sglang 0.4.1.post3__py3-none-any.whl → 0.4.1.post5__py3-none-any.whl

sglang/srt/layers/sampler.py

@@ -1,5 +1,5 @@
 import logging
-from typing import Union
+from typing import List
 
 import torch
 from torch import nn
@@ -28,13 +28,12 @@ class Sampler(nn.Module):
 
     def forward(
         self,
-        logits: Union[torch.Tensor, LogitsProcessorOutput],
+        logits_output: LogitsProcessorOutput,
         sampling_info: SamplingBatchInfo,
+        return_logprob: bool,
+        top_logprobs_nums: List[int],
     ):
-        if isinstance(logits, LogitsProcessorOutput):
-            logits = logits.next_token_logits
-
-        logits = logits.contiguous()
+        logits = logits_output.next_token_logits
 
         if self.use_nan_detectioin and torch.any(torch.isnan(logits)):
             logger.warning("Detected errors during sampling! NaN in the logits.")
@@ -47,6 +46,8 @@ class Sampler(nn.Module):
         if sampling_info.is_all_greedy:
             # Use torch.argmax if all requests use greedy sampling
             batch_next_token_ids = torch.argmax(logits, -1)
+            if return_logprob:
+                logprobs = torch.nn.functional.log_softmax(logits, dim=-1)
         else:
             # Post process logits
             logits.div_(sampling_info.temperatures)
@@ -54,6 +55,14 @@ class Sampler(nn.Module):
             del logits
 
             if global_server_args_dict["sampling_backend"] == "flashinfer":
+                if return_logprob:
+                    # NOTE: the top_p_renorm_prob from flashinfer has numerical problems,
+                    # https://github.com/flashinfer-ai/flashinfer/issues/708
+                    # so we use the torch implementation.
+                    logprobs = torch.log(
+                        top_p_normalize_probs_torch(probs, sampling_info.top_ps)
+                    )
+
                 max_top_k_round, batch_size = 32, probs.shape[0]
                 uniform_samples = torch.rand(
                     (max_top_k_round, batch_size), device=probs.device
@@ -76,6 +85,7 @@ class Sampler(nn.Module):
                 if self.use_nan_detectioin and not torch.all(success):
                     logger.warning("Detected errors during sampling!")
                     batch_next_token_ids = torch.zeros_like(batch_next_token_ids)
+
             elif global_server_args_dict["sampling_backend"] == "pytorch":
                 # A slower fallback implementation with torch native operations.
                 batch_next_token_ids = top_k_top_p_min_p_sampling_from_probs_torch(
@@ -85,12 +95,31 @@ class Sampler(nn.Module):
                     sampling_info.min_ps,
                     sampling_info.need_min_p_sampling,
                 )
+                if return_logprob:
+                    logprobs = torch.log(
+                        top_p_normalize_probs_torch(probs, sampling_info.top_ps)
+                    )
             else:
                 raise ValueError(
                     f"Invalid sampling backend: {global_server_args_dict['sampling_backend']}"
                 )
 
-        return batch_next_token_ids.to(torch.int32)
+        batch_next_token_ids = batch_next_token_ids.to(torch.int32)
+
+        # Attach logprobs to logits_output (in-place modification)
+        if return_logprob:
+            if any(x > 0 for x in top_logprobs_nums):
+                (
+                    logits_output.next_token_top_logprobs_val,
+                    logits_output.next_token_top_logprobs_idx,
+                ) = get_top_logprobs(logprobs, top_logprobs_nums)
+
+            logits_output.next_token_logprobs = logprobs[
+                torch.arange(len(batch_next_token_ids), device=sampling_info.device),
+                batch_next_token_ids,
+            ]
+
+        return batch_next_token_ids
 
 
 def top_k_top_p_min_p_sampling_from_probs_torch(
@@ -120,20 +149,27 @@ def top_k_top_p_min_p_sampling_from_probs_torch(
     return batch_next_token_ids
 
 
-def top_p_normalize_probs(
+def top_p_normalize_probs_torch(
     probs: torch.Tensor,
     top_ps: torch.Tensor,
 ):
-    if global_server_args_dict["sampling_backend"] == "flashinfer":
-        return top_p_renorm_prob(probs, top_ps)
-    elif global_server_args_dict["sampling_backend"] == "pytorch":
-        # See also top_k_top_p_min_p_sampling_from_probs_torch
-        probs_sort, probs_idx = probs.sort(dim=-1, descending=True)
-        probs_sum = torch.cumsum(probs_sort, dim=-1)
-        probs_sort[(probs_sum - probs_sort) > top_ps.view(-1, 1)] = 0.0
-        probs_sort.div_(probs_sort.sum(dim=-1, keepdim=True))
-        return torch.zeros_like(probs_sort).scatter_(-1, probs_idx, probs_sort)
-    else:
-        raise ValueError(
-            f"Invalid sampling backend: {global_server_args_dict['sampling_backend']}"
-        )
+    # See also top_k_top_p_min_p_sampling_from_probs_torch
+    probs_sort, probs_idx = probs.sort(dim=-1, descending=True)
+    probs_sum = torch.cumsum(probs_sort, dim=-1)
+    probs_sort[(probs_sum - probs_sort) > top_ps.view(-1, 1)] = 0.0
+    probs_sort.div_(probs_sort.sum(dim=-1, keepdim=True))
+    return torch.zeros_like(probs_sort).scatter_(-1, probs_idx, probs_sort)
+
+
+def get_top_logprobs(logprobs: torch.Tensor, top_logprobs_nums: List[int]):
+    max_k = max(top_logprobs_nums)
+    ret = logprobs.topk(max_k, dim=1)
+    values = ret.values.tolist()
+    indices = ret.indices.tolist()
+
+    output_top_logprobs_val = []
+    output_top_logprobs_idx = []
+    for i, k in enumerate(top_logprobs_nums):
+        output_top_logprobs_val.append(values[i][:k])
+        output_top_logprobs_idx.append(indices[i][:k])
+    return output_top_logprobs_val, output_top_logprobs_idx

sglang/srt/layers/torchao_utils.py

@@ -11,6 +11,22 @@ import torch
 logger = logging.getLogger(__name__)
 
 
+def get_gemlite_cache_path() -> str:
+    return f"/tmp/{pwd.getpwuid(os.getuid()).pw_gecos}_gemlite.json"
+
+
+def save_gemlite_cache(print_error: bool = False) -> bool:
+    try:
+        from gemlite.core import GemLiteLinearTriton
+
+        GemLiteLinearTriton.cache_config(get_gemlite_cache_path())
+    except Exception:
+        if print_error:
+            logger.error("Failed to save the GemLite cache.")
+        return False
+    return True
+
+
 def apply_torchao_config_to_model(
     model: torch.nn.Module, torchao_config: str, filter_fn=None
 ):
@@ -74,9 +90,7 @@ def apply_torchao_config_to_model(
         )
 
         # try to load gemlite kernel config
-        GemLiteLinearTriton.load_config(
-            f"/tmp/{pwd.getpwuid(os.getuid()).pw_gecos}_gemlite.json"
-        )
+        GemLiteLinearTriton.load_config(get_gemlite_cache_path())
 
     elif "fp8wo" in torchao_config:
         # this requires newer hardware

sglang/srt/layers/vocab_parallel_embedding.py

@@ -12,8 +12,8 @@ from vllm.distributed import (
     get_tensor_model_parallel_world_size,
     tensor_model_parallel_all_reduce,
 )
-from vllm.model_executor.parameter import BasevLLMParameter
 
+from sglang.srt.layers.parameter import BasevLLMParameter
 from sglang.srt.layers.quantization.base_config import (
     QuantizationConfig,
     QuantizeMethodBase,

sglang/srt/managers/cache_controller.py

@@ -0,0 +1,307 @@
+from __future__ import annotations
+
+"""
+Copyright 2023-2025 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.
+"""
+
+import logging
+import threading
+from queue import PriorityQueue, Queue
+from typing import Optional
+
+import torch
+
+from sglang.srt.mem_cache.memory_pool import MHATokenToKVPool, MLATokenToKVPoolHost
+
+logger = logging.getLogger(__name__)
+
+
+class CacheOperation:
+
+    counter = 0
+
+    def __init__(
+        self,
+        host_indices: torch.Tensor,
+        device_indices: torch.Tensor,
+        node_id: int,
+        priority: Optional[int] = None,
+    ):
+        self.host_indices = host_indices
+        self.device_indices = device_indices
+        self.node_ids = [node_id]
+        self.data = None
+
+        self.id = CacheOperation.counter
+        CacheOperation.counter += 1
+        # default priority is the order of creation
+        self.priority = priority if priority is not None else self.id
+
+    def merge(self, other: "CacheOperation") -> None:
+        # multiple operations can be merged into a single operation for batch processing
+        self.host_indices = torch.cat([self.host_indices, other.host_indices])
+        self.device_indices = torch.cat([self.device_indices, other.device_indices])
+        self.priority = min(self.priority, other.priority)
+        self.node_ids.extend(other.node_ids)
+
+    def __lt__(self, other: "CacheOperation"):
+        return self.priority < other.priority
+
+
+class TransferBuffer:
+    """
+    Overlapping buffer preparation and transfer operations to improve throughput.
+    """
+
+    def __init__(self, buffer_count: int = 3, max_buffer_size: int = 1000) -> None:
+        self.buffers = Queue(maxsize=buffer_count)
+        # todo: adjust the buffer size based on throughput profile of the system
+        self.max_buffer_size = max_buffer_size
+
+    def full(self) -> bool:
+        return self.buffers.full()
+
+    def empty(self) -> bool:
+        return self.buffers.empty()
+
+    def put(self, item, block=True) -> None:
+        self.buffers.put(item, block=block)
+
+    def get(self, block=True) -> Optional[CacheOperation]:
+        try:
+            return self.buffers.get(block=block)
+        except Exception as e:
+            logger.error(e)
+
+
+class HiCacheController:
+
+    def __init__(
+        self,
+        mem_pool_device: MHATokenToKVPool,
+        mem_pool_host: MLATokenToKVPoolHost,
+        write_policy: str = "write_through_selective",
+    ):
+
+        self.mem_pool_device = mem_pool_device
+        self.mem_pool_host = mem_pool_host
+        self.write_policy = write_policy
+
+        if write_policy not in [
+            "write_through",
+            "write_through_selective",
+            "write_back",
+        ]:
+            raise ValueError(f"Invalid write policy: {write_policy}")
+
+        self.write_queue = PriorityQueue()
+        self.load_queue = PriorityQueue()
+
+        self.ack_write_queue = Queue()
+        self.ack_load_queue = Queue()
+
+        self.write_buffer = TransferBuffer()
+        self.load_buffer = TransferBuffer()
+
+        self.write_stream = torch.cuda.Stream()
+        self.load_stream = torch.cuda.Stream()
+
+        self.write_thread = threading.Thread(
+            target=self.write_thread_func_buffer, daemon=True
+        )
+        self.load_thread = threading.Thread(
+            target=self.load_thread_func_buffer, daemon=True
+        )
+        self.write_thread.start()
+        self.load_thread.start()
+
+    def write(
+        self,
+        device_indices: torch.Tensor,
+        priority: Optional[int] = None,
+        node_id: int = 0,
+    ) -> Optional[torch.Tensor]:
+        """
+        Back up KV caches from device memory to host memory.
+        """
+        host_indices = self.mem_pool_host.alloc(len(device_indices))
+        if host_indices is None:
+            return None
+        self.write_queue.put(
+            CacheOperation(host_indices, device_indices, node_id, priority)
+        )
+        self.mem_pool_host.protect_write(host_indices)
+        return host_indices
+
+    def load(
+        self,
+        host_indices: torch.Tensor,
+        priority: Optional[int] = None,
+        node_id: int = 0,
+    ) -> Optional[torch.Tensor]:
+        """
+        Load KV caches from host memory to device memory.
+        """
+        device_indices = self.mem_pool_device.alloc(len(host_indices))
+        if device_indices is None:
+            return None
+        self.load_queue.put(
+            CacheOperation(host_indices, device_indices, node_id, priority)
+        )
+        self.mem_pool_host.protect_load(host_indices)
+        return device_indices
+
+    def write_thread_func_direct(self):
+        """
+        Directly write through KV caches to host memory without buffering.
+        """
+        with torch.cuda.stream(self.write_stream):
+            while True:
+                try:
+                    operation = self.write_queue.get(block=True)
+                    operation.data = self.mem_pool_device.get_flat_data(
+                        operation.device_indices
+                    )
+                    self.mem_pool_host.transfer(operation.host_indices, operation.data)
+                    self.mem_pool_host.complete_io(operation.host_indices)
+                    for node_id in operation.node_ids:
+                        self.ack_write_queue.put(node_id)
+                except Exception as e:
+                    logger.error(e)
+
+    def load_thread_func_direct(self):
+        """
+        Directly load KV caches from host memory to device memory without buffering.
+        """
+        with torch.cuda.stream(self.load_stream):
+            while True:
+                try:
+                    operation = self.load_queue.get(block=True)
+                    operation.data = self.mem_pool_host.get_flat_data(
+                        operation.host_indices
+                    )
+                    self.mem_pool_device.transfer(
+                        operation.device_indices, operation.data
+                    )
+                    self.mem_pool_host.complete_io(operation.host_indices)
+                    for node_id in operation.node_ids:
+                        self.ack_load_queue.put(node_id)
+                except Exception as e:
+                    logger.error(e)
+
+    def write_aux_func(self, no_wait=False):
+        """
+        Auxiliary function to prepare the buffer for write operations.
+        """
+        buffer = None
+        while True:
+            try:
+                operation = self.write_queue.get(block=True)
+                if buffer is None:
+                    buffer = operation
+                else:
+                    buffer.merge(operation)
+                if (
+                    no_wait
+                    or len(buffer.host_indices) >= self.write_buffer.max_buffer_size
+                    or self.write_queue.empty()
+                    or self.write_buffer.empty()
+                ):
+                    assert (
+                        buffer.device_indices.is_cuda
+                    ), "Device indices should be on GPU"
+                    buffer.data = self.mem_pool_device.get_flat_data(
+                        buffer.device_indices
+                    ).contiguous()
+                    self.write_buffer.put(buffer, block=True)
+                    buffer = None
+            except Exception as e:
+                logger.error(e)
+
+    def load_aux_func(self):
+        """
+        Auxiliary function to prepare the buffer for load operations.
+        """
+        buffer = None
+        while True:
+            try:
+                operation = self.load_queue.get(block=True)
+                if buffer is None:
+                    buffer = operation
+                else:
+                    buffer.merge(operation)
+                if (
+                    len(buffer.host_indices) >= self.load_buffer.max_buffer_size
+                    or self.load_queue.empty()
+                    or self.load_buffer.empty()
+                ):
+                    buffer.data = (
+                        self.mem_pool_host.get_flat_data(buffer.host_indices)
+                        .contiguous()
+                        .pin_memory()
+                    )
+                    self.load_buffer.put(buffer, block=True)
+                    buffer = None
+            except Exception as e:
+                logger.error(e)
+
+    def write_thread_func_buffer(self):
+        aux_thread = threading.Thread(target=self.write_aux_func, daemon=True)
+        aux_thread.start()
+        with torch.cuda.stream(self.write_stream):
+            while True:
+                operation = self.write_buffer.get()
+                if operation is None:
+                    continue
+                self.mem_pool_host.transfer(operation.host_indices, operation.data)
+                self.mem_pool_host.complete_io(operation.host_indices)
+                for node_id in operation.node_ids:
+                    self.ack_write_queue.put(node_id)
+
+    def load_thread_func_buffer(self):
+        aux_thread = threading.Thread(target=self.load_aux_func, daemon=True)
+        aux_thread.start()
+        with torch.cuda.stream(self.load_stream):
+            while True:
+                operation = self.load_buffer.get()
+                if operation is None:
+                    continue
+                self.mem_pool_device.transfer(operation.device_indices, operation.data)
+                self.mem_pool_host.complete_io(operation.host_indices)
+                for node_id in operation.node_ids:
+                    self.ack_load_queue.put(node_id)
+
+    def evict_device(
+        self, device_indices: torch.Tensor, host_indices: torch.Tensor
+    ) -> int:
+        if self.mem_pool_host.is_synced(host_indices):
+            self.mem_pool_device.free(device_indices)
+            self.mem_pool_host.update_backup(host_indices)
+            return len(device_indices)
+        else:
+            raise ValueError(
+                f"Inconsistent states: {self.mem_pool_host.get_state(host_indices)}"
+            )
+
+    def evict_host(self, host_indices: torch.Tensor, backup_only: bool = True) -> int:
+        if not backup_only:
+            raise ValueError("Other eviction policies are not supported yet.")
+
+        if self.mem_pool_host.is_backup(host_indices):
+            self.mem_pool_host.free(host_indices)
+            return len(host_indices)
+        else:
+            raise ValueError(
+                f"Inconsistent states: {self.mem_pool_host.get_state(host_indices)}"
+            )

sglang/srt/managers/data_parallel_controller.py

@@ -20,6 +20,7 @@ import threading
 from enum import Enum, auto
 
 import psutil
+import setproctitle
 import zmq
 
 from sglang.srt.managers.io_struct import (
@@ -230,6 +231,7 @@ def run_data_parallel_controller_process(
     port_args: PortArgs,
     pipe_writer,
 ):
+    setproctitle.setproctitle("sglang::data_parallel_controller")
     configure_logger(server_args)
     parent_process = psutil.Process().parent()
 

sglang/srt/managers/io_struct.py

@@ -426,8 +426,7 @@ class UpdateWeightsFromDistributedReqOutput:
 
 @dataclass
 class UpdateWeightsFromTensorReqInput:
-    name: str
-    tensor: torch.Tensor
+    serialized_named_tensors: bytes  # indeed Dict[str, torch.Tensor]
 
 
 @dataclass

sglang/srt/managers/schedule_batch.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 # 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.
@@ -29,7 +31,7 @@ ScheduleBatch -> ModelWorkerBatch -> ForwardBatch
 
 import dataclasses
 import logging
-from typing import List, Optional, Set, Tuple, Union
+from typing import TYPE_CHECKING, List, Optional, Set, Tuple, Union
 
 import numpy as np
 import torch
@@ -42,11 +44,15 @@ from sglang.srt.constrained.base_grammar_backend import BaseGrammarObject
 from sglang.srt.mem_cache.base_prefix_cache import BasePrefixCache
 from sglang.srt.mem_cache.chunk_cache import ChunkCache
 from sglang.srt.mem_cache.memory_pool import BaseTokenToKVPool, ReqToTokenPool
-from sglang.srt.model_executor.forward_batch_info import ForwardMode
+from sglang.srt.model_executor.forward_batch_info import CaptureHiddenMode, ForwardMode
 from sglang.srt.sampling.sampling_batch_info import SamplingBatchInfo
 from sglang.srt.sampling.sampling_params import SamplingParams
 from sglang.srt.server_args import ServerArgs
 
+if TYPE_CHECKING:
+    from sglang.srt.speculative.spec_info import SpecInfo, SpeculativeAlgorithm
+
+
 INIT_INCREMENTAL_DETOKENIZATION_OFFSET = 5
 
 # Put some global args for easy access
@@ -565,9 +571,13 @@ class ScheduleBatch:
     # Has grammar
     has_grammar: bool = False
 
-    # device
+    # Device
     device: str = "cuda"
 
+    # Speculative decoding
+    spec_algorithm: SpeculativeAlgorithm = None
+    spec_info: Optional[SpecInfo] = None
+
     @classmethod
     def init_new(
         cls,
@@ -577,6 +587,7 @@ class ScheduleBatch:
         tree_cache: BasePrefixCache,
         model_config: ModelConfig,
         enable_overlap: bool,
+        spec_algorithm: SpeculativeAlgorithm,
     ):
         return cls(
             reqs=reqs,
@@ -589,6 +600,7 @@ class ScheduleBatch:
             has_stream=any(req.stream for req in reqs),
             has_grammar=any(req.grammar for req in reqs),
             device=req_to_token_pool.device,
+            spec_algorithm=spec_algorithm,
         )
 
     def batch_size(self):
@@ -998,6 +1010,8 @@ class ScheduleBatch:
 
     def prepare_for_decode(self):
         self.forward_mode = ForwardMode.DECODE
+        if self.spec_algorithm.is_eagle():
+            return
 
         self.input_ids = self.output_ids
         self.output_ids = None
@@ -1103,6 +1117,9 @@ class ScheduleBatch:
         self.has_stream |= other.has_stream
         self.has_grammar |= other.has_grammar
 
+        if self.spec_info:
+            self.spec_info.merge_batch(other.spec_info)
+
     def get_model_worker_batch(self):
         if self.forward_mode.is_decode() or self.forward_mode.is_idle():
             extend_seq_lens = extend_prefix_lens = extend_logprob_start_lens = None
@@ -1144,6 +1161,13 @@ class ScheduleBatch:
             lora_paths=[req.lora_path for req in self.reqs],
             sampling_info=self.sampling_info,
             input_embeds=self.input_embeds,
+            spec_algorithm=self.spec_algorithm,
+            spec_info=self.spec_info,
+            capture_hidden_mode=(
+                getattr(self.spec_info, "capture_hidden_mode", CaptureHiddenMode.NULL)
+                if self.spec_info
+                else CaptureHiddenMode.NULL
+            ),
         )
 
     def copy(self):
@@ -1155,6 +1179,7 @@ class ScheduleBatch:
             out_cache_loc=self.out_cache_loc,
             return_logprob=self.return_logprob,
             decoding_reqs=self.decoding_reqs,
+            spec_algorithm=self.spec_algorithm,
         )
 
     def __str__(self):
@@ -1214,6 +1239,11 @@ class ModelWorkerBatch:
     # The input Embeds
     input_embeds: Optional[torch.tensor] = None
 
+    # Speculative decoding
+    spec_algorithm: SpeculativeAlgorithm = None
+    spec_info: Optional[SpecInfo] = None
+    capture_hidden_mode: CaptureHiddenMode = None
+
 
 @triton.jit
 def write_req_to_token_pool_triton(