sglang 0.5.3__py3-none-any.whl → 0.5.3.post1__py3-none-any.whl

sglang/srt/layers/attention/mamba/mamba_utils.py

@@ -1,81 +0,0 @@
-# Adapted from: https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/mamba/mamba_utils.py
-from sglang.srt.distributed.utils import divide
-
-
-class MambaStateShapeCalculator:
-
-    @classmethod
-    def linear_attention_state_shape(
-        cls,
-        num_heads: int,
-        tp_size: int,
-        head_dim: int,
-    ) -> tuple[tuple[int, int, int], ...]:
-
-        state_shape = (num_heads // tp_size, head_dim, head_dim)
-        return (state_shape,)
-
-    @classmethod
-    def mamba1_state_shape(
-        cls,
-        tp_world_size: int,
-        intermediate_size: int,
-        state_size: int,
-        conv_kernel: int,
-    ) -> tuple[tuple[int, int], tuple[int, int]]:
-        conv_state_shape = (divide(intermediate_size, tp_world_size), conv_kernel - 1)
-
-        temporal_state_shape = (divide(intermediate_size, tp_world_size), state_size)
-
-        conv_state_shape = conv_state_shape[1], conv_state_shape[0]
-
-        return conv_state_shape, temporal_state_shape
-
-    @classmethod
-    def mamba2_state_shape(
-        cls,
-        tp_world_size: int,
-        intermediate_size: int,
-        n_groups: int,
-        num_heads: int,
-        head_dim: int,
-        state_size: int,
-        conv_kernel: int,
-    ) -> tuple[tuple[int, int], tuple[int, int, int]]:
-        # if n_groups is not divisible by world_size, need to extend the shards
-        # to ensure all groups needed by a head is sharded along with it
-        n_groups = n_groups + cls.extra_groups_for_head_shards(n_groups, tp_world_size)
-        # heads and n_groups are TP-ed
-        conv_dim = intermediate_size + 2 * n_groups * state_size
-
-        # contiguous along 'dim' axis
-        conv_state_shape = (conv_kernel - 1, divide(conv_dim, tp_world_size))
-
-        # These are not TP-ed as they depend on A, dt_bias, D
-        # - they are typically small
-        #   e.g., (h_heads, head_dim, state_size) = (128, 64, 128)
-        temporal_state_shape = (divide(num_heads, tp_world_size), head_dim, state_size)
-        return conv_state_shape, temporal_state_shape
-
-    @classmethod
-    def short_conv_state_shape(
-        cls,
-        tp_world_size: int,
-        intermediate_size: int,
-        conv_kernel: int,
-    ) -> tuple[tuple[int, int]]:
-        conv_dim = divide(intermediate_size, tp_world_size)
-        conv_state_shape = (conv_kernel - 1, conv_dim)
-        return (conv_state_shape,)
-
-    @classmethod
-    def extra_groups_for_head_shards(cls, ngroups: int, tp_size: int):
-        """Compute the increase in group numbers to account for
-        replication in order to accompany the head shards."""
-
-        # in the case ngoups % tp_size == 0, this will be zero
-        if ngroups % tp_size == 0:
-            return 0
-
-        # for n_groups == 1, this is exactly tp_size - n_groups
-        return tp_size - ngroups

sglang/srt/managers/tp_worker_overlap_thread.py

@@ -1,311 +0,0 @@
-# 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.
-# ==============================================================================
-"""A tensor parallel worker."""
-from __future__ import annotations
-
-import dataclasses
-import logging
-import signal
-import threading
-from queue import Queue
-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
-from sglang.utils import get_exception_traceback
-
-if TYPE_CHECKING:
-    from sglang.srt.managers.cache_controller import LayerDoneCounter
-
-logger = logging.getLogger(__name__)
-
-
-class TpModelWorkerClient:
-    """A tensor parallel model worker."""
-
-    def __init__(
-        self,
-        server_args: ServerArgs,
-        gpu_id: int,
-        tp_rank: int,
-        moe_ep_rank: int,
-        pp_rank: int,
-        dp_rank: Optional[int],
-        nccl_port: int,
-    ):
-        # Load the model
-        self.worker = TpModelWorker(
-            server_args, gpu_id, tp_rank, moe_ep_rank, pp_rank, dp_rank, nccl_port
-        )
-        self.max_running_requests = self.worker.max_running_requests
-        self.device = self.worker.device
-        self.gpu_id = gpu_id
-
-        # Init future mappings
-        self.future_map = FutureMap(self.max_running_requests, self.device)
-
-        # Launch threads
-        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(
-            target=self.forward_thread_func,
-        )
-        self.forward_thread.start()
-        self.parent_process = psutil.Process().parent()
-        self.scheduler_stream = torch.get_device_module(self.device).current_stream()
-        if self.device == "cpu":
-            self.scheduler_stream.synchronize = lambda: None  # No-op for CPU
-
-        self.hicache_layer_transfer_counter = None
-
-    def register_hicache_layer_transfer_counter(self, counter: LayerDoneCounter):
-        self.hicache_layer_transfer_counter = counter
-
-    def get_worker_info(self):
-        return self.worker.get_worker_info()
-
-    def get_tokens_per_layer_info(self):
-        return self.worker.get_tokens_per_layer_info()
-
-    @property
-    def sliding_window_size(self) -> Optional[int]:
-        return self.worker.sliding_window_size
-
-    @property
-    def is_hybrid(self) -> bool:
-        return self.worker.is_hybrid
-
-    def get_pad_input_ids_func(self):
-        return self.worker.get_pad_input_ids_func()
-
-    def get_tp_group(self):
-        return self.worker.get_tp_group()
-
-    def get_attention_tp_group(self):
-        return self.worker.get_attention_tp_group()
-
-    def get_attention_tp_cpu_group(self):
-        return self.worker.get_attention_tp_cpu_group()
-
-    def get_memory_pool(self):
-        return (
-            self.worker.model_runner.req_to_token_pool,
-            self.worker.model_runner.token_to_kv_pool_allocator,
-        )
-
-    def get_kv_cache(self):
-        return self.worker.model_runner.token_to_kv_pool
-
-    def forward_thread_func(self):
-        try:
-            with torch.get_device_module(self.device).stream(self.forward_stream):
-                self.forward_thread_func_()
-        except Exception:
-            traceback = get_exception_traceback()
-            logger.error(f"TpModelWorkerClient hit an exception: {traceback}")
-            self.parent_process.send_signal(signal.SIGQUIT)
-
-    @DynamicGradMode()
-    def forward_thread_func_(self):
-        batch_pt = 0
-        batch_lists: List = [None] * 2
-
-        while True:
-            model_worker_batch, future_map_ct, sync_event = self.input_queue.get()
-            if not model_worker_batch:
-                break
-
-            sync_event.wait()
-
-            # Keep a reference of model_worker_batch by storing it into a list.
-            # Otherwise, the tensor members of model_worker_batch will be released
-            # by pytorch and cause CUDA illegal memory access errors.
-            batch_lists[batch_pt % 2] = model_worker_batch
-            batch_pt += 1
-
-            # Create event
-            copy_done = torch.get_device_module(self.device).Event()
-
-            # Resolve future tokens in the input
-            self.future_map.resolve_future(model_worker_batch)
-
-            # 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 = (
-                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)
-            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:
-                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)
-                    )
-            if logits_output.hidden_states is not None:
-                logits_output.hidden_states = logits_output.hidden_states.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(
-                (copy_done, logits_output, next_token_ids, can_run_cuda_graph)
-            )
-
-    def resolve_last_batch_result(self, launch_done: Optional[threading.Event] = None):
-        """
-        This function is called to resolve the last batch result and
-        wait for the current batch to be launched. Used in overlap mode.
-        """
-        copy_done, logits_output, next_token_ids, can_run_cuda_graph = (
-            self.output_queue.get()
-        )
-
-        if launch_done is not None:
-            launch_done.wait()
-        copy_done.synchronize()
-
-        if logits_output.next_token_logprobs is not None:
-            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()
-            )
-        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
-    ) -> 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()
-        model_worker_batch.sampling_info = self.cur_sampling_info = dataclasses.replace(
-            sampling_info,
-            sampling_info_done=threading.Event(),
-            penalizer_orchestrator=None,
-        )
-
-        # A cuda stream sync here to avoid the cuda illegal memory access error.
-        sync_event = torch.get_device_module(self.device).Event()
-        sync_event.record(self.scheduler_stream)
-
-        # Push a new batch to the queue
-        bs = len(model_worker_batch.seq_lens)
-        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,
-        )
-
-    def update_weights_from_disk(self, recv_req: UpdateWeightFromDiskReqInput):
-        success, message = self.worker.update_weights_from_disk(recv_req)
-        return success, message
-
-    def init_weights_update_group(self, recv_req: InitWeightsUpdateGroupReqInput):
-        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
-    ):
-        success, message = self.worker.update_weights_from_distributed(recv_req)
-        return success, message
-
-    def update_weights_from_tensor(self, recv_req: UpdateWeightsFromTensorReqInput):
-        success, message = self.worker.update_weights_from_tensor(recv_req)
-        return success, message
-
-    def get_weights_by_name(self, recv_req: GetWeightsByNameReqInput):
-        return self.worker.get_weights_by_name(recv_req)
-
-    def load_lora_adapter(self, recv_req: LoadLoRAAdapterReqInput):
-        return self.worker.load_lora_adapter(recv_req)
-
-    def unload_lora_adapter(self, recv_req: UnloadLoRAAdapterReqInput):
-        return self.worker.unload_lora_adapter(recv_req)
-
-    def can_run_lora_batch(self, lora_ids: list[str]) -> bool:
-        return self.worker.can_run_lora_batch(lora_ids)
-
-    def __delete__(self):
-        self.input_queue.put((None, None))
-        self.copy_queue.put((None, None, None))