sglang 0.4.6.post3__py3-none-any.whl → 0.4.6.post5__py3-none-any.whl

sglang/srt/layers/communicator.py

@@ -0,0 +1,451 @@
+# Copyright 2023-2024 SGLang Team
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+from dataclasses import dataclass
+from enum import Enum, auto
+from typing import Dict, Optional, Tuple
+
+import torch.distributed
+
+from sglang.srt.distributed import (
+    get_tensor_model_parallel_world_size,
+    tensor_model_parallel_all_reduce,
+)
+from sglang.srt.layers.dp_attention import (
+    attn_tp_all_gather,
+    attn_tp_reduce_scatter,
+    dp_gather_partial,
+    dp_scatter,
+    get_attention_tp_rank,
+    get_attention_tp_size,
+    get_local_attention_dp_size,
+)
+from sglang.srt.managers.schedule_batch import global_server_args_dict
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+
+
+class ScatterMode(Enum):
+    SCATTERED = auto()
+    TP_ATTN_FULL = auto()
+    FULL = auto()
+
+
+@dataclass
+class _LayerModeComputationContext:
+    num_layers: int
+    layer_id: int
+    is_layer_sparse: bool
+    is_previous_layer_sparse: Optional[bool]
+
+    def previous_layer(self):
+        assert self.is_previous_layer_sparse is not None
+        return _LayerModeComputationContext(
+            layer_id=self.layer_id - 1,
+            is_layer_sparse=self.is_previous_layer_sparse,
+            is_previous_layer_sparse=None,
+            num_layers=self.num_layers,
+        )
+
+
+@dataclass
+class LayerScatterModes:
+    layer_input_mode: ScatterMode
+    attn_mode: ScatterMode
+    # Can be further split into e.g. mlp_input_mode and mlp_output_mode if needed
+    mlp_mode: ScatterMode
+    middle_residual_mode: ScatterMode
+    layer_output_mode: ScatterMode
+
+    @classmethod
+    def init_new(cls, **kwargs):
+        context = _LayerModeComputationContext(**kwargs)
+        return cls(
+            layer_input_mode=cls._compute_layer_input_mode(context),
+            attn_mode=ScatterMode.TP_ATTN_FULL,
+            mlp_mode=cls._compute_mlp_mode(context),
+            middle_residual_mode=cls._compute_middle_residual_mode(context),
+            layer_output_mode=cls._compute_layer_output_mode(context),
+        )
+
+    @classmethod
+    def _compute_layer_input_mode(cls, context: _LayerModeComputationContext):
+        if context.layer_id == 0:
+            return ScatterMode.TP_ATTN_FULL
+        return cls._compute_layer_output_mode(context.previous_layer())
+
+    @classmethod
+    def _compute_mlp_mode(cls, context: _LayerModeComputationContext):
+        if context.is_layer_sparse:
+            return (
+                ScatterMode.SCATTERED
+                if global_server_args_dict["enable_deepep_moe"]
+                else ScatterMode.FULL
+            )
+        else:
+            return (
+                ScatterMode.SCATTERED
+                if enable_moe_dense_fully_dp()
+                else ScatterMode.FULL
+            )
+
+    @classmethod
+    def _compute_middle_residual_mode(cls, context: _LayerModeComputationContext):
+        mlp_mode = cls._compute_mlp_mode(context)
+        if mlp_mode == ScatterMode.SCATTERED:
+            return ScatterMode.SCATTERED
+        if mlp_mode == ScatterMode.FULL:
+            return ScatterMode.TP_ATTN_FULL
+        raise NotImplementedError
+
+    @classmethod
+    def _compute_layer_output_mode(cls, context: _LayerModeComputationContext):
+        mlp_mode = cls._compute_mlp_mode(context)
+        if context.layer_id == context.num_layers - 1:
+            return ScatterMode.TP_ATTN_FULL
+        if mlp_mode == ScatterMode.SCATTERED:
+            return ScatterMode.SCATTERED
+        if mlp_mode == ScatterMode.FULL:
+            return ScatterMode.TP_ATTN_FULL
+        raise NotImplementedError
+
+
+def enable_moe_dense_fully_dp():
+    return global_server_args_dict["moe_dense_tp_size"] == 1
+
+
+class LayerCommunicator:
+    def __init__(
+        self,
+        layer_scatter_modes: LayerScatterModes,
+        input_layernorm: torch.nn.Module,
+        post_attention_layernorm: torch.nn.Module,
+    ):
+        self.layer_scatter_modes = layer_scatter_modes
+        self.input_layernorm = input_layernorm
+        self.post_attention_layernorm = post_attention_layernorm
+
+        self.attn_tp_rank = get_attention_tp_rank()
+        self.attn_tp_size = get_attention_tp_size()
+        self.local_attn_dp_size = get_local_attention_dp_size()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.process_group_sizes = {
+            ScatterMode.SCATTERED: 1,
+            ScatterMode.TP_ATTN_FULL: self.attn_tp_size,
+            ScatterMode.FULL: self.tp_size,
+        }
+
+    def prepare_attn(
+        self,
+        hidden_states: torch.Tensor,
+        residual: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ):
+        if hidden_states.shape[0] == 0:
+            residual = hidden_states
+        else:
+            if residual is None:
+                residual = hidden_states
+                hidden_states = self.input_layernorm(hidden_states)
+            else:
+                hidden_states, residual = self.input_layernorm(hidden_states, residual)
+
+        hidden_states = _communicate_simple(
+            hidden_states=hidden_states,
+            forward_batch=forward_batch,
+            input_mode=self.layer_scatter_modes.layer_input_mode,
+            output_mode=self.layer_scatter_modes.attn_mode,
+            context=self._compute_context(forward_batch),
+        )
+
+        return hidden_states, residual
+
+    def prepare_mlp(
+        self,
+        hidden_states: torch.Tensor,
+        residual: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ):
+        return _communicate_with_all_reduce_and_layer_norm(
+            hidden_states=hidden_states,
+            residual=residual,
+            forward_batch=forward_batch,
+            hidden_states_input_mode=self.layer_scatter_modes.attn_mode,
+            residual_input_mode=self.layer_scatter_modes.layer_input_mode,
+            hidden_states_output_mode=self.layer_scatter_modes.mlp_mode,
+            residual_output_mode=self.layer_scatter_modes.middle_residual_mode,
+            layernorm=self.post_attention_layernorm,
+            context=self._compute_context(forward_batch),
+        )
+
+    def postprocess_layer(
+        self,
+        hidden_states: torch.Tensor,
+        residual: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ):
+        return _communicate_summable_tensor_pair(
+            hidden_states=hidden_states,
+            residual=residual,
+            forward_batch=forward_batch,
+            hidden_states_input_mode=self.layer_scatter_modes.mlp_mode,
+            residual_input_mode=self.layer_scatter_modes.middle_residual_mode,
+            output_mode=self.layer_scatter_modes.layer_output_mode,
+            context=self._compute_context(forward_batch),
+        )
+
+    def _compute_context(self, forward_batch: ForwardBatch):
+        return _Context(
+            num_tokens_of_mode=_compute_num_tokens_of_mode(
+                forward_batch,
+                attn_tp_rank=self.attn_tp_rank,
+                attn_tp_size=self.attn_tp_size,
+            ),
+            process_group_sizes=self.process_group_sizes,
+            attn_tp_rank=self.attn_tp_rank,
+            attn_tp_size=self.attn_tp_size,
+            local_attn_dp_size=self.local_attn_dp_size,
+            tp_size=self.tp_size,
+        )
+
+
+def _compute_num_tokens_of_mode(
+    forward_batch: ForwardBatch, attn_tp_rank: int, attn_tp_size: int
+):
+    tp_attn_full_num_tokens = forward_batch.input_ids.shape[0]
+    return {
+        ScatterMode.SCATTERED: _torch_tensor_split_len(
+            tp_attn_full_num_tokens, attn_tp_size, attn_tp_rank
+        ),
+        ScatterMode.TP_ATTN_FULL: tp_attn_full_num_tokens,
+        ScatterMode.FULL: (
+            forward_batch.gathered_buffer.shape[0]
+            if global_server_args_dict["enable_dp_attention"]
+            else forward_batch.input_ids.shape[0]
+        ),
+    }
+
+
+def _torch_tensor_split_len(tensor_len: int, n: int, output_index: int):
+    if output_index < int(tensor_len % n):
+        return int(tensor_len / n) + 1
+    else:
+        return int(tensor_len / n)
+
+
+@dataclass
+class _Context:
+    num_tokens_of_mode: Dict["ScatterMode", int]
+    process_group_sizes: Dict["ScatterMode", int]
+    attn_tp_rank: int
+    attn_tp_size: int
+    local_attn_dp_size: int
+    tp_size: int
+
+    def is_same_group_size(self, a: "ScatterMode", b: "ScatterMode"):
+        return self.process_group_sizes[a] == self.process_group_sizes[b]
+
+    def check_shape(self, x: torch.Tensor, mode: ScatterMode):
+        if x is None:
+            return
+
+        actual_num_tokens = x.shape[0]
+        expect_num_tokens = self.num_tokens_of_mode[mode]
+        assert (
+            actual_num_tokens == expect_num_tokens
+        ), f"{actual_num_tokens=} {expect_num_tokens=} {mode=} {x.shape=} {self.num_tokens_of_mode=} {self.process_group_sizes=}"
+        return x
+
+    def check_shapes(
+        self, xs: Tuple[torch.Tensor, ...], modes: Tuple[ScatterMode, ...]
+    ) -> Tuple[torch.Tensor, ...]:
+        return tuple(
+            [self.check_shape(x, mode) for x, mode in zip(xs, modes, strict=True)]
+        )
+
+
+def _communicate_simple(
+    hidden_states: torch.Tensor,
+    forward_batch: ForwardBatch,
+    input_mode: ScatterMode,
+    output_mode: ScatterMode,
+    context: _Context,
+) -> torch.Tensor:
+    def _inner():
+        nonlocal hidden_states
+
+        if context.is_same_group_size(input_mode, output_mode):
+            return hidden_states
+
+        if (input_mode == ScatterMode.SCATTERED) and (
+            output_mode == ScatterMode.TP_ATTN_FULL
+        ):
+            hidden_states, local_hidden_states = (
+                forward_batch.gathered_buffer[: forward_batch.input_ids.shape[0]],
+                hidden_states,
+            )
+            attn_tp_all_gather(
+                list(hidden_states.tensor_split(context.attn_tp_size)),
+                local_hidden_states,
+            )
+            return hidden_states
+
+        raise NotImplementedError(f"{input_mode=} {output_mode=}")
+
+    context.check_shape(hidden_states, input_mode)
+    return context.check_shape(_inner(), output_mode)
+
+
+def _communicate_with_all_reduce_and_layer_norm(
+    hidden_states: torch.Tensor,
+    residual: torch.Tensor,
+    hidden_states_input_mode: ScatterMode,
+    residual_input_mode: ScatterMode,
+    hidden_states_output_mode: ScatterMode,
+    residual_output_mode: ScatterMode,
+    forward_batch: ForwardBatch,
+    layernorm: torch.nn.Module,
+    context: _Context,
+):
+    """Besides communication, needs to
+    1. All reduce in tp_attn_group on hidden_states
+    2. Apply layer norm
+    """
+
+    def _inner():
+        nonlocal hidden_states, residual
+
+        if (
+            context.is_same_group_size(
+                hidden_states_input_mode, hidden_states_output_mode
+            )
+            and context.is_same_group_size(residual_input_mode, residual_output_mode)
+            and context.attn_tp_size == 1
+        ):
+            # TODO move these `if shape != 0` into LayerNorm itself
+            if hidden_states.shape[0] != 0:
+                hidden_states, residual = layernorm(hidden_states, residual)
+            return hidden_states, residual
+
+        if (
+            (hidden_states_input_mode == ScatterMode.TP_ATTN_FULL)
+            and (residual_input_mode == ScatterMode.TP_ATTN_FULL)
+            and (hidden_states_output_mode == ScatterMode.FULL)
+            and (residual_output_mode == ScatterMode.TP_ATTN_FULL)
+        ):
+            if context.local_attn_dp_size != 1:
+                if context.attn_tp_rank == 0:
+                    hidden_states += residual
+                hidden_states, local_hidden_states = (
+                    forward_batch.gathered_buffer,
+                    hidden_states,
+                )
+                dp_gather_partial(hidden_states, local_hidden_states, forward_batch)
+                dp_scatter(residual, hidden_states, forward_batch)
+                if hidden_states.shape[0] != 0:
+                    hidden_states = layernorm(hidden_states)
+            else:
+                hidden_states = tensor_model_parallel_all_reduce(hidden_states)
+                hidden_states, residual = layernorm(hidden_states, residual)
+            return hidden_states, residual
+
+        if (
+            (hidden_states_input_mode == ScatterMode.TP_ATTN_FULL)
+            and (
+                residual_input_mode in [ScatterMode.SCATTERED, ScatterMode.TP_ATTN_FULL]
+            )
+            and (hidden_states_output_mode == ScatterMode.SCATTERED)
+            and (residual_output_mode == ScatterMode.SCATTERED)
+        ):
+            tensor_list = list(hidden_states.tensor_split(context.attn_tp_size))
+            hidden_states = tensor_list[context.attn_tp_rank]
+            attn_tp_reduce_scatter(hidden_states, tensor_list)
+            if residual_input_mode == ScatterMode.TP_ATTN_FULL:
+                residual = residual.tensor_split(context.attn_tp_size)[
+                    context.attn_tp_rank
+                ]
+            if hidden_states.shape[0] != 0:
+                hidden_states, residual = layernorm(hidden_states, residual)
+            return hidden_states, residual
+
+        raise NotImplementedError(
+            f"{hidden_states_input_mode=} {residual_input_mode=} {residual_output_mode=} {residual_output_mode=}"
+        )
+
+    context.check_shapes(
+        (hidden_states, residual), (hidden_states_input_mode, residual_input_mode)
+    )
+    return context.check_shapes(
+        _inner(), (hidden_states_output_mode, residual_output_mode)
+    )
+
+
+def _communicate_summable_tensor_pair(
+    hidden_states: torch.Tensor,
+    residual: torch.Tensor,
+    forward_batch: ForwardBatch,
+    hidden_states_input_mode: ScatterMode,
+    residual_input_mode: ScatterMode,
+    output_mode: ScatterMode,
+    context: _Context,
+):
+    """It is allowed to make (hidden_states, residual) := (hidden_states + residual, None) if needed."""
+
+    def _inner():
+        nonlocal hidden_states, residual
+
+        if context.is_same_group_size(
+            hidden_states_input_mode, output_mode
+        ) and context.is_same_group_size(residual_input_mode, output_mode):
+            return hidden_states, residual
+
+        if (
+            (hidden_states_input_mode == ScatterMode.FULL)
+            and (residual_input_mode == ScatterMode.TP_ATTN_FULL)
+            and (output_mode == ScatterMode.TP_ATTN_FULL)
+        ):
+            # TODO(ch-wan): use reduce-scatter in MLP to avoid this scatter
+            # important: forward batch.gathered_buffer is used both after scatter and after gather.
+            # be careful about this!
+            hidden_states, global_hidden_states = (
+                forward_batch.gathered_buffer[: forward_batch.input_ids.shape[0]],
+                hidden_states,
+            )
+            dp_scatter(hidden_states, global_hidden_states, forward_batch)
+            return hidden_states, residual
+
+        if (
+            (hidden_states_input_mode == ScatterMode.SCATTERED)
+            and (residual_input_mode == ScatterMode.SCATTERED)
+            and (output_mode == ScatterMode.TP_ATTN_FULL)
+        ):
+            hidden_states += residual
+            residual = None
+            hidden_states, local_hidden_states = (
+                forward_batch.gathered_buffer[: forward_batch.input_ids.shape[0]],
+                hidden_states,
+            )
+            attn_tp_all_gather(
+                list(hidden_states.tensor_split(context.attn_tp_size)),
+                local_hidden_states,
+            )
+            return hidden_states, residual
+
+        raise NotImplementedError(
+            f"{hidden_states_input_mode=} {residual_input_mode=} {output_mode=}"
+        )
+
+    context.check_shapes(
+        (hidden_states, residual), (hidden_states_input_mode, residual_input_mode)
+    )
+    return context.check_shapes(_inner(), (output_mode, output_mode))

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 +123,23 @@ 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
 
 
 @contextmanager
@@ -112,19 +152,19 @@ 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)
@@ -201,7 +241,7 @@ def _dp_gather(
             global_tokens, local_tokens, 0, local_start_pos, local_num_tokens, False
         )
 
-    # Input IDs are in int 32. We should use inplace_all_reduce for local case becaues of custom all reduce.
+    # Input IDs are in int 32. We should use inplace_all_reduce for local case because of custom all reduce.
     NUM_GPUS_PER_NODE = 8
     if (
         not local_tokens.dtype.is_floating_point
@@ -252,12 +292,12 @@ def dp_scatter(
         )
 
 
-def tp_reduce_scatter(
+def attn_tp_reduce_scatter(
     output: torch.Tensor,
     input_list: List[torch.Tensor],
 ):
     return get_attention_tp_group().reduce_scatter(output, input_list)
 
 
-def tp_all_gather(output_list: List[torch.Tensor], input_: torch.Tensor):
+def attn_tp_all_gather(output_list: List[torch.Tensor], input_: torch.Tensor):
     return get_attention_tp_group().all_gather(input_, tensor_list=output_list)

sglang/srt/layers/layernorm.py

@@ -76,7 +76,7 @@ class RMSNorm(CustomOp):
         residual: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
         if not x.is_contiguous():
-            # NOTE: Romove this if aiter kernel supports discontinuous input
+            # NOTE: Remove this if aiter kernel supports discontinuous input
             x = x.contiguous()
         if residual is not None:
             fused_add_rms_norm(x, residual, self.weight.data, self.variance_epsilon)