sglang 0.4.6.post5__py3-none-any.whl → 0.4.7.post1__py3-none-any.whl

sglang/srt/model_executor/expert_location_updater.py

@@ -12,8 +12,10 @@
 # limitations under the License.
 # ==============================================================================
 import logging
-from typing import Dict, List, Tuple
+from collections import defaultdict
+from typing import Dict, List, Optional, Tuple
 
+import einops
 import torch
 import torch.distributed
 from torch.distributed import P2POp
@@ -22,58 +24,131 @@ from sglang.srt.managers.expert_location import (
     ExpertLocationMetadata,
     get_global_expert_location_metadata,
 )
+from sglang.srt.managers.schedule_batch import global_server_args_dict
+from sglang.srt.utils import get_bool_env_var
 
 logger = logging.getLogger(__name__)
 
 
-def update_expert_location(
+class ExpertLocationUpdater:
+    def __init__(self):
+        self._first_execution = True
+
+    def update(
+        self,
+        routed_experts_weights_of_layer: Dict[int, List[torch.Tensor]],
+        new_expert_location_metadata: ExpertLocationMetadata,
+        update_layer_ids: List[int],
+        nnodes: int,
+        rank: int,
+    ):
+        if self._first_execution:
+            self._first_execution = False
+            torch.cuda.empty_cache()
+
+        old_expert_location_metadata = get_global_expert_location_metadata()
+        _update_expert_weights(
+            routed_experts_weights_of_layer=routed_experts_weights_of_layer,
+            old_expert_location_metadata=old_expert_location_metadata,
+            new_expert_location_metadata=new_expert_location_metadata,
+            update_layer_ids=update_layer_ids,
+            nnodes=nnodes,
+            rank=rank,
+        )
+        old_expert_location_metadata.update(
+            new_expert_location_metadata,
+            update_layer_ids=update_layer_ids,
+        )
+
+
+def _update_expert_weights(**kwargs):
+    if get_bool_env_var("SGLANG_EXPERT_LOCATION_UPDATER_CANARY"):
+        return _update_expert_weights_with_canary(**kwargs)
+    else:
+        return _update_expert_weights_raw(**kwargs)
+
+
+# can add watchdog as well
+def _update_expert_weights_with_canary(
     routed_experts_weights_of_layer: Dict[int, List[torch.Tensor]],
+    old_expert_location_metadata: ExpertLocationMetadata,
     new_expert_location_metadata: ExpertLocationMetadata,
+    update_layer_ids: List[int],
     nnodes: int,
     rank: int,
 ):
-    old_expert_location_metadata = get_global_expert_location_metadata()
-    _update_expert_weights(
-        routed_experts_weights_of_layer,
-        old_expert_location_metadata,
-        new_expert_location_metadata,
-        nnodes,
-        rank,
+    num_local_physical_experts = old_expert_location_metadata.num_local_physical_experts
+
+    def _get_canary_value(meta: ExpertLocationMetadata, layer_id: int):
+        return meta.physical_to_logical_map_cpu[
+            layer_id,
+            num_local_physical_experts * rank : num_local_physical_experts * (rank + 1),
+        ]
+
+    routed_experts_weights_of_layer = {
+        k: [x for x in v] for k, v in routed_experts_weights_of_layer.items()
+    }
+    for layer_id in update_layer_ids:
+        canary_tensor = (
+            _get_canary_value(old_expert_location_metadata, layer_id)
+            .clone()
+            .to(device=global_server_args_dict["device"], non_blocking=True)
+        )
+        routed_experts_weights_of_layer[layer_id].append(canary_tensor)
+
+    _update_expert_weights_raw(
+        routed_experts_weights_of_layer=routed_experts_weights_of_layer,
+        old_expert_location_metadata=old_expert_location_metadata,
+        new_expert_location_metadata=new_expert_location_metadata,
+        update_layer_ids=update_layer_ids,
+        nnodes=nnodes,
+        rank=rank,
     )
-    old_expert_location_metadata.update(new_expert_location_metadata)
+
+    for layer_id in update_layer_ids:
+        # can optimize speed if needed
+        expect_value = _get_canary_value(new_expert_location_metadata, layer_id)
+        actual_value = routed_experts_weights_of_layer[layer_id][-1].cpu()
+        assert torch.all(expect_value == actual_value), (
+            f"{expect_value=} {actual_value=} {layer_id=} "
+            f"{old_expert_location_metadata.physical_to_logical_map_cpu.tolist()=} "
+            f"{new_expert_location_metadata.physical_to_logical_map_cpu.tolist()=} "
+        )
 
 
-def _update_expert_weights(
+def _update_expert_weights_raw(
     routed_experts_weights_of_layer: Dict[int, List[torch.Tensor]],
     old_expert_location_metadata: ExpertLocationMetadata,
     new_expert_location_metadata: ExpertLocationMetadata,
+    update_layer_ids: List[int],
     nnodes: int,
     rank: int,
 ):
+    log_metrics = get_bool_env_var("SGLANG_EXPERT_LOCATION_UPDATER_LOG_METRICS")
+
     temp_buffers = create_temp_buffers(
-        next(iter(routed_experts_weights_of_layer.values()))
+        routed_experts_weights_of_layer[update_layer_ids[0]]
     )
 
     world_size = torch.distributed.get_world_size()
     num_local_physical_experts = old_expert_location_metadata.num_local_physical_experts
     num_gpu_per_node = world_size // nnodes
 
-    old_physical_to_logical_map = (
-        old_expert_location_metadata.physical_to_logical_map.tolist()
-    )
-    new_physical_to_logical_map = (
-        new_expert_location_metadata.physical_to_logical_map.tolist()
-    )
-
-    for layer_id in sorted(routed_experts_weights_of_layer.keys()):
+    for layer_id in update_layer_ids:
         update_expert_weights_single_layer(
             routed_experts_weights=routed_experts_weights_of_layer[layer_id],
             temp_buffers=temp_buffers,
-            old_physical_to_logical_map=old_physical_to_logical_map[layer_id],
-            new_physical_to_logical_map=new_physical_to_logical_map[layer_id],
+            old_physical_to_logical_map=old_expert_location_metadata.physical_to_logical_map_cpu[
+                layer_id
+            ].tolist(),
+            new_physical_to_logical_map=new_expert_location_metadata.physical_to_logical_map_cpu[
+                layer_id
+            ].tolist(),
             num_local_physical_experts=num_local_physical_experts,
             num_gpu_per_node=num_gpu_per_node,
             rank=rank,
+            world_size=world_size,
+            log_metrics=log_metrics,
         )
 
 
@@ -89,7 +164,9 @@ def update_expert_weights_single_layer(
     num_local_physical_experts: int,
     num_gpu_per_node: int,
     rank: int,
+    world_size: Optional[int] = None,
     debug: bool = False,
+    log_metrics: bool = False,
 ):
     assert all(
         tensor.shape[0] == num_local_physical_experts
@@ -121,6 +198,14 @@ def update_expert_weights_single_layer(
         _execute_p2p_ops(p2p_op_infos)
         _execute_buffer2weight_copies(buffer2weight_copy_infos)
 
+        if log_metrics:
+            _log_p2p_op_metrics(
+                p2p_op_infos,
+                world_size=world_size,
+                num_gpu_per_node=num_gpu_per_node,
+                self_node_id=self_node_id,
+            )
+
         if debug:
             output_logs.append(f"{p2p_op_infos=}")
             output_logs.append(f"{buffer2weight_copy_infos=}")
@@ -420,3 +505,53 @@ def _deduplicate_ordered(arr: List[int]):
         if len(output) == 0 or item != output[-1]:
             output.append(item)
     return output
+
+
+def _log_p2p_op_metrics(
+    p2p_op_infos: List[Tuple[int, List[P2POp]]],
+    num_gpu_per_node: int,
+    world_size: int,
+    self_node_id: int,
+):
+    text = ""
+    all_ops = [op for _, ops in p2p_op_infos for op in ops]
+
+    for direction, ops in _group_by(all_ops, _get_direction_from_op).items():
+        nbytes_of_gpu = [0] * world_size
+        for op in ops:
+            nbytes_of_gpu[op.peer] += op.tensor.nbytes
+        nbytes_of_gpu = torch.tensor(nbytes_of_gpu, dtype=torch.int64)
+
+        nbytes_of_node = einops.reduce(
+            nbytes_of_gpu,
+            "(num_nodes num_gpu_per_node) -> num_nodes",
+            num_gpu_per_node=num_gpu_per_node,
+            reduction="sum",
+        )
+
+        nbytes_curr_node = nbytes_of_node[self_node_id]
+        nbytes_cross_node = torch.sum(nbytes_of_node) - nbytes_curr_node
+
+        text += (
+            f"{direction}_nbytes_of_gpu={nbytes_of_gpu.tolist()} "
+            f"{direction}_nbytes_of_node={nbytes_of_node.tolist()} "
+            f"{direction}_nbytes_curr_node={nbytes_curr_node.item()} "
+            f"{direction}_nbytes_cross_node={nbytes_cross_node.item()} "
+        )
+
+    logger.info(f"[ExpertLocationUpdater] {text}")
+
+
+def _get_direction_from_op(op: P2POp):
+    if op.op == torch.distributed.isend:
+        return "isend"
+    if op.op == torch.distributed.irecv:
+        return "irecv"
+    raise NotImplementedError
+
+
+def _group_by(items, keyfunc):
+    ans = defaultdict(list)
+    for item in items:
+        ans[keyfunc(item)].append(item)
+    return dict(ans)

sglang/srt/model_executor/forward_batch_info.py

@@ -31,14 +31,15 @@ from __future__ import annotations
 
 from dataclasses import dataclass
 from enum import IntEnum, auto
-from typing import TYPE_CHECKING, Dict, List, Optional, Union
+from functools import total_ordering
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
 
 import torch
 import triton
 import triton.language as tl
 
 from sglang.srt.layers.rotary_embedding import MRotaryEmbedding
-from sglang.srt.utils import flatten_nested_list, get_compiler_backend
+from sglang.srt.utils import flatten_nested_list, get_compiler_backend, support_triton
 
 if TYPE_CHECKING:
     from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
@@ -117,12 +118,14 @@ class ForwardMode(IntEnum):
         return self == ForwardMode.DECODE or self == ForwardMode.IDLE
 
 
+@total_ordering
 class CaptureHiddenMode(IntEnum):
-    NULL = auto()
-    # Capture hidden states of all tokens.
-    FULL = auto()
+    # Do not capture anything.
+    NULL = 0
     # Capture a hidden state of the last token.
-    LAST = auto()
+    LAST = 1
+    # Capture hidden states of all tokens.
+    FULL = 2
 
     def need_capture(self):
         return self != CaptureHiddenMode.NULL
@@ -133,6 +136,9 @@ class CaptureHiddenMode(IntEnum):
     def is_last(self):
         return self == CaptureHiddenMode.LAST
 
+    def __lt__(self, other):
+        return self.value < other.value
+
 
 @dataclass
 class ForwardBatch:
@@ -218,6 +224,9 @@ class ForwardBatch:
     # For input embeddings
     input_embeds: Optional[torch.tensor] = None
 
+    # For cross-encoder model
+    token_type_ids: Optional[torch.Tensor] = None
+
     # Sampling info
     sampling_info: SamplingBatchInfo = None
 
@@ -239,6 +248,7 @@ class ForwardBatch:
     dp_local_num_tokens: Optional[torch.Tensor] = None  # cached info at runtime
     gathered_buffer: Optional[torch.Tensor] = None
     can_run_dp_cuda_graph: bool = False
+    global_forward_mode: Optional[ForwardMode] = None
 
     # Speculative decoding
     spec_info: Optional[Union[EagleVerifyInput, EagleDraftInput]] = None
@@ -252,18 +262,19 @@ class ForwardBatch:
     # For Qwen2-VL
     mrope_positions: torch.Tensor = None
 
+    # For two-batch overlap
+    tbo_split_seq_index: Optional[int] = None
+    tbo_parent_token_range: Optional[Tuple[int, int]] = None
+    tbo_children: Optional[List["ForwardBatch"]] = None
+
     @classmethod
     def init_new(
         cls,
         batch: ModelWorkerBatch,
         model_runner: ModelRunner,
     ):
-        device = model_runner.device
-        extend_input_logprob_token_ids_gpu = None
-        if batch.extend_input_logprob_token_ids is not None:
-            extend_input_logprob_token_ids_gpu = (
-                batch.extend_input_logprob_token_ids.to(device, non_blocking=True)
-            )
+        from sglang.srt.two_batch_overlap import TboForwardBatchPreparer
+
         ret = cls(
             forward_mode=batch.forward_mode,
             batch_size=len(batch.seq_lens),
@@ -277,10 +288,12 @@ class ForwardBatch:
             encoder_lens_cpu=batch.encoder_lens_cpu,
             encoder_out_cache_loc=batch.encoder_out_cache_loc,
             seq_lens_sum=batch.seq_lens_sum,
+            seq_lens_cpu=batch.seq_lens_cpu,
             return_logprob=batch.return_logprob,
             top_logprobs_nums=batch.top_logprobs_nums,
             token_ids_logprobs=batch.token_ids_logprobs,
             can_run_dp_cuda_graph=batch.can_run_dp_cuda_graph,
+            global_forward_mode=batch.global_forward_mode,
             lora_paths=batch.lora_paths,
             sampling_info=batch.sampling_info,
             req_to_token_pool=model_runner.req_to_token_pool,
@@ -290,11 +303,20 @@ class ForwardBatch:
             spec_info=batch.spec_info,
             capture_hidden_mode=batch.capture_hidden_mode,
             input_embeds=batch.input_embeds,
-            extend_input_logprob_token_ids_gpu=extend_input_logprob_token_ids_gpu,
-            num_token_non_padded=torch.tensor(
-                len(batch.input_ids), dtype=torch.int32
-            ).to(device, non_blocking=True),
+            token_type_ids=batch.token_type_ids,
+            tbo_split_seq_index=batch.tbo_split_seq_index,
         )
+        device = model_runner.device
+
+        if batch.extend_input_logprob_token_ids is not None:
+            ret.extend_input_logprob_token_ids_gpu = (
+                batch.extend_input_logprob_token_ids.to(device, non_blocking=True)
+            )
+
+        if enable_num_token_non_padded(model_runner.server_args):
+            ret.num_token_non_padded = torch.tensor(
+                len(batch.input_ids), dtype=torch.int32
+            ).to(device, non_blocking=True)
 
         # For DP attention
         if batch.global_num_tokens is not None:
@@ -314,8 +336,10 @@ class ForwardBatch:
                 dtype=model_runner.dtype,
                 device=device,
             )
+
         if ret.forward_mode.is_idle():
             ret.positions = torch.empty((0,), device=device)
+            TboForwardBatchPreparer.prepare(ret)
             return ret
 
         # Override the positions with spec_info
@@ -325,10 +349,6 @@ class ForwardBatch:
         ):
             ret.positions = ret.spec_info.positions
 
-        # Get seq_lens_cpu if needed
-        if ret.seq_lens_cpu is None:
-            ret.seq_lens_cpu = batch.seq_lens_cpu
-
         # Init position information
         if ret.forward_mode.is_decode():
             if ret.positions is None:
@@ -340,8 +360,8 @@ class ForwardBatch:
             ret.extend_prefix_lens = torch.tensor(
                 batch.extend_prefix_lens, dtype=torch.int32
             ).to(device, non_blocking=True)
-            if model_runner.server_args.attention_backend != "torch_native":
-                ret.extend_num_tokens = batch.extend_num_tokens
+            ret.extend_num_tokens = batch.extend_num_tokens
+            if support_triton(model_runner.server_args.attention_backend):
                 positions, ret.extend_start_loc = compute_position_triton(
                     ret.extend_prefix_lens,
                     ret.extend_seq_lens,
@@ -364,6 +384,8 @@ class ForwardBatch:
         if model_runner.server_args.lora_paths is not None:
             model_runner.lora_manager.prepare_lora_batch(ret)
 
+        TboForwardBatchPreparer.prepare(ret)
+
         return ret
 
     def merge_mm_inputs(self) -> Optional[MultimodalInputs]:
@@ -588,6 +610,14 @@ class ForwardBatch:
         # Precompute the kv indices for each chunk
         self.prepare_chunked_kv_indices(device)
 
+    @property
+    def can_run_tbo(self):
+        return self.tbo_split_seq_index is not None
+
+
+def enable_num_token_non_padded(server_args):
+    return server_args.enable_ep_moe or server_args.enable_deepep_moe
+
 
 class PPProxyTensors:
     # adapted from https://github.com/vllm-project/vllm/blob/d14e98d924724b284dc5eaf8070d935e214e50c0/vllm/sequence.py#L1103