File
@@ -128,12 +128,19 @@ class TpModelWorker:
def get_tp_cpu_group(self):
return self.model_runner.tp_group.cpu_group
def get_tp_device_group(self):
return self.model_runner.tp_group.device_group
def get_memory_pool(self):
return (
self.model_runner.req_to_token_pool,
self.model_runner.token_to_kv_pool,
)
def forward_batch_idle(self, model_worker_batch: ModelWorkerBatch):
forward_batch = ForwardBatch.init_new(model_worker_batch, self.model_runner)
self.model_runner.forward(forward_batch)
def forward_batch_generation(self, model_worker_batch: ModelWorkerBatch):
forward_batch = ForwardBatch.init_new(model_worker_batch, self.model_runner)
logits_output = self.model_runner.forward(forward_batch)
python/sglang/srt/managers/tp_worker_overlap_thread.py
+3
View File
@@ -88,6 +88,9 @@ class TpModelWorkerClient:
def get_tp_cpu_group(self):
return self.worker.get_tp_cpu_group()
def get_tp_device_group(self):
return self.worker.get_tp_device_group()
def get_memory_pool(self):
return (
self.worker.model_runner.req_to_token_pool,