sglang 0.5.2rc2__py3-none-any.whl → 0.5.3.post1__py3-none-any.whl

sglang/srt/managers/tp_worker.py

@@ -12,38 +12,44 @@
 # 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
 
 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,
     UpdateWeightsFromTensorReqInput,
 )
 from sglang.srt.managers.schedule_batch import ModelWorkerBatch, global_server_args_dict
+from sglang.srt.managers.scheduler import GenerationBatchResult
 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.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 +84,11 @@ 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 +148,8 @@ 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 +173,10 @@ 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)
 
@@ -220,11 +231,11 @@ class TpModelWorker:
     def forward_batch_generation(
         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,
+    ) -> GenerationBatchResult:
+        # 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
@@ -239,23 +250,51 @@ class TpModelWorker:
             logits_output, can_run_cuda_graph = self.model_runner.forward(
                 forward_batch, pp_proxy_tensors=pp_proxy_tensors
             )
-            if launch_done is not None:
-                launch_done.set()
+            batch_result = GenerationBatchResult(
+                logits_output=logits_output,
+                can_run_cuda_graph=can_run_cuda_graph,
+            )
 
-            if skip_sample:
-                next_token_ids = None
+            if is_verify:
+                # Skip sampling and return logits for target forward
+                return batch_result
+
+            if model_worker_batch.delay_sample_launch:
+                batch_result.delay_sample_launch = True
+                batch_result.forward_batch = forward_batch
+                return batch_result
+
+            if model_worker_batch.is_prefill_only:
+                # For prefill-only requests, create dummy token IDs on CPU
+                # The size should match the batch size (number of sequences), not total tokens
+                batch_result.next_token_ids = torch.zeros(
+                    len(model_worker_batch.seq_lens),
+                    dtype=torch.long,
+                    device=model_worker_batch.input_ids.device,
+                )
+                if (
+                    model_worker_batch.return_logprob
+                    and logits_output.next_token_logits is not None
+                ):
+                    # NOTE: Compute logprobs without full sampling
+                    self.model_runner.compute_logprobs_only(
+                        logits_output, model_worker_batch
+                    )
             else:
-                next_token_ids = self.model_runner.sample(
-                    logits_output, model_worker_batch
+                batch_result.next_token_ids = self.model_runner.sample(
+                    logits_output, forward_batch
                 )
 
-            return logits_output, next_token_ids, can_run_cuda_graph
+            return batch_result
         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 GenerationBatchResult(
+                pp_hidden_states_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 +319,37 @@ 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/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
@@ -274,10 +274,15 @@ class SWATokenToKVPoolAllocator(BaseTokenToKVPoolAllocator):
         self.full_to_swa_index_mapping[free_index] = 0
 
     def backup_state(self):
-        raise NotImplementedError
+        return [
+            self.full_attn_allocator.backup_state(),
+            self.swa_attn_allocator.backup_state(),
+        ]
 
     def restore_state(self, state):
-        raise NotImplementedError
+        assert len(state) == 2
+        self.full_attn_allocator.restore_state(state[0])
+        self.swa_attn_allocator.restore_state(state[1])
 
     def clear(self):
         self.swa_attn_allocator.clear()
@@ -294,7 +299,6 @@ 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 +327,6 @@ 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 +378,6 @@ def alloc_decode_kernel(
     last_loc_ptr,
     free_page_ptr,
     out_indices,
-    ret_values,
     bs_upper: tl.constexpr,
     page_size: tl.constexpr,
 ):
@@ -404,10 +400,6 @@ 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 +430,6 @@ 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 +459,9 @@ 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 +490,6 @@ 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 +498,11 @@ 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 +512,7 @@ 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 +530,6 @@ class PagedTokenToKVPoolAllocator(BaseTokenToKVPoolAllocator):
             last_loc,
             self.free_pages,
             out_indices,
-            self.ret_values,
             next_power_of_2(bs),
             self.page_size,
         )
@@ -542,7 +537,11 @@ 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
 

sglang/srt/mem_cache/allocator_ascend.py

@@ -1,13 +1,9 @@
 from __future__ import annotations
 
-from typing import TYPE_CHECKING
-
 import torch
 
 from sglang.srt.mem_cache.allocator import PagedTokenToKVPoolAllocator
-
-if TYPE_CHECKING:
-    from sglang.srt.mem_cache.memory_pool import KVCache
+from sglang.srt.utils import get_num_new_pages
 
 
 def alloc_extend_kernel_ascend(
@@ -69,7 +65,9 @@ class AscendPagedTokenToKVPoolAllocator(PagedTokenToKVPoolAllocator):
     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,
     ):
@@ -79,42 +77,54 @@ class AscendPagedTokenToKVPoolAllocator(PagedTokenToKVPoolAllocator):
             )
 
         num_new_pages = (
-            (
-                (seq_lens + self.page_size - 1) // self.page_size
-                - (prefix_lens + self.page_size - 1) // self.page_size
-            )
-            .sum()
-            .item()
-        )
-        if self.need_sort and num_new_pages > len(self.free_pages):
+            (seq_lens + self.page_size - 1) // self.page_size
+            - (prefix_lens + self.page_size - 1) // self.page_size
+        ).sum()
+        num_new_pages_item = num_new_pages.item()
+        if self.need_sort and num_new_pages_item > len(self.free_pages):
             self.merge_and_sort_free()
 
-        if num_new_pages > len(self.free_pages):
+        if num_new_pages_item > len(self.free_pages):
             return None
 
         out_indices = torch.empty(
-            (extend_num_tokens,), dtype=torch.int32, device=self.device
+            (extend_num_tokens,), dtype=torch.int64, device=self.device
         )
 
-        alloc_extend_kernel_ascend(
-            prefix_lens,
-            seq_lens,
-            last_loc,
-            self.free_pages,
-            out_indices,
-            self.page_size,
-            self.device,
-        )
+        if num_new_pages_item < 200:
+            import sgl_kernel_npu
+
+            torch.ops.npu.alloc_extend(
+                prefix_lens,
+                seq_lens,
+                last_loc,
+                self.free_pages,
+                self.page_size,
+                out_indices,
+                num_new_pages,
+            )
+
+        else:
+            alloc_extend_kernel_ascend(
+                prefix_lens,
+                seq_lens,
+                last_loc,
+                self.free_pages,
+                out_indices,
+                self.page_size,
+                self.device,
+            )
 
         if self.debug_mode:
             assert len(torch.unique(out_indices)) == len(out_indices)
 
-        self.free_pages = self.free_pages[num_new_pages:]
+        self.free_pages = self.free_pages[num_new_pages_item:]
         return out_indices
 
     def alloc_decode(
         self,
         seq_lens: torch.Tensor,
+        seq_lens_cpu: torch.Tensor,
         last_loc: torch.Tensor,
     ):
         if self.debug_mode:
@@ -122,8 +132,11 @@ class AscendPagedTokenToKVPoolAllocator(PagedTokenToKVPoolAllocator):
                 (last_loc + 2) % self.page_size == seq_lens % self.page_size
             )
 
-        need_new_pages = (seq_lens % self.page_size == 1).int()
-        num_new_pages = need_new_pages.sum().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):
             self.merge_and_sort_free()
@@ -131,6 +144,7 @@ class AscendPagedTokenToKVPoolAllocator(PagedTokenToKVPoolAllocator):
         if num_new_pages > len(self.free_pages):
             return None
 
+        need_new_pages = (seq_lens % self.page_size == 1).int()
         end_new_pages = torch.cumsum(need_new_pages, 0)
         start_new_pages = end_new_pages - need_new_pages
         if num_new_pages == 0:

sglang/srt/mem_cache/base_prefix_cache.py

@@ -36,7 +36,7 @@ class BasePrefixCache(ABC):
         pass
 
     @abstractmethod
-    def match_prefix(self, key: List[int], **kwargs) -> MatchResult:
+    def match_prefix(self, key: Any, **kwargs) -> MatchResult:
         pass
 
     @abstractmethod

sglang/srt/mem_cache/chunk_cache.py

@@ -28,6 +28,13 @@ class ChunkCache(BasePrefixCache):
         self.token_to_kv_pool_allocator = token_to_kv_pool_allocator
         self.page_size = page_size
 
+    # NOTE (csy): this is to determine if a cache has prefix matching feature.
+    # Chunk cache always return True to indicate no prefix matching.
+    # TODO (csy): Using a prefix cache trait to replace this
+    @property
+    def disable(self):
+        return True
+
     def reset(self):
         pass
 
@@ -38,7 +45,7 @@ class ChunkCache(BasePrefixCache):
             last_host_node=None,
         )
 
-    def cache_finished_req(self, req: Req):
+    def cache_finished_req(self, req: Req, insert: bool = True):
         kv_indices = self.req_to_token_pool.req_to_token[
             req.req_pool_idx,
             # For decode server: if req.output_ids is empty, we want to free all req.origin_input_ids
@@ -53,7 +60,7 @@ class ChunkCache(BasePrefixCache):
         ]
 
         # `req.prefix_indices` will be used in `PrefillAdder::add_chunked_req` later
-        req.prefix_indices = kv_indices
+        req.prefix_indices = kv_indices.to(dtype=torch.int64, copy=True)
 
     def evict(self, num_tokens: int):
         pass

sglang/srt/mem_cache/evict_policy.py

@@ -0,0 +1,23 @@
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from typing import TYPE_CHECKING, List, Tuple, Union
+
+if TYPE_CHECKING:
+    from sglang.srt.mem_cache.radix_cache import TreeNode
+
+
+class EvictionStrategy(ABC):
+    @abstractmethod
+    def get_priority(self, node: "TreeNode") -> Union[float, Tuple]:
+        pass
+
+
+class LRUStrategy(EvictionStrategy):
+    def get_priority(self, node: "TreeNode") -> float:
+        return node.last_access_time
+
+
+class LFUStrategy(EvictionStrategy):
+    def get_priority(self, node: "TreeNode") -> Tuple[int, float]:
+        return (node.hit_count, node.last_access_time)