Fix two issues related to `--moe-dense-tp-size=1` (#5657)

Co-authored-by: liusy58 <liusy58@linux.alibaba.com>
Co-authored-by: 颉沆 <xiehang.lsy@alibaba-inc.com>

python/sglang/srt/layers/dp_attention.py

@@ -24,8 +24,10 @@ if TYPE_CHECKING:
 _ATTN_TP_GROUP = None
 _ATTN_TP_RANK = None
 _ATTN_TP_SIZE = None
-_DP_RANK = None
-_DP_SIZE = None
+_ATTN_DP_RANK = None
+_ATTN_DP_SIZE = None
+_LOCAL_ATTN_DP_SIZE = None
+_LOCAL_ATTN_DP_RANK = None
 
 
 def compute_dp_attention_world_info(enable_dp_attention, tp_rank, tp_size, dp_size):
@@ -33,9 +35,27 @@ def compute_dp_attention_world_info(enable_dp_attention, tp_rank, tp_size, dp_si
         return tp_rank, tp_size, 0
 
     attn_tp_size = tp_size // dp_size
-    dp_rank = tp_rank // attn_tp_size
+    attn_dp_rank = tp_rank // attn_tp_size
     attn_tp_rank = tp_rank % attn_tp_size
-    return attn_tp_rank, attn_tp_size, dp_rank
+
+    return attn_tp_rank, attn_tp_size, attn_dp_rank
+
+
+def compute_dp_attention_local_info(
+    enable_dp_attention, tp_rank, tp_size, dp_size, moe_dense_tp_size
+):
+    if not enable_dp_attention:
+        return tp_rank, tp_size, 0
+
+    local_tp_size = moe_dense_tp_size if moe_dense_tp_size else tp_size
+    local_tp_rank = tp_rank % local_tp_size
+    local_dp_size = max(1, dp_size // (tp_size // local_tp_size))
+
+    local_attn_tp_size = local_tp_size // local_dp_size
+    local_attn_dp_rank = local_tp_rank // local_attn_tp_size
+    local_attn_tp_rank = local_tp_rank % local_attn_tp_size
+
+    return local_attn_tp_rank, local_attn_tp_size, local_attn_dp_rank
 
 
 def initialize_dp_attention(
@@ -43,22 +63,32 @@ def initialize_dp_attention(
     tp_rank: int,
     tp_size: int,
     dp_size: int,
+    moe_dense_tp_size: int,
     pp_size: int,
 ):
-    global _ATTN_TP_GROUP, _ATTN_TP_RANK, _ATTN_TP_SIZE, _DP_RANK, _DP_SIZE
+    global _ATTN_TP_GROUP, _ATTN_TP_RANK, _ATTN_TP_SIZE, _ATTN_DP_RANK, _ATTN_DP_SIZE
+    global _LOCAL_ATTN_DP_SIZE, _LOCAL_ATTN_DP_RANK
 
     from sglang.srt.layers.sampler import SYNC_TOKEN_IDS_ACROSS_TP
 
-    _ATTN_TP_RANK, _ATTN_TP_SIZE, _DP_RANK = compute_dp_attention_world_info(
+    _ATTN_TP_RANK, _ATTN_TP_SIZE, _ATTN_DP_RANK = compute_dp_attention_world_info(
         enable_dp_attention, tp_rank, tp_size, dp_size
     )
+    _, _, _LOCAL_ATTN_DP_RANK = compute_dp_attention_local_info(
+        enable_dp_attention, tp_rank, tp_size, dp_size, moe_dense_tp_size
+    )
 
     if enable_dp_attention:
         local_rank = tp_rank % (tp_size // dp_size)
-        _DP_SIZE = dp_size
+        _ATTN_DP_SIZE = dp_size
+        if moe_dense_tp_size is None:
+            _LOCAL_ATTN_DP_SIZE = _ATTN_DP_SIZE
+        else:
+            _LOCAL_ATTN_DP_SIZE = max(1, dp_size // (tp_size // moe_dense_tp_size))
     else:
         local_rank = tp_rank
-        _DP_SIZE = 1
+        _ATTN_DP_SIZE = 1
+        _LOCAL_ATTN_DP_SIZE = 1
 
     tp_group = get_tp_group()
     _ATTN_TP_GROUP = GroupCoordinator(
@@ -93,13 +123,33 @@ def get_attention_tp_size():
 
 
 def get_attention_dp_rank():
-    assert _DP_RANK is not None, "dp attention not initialized!"
-    return _DP_RANK
+    assert _ATTN_DP_RANK is not None, "dp attention not initialized!"
+    return _ATTN_DP_RANK
 
 
 def get_attention_dp_size():
-    assert _DP_SIZE is not None, "dp attention not initialized!"
-    return _DP_SIZE
+    assert _ATTN_DP_SIZE is not None, "dp attention not initialized!"
+    return _ATTN_DP_SIZE
+
+
+def get_local_attention_dp_rank():
+    assert _LOCAL_ATTN_DP_RANK is not None, "dp attention not initialized!"
+    return _LOCAL_ATTN_DP_RANK
+
+
+def get_local_attention_dp_size():
+    assert _LOCAL_ATTN_DP_SIZE is not None, "dp attention not initialized!"
+    return _LOCAL_ATTN_DP_SIZE
+
+
+def get_local_attention_dp_rank():
+    assert _LOCAL_ATTN_DP_RANK is not None, "dp attention not initialized!"
+    return _LOCAL_ATTN_DP_RANK
+
+
+def get_local_attention_dp_size():
+    assert _LOCAL_ATTN_DP_SIZE is not None, "dp attention not initialized!"
+    return _LOCAL_ATTN_DP_SIZE
 
 
 @contextmanager
@@ -112,19 +162,19 @@ def disable_dp_size():
     Args:
         tp_group (GroupCoordinator): the tp group coordinator
     """
-    global _DP_SIZE
-    assert _DP_SIZE is not None, "dp attention not initialized!"
+    global _ATTN_DP_SIZE
+    assert _ATTN_DP_SIZE is not None, "dp attention not initialized!"
 
-    old_dp_size = _DP_SIZE
-    _DP_SIZE = 1
+    old_dp_size = _ATTN_DP_SIZE
+    _ATTN_DP_SIZE = 1
     try:
         yield
     finally:
-        _DP_SIZE = old_dp_size
+        _ATTN_DP_SIZE = old_dp_size
 
 
 def get_dp_local_info(forward_batch: ForwardBatch):
-    dp_rank = get_attention_dp_rank()
+    dp_rank = get_local_attention_dp_rank()
 
     if forward_batch.dp_local_start_pos is None:
         cumtokens = torch.cumsum(forward_batch.global_num_tokens_gpu, dim=0)

python/sglang/srt/layers/logits_processor.py

@@ -30,9 +30,10 @@ from sglang.srt.layers.dp_attention import (
     attn_tp_all_gather,
     dp_gather_replicate,
     dp_scatter,
-    get_attention_dp_rank,
     get_attention_dp_size,
     get_attention_tp_size,
+    get_local_attention_dp_rank,
+    get_local_attention_dp_size,
 )
 from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding
 from sglang.srt.managers.schedule_batch import global_server_args_dict
@@ -46,6 +47,18 @@ from sglang.srt.utils import dump_to_file
 logger = logging.getLogger(__name__)
 
 
+from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding
+from sglang.srt.managers.schedule_batch import global_server_args_dict
+from sglang.srt.model_executor.forward_batch_info import (
+    CaptureHiddenMode,
+    ForwardBatch,
+    ForwardMode,
+)
+from sglang.srt.utils import dump_to_file
+
+logger = logging.getLogger(__name__)
+
+
 @dataclasses.dataclass
 class LogitsProcessorOutput:
     ## Part 1: This part will be assigned in python/sglang/srt/layers/logits_processor.py::LogitsProcessor
@@ -170,7 +183,7 @@ class LogitsMetadata:
             return
 
         cumtokens = torch.cumsum(self.global_num_tokens_for_logprob_gpu, dim=0)
-        dp_rank = get_attention_dp_rank()
+        dp_rank = get_local_attention_dp_rank()
         if dp_rank == 0:
             dp_local_start_pos = torch.zeros_like(
                 self.global_num_tokens_for_logprob_gpu[0]
@@ -324,7 +337,8 @@ class LogitsProcessor(nn.Module):
 
         if self.debug_tensor_dump_output_folder:
             assert (
-                not self.do_tensor_parallel_all_gather or get_attention_dp_size() == 1
+                not self.do_tensor_parallel_all_gather
+                or get_local_attention_dp_size() == 1
             ), "dp attention + sharded lm_head doesn't support full logits"
             full_logits = self._get_logits(hidden_states, lm_head, logits_metadata)
             dump_to_file(self.debug_tensor_dump_output_folder, "logits", full_logits)

python/sglang/srt/managers/scheduler.py

@@ -207,7 +207,8 @@ class Scheduler(
         self.page_size = server_args.page_size
 
         # Distributed rank info
-        self.attn_tp_rank, self.attn_tp_size, self.dp_rank = (
+        self.dp_size = server_args.dp_size
+        self.attn_tp_rank, self.attn_tp_size, self.attn_dp_rank = (
             compute_dp_attention_world_info(
                 server_args.enable_dp_attention,
                 self.tp_rank,
@@ -768,7 +769,7 @@ class Scheduler(
                         )
 
                     # send out reqs to the next stage
-                    dp_offset = self.dp_rank * self.attn_tp_size
+                    dp_offset = self.attn_dp_rank * self.attn_tp_size
                     if self.attn_tp_rank == 0:
                         point_to_point_pyobj(
                             recv_reqs,
@@ -815,7 +816,7 @@ class Scheduler(
                 recv_reqs = None
         else:
             if self.attn_tp_rank == 0:
-                dp_offset = self.dp_rank * self.attn_tp_size
+                dp_offset = self.attn_dp_rank * self.attn_tp_size
                 recv_reqs = point_to_point_pyobj(
                     [],
                     self.pp_rank * self.tp_size + dp_offset,
@@ -1610,6 +1611,7 @@ class Scheduler(
             local_batch,
             dp_size=self.server_args.dp_size,
             attn_tp_size=self.attn_tp_size,
+            moe_dense_tp_size=self.server_args.moe_dense_tp_size,
             tp_cpu_group=self.tp_cpu_group,
             get_idle_batch=self.get_idle_batch,
             disable_cuda_graph=self.server_args.disable_cuda_graph,
@@ -1622,6 +1624,7 @@ class Scheduler(
         local_batch: ScheduleBatch,
         dp_size,
         attn_tp_size: int,
+        moe_dense_tp_size: Optional[int],
         tp_cpu_group,
         get_idle_batch,
         disable_cuda_graph: bool,
@@ -1631,15 +1634,15 @@ class Scheduler(
         # Check if other DP workers have running batches
         if local_batch is None:
             num_tokens = 0
-            global_num_tokens_for_logprob = 0
+            num_tokens_for_logprob = 0
         elif local_batch.forward_mode.is_decode():
             num_tokens = local_batch.batch_size()
             if not spec_algorithm.is_none() and spec_algorithm.is_eagle():
                 num_tokens = num_tokens * speculative_num_draft_tokens
-            global_num_tokens_for_logprob = num_tokens
+            num_tokens_for_logprob = num_tokens
         else:
             num_tokens = local_batch.extend_num_tokens
-            global_num_tokens_for_logprob = sum(
+            num_tokens_for_logprob = sum(
                 [
                     # We should have at least 1 token for sample in every case.
                     max(extend_len - logprob_start_len, 1)
@@ -1666,7 +1669,7 @@ class Scheduler(
             [
                 num_tokens,
                 can_cuda_graph,
-                global_num_tokens_for_logprob,
+                num_tokens_for_logprob,
                 is_extend_in_batch,
             ],
             dtype=torch.int64,
@@ -1689,8 +1692,15 @@ class Scheduler(
             local_batch = get_idle_batch()
 
         if local_batch is not None:
-            local_batch.global_num_tokens = global_num_tokens
-            local_batch.global_num_tokens_for_logprob = global_num_tokens_for_logprob
+            # TODO: handle the case when moe_dense_tp_size != 1
+            if moe_dense_tp_size == 1 and global_server_args_dict["enable_dp_lm_head"]:
+                local_batch.global_num_tokens = [num_tokens]
+                local_batch.global_num_tokens_for_logprob = [num_tokens_for_logprob]
+            else:
+                local_batch.global_num_tokens = global_num_tokens
+                local_batch.global_num_tokens_for_logprob = (
+                    global_num_tokens_for_logprob
+                )
 
             # Check forward mode for cuda graph
             if not disable_cuda_graph:
@@ -2177,8 +2187,8 @@ class Scheduler(
 
     def get_print_prefix(self):
         prefix = ""
-        if self.dp_rank is not None:
-            prefix += f" DP{self.dp_rank}"
+        if self.attn_dp_rank is not None:
+            prefix += f" DP{self.attn_dp_rank}"
         if self.server_args.tp_size > 1:
             prefix += f" TP{self.tp_rank}"
         if self.pp_size > 1:

python/sglang/srt/model_executor/model_runner.py

@@ -401,6 +401,7 @@ class ModelRunner:
                 tp_rank=self.tp_rank,
                 tp_size=self.tp_size,
                 dp_size=self.server_args.dp_size,
+                moe_dense_tp_size=self.server_args.moe_dense_tp_size,
                 pp_size=self.server_args.pp_size,
             )
 

python/sglang/srt/models/deepseek_v2.py

@@ -40,9 +40,9 @@ from sglang.srt.layers.dp_attention import (
     attn_tp_reduce_scatter,
     dp_gather_partial,
     dp_scatter,
-    get_attention_dp_size,
     get_attention_tp_rank,
     get_attention_tp_size,
+    get_local_attention_dp_size,
 )
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
@@ -438,7 +438,6 @@ class DeepseekV2AttentionMLA(nn.Module):
         self.v_head_dim = v_head_dim
         self.q_lora_rank = q_lora_rank
         self.kv_lora_rank = kv_lora_rank
-        self.dp_size = get_attention_dp_size()
         attn_tp_rank = get_attention_tp_rank()
         attn_tp_size = get_attention_tp_size()
 
@@ -1133,7 +1132,7 @@ class DeepseekV2DecoderLayer(nn.Module):
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         self.enable_dp_attention = global_server_args_dict["enable_dp_attention"]
         self.layer_id = layer_id
-        self.dp_size = get_attention_dp_size()
+        self.local_dp_size = get_local_attention_dp_size()
         self.attn_tp_size = get_attention_tp_size()
         self.attn_tp_rank = get_attention_tp_rank()
         self.self_attn = DeepseekV2AttentionMLA(
@@ -1184,7 +1183,8 @@ class DeepseekV2DecoderLayer(nn.Module):
             )
 
         self.input_is_scattered = (
-            previous_layer_info.ffn_input_mode == _FFNInputMode.SCATTERED
+            layer_id > 0
+            and previous_layer_info.ffn_input_mode == _FFNInputMode.SCATTERED
         )
         self.is_last_layer = self.layer_id == config.num_hidden_layers - 1
 
@@ -1264,7 +1264,7 @@ class DeepseekV2DecoderLayer(nn.Module):
         # Gather
         if get_tensor_model_parallel_world_size() > 1:
             # all gather and all reduce
-            if self.dp_size != 1:
+            if self.local_dp_size != 1:
                 if self.attn_tp_rank == 0:
                     hidden_states += residual
                 hidden_states, local_hidden_states = (
@@ -1289,7 +1289,7 @@ class DeepseekV2DecoderLayer(nn.Module):
 
         # TODO(ch-wan): use reduce-scatter in MLP to avoid this scatter
         # Scatter
-        if self.dp_size != 1:
+        if self.local_dp_size != 1:
             # important: forward batch.gathered_buffer is used both after scatter and after gather.
             # be careful about this!
             hidden_states, global_hidden_states = (
@@ -1413,7 +1413,7 @@ class DeepseekV2Model(nn.Module):
         )
         self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 
-        self.dp_size = get_attention_dp_size()
+        self.dp_size = get_local_attention_dp_size()
 
     def get_input_embeddings(self) -> torch.Tensor:
         return self.embed_tokens
@@ -1478,7 +1478,7 @@ class DeepseekV2ForCausalLM(nn.Module):
             use_attn_tp_group=global_server_args_dict["enable_dp_lm_head"],
         )
         self.logits_processor = LogitsProcessor(config)
-        self.dp_size = get_attention_dp_size()
+        self.dp_size = get_local_attention_dp_size()
 
     def determine_n_share_experts_fusion(
         self, architecture: str = "DeepseekV3ForCausalLM"

python/sglang/srt/models/llama4.py

@@ -30,9 +30,9 @@ from sglang.srt.distributed import (
 from sglang.srt.layers.dp_attention import (
     dp_gather_partial,
     dp_scatter,
-    get_attention_dp_size,
     get_attention_tp_rank,
     get_attention_tp_size,
+    get_local_attention_dp_size,
 )
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
@@ -198,7 +198,6 @@ class Llama4Attention(nn.Module):
         self.use_rope = int((layer_id + 1) % 4 != 0)
         self.use_qk_norm = config.use_qk_norm and self.use_rope
 
-        self.dp_size = get_attention_dp_size()
         attn_tp_rank = get_attention_tp_rank()
         attn_tp_size = get_attention_tp_size()
 
@@ -342,7 +341,7 @@ class Llama4DecoderLayer(nn.Module):
         rope_theta = config.rope_theta
         rope_scaling = config.rope_scaling
         max_position_embeddings = config.max_position_embeddings
-        self.dp_size = get_attention_dp_size()
+        self.local_dp_size = get_local_attention_dp_size()
         self.attn_tp_size = get_attention_tp_size()
         self.attn_tp_rank = get_attention_tp_rank()
 
@@ -405,7 +404,7 @@ class Llama4DecoderLayer(nn.Module):
         # Gather
         if get_tensor_model_parallel_world_size() > 1:
             # all gather and all reduce
-            if self.dp_size != 1:
+            if self.local_dp_size != 1:
                 if self.attn_tp_rank == 0:
                     hidden_states += residual
                 hidden_states, local_hidden_states = (
@@ -430,7 +429,7 @@ class Llama4DecoderLayer(nn.Module):
 
         # TODO(ch-wan): use reduce-scatter in MLP to avoid this scatter
         # Scatter
-        if self.dp_size != 1:
+        if self.local_dp_size != 1:
             # important: forward batch.gathered_buffer is used both after scatter and after gather.
             # be careful about this!
             hidden_states, global_hidden_states = (