Support DP MLA (#1970)

.github/workflows/pr-test.yml

@@ -244,6 +244,7 @@ jobs:
           cd test/srt
           python3 test_mla.py
           python3 test_mla_fp8.py
+          python3 test_dp_attention.py
 
       - name: Evaluate data parallelism accuracy (DP=2)
         timeout-minutes: 10

python/sglang/srt/layers/attention/triton_backend.py

@@ -28,9 +28,13 @@ def __init__(self, model_runner: ModelRunner):
 
         self.decode_attention_fwd = decode_attention_fwd
         self.extend_attention_fwd = extend_attention_fwd
-        self.num_head = (
-            model_runner.model_config.num_attention_heads // model_runner.tp_size
-        )
+
+        if model_runner.server_args.enable_dp_attention:
+            self.num_head = model_runner.model_config.num_attention_heads
+        else:
+            self.num_head = (
+                model_runner.model_config.num_attention_heads // model_runner.tp_size
+            )
 
         if global_server_args_dict.get("triton_attention_reduce_in_fp32", False):
             self.reduce_dtype = torch.float32

python/sglang/srt/managers/data_parallel_controller.py

@@ -81,20 +81,34 @@ def __init__(self, server_args, port_args) -> None:
         # 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 @@ def launch_tensor_parallel_group(
 
         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

@@ -56,6 +56,7 @@
     "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 @@ def prepare_encoder_info_decode(self):
         # 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 @@ def merge_batch(self, other: "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 @@ def get_model_worker_batch(self):
             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

@@ -110,7 +110,7 @@ def __init__(
         # 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 @@ def event_loop_normal(self):
             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 @@ def event_loop_normal(self):
                         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 @@ def event_loop_overlap(self):
 
             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 @@ def recv_requests(self):
         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 @@ def run_batch(self, batch: ScheduleBatch):
                 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 @@ def run_batch(self, batch: ScheduleBatch):
         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

@@ -128,12 +128,19 @@ def get_pad_input_ids_func(self):
     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

@@ -88,6 +88,9 @@ def get_pad_input_ids_func(self):
     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,

python/sglang/srt/model_executor/forward_batch_info.py

@@ -56,6 +56,8 @@ class ForwardMode(IntEnum):
     DECODE = auto()
     # Contains both EXTEND and DECODE.
     MIXED = auto()
+    # No sequence to forward. For data parallel attention, some workers wil be IDLE if no sequence allocated.
+    IDLE = auto()
 
     def is_prefill(self):
         return self == ForwardMode.PREFILL
@@ -69,6 +71,9 @@ def is_decode(self):
     def is_mixed(self):
         return self == ForwardMode.MIXED
 
+    def is_idle(self):
+        return self == ForwardMode.IDLE
+
 
 @dataclass
 class ForwardBatch:
@@ -128,6 +133,10 @@ class ForwardBatch:
     # For Qwen2-VL
     mrope_positions: torch.Tensor = None
 
+    # For DP attention
+    global_num_tokens: Optional[List[int]] = None
+    gathered_buffer: Optional[torch.Tensor] = None
+
     def compute_mrope_positions(
         self, model_runner: ModelRunner, batch: ModelWorkerBatch
     ):
@@ -209,10 +218,22 @@ def init_new(
             seq_lens_sum=batch.seq_lens_sum,
             return_logprob=batch.return_logprob,
             top_logprobs_nums=batch.top_logprobs_nums,
+            global_num_tokens=batch.global_num_tokens,
             lora_paths=batch.lora_paths,
             sampling_info=batch.sampling_info,
         )
 
+        if ret.global_num_tokens is not None:
+            max_len = max(ret.global_num_tokens)
+            ret.gathered_buffer = torch.zeros(
+                (max_len * model_runner.tp_size, model_runner.model_config.hidden_size),
+                dtype=model_runner.dtype,
+                device=device,
+            )
+
+        if ret.forward_mode.is_idle():
+            return ret
+
         # Init position information
         if not ret.forward_mode.is_decode():
             ret.positions = torch.concat(

python/sglang/srt/model_executor/model_runner.py

@@ -141,6 +141,7 @@ def __init__(
                 "torchao_config": server_args.torchao_config,
                 "disable_penalizer": server_args.disable_penalizer,
                 "disable_nan_detection": server_args.disable_nan_detection,
+                "enable_dp_attention": server_args.enable_dp_attention,
             }
         )
 
@@ -592,11 +593,18 @@ def forward_extend(self, forward_batch: ForwardBatch):
                 get_embedding=True,
             )
 
+    def forward_idle(self, forward_batch: ForwardBatch):
+        return self.model.forward(
+            forward_batch.input_ids, forward_batch.positions, forward_batch
+        )
+
     def forward(self, forward_batch: ForwardBatch) -> LogitsProcessorOutput:
         if forward_batch.forward_mode.is_decode():
             return self.forward_decode(forward_batch)
         elif forward_batch.forward_mode.is_extend():
             return self.forward_extend(forward_batch)
+        elif forward_batch.forward_mode.is_idle():
+            return self.forward_idle(forward_batch)
         else:
             raise ValueError(f"Invaid forward mode: {forward_batch.forward_mode}")