sglang 0.5.2rc2__py3-none-any.whl → 0.5.3rc2__py3-none-any.whl

sglang/srt/managers/tp_worker.py

@@ -12,25 +12,24 @@
 # limitations under the License.
 # ==============================================================================
 """A tensor parallel worker."""
+from __future__ import annotations
 
 import logging
 import threading
-from typing import Optional, Tuple, Union
+from typing import TYPE_CHECKING, Optional, Tuple, Union
 
 import torch
 
 from sglang.srt.configs.model_config import ModelConfig
 from sglang.srt.distributed import get_pp_group, get_world_group
-from sglang.srt.hf_transformers_utils import (
-    get_processor,
-    get_tokenizer,
-    get_tokenizer_from_processor,
-)
 from sglang.srt.layers.logits_processor import LogitsProcessorOutput
 from sglang.srt.managers.io_struct import (
+    DestroyWeightsUpdateGroupReqInput,
     GetWeightsByNameReqInput,
+    InitWeightsSendGroupForRemoteInstanceReqInput,
     InitWeightsUpdateGroupReqInput,
     LoadLoRAAdapterReqInput,
+    SendWeightsToRemoteInstanceReqInput,
     UnloadLoRAAdapterReqInput,
     UpdateWeightFromDiskReqInput,
     UpdateWeightsFromDistributedReqInput,
@@ -39,11 +38,23 @@ from sglang.srt.managers.io_struct import (
 from sglang.srt.managers.schedule_batch import ModelWorkerBatch, global_server_args_dict
 from sglang.srt.mem_cache.allocator import BaseTokenToKVPoolAllocator
 from sglang.srt.mem_cache.memory_pool import ReqToTokenPool
-from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
+from sglang.srt.model_executor.forward_batch_info import (
+    ForwardBatch,
+    ForwardBatchOutput,
+    PPProxyTensors,
+)
 from sglang.srt.model_executor.model_runner import ModelRunner
-from sglang.srt.patch_torch import monkey_patch_torch_reductions
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import MultiprocessingSerializer, broadcast_pyobj, set_random_seed
+from sglang.srt.utils.hf_transformers_utils import (
+    get_processor,
+    get_tokenizer,
+    get_tokenizer_from_processor,
+)
+from sglang.srt.utils.patch_torch import monkey_patch_torch_reductions
+
+if TYPE_CHECKING:
+    from sglang.srt.managers.cache_controller import LayerDoneCounter
 
 logger = logging.getLogger(__name__)
 
@@ -78,6 +89,11 @@ class TpModelWorker:
                 if not is_draft_worker
                 else server_args.speculative_draft_model_path
             ),
+            model_revision=(
+                server_args.revision
+                if not is_draft_worker
+                else server_args.speculative_draft_model_revision
+            ),
             is_draft_model=is_draft_worker,
         )
 
@@ -137,8 +153,8 @@ class TpModelWorker:
         assert self.max_running_requests > 0, "max_running_request is zero"
         self.max_queued_requests = server_args.max_queued_requests
         assert (
-            self.max_running_requests > 0
-        ), "max_queued_requests is zero. We need to be at least 1 to schedule a request."
+            self.max_queued_requests is None or self.max_queued_requests >= 1
+        ), "If configured, max_queued_requests must be at least 1 for any work to be scheduled."
         self.max_req_len = min(
             self.model_config.context_len - 1,
             self.max_total_num_tokens - 1,
@@ -162,10 +178,10 @@ class TpModelWorker:
 
         self.hicache_layer_transfer_counter = None
 
-    def register_hicache_layer_transfer_counter(self, counter):
+    def register_hicache_layer_transfer_counter(self, counter: LayerDoneCounter):
         self.hicache_layer_transfer_counter = counter
 
-    def set_hicache_consumer(self, consumer_index):
+    def set_hicache_consumer(self, consumer_index: int):
         if self.hicache_layer_transfer_counter is not None:
             self.hicache_layer_transfer_counter.set_consumer(consumer_index)
 
@@ -221,10 +237,11 @@ class TpModelWorker:
         self,
         model_worker_batch: ModelWorkerBatch,
         launch_done: Optional[threading.Event] = None,
-        skip_sample: bool = False,
-    ) -> Tuple[
-        Union[LogitsProcessorOutput, torch.Tensor], Optional[torch.Tensor], bool
-    ]:
+        is_verify: bool = False,
+    ) -> ForwardBatchOutput:
+        # update the consumer index of hicache to the running batch
+        self.set_hicache_consumer(model_worker_batch.hicache_consumer_index)
+
         forward_batch = ForwardBatch.init_new(model_worker_batch, self.model_runner)
 
         pp_proxy_tensors = None
@@ -242,20 +259,31 @@ class TpModelWorker:
             if launch_done is not None:
                 launch_done.set()
 
-            if skip_sample:
-                next_token_ids = None
-            else:
-                next_token_ids = self.model_runner.sample(
+            skip_sample = is_verify or model_worker_batch.is_prefill_only
+            next_token_ids = None
+
+            if not skip_sample:
+                next_token_ids = self.model_runner.sample(logits_output, forward_batch)
+            elif model_worker_batch.return_logprob and not is_verify:
+                # NOTE: Compute logprobs without full sampling
+                self.model_runner.compute_logprobs_only(
                     logits_output, model_worker_batch
                 )
 
-            return logits_output, next_token_ids, can_run_cuda_graph
+            return ForwardBatchOutput(
+                logits_output=logits_output,
+                next_token_ids=next_token_ids,
+                can_run_cuda_graph=can_run_cuda_graph,
+            )
         else:
             pp_proxy_tensors, can_run_cuda_graph = self.model_runner.forward(
                 forward_batch,
                 pp_proxy_tensors=pp_proxy_tensors,
             )
-            return pp_proxy_tensors.tensors, None, can_run_cuda_graph
+            return ForwardBatchOutput(
+                pp_proxy_tensors=pp_proxy_tensors,
+                can_run_cuda_graph=can_run_cuda_graph,
+            )
 
     def forward_batch_embedding(self, model_worker_batch: ModelWorkerBatch):
         forward_batch = ForwardBatch.init_new(model_worker_batch, self.model_runner)
@@ -280,6 +308,37 @@ class TpModelWorker:
         )
         return success, message
 
+    def destroy_weights_update_group(self, recv_req: DestroyWeightsUpdateGroupReqInput):
+        success, message = self.model_runner.destroy_weights_update_group(
+            recv_req.group_name,
+        )
+        return success, message
+
+    def init_weights_send_group_for_remote_instance(
+        self, recv_req: InitWeightsSendGroupForRemoteInstanceReqInput
+    ):
+        success, message = (
+            self.model_runner.init_weights_send_group_for_remote_instance(
+                recv_req.master_address,
+                recv_req.ports,
+                recv_req.group_rank,
+                recv_req.world_size,
+                recv_req.group_name,
+                recv_req.backend,
+            )
+        )
+        return success, message
+
+    def send_weights_to_remote_instance(
+        self, recv_req: SendWeightsToRemoteInstanceReqInput
+    ):
+        success, message = self.model_runner.send_weights_to_remote_instance(
+            recv_req.master_address,
+            recv_req.ports,
+            recv_req.group_name,
+        )
+        return success, message
+
     def update_weights_from_distributed(
         self, recv_req: UpdateWeightsFromDistributedReqInput
     ):

sglang/srt/managers/tp_worker_overlap_thread.py

@@ -12,42 +12,42 @@
 # limitations under the License.
 # ==============================================================================
 """A tensor parallel worker."""
+from __future__ import annotations
 
 import dataclasses
 import logging
 import signal
 import threading
 from queue import Queue
-from typing import Optional, Tuple
+from typing import TYPE_CHECKING, List, Optional, Tuple
 
 import psutil
 import torch
 
 from sglang.srt.managers.io_struct import (
+    DestroyWeightsUpdateGroupReqInput,
     GetWeightsByNameReqInput,
+    InitWeightsSendGroupForRemoteInstanceReqInput,
     InitWeightsUpdateGroupReqInput,
     LoadLoRAAdapterReqInput,
+    SendWeightsToRemoteInstanceReqInput,
     UnloadLoRAAdapterReqInput,
     UpdateWeightFromDiskReqInput,
     UpdateWeightsFromDistributedReqInput,
     UpdateWeightsFromTensorReqInput,
 )
+from sglang.srt.managers.overlap_utils import FutureMap
 from sglang.srt.managers.schedule_batch import ModelWorkerBatch
 from sglang.srt.managers.tp_worker import TpModelWorker
+from sglang.srt.model_executor.forward_batch_info import ForwardBatchOutput
 from sglang.srt.server_args import ServerArgs
-from sglang.srt.utils import DynamicGradMode, get_compiler_backend
+from sglang.srt.utils import DynamicGradMode
 from sglang.utils import get_exception_traceback
 
-logger = logging.getLogger(__name__)
-
+if TYPE_CHECKING:
+    from sglang.srt.managers.cache_controller import LayerDoneCounter
 
-@torch.compile(dynamic=True, backend=get_compiler_backend())
-def resolve_future_token_ids(input_ids, future_token_ids_map):
-    input_ids[:] = torch.where(
-        input_ids < 0,
-        future_token_ids_map[torch.clamp(-input_ids, min=0)],
-        input_ids,
-    )
+logger = logging.getLogger(__name__)
 
 
 class TpModelWorkerClient:
@@ -72,14 +72,10 @@ class TpModelWorkerClient:
         self.gpu_id = gpu_id
 
         # Init future mappings
-        self.future_token_ids_ct = 0
-        self.future_token_ids_limit = self.max_running_requests * 3
-        self.future_token_ids_map = torch.empty(
-            (self.max_running_requests * 5,), dtype=torch.int64, device=self.device
-        )
+        self.future_map = FutureMap(self.max_running_requests, self.device)
 
         # Launch threads
-        self.input_queue = Queue()
+        self.input_queue = Queue[Tuple[ModelWorkerBatch, int, torch.Event]]()
         self.output_queue = Queue()
         self.forward_stream = torch.get_device_module(self.device).Stream()
         self.forward_thread = threading.Thread(
@@ -93,13 +89,9 @@ class TpModelWorkerClient:
 
         self.hicache_layer_transfer_counter = None
 
-    def register_hicache_layer_transfer_counter(self, counter):
+    def register_hicache_layer_transfer_counter(self, counter: LayerDoneCounter):
         self.hicache_layer_transfer_counter = counter
 
-    def set_hicache_consumer(self, consumer_index):
-        if self.hicache_layer_transfer_counter is not None:
-            self.hicache_layer_transfer_counter.set_consumer(consumer_index)
-
     def get_worker_info(self):
         return self.worker.get_worker_info()
 
@@ -147,10 +139,10 @@ class TpModelWorkerClient:
     @DynamicGradMode()
     def forward_thread_func_(self):
         batch_pt = 0
-        batch_lists = [None] * 2
+        batch_lists: List = [None] * 2
 
         while True:
-            model_worker_batch, future_token_ids_ct, sync_event = self.input_queue.get()
+            model_worker_batch, future_map_ct, sync_event = self.input_queue.get()
             if not model_worker_batch:
                 break
 
@@ -166,29 +158,35 @@ class TpModelWorkerClient:
             copy_done = torch.get_device_module(self.device).Event()
 
             # Resolve future tokens in the input
-            input_ids = model_worker_batch.input_ids
-            resolve_future_token_ids(input_ids, self.future_token_ids_map)
+            self.future_map.resolve_future(model_worker_batch)
 
-            # update the consumer index of hicache to the running batch
-            self.set_hicache_consumer(model_worker_batch.hicache_consumer_index)
             # Run forward
+            forward_batch_output = self.worker.forward_batch_generation(
+                model_worker_batch,
+                model_worker_batch.launch_done,
+            )
+
             logits_output, next_token_ids, can_run_cuda_graph = (
-                self.worker.forward_batch_generation(
-                    model_worker_batch, model_worker_batch.launch_done
-                )
+                forward_batch_output.logits_output,
+                forward_batch_output.next_token_ids,
+                forward_batch_output.can_run_cuda_graph,
             )
 
             # Update the future token ids map
             bs = len(model_worker_batch.seq_lens)
-            self.future_token_ids_map[
-                future_token_ids_ct + 1 : future_token_ids_ct + bs + 1
-            ] = next_token_ids
+            if model_worker_batch.is_prefill_only:
+                # For prefill-only requests, create dummy token IDs on CPU
+                next_token_ids = torch.zeros(bs, dtype=torch.long)
+
+            # store the future indices into future map
+            self.future_map.store_to_map(future_map_ct, bs, next_token_ids)
 
             # Copy results to the CPU
             if model_worker_batch.return_logprob:
-                logits_output.next_token_logprobs = (
-                    logits_output.next_token_logprobs.to("cpu", non_blocking=True)
-                )
+                if logits_output.next_token_logprobs is not None:
+                    logits_output.next_token_logprobs = (
+                        logits_output.next_token_logprobs.to("cpu", non_blocking=True)
+                    )
                 if logits_output.input_token_logprobs is not None:
                     logits_output.input_token_logprobs = (
                         logits_output.input_token_logprobs.to("cpu", non_blocking=True)
@@ -197,7 +195,9 @@ class TpModelWorkerClient:
                 logits_output.hidden_states = logits_output.hidden_states.to(
                     "cpu", non_blocking=True
                 )
-            next_token_ids = next_token_ids.to("cpu", non_blocking=True)
+            # Only copy to CPU if not already on CPU
+            if next_token_ids.device.type != "cpu":
+                next_token_ids = next_token_ids.to("cpu", non_blocking=True)
             copy_done.record()
 
             self.output_queue.put(
@@ -221,16 +221,16 @@ class TpModelWorkerClient:
             logits_output.next_token_logprobs = (
                 logits_output.next_token_logprobs.tolist()
             )
-            if logits_output.input_token_logprobs is not None:
-                logits_output.input_token_logprobs = tuple(
-                    logits_output.input_token_logprobs.tolist()
-                )
+        if logits_output.input_token_logprobs is not None:
+            logits_output.input_token_logprobs = tuple(
+                logits_output.input_token_logprobs.tolist()
+            )
         next_token_ids = next_token_ids.tolist()
         return logits_output, next_token_ids, can_run_cuda_graph
 
     def forward_batch_generation(
         self, model_worker_batch: ModelWorkerBatch
-    ) -> Tuple[None, torch.Tensor, bool]:
+    ) -> ForwardBatchOutput:
         # Create a new copy of sampling_info because it will be updated in-place by the scheduler for the next batch.
         sampling_info = model_worker_batch.sampling_info
         sampling_info.update_penalties()
@@ -245,21 +245,18 @@ class TpModelWorkerClient:
         sync_event.record(self.scheduler_stream)
 
         # Push a new batch to the queue
-        self.input_queue.put((model_worker_batch, self.future_token_ids_ct, sync_event))
-
-        # Allocate output future objects
         bs = len(model_worker_batch.seq_lens)
-        future_next_token_ids = torch.arange(
-            -(self.future_token_ids_ct + 1),
-            -(self.future_token_ids_ct + 1 + bs),
-            -1,
-            dtype=torch.int64,
-            device=self.device,
+        cur_future_map_ct = self.future_map.update_ct(bs)
+        self.input_queue.put((model_worker_batch, cur_future_map_ct, sync_event))
+
+        # get this forward batch's future token ids
+        future_next_token_ids = self.future_map.update_next_future(
+            cur_future_map_ct, bs
+        )
+        return ForwardBatchOutput(
+            next_token_ids=future_next_token_ids,
+            can_run_cuda_graph=False,
         )
-        self.future_token_ids_ct = (
-            self.future_token_ids_ct + bs
-        ) % self.future_token_ids_limit
-        return None, future_next_token_ids, False
 
     def update_weights_from_disk(self, recv_req: UpdateWeightFromDiskReqInput):
         success, message = self.worker.update_weights_from_disk(recv_req)
@@ -269,6 +266,24 @@ class TpModelWorkerClient:
         success, message = self.worker.init_weights_update_group(recv_req)
         return success, message
 
+    def destroy_weights_update_group(self, recv_req: DestroyWeightsUpdateGroupReqInput):
+        success, message = self.worker.destroy_weights_update_group(recv_req)
+        return success, message
+
+    def init_weights_send_group_for_remote_instance(
+        self, recv_req: InitWeightsSendGroupForRemoteInstanceReqInput
+    ):
+        success, message = self.worker.init_weights_send_group_for_remote_instance(
+            recv_req
+        )
+        return success, message
+
+    def send_weights_to_remote_instance(
+        self, recv_req: SendWeightsToRemoteInstanceReqInput
+    ):
+        success, message = self.worker.send_weights_to_remote_instance(recv_req)
+        return success, message
+
     def update_weights_from_distributed(
         self, recv_req: UpdateWeightsFromDistributedReqInput
     ):

sglang/srt/managers/utils.py

@@ -2,11 +2,10 @@ from __future__ import annotations
 
 import logging
 import multiprocessing as mp
-from http import HTTPStatus
 from typing import TYPE_CHECKING, Dict, List, Optional
 
 from sglang.srt.layers.logits_processor import LogitsProcessorOutput
-from sglang.srt.managers.schedule_batch import FINISH_ABORT, Req
+from sglang.srt.managers.schedule_batch import Req
 from sglang.srt.model_executor.forward_batch_info import PPProxyTensors
 
 if TYPE_CHECKING:
@@ -97,46 +96,3 @@ def get_logprob_from_pp_outputs(
     ]
 
     return logits_output, extend_input_len_per_req, extend_logprob_start_len_per_req
-
-
-class DPBalanceMeta:
-    """
-    This class will be use in scheduler and dp controller
-    """
-
-    def __init__(self, num_workers: int):
-        self.num_workers = num_workers
-        self._manager = mp.Manager()
-        self.mutex = self._manager.Lock()
-
-        init_local_tokens = [0] * self.num_workers
-        init_onfly_info = [self._manager.dict() for _ in range(self.num_workers)]
-
-        self.shared_state = self._manager.Namespace()
-        self.shared_state.local_tokens = self._manager.list(init_local_tokens)
-        self.shared_state.onfly_info = self._manager.list(init_onfly_info)
-
-    def destructor(self):
-        # we must destructor this class manually
-        self._manager.shutdown()
-
-    def get_shared_onfly(self) -> List[Dict[int, int]]:
-        return [dict(d) for d in self.shared_state.onfly_info]
-
-    def set_shared_onfly_info(self, data: List[Dict[int, int]]):
-        self.shared_state.onfly_info = data
-
-    def get_shared_local_tokens(self) -> List[int]:
-        return list(self.shared_state.local_tokens)
-
-    def set_shared_local_tokens(self, data: List[int]):
-        self.shared_state.local_tokens = data
-
-    def __getstate__(self):
-        state = self.__dict__.copy()
-        del state["_manager"]
-        return state
-
-    def __setstate__(self, state):
-        self.__dict__.update(state)
-        self._manager = None

sglang/srt/mem_cache/allocator.py

@@ -27,7 +27,7 @@ import triton
 import triton.language as tl
 
 from sglang.srt.mem_cache.memory_pool import SWAKVPool
-from sglang.srt.utils import get_bool_env_var, next_power_of_2
+from sglang.srt.utils import get_bool_env_var, get_num_new_pages, next_power_of_2
 
 if TYPE_CHECKING:
     from sglang.srt.mem_cache.memory_pool import KVCache
@@ -294,7 +294,6 @@ def alloc_extend_kernel(
     last_loc_ptr,
     free_page_ptr,
     out_indices,
-    ret_values,
     bs_upper: tl.constexpr,
     page_size: tl.constexpr,
     max_num_extend_tokens: tl.constexpr,
@@ -323,13 +322,6 @@ def alloc_extend_kernel(
     sum_num_new_pages = tl.sum(num_new_pages)
     new_page_start_loc = sum_num_new_pages - num_page_start_loc_self
 
-    # Return value
-    if pid == tl.num_programs(0) - 1:
-        merged_value = (sum_num_new_pages.to(tl.int64)) << 32 | sum_extend_lens.to(
-            tl.int64
-        )
-        tl.store(ret_values, merged_value)
-
     # Part 1: fill the old partial page
     last_loc = tl.load(last_loc_ptr + pid)
     num_part1 = (
@@ -381,7 +373,6 @@ def alloc_decode_kernel(
     last_loc_ptr,
     free_page_ptr,
     out_indices,
-    ret_values,
     bs_upper: tl.constexpr,
     page_size: tl.constexpr,
 ):
@@ -404,10 +395,6 @@ def alloc_decode_kernel(
     sum_num_new_pages = tl.sum(num_new_pages)
     new_page_start_loc = sum_num_new_pages - num_page_start_loc_self
 
-    # Return value
-    if pid == tl.num_programs(0) - 1:
-        tl.store(ret_values, sum_num_new_pages)
-
     if num_page_start_loc_self == 0:
         last_loc = tl.load(last_loc_ptr + pid)
         tl.store(out_indices + pid, last_loc + 1)
@@ -438,7 +425,6 @@ class PagedTokenToKVPoolAllocator(BaseTokenToKVPoolAllocator):
         super().__init__(size, page_size, dtype, device, kvcache, need_sort)
         self.num_pages = size // page_size
         self.debug_mode = get_bool_env_var("SGLANG_DEBUG_MEMORY_POOL")
-        self.ret_values = torch.empty((), dtype=torch.int64, device=self.device)
         self.seen_max_num_extend_tokens_next_power_of_2 = 1
         self.clear()
 
@@ -468,7 +454,9 @@ class PagedTokenToKVPoolAllocator(BaseTokenToKVPoolAllocator):
     def alloc_extend(
         self,
         prefix_lens: torch.Tensor,
+        prefix_lens_cpu: torch.Tensor,
         seq_lens: torch.Tensor,
+        seq_lens_cpu: torch.Tensor,
         last_loc: torch.Tensor,
         extend_num_tokens: int,
     ):
@@ -497,7 +485,6 @@ class PagedTokenToKVPoolAllocator(BaseTokenToKVPoolAllocator):
             last_loc,
             self.free_pages,
             out_indices,
-            self.ret_values,
             next_power_of_2(bs),
             self.page_size,
             self.seen_max_num_extend_tokens_next_power_of_2,
@@ -506,8 +493,11 @@ class PagedTokenToKVPoolAllocator(BaseTokenToKVPoolAllocator):
         if self.debug_mode:
             assert len(torch.unique(out_indices)) == len(out_indices)
 
-        merged_value = self.ret_values.item()
-        num_new_pages = merged_value >> 32
+        num_new_pages = get_num_new_pages(
+            seq_lens=seq_lens_cpu,
+            page_size=self.page_size,
+            prefix_lens=prefix_lens_cpu,
+        )
         if num_new_pages > len(self.free_pages):
             return None
 
@@ -517,6 +507,7 @@ class PagedTokenToKVPoolAllocator(BaseTokenToKVPoolAllocator):
     def alloc_decode(
         self,
         seq_lens: torch.Tensor,
+        seq_lens_cpu: torch.Tensor,
         last_loc: torch.Tensor,
     ):
         if self.debug_mode:
@@ -534,7 +525,6 @@ class PagedTokenToKVPoolAllocator(BaseTokenToKVPoolAllocator):
             last_loc,
             self.free_pages,
             out_indices,
-            self.ret_values,
             next_power_of_2(bs),
             self.page_size,
         )
@@ -542,7 +532,11 @@ class PagedTokenToKVPoolAllocator(BaseTokenToKVPoolAllocator):
         if self.debug_mode:
             assert len(torch.unique(out_indices)) == len(out_indices)
 
-        num_new_pages = self.ret_values.item()
+        num_new_pages = get_num_new_pages(
+            seq_lens=seq_lens_cpu,
+            page_size=self.page_size,
+            decode=True,
+        )
         if num_new_pages > len(self.free_pages):
             return None