Files
enginex-ascend-910-vllm/vllm_ascend/worker/worker.py
Sun Ruoxi 7f8a1b1f7a init v0.23.0
Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>
2026-08-27 15:11:51 +08:00

1047 lines
47 KiB
Python

#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2023 The vLLM 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.
# This file is a part of the vllm-ascend project.
# Adapted from vllm-project/vllm/vllm/worker/gpu_worker.py
#
import copy
import gc
import logging
from types import NoneType
import torch
import torch.nn as nn
import torch_npu
from torch_npu.op_plugin.atb._atb_ops import _register_atb_extensions
from torch_npu.profiler import dynamic_profile as dp
from vllm.config import CUDAGraphMode, VllmConfig, set_current_vllm_config
from vllm.distributed import ensure_model_parallel_initialized, get_pcp_group, init_distributed_environment
from vllm.distributed.ec_transfer import ensure_ec_transfer_initialized
from vllm.distributed.kv_transfer import (
ensure_kv_transfer_initialized,
ensure_kv_transfer_shutdown,
get_kv_transfer_group,
has_kv_transfer_group,
)
from vllm.distributed.kv_transfer.kv_connector.v1.base import KVConnectorHandshakeMetadata
from vllm.distributed.parallel_state import Handle, get_pp_group, get_tp_group
from vllm.logger import logger
from vllm.lora.request import LoRARequest
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from vllm.tasks import SupportedTask
from vllm.utils.mem_constants import GiB_bytes
from vllm.utils.mem_utils import MemorySnapshot, format_gib, memory_profiling
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec
from vllm.v1.outputs import EMPTY_MODEL_RUNNER_OUTPUT, AsyncModelRunnerOutput, DraftTokenIds, ModelRunnerOutput
from vllm.v1.utils import report_usage_stats
from vllm.v1.worker.gpu_worker import AsyncIntermediateTensors
from vllm.v1.worker.worker_base import CompilationTimes, WorkerBase
from vllm.v1.worker.workspace import init_workspace_manager
import vllm_ascend.envs as envs_ascend
from vllm_ascend.ascend_config import get_ascend_config, init_ascend_config
from vllm_ascend.batch_invariant import init_batch_invariance
from vllm_ascend.cpu_binding import bind_cpus
from vllm_ascend.device_allocator.camem import CaMemAllocator
from vllm_ascend.device_allocator.sleep_mem_optimized import SleepWakeupManager
from vllm_ascend.distributed.parallel_state import init_ascend_model_parallel
from vllm_ascend.ops.triton.triton_utils import init_device_properties_triton
from vllm_ascend.profiler.torch_npu_profiler import TorchNPUProfilerWrapper
from vllm_ascend.utils import (
AscendDeviceType,
check_ascend_device_type,
enable_sp,
get_ascend_device_type,
register_ascend_customop,
setup_ascend_local_comm_res,
vllm_version_is,
)
from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
torch._dynamo.trace_rules.clear_lru_cache() # noqa: E402
from torch._dynamo.variables import TorchInGraphFunctionVariable # noqa: E402
from vllm.utils.torch_utils import set_random_seed # noqa: E402
torch_non_c_binding_in_graph_functions_npu = dict.fromkeys(
["torch.npu.current_stream"],
TorchInGraphFunctionVariable,
) # noqa: E402
torch_non_c_binding_in_graph_functions_npu["torch.npu.stream"] = TorchInGraphFunctionVariable # noqa: E402
torch._dynamo.trace_rules.torch_name_rule_map.append(torch_non_c_binding_in_graph_functions_npu) # noqa: E402
class NPUWorker(WorkerBase):
def __init__(
self,
vllm_config: VllmConfig,
local_rank: int,
rank: int,
distributed_init_method: str,
is_driver_worker: bool = False,
# Additional parameters for compatibility with vllm
**kwargs,
):
"""Initialize the worker for Ascend."""
if not envs_ascend.COMPILE_CUSTOM_KERNELS:
logger.warning(
"COMPILE_CUSTOM_KERNELS is set to False. "
"In most scenarios, without custom kernels, vllm-ascend will not function correctly."
)
# register patch for vllm
from vllm_ascend.utils import adapt_patch
adapt_patch()
# Register ops when worker init.
from vllm_ascend import ops
ops.register_dummy_fusion_op()
if get_ascend_device_type() != AscendDeviceType.A5:
_register_atb_extensions()
register_ascend_customop(vllm_config)
# init ascend config and soc version
init_ascend_config(vllm_config)
from vllm_ascend.logger import configure_ascend_file_logging
configure_ascend_file_logging()
check_ascend_device_type()
super().__init__(
vllm_config=vllm_config,
local_rank=local_rank,
rank=rank,
distributed_init_method=distributed_init_method,
is_driver_worker=is_driver_worker,
)
if self.cache_config.cache_dtype == "auto":
self.cache_dtype = self.model_config.dtype
else:
self.cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[self.cache_config.cache_dtype]
# Profiler is lazily initialized on first profile(is_start=True) call (RFC #6954)
self.profiler_config = vllm_config.profiler_config
self.profiler: TorchNPUProfilerWrapper | None = None
self.npugraph_memory_bytes = 0
if vllm_config.model_config and vllm_config.model_config.enable_sleep_mode:
# Buffers saved before sleep
self._sleep_saved_buffers: dict[str, torch.Tensor] = {}
self.sleep_wakeup_manager = SleepWakeupManager(vllm_config, self, lambda: getattr(self, "model_runner", None))
# Weight transfer engine is created in `load_model` once the model
# is available, since the engine needs a reference to the model.
self.weight_transfer_engine = None
self._weight_update_active = False
self._is_checkpoint_format = True
# FixMe: this is a patch to fix the issue cause by https://github.com/vllm-project/vllm/commit/de94289a98d7ec52a5ef02719e01a1db8b505170
from vllm.model_executor.layers.linear import WEIGHT_LOADER_V2_SUPPORTED
if "UnquantizedLinearMethod" in WEIGHT_LOADER_V2_SUPPORTED:
WEIGHT_LOADER_V2_SUPPORTED.remove("UnquantizedLinearMethod")
self.use_v2_model_runner = self.vllm_config.use_v2_model_runner
if self.use_v2_model_runner and vllm_version_is("0.23.0"):
logger.warning("VLLM_USE_V2_MODEL_RUNNER is not supported on vllm 0.23.0; falling back to v1 model runner.")
self.use_v2_model_runner = False
self._pp_send_work: list[Handle] = []
ascend_compilation_config = get_ascend_config().ascend_compilation_config
if ascend_compilation_config.enable_npugraph_ex and ascend_compilation_config.enable_static_kernel:
# Prevent duplicate triggers, execute the exit logic only once
shutdown_request = False
def signal_handler(signum, frame):
nonlocal shutdown_request
if not shutdown_request:
shutdown_request = True
self.uninstall_static_kernel()
raise SystemExit()
# Either SIGTERM or SIGINT will terminate the worker
import signal
signal.signal(signal.SIGTERM, signal_handler)
signal.signal(signal.SIGINT, signal_handler)
def uninstall_static_kernel(self):
import fcntl
import os
import subprocess
ascend_home_path = os.environ["ASCEND_HOME_PATH"]
static_kernel_dir_path = os.path.join(ascend_home_path, "opp/static_kernel")
uninstall_script_path = os.path.join(static_kernel_dir_path, "ai_core/uninstall.sh")
lock_file_path = os.path.join(static_kernel_dir_path, "uninstall.lock")
if not os.path.exists(uninstall_script_path):
return
with open(lock_file_path, "w") as lock_fd:
try:
fcntl.flock(lock_fd, fcntl.LOCK_EX | fcntl.LOCK_NB)
subprocess.Popen(
["bash", uninstall_script_path],
stdin=subprocess.DEVNULL,
stdout=subprocess.DEVNULL,
stderr=subprocess.DEVNULL,
start_new_session=True,
)
except (BlockingIOError, OSError):
return
finally:
try:
fcntl.flock(lock_fd, fcntl.LOCK_UN)
if os.path.exists(lock_file_path):
os.remove(lock_file_path)
except Exception:
return
def sleep(self, level: int = 1) -> None:
free_bytes_before_sleep = torch.npu.mem_get_info()[0]
# Save the buffers before level 2 sleep
if level == 2:
model = self.model_runner.model
self._sleep_saved_buffers = {name: buffer.cpu().clone() for name, buffer in model.named_buffers()}
cleanup_enabled = getattr(get_ascend_config(), "enable_sleep_mode_extra_cleanup", False)
if cleanup_enabled:
self.sleep_wakeup_manager.sleep()
allocator = CaMemAllocator.get_instance()
allocator.sleep(offload_tags=("weights",) if level == 1 else tuple())
free_bytes_after_sleep, total = torch.npu.mem_get_info()
freed_bytes = free_bytes_after_sleep - free_bytes_before_sleep
used_bytes = total - free_bytes_after_sleep
assert freed_bytes >= 0, "Memory usage increased after sleeping."
logger.info(
"Sleep mode (level=%s) freed %.2f GiB memory, %.2f GiB memory is still in use.",
level,
freed_bytes / GiB_bytes,
used_bytes / GiB_bytes,
)
def wake_up(self, tags: list[str] | None = None) -> None:
nz_mode = get_ascend_config().weight_nz_mode
if nz_mode:
raise ValueError(
"FRACTAL_NZ mode is enabled. This may cause model parameter precision issues "
"in the RL scenarios. Please set weight_nz_mode=0 via --additional-config."
)
allocator = CaMemAllocator.get_instance()
allocator.wake_up(tags=tags)
hidden_size = self.vllm_config.model_config.hf_text_config.hidden_size
model = self.model_runner.model
if self.vllm_config.quant_config is None and (tags is None or "weights" in tags):
for name, param in model.named_parameters():
if "w2_weight" in name and param.shape[2] == hidden_size:
parts = name.split(".")
param_name = parts[-1]
parent_module = model.get_submodule(".".join(parts[:-1]))
w2_data = param.transpose(1, 2)
w2_data = torch.nn.Parameter(w2_data, requires_grad=False)
setattr(parent_module, param_name, w2_data)
elif "w13_weight" in name and param.shape[1] == hidden_size:
parts = name.split(".")
param_name = parts[-1]
parent_module = model.get_submodule(".".join(parts[:-1]))
w13_data = param.transpose(1, 2)
w13_data = torch.nn.Parameter(w13_data, requires_grad=False)
setattr(parent_module, param_name, w13_data)
# Restore the buffers after level 2 sleep
if len(self._sleep_saved_buffers):
for name, buffer in model.named_buffers():
if name in self._sleep_saved_buffers:
buffer.data.copy_(self._sleep_saved_buffers[name].data)
self._sleep_saved_buffers = {}
cleanup_enabled = getattr(get_ascend_config(), "enable_sleep_mode_extra_cleanup", False)
if cleanup_enabled:
self.sleep_wakeup_manager.wakeup(tags)
def _check_weight_transfer_engine(self) -> None:
if self.weight_transfer_engine is None:
raise RuntimeError(
"Weight transfer not configured. Please set weight_transfer_config to enable weight transfer."
)
def init_weight_transfer_engine(self, init_info: dict) -> None:
"""Initialize the HCCL weight transfer process group with the trainer."""
self._check_weight_transfer_engine()
assert self.weight_transfer_engine is not None
typed_init_info = self.weight_transfer_engine.parse_init_info(init_info)
self.weight_transfer_engine.init_transfer_engine(typed_init_info)
def _check_nz_disabled(self) -> None:
if envs_ascend.VLLM_ASCEND_ENABLE_NZ:
raise ValueError(
"FRACTAL_NZ mode is enabled. This may cause model parameter "
"precision issues in the RL scenarios. Please set "
"VLLM_ASCEND_ENABLE_NZ=0."
)
def start_weight_update(self, is_checkpoint_format: bool = True) -> None:
"""Begin a new weight update; prepares the model for layerwise reload."""
self._check_weight_transfer_engine()
if self._weight_update_active:
raise RuntimeError(
"start_weight_update called while a weight update is already active. Call finish_weight_update first."
)
self._check_nz_disabled()
if is_checkpoint_format:
from vllm.model_executor.model_loader.reload import initialize_layerwise_reload
model = self.model_runner.model
with torch.device(self.device):
initialize_layerwise_reload(model)
self._is_checkpoint_format = is_checkpoint_format
self._weight_update_active = True
def update_weights(self, update_info: dict) -> None:
"""Receive a chunk of weights from the trainer and load them in place."""
self._check_weight_transfer_engine()
assert self.weight_transfer_engine is not None
typed_update_info = self.weight_transfer_engine.parse_update_info(update_info)
model = self.model_runner.model
# state machine driven by start/finish.
if not self._weight_update_active:
raise RuntimeError("start_weight_update must be called before update_weights.")
with torch.device(self.device):
if self._is_checkpoint_format:
self.weight_transfer_engine.receive_weights(
typed_update_info,
load_weights=model.load_weights,
)
else:
def load_weights_direct(weights: list[tuple[str, torch.Tensor]]) -> None:
with torch.no_grad():
for name, weight in weights:
param = model.get_parameter(name)
param.copy_(weight)
self.weight_transfer_engine.receive_weights(
typed_update_info,
load_weights=load_weights_direct,
)
# HCCL broadcast / packed paths are asynchronous.
# Sync so the next step uses the new weights.
torch.npu.synchronize()
def finish_weight_update(self) -> None:
"""Finish the current weight update; runs layerwise postprocessing."""
self._check_weight_transfer_engine()
if not self._weight_update_active:
raise RuntimeError("start_weight_update must be called before finish_weight_update.")
if self._is_checkpoint_format:
from vllm.model_executor.model_loader.reload import finalize_layerwise_reload
model = self.model_runner.model
with torch.device(self.device):
finalize_layerwise_reload(model, self.model_config)
self._weight_update_active = False
self._is_checkpoint_format = True
def shutdown(self) -> None:
if ensure_kv_transfer_shutdown is not None:
ensure_kv_transfer_shutdown()
if self.profiler is not None:
self.profiler.shutdown()
if weight_transfer_engine := getattr(self, "weight_transfer_engine", None):
weight_transfer_engine.shutdown()
if model_runner := getattr(self, "model_runner", None):
shutdown_fn = getattr(model_runner, "shutdown", None)
if callable(shutdown_fn):
shutdown_fn()
def initialize_cache(self, num_gpu_blocks: int, num_cpu_blocks: int) -> None:
self.cache_config.num_gpu_blocks = num_gpu_blocks
self.cache_config.num_cpu_blocks = num_cpu_blocks
def _init_device(self):
if not vllm_version_is("0.23.0"):
# vLLM v0.24.0 (PR #45026) removed automatic per-process device
# isolation for DP workers. Mirror gpu_worker.py::init_device:
# shift self.local_rank by dp_local_rank * tp_pp_world_size so
# that each DP group binds to a distinct set of NPUs.
parallel_config = self.parallel_config
if (
parallel_config.distributed_executor_backend not in ("ray", "external_launcher")
and parallel_config.data_parallel_backend != "ray"
and parallel_config.nnodes_within_dp == 1
# vllm-ascend: when the user pre-shards devices via
# --device-ids (which becomes assigned_physical_gpu_ids),
# each child process already binds to its own NPU(s); the
# DP local_rank shift below would push local_rank past the
# length of the per-rank device list and trip the assert
# in this same method. Skip the shift in that case.
and parallel_config.assigned_physical_gpu_ids is None
):
dp_local_rank = parallel_config.data_parallel_rank_local
if dp_local_rank is None:
dp_local_rank = parallel_config.data_parallel_index
tp_pp_world_size = parallel_config.pipeline_parallel_size * parallel_config.tensor_parallel_size
self.local_rank += dp_local_rank * tp_pp_world_size
# Publish the logical-to-physical mapping for topology queries.
assigned_physical_gpu_ids = parallel_config.assigned_physical_gpu_ids
if assigned_physical_gpu_ids is not None:
from vllm.platforms.interface import set_assigned_physical_gpu_ids
set_assigned_physical_gpu_ids(assigned_physical_gpu_ids)
assert self.local_rank < len(assigned_physical_gpu_ids), (
f"local_rank {self.local_rank} is out of bounds for "
f"assigned_physical_gpu_ids {assigned_physical_gpu_ids}"
)
if parallel_config.distributed_executor_backend not in ("ray", "external_launcher"):
assert parallel_config.local_world_size <= len(assigned_physical_gpu_ids), (
f"local_world_size ({parallel_config.local_world_size}) "
f"exceeds assigned_physical_gpu_ids count "
f"({len(assigned_physical_gpu_ids)})"
)
else:
visible_device_count = torch.npu.device_count() if torch.npu.is_available() else 0
assert self.local_rank < visible_device_count, (
f"DP adjusted local rank {self.local_rank} is out of bounds for {visible_device_count} devices."
)
visible_device_index = current_platform.logical_device_id_to_visible_device_id(self.local_rank)
device = torch.device(f"{current_platform.device_type}:{visible_device_index}")
else:
device = torch.device(f"npu:{self.local_rank}")
torch.npu.set_device(device)
# Import _inductor for graph mode execution with triton
# This lazy import avoids torch_npu re-initialization in patch
# Note that this should be imported after torch.npu.set_device
# to avoid repeated set_device in extra processes
from vllm.triton_utils import HAS_TRITON
if HAS_TRITON:
import torch_npu._inductor # noqa: F401
gc.collect()
torch.npu.empty_cache()
if get_ascend_device_type() == AscendDeviceType.A5:
setup_ascend_local_comm_res(self.local_rank, self.vllm_config.kv_transfer_config)
# take current memory snapshot
if vllm_version_is("0.23.0"):
self.init_snapshot = MemorySnapshot()
else:
self.init_snapshot = MemorySnapshot(device=device)
self.requested_memory = self.init_snapshot.total_memory * self.cache_config.gpu_memory_utilization
if self.init_snapshot.free_memory < self.requested_memory:
GiB = lambda b: round(b / GiB_bytes, 2)
raise ValueError(
f"Free memory on device "
f"({GiB(self.init_snapshot.free_memory)}/"
f"{GiB(self.init_snapshot.total_memory)} GiB) on startup "
f"is less than desired GPU memory utilization "
f"({self.cache_config.gpu_memory_utilization}, "
f"{GiB(self.requested_memory)} GiB). Decrease GPU memory "
f"utilization or reduce GPU memory used by other processes."
)
if (
self.parallel_config.data_parallel_size > 1
and self.parallel_config.data_parallel_size_local > 0
and self.parallel_config.distributed_executor_backend not in ["ray", "external_launcher"]
and self.vllm_config.parallel_config.data_parallel_backend != "ray"
and self.vllm_config.parallel_config.nnodes_within_dp == 1
):
visible_device_count = torch.npu.device_count() if torch.npu.is_available() else 0
assert self.parallel_config.local_world_size <= visible_device_count, (
f"local_world_size ({self.parallel_config.local_world_size}) must "
f"be less than or equal to the number of visible devices "
f"({visible_device_count})."
)
# Initialize the distributed environment.
self._init_worker_distributed_environment()
# Set random seed.
set_random_seed(self.model_config.seed)
# Initialize device properties used by triton kernels.
init_device_properties_triton()
return device
def init_device(self):
# NOTE: KEEP device the member of `NPUWorker`, as it will be checked
# in ray scenario. see https://github.com/vllm-project/vllm/pull/26845
# for more details
self.device = self._init_device()
# Initialize workspace manager
num_ubatches = 1
init_workspace_manager(self.device, num_ubatches)
# Init ModelRunner here, so that we have access to self.device.
if self.use_v2_model_runner:
logger.warning("npu model runner v2 is in developing, some features doesn't work for now.")
from vllm_ascend.worker.v2.model_runner import NPUModelRunner as NPUModelRunnerV2
self.model_runner = NPUModelRunnerV2(self.vllm_config, self.device)
else:
self.model_runner = NPUModelRunner(self.vllm_config, self.device)
if self.rank == 0:
# If usage stat is enabled, collect relevant info.
report_usage_stats(self.vllm_config)
@torch.inference_mode()
def determine_available_memory(self) -> int:
"""Profiles the peak memory usage of the model to determine how much
memory can be used for KV cache without OOMs.
The engine will first conduct a profiling of the existing memory usage.
Then, it calculates the free memory that can be used for KV cache in
bytes.
"""
GiB = lambda b: b / GiB_bytes
# Fast path: user has explicitly specified KV cache size via
# --kv-cache-memory. Still run profile_run() to compile the model,
# but skip the memory profiling calculation entirely.
if kv_cache_memory_bytes := self.cache_config.kv_cache_memory_bytes:
self.model_runner.profile_run()
logger.info(
"Initial free memory %.2f GiB, reserved %.2f GiB for KV Cache "
"as specified by kv_cache_memory_bytes, skipping memory profiling. "
"This does not respect the gpu_memory_utilization config. "
"Only use kv_cache_memory_bytes when you want manual control of "
"KV cache memory size. If OOM'ed, check the difference of initial "
"free memory between the current run and the previous run where "
"kv_cache_memory_bytes is suggested and update it correspondingly.",
GiB(self.init_snapshot.free_memory),
GiB(kv_cache_memory_bytes),
)
return kv_cache_memory_bytes
# Execute a forward pass with dummy inputs to profile the memory usage
# of the model.
with memory_profiling(
self.init_snapshot,
weights_memory=int(self.model_runner.model_memory_usage),
) as profile_result:
self.model_runner.profile_run()
# Record torch peak INSIDE the context and BEFORE graph capture,
# so that graph pool allocations don't inflate the activation peak.
# The memory_profiling context will also compute torch_peak_increase
# on exit, but we override it below with this pre-graph value.
profile_torch_peak = torch.npu.memory_stats(self.device).get("allocated_bytes.all.peak", 0)
# Override torch_peak_increase with the pre-graph-capture value to
# avoid double-counting graph pool memory as activation memory.
profile_result.torch_peak_increase = profile_torch_peak - profile_result.before_profile.torch_peak
profile_result.non_kv_cache_memory = (
profile_result.non_torch_increase + profile_result.torch_peak_increase + profile_result.weights_memory
)
# Save per-category memory for use in compile_or_warm_up_model() (step 5).
self.peak_activation_memory = profile_result.torch_peak_increase
self.non_torch_memory = profile_result.non_torch_increase
free_gpu_memory = profile_result.after_profile.free_memory
assert self.init_snapshot.free_memory > free_gpu_memory, (
"Error in memory profiling. "
f"Initial free memory {GiB(self.init_snapshot.free_memory)} GiB, "
f"current free memory {GiB(free_gpu_memory)} GiB. "
"This happens when other processes sharing the same container "
"release GPU memory while vLLM is profiling during initialization. "
"To fix this, ensure consistent GPU memory allocation or "
"isolate vLLM in its own container."
)
self.available_kv_cache_memory_bytes = self.requested_memory - profile_result.non_kv_cache_memory
logger.debug(profile_result)
logger.info_once(
"Available KV cache memory: %.2f GiB", GiB(self.available_kv_cache_memory_bytes), scope="local"
)
return int(self.available_kv_cache_memory_bytes)
def log_memory_stats(self) -> None:
"""Profiles the torch reserved memory, torch allocated memory in execute_model()."""
if not logger.isEnabledFor(logging.DEBUG):
return
self.torch_reserved = torch.npu.memory_reserved()
self.torch_allocated = torch.npu.memory_allocated()
logger.debug(
"torch reserved memory: %.2f GiB, torch allocated memory: %.2f GiB",
self.torch_reserved / GiB_bytes,
self.torch_allocated / GiB_bytes,
)
def execute_model(
self,
scheduler_output: "SchedulerOutput",
) -> ModelRunnerOutput | AsyncModelRunnerOutput | None:
self.log_memory_stats()
# enable msMonitor to monitor the performance of vllm-ascend
if get_ascend_config().msmonitor_use_daemon:
dp.step()
if self._pp_send_work:
for handle in self._pp_send_work:
handle.wait()
self._pp_send_work = []
intermediate_tensors = None
forward_pass = scheduler_output.total_num_scheduled_tokens > 0
if forward_pass and not get_pp_group().is_first_rank:
# If flashcomm1 is used, this all_gather_group parameter needs to be removed, otherwise
# it will conflict with the all-gather operation in flashcomm1.
if enable_sp():
all_gather_group = None
else:
all_gather_group = get_tp_group()
tensor_dict, comm_handles, comm_postprocess = get_pp_group().irecv_tensor_dict(
all_gather_group=all_gather_group
)
assert tensor_dict is not None
intermediate_tensors = AsyncIntermediateTensors(
tensor_dict,
comm_handles=comm_handles,
comm_postprocess=comm_postprocess,
)
if self.profiler is not None:
self.profiler.step()
output = self.model_runner.execute_model(scheduler_output, intermediate_tensors)
if isinstance(output, (ModelRunnerOutput, AsyncModelRunnerOutput, NoneType)):
return output
assert isinstance(output, IntermediateTensors)
parallel_config = self.vllm_config.parallel_config
assert parallel_config.distributed_executor_backend != ("external_launcher") and not get_pp_group().is_last_rank
# If flashcomm1 is used, this all_gather_group parameter needs to be removed, otherwise
# it will conflict with the all-gather operation in flashcomm1.
if enable_sp():
all_gather_group = None
else:
all_gather_group = get_tp_group()
self._pp_send_work = get_pp_group().isend_tensor_dict(
output.tensors,
all_gather_group=all_gather_group,
)
kv_connector_output = output.kv_connector_output
if not kv_connector_output:
return None
# In case of PP with kv transfer, we need to pass through the
# kv_connector_output
if not kv_connector_output.finished_sending and not kv_connector_output.finished_recving:
return EMPTY_MODEL_RUNNER_OUTPUT
output = copy.copy(EMPTY_MODEL_RUNNER_OUTPUT)
output.kv_connector_output = kv_connector_output
return output
@torch.inference_mode()
def sample_tokens(self, grammar_output: "GrammarOutput") -> ModelRunnerOutput | AsyncModelRunnerOutput:
return self.model_runner.sample_tokens(grammar_output)
def load_model(self) -> None:
if self.vllm_config.model_config.enable_sleep_mode:
allocator = CaMemAllocator.get_instance()
assert allocator.get_current_usage() == 0, "Sleep mode can only be used for one instance per process."
context = allocator.use_memory_pool(tag="weights")
else:
from contextlib import nullcontext
context = nullcontext() # type: ignore
with context, set_current_vllm_config(self.vllm_config):
self.model_runner.load_model()
if self.vllm_config.weight_transfer_config is not None:
from vllm.distributed.weight_transfer.factory import (
WeightTransferEngineFactory,
)
# main: create_engine takes (config, parallel_config, model)
self.weight_transfer_engine = WeightTransferEngineFactory.create_engine(
self.vllm_config.weight_transfer_config,
self.vllm_config.parallel_config,
self.model_runner.get_model(),
)
def compile_or_warm_up_model(self) -> CompilationTimes:
# Note: need to adapt for graph mode.
warmup_sizes = (self.vllm_config.compilation_config.compile_sizes or []).copy()
if not self.model_config.enforce_eager:
cg_capture_sizes: list[int] = []
if self.vllm_config.compilation_config.cudagraph_mode != CUDAGraphMode.NONE:
cg_sizes = self.vllm_config.compilation_config.cudagraph_capture_sizes
cg_capture_sizes = [] if cg_sizes is None else cg_sizes
warmup_sizes = [x for x in warmup_sizes if x not in cg_capture_sizes]
compile_ranges = self.vllm_config.compilation_config.get_compile_ranges()
# For each compile_range, if none of the batch sizes
# in warmup_sizes or cudagraph_capture_sizes are in the range,
# add the end of the range to ensure compilation/warmup.
all_sizes = set(cg_capture_sizes)
all_sizes.update([x for x in warmup_sizes if isinstance(x, int)])
for compile_range in compile_ranges:
if not any(x in compile_range for x in all_sizes):
warmup_sizes.append(compile_range.end)
for size in sorted(warmup_sizes, reverse=True):
logger.info("Compile and warming up model for size %d", size)
self.model_runner._dummy_run(size)
npugraph_memory_bytes = 0
if not self.model_config.enforce_eager:
npugraph_memory_bytes = self.model_runner.capture_model()
# Suggest an optimal --kv-cache-memory value for future runs.
# Only emitted when we ran full profiling (kv_cache_memory_bytes was not
# pre-specified) so that peak_activation_memory etc. are available.
# non_kv_memory already includes NPU graph memory, so the suggestion
# accounts for all measured memory categories. A 150 MiB buffer is kept
# because memory_profiling may slightly underestimate non-torch
# allocations (ACL context, HCCL buffers, driver layer, etc.).
if self.cache_config.kv_cache_memory_bytes is None and hasattr(self, "peak_activation_memory"):
redundancy_buffer = 150 * (1 << 20) # 150 MiB safety margin
non_kv_memory = (
self.model_runner.model_memory_usage
+ self.peak_activation_memory
+ self.non_torch_memory
+ npugraph_memory_bytes
)
self.npugraph_memory_bytes = npugraph_memory_bytes
suggested_to_requested = int(self.requested_memory) - non_kv_memory - redundancy_buffer
suggested_to_gpu_limit = int(self.init_snapshot.free_memory) - non_kv_memory - redundancy_buffer
msg = (
f"Free memory on device "
f"({format_gib(self.init_snapshot.free_memory)}/"
f"{format_gib(self.init_snapshot.total_memory)} GiB) on startup. "
f"Desired GPU memory utilization is "
f"({self.cache_config.gpu_memory_utilization}, "
f"{format_gib(self.requested_memory)} GiB). "
f"Actual usage: {format_gib(self.model_runner.model_memory_usage)} GiB "
f"for weights, {format_gib(self.peak_activation_memory)} GiB for peak "
f"activation, {format_gib(self.non_torch_memory)} GiB for non-torch "
f"memory, {format_gib(npugraph_memory_bytes)} GiB for NPU graph memory. "
f"Replace gpu_memory_utilization with "
f"`--kv-cache-memory={suggested_to_requested}` "
f"({format_gib(suggested_to_requested)} GiB) to fit into requested "
f"memory, or `--kv-cache-memory={suggested_to_gpu_limit}` "
f"({format_gib(suggested_to_gpu_limit)} GiB) to fully utilize NPU "
f"free memory. Current KV cache memory: "
f"{format_gib(self.available_kv_cache_memory_bytes)} GiB."
)
logger.info(msg)
# Call ATB matmul to warm up; otherwise, the first operation (ReshapeAndCache)
# may cause performance degradation at runtime.
if get_ascend_device_type() != AscendDeviceType.A5:
self._warm_up_atb()
# Bind after warmup so hot allocations are already materialized on the
# worker process before migratepages/taskset run.
if get_ascend_config().enable_cpu_binding:
try:
bind_cpus(self.local_rank)
except Exception as e:
logger.warning("Bind cpus failed in rank%s: %s Skip binding cpu.", self.local_rank, e)
# Reset the seed to ensure that the random state is not affected by
# the model initialization and profiling.
set_random_seed(self.model_config.seed)
return CompilationTimes(
language_model=self.vllm_config.compilation_config.compilation_time,
# `encoder_compilation_time` was added after v0.19.1 (vLLM #39240); fall
# back to 0.0 so the older release still constructs CompilationTimes.
encoder=getattr(
self.vllm_config.compilation_config,
"encoder_compilation_time",
0.0,
),
)
def _warm_up_atb(self):
x = torch.rand((2, 4), dtype=torch.float16).npu()
weight = torch.rand((2, 4), dtype=torch.float16).npu()
c = torch.rand((4, 4), dtype=torch.float32).npu()
torch_npu._npu_matmul_add_fp32(x, weight, c)
def get_model(self) -> nn.Module:
return self.model_runner.get_model()
@torch.inference_mode()
def profile_prefill_latency(self, num_tokens: int) -> float:
"""
Profile prefill latency for a given number of tokens.
This runs a real model forward pass and measures the execution time.
Used for profiling-based dynamic chunk sizing.
In PP (Pipeline Parallelism) mode:
- All workers execute the forward pass to stay synchronized
- Only the timing from PP0 (first rank) is meaningful for scheduling
- PP0 includes all the pipeline stages' latency when using async scheduling
Args:
num_tokens: Number of tokens to profile
Returns:
Latency in milliseconds
"""
import time
# Clamp to valid range
num_tokens = min(num_tokens, self.scheduler_config.max_num_batched_tokens)
num_tokens = max(num_tokens, 1)
# Synchronize all devices before timing
# This ensures clean measurement in PP/TP scenarios
torch.npu.synchronize()
# In PP mode, we still run on all ranks to keep them synchronized
# but only the first rank's timing is used for scheduling decisions
is_first_pp_rank = get_pp_group().is_first_rank
start = time.perf_counter()
# Run real model forward with force_attention=True
# This ensures attention is actually executed, not skipped.
# Without force_attention, attn_metadata may be None and attention
# won't run, making profiling results inaccurate.
# _dummy_run handles PP internally (intermediate tensors, etc.)
self.model_runner._dummy_run(
num_tokens=num_tokens,
force_attention=True, # Critical: ensure attention is executed
profile_cpp=True,
)
# Synchronize after forward to ensure NPU operations complete
torch.npu.synchronize()
latency_ms = (time.perf_counter() - start) * 1000
# Log for debugging in PP mode
if not is_first_pp_rank:
if logger.isEnabledFor(logging.DEBUG):
logger.debug(
"[ProfilingChunk] PP rank %s: profiled %s tokens, latency=%.2f ms (not used)",
get_pp_group().rank_in_group,
num_tokens,
latency_ms,
)
return latency_ms
def get_kv_connector_handshake_metadata(
self,
) -> dict[tuple[int, ...], KVConnectorHandshakeMetadata] | None:
"""Get KV connector metadata from this worker if available."""
if not has_kv_transfer_group():
return None
connector = get_kv_transfer_group()
# Return None for connectors that don't need to exchange handshake
# metadata across workers.
if (metadata := connector.get_handshake_metadata()) is None:
return None
tp_rank = get_tp_group().rank_in_group
pp_rank = get_pp_group().rank_in_group
pcp_size = get_pcp_group().world_size
if pcp_size > 1:
pcp_rank = get_pcp_group().rank_in_group
return {(pp_rank, pcp_rank, tp_rank): metadata}
return {(pp_rank, tp_rank): metadata}
def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
return self.model_runner.get_kv_cache_spec()
def update_max_model_len(self, max_model_len: int) -> None:
"""Update max_model_len after auto-fit to NPU memory.
This is called when max_model_len=-1 is used and the engine
automatically determines the maximum context length that fits
in GPU memory. Workers need to update their cached max_model_len
to match the engine's decision.
"""
self.model_config.max_model_len = max_model_len
if self.model_runner is not None:
self.model_runner.update_max_model_len(max_model_len)
logger.debug("Updated max_model_len to %s", max_model_len)
def initialize_from_config(self, kv_cache_config: KVCacheConfig) -> None:
"""Allocate NPU KV cache with the specified kv_cache_config."""
ensure_kv_transfer_initialized(self.vllm_config, kv_cache_config)
if self.vllm_config.model_config.enable_sleep_mode:
allocator = CaMemAllocator.get_instance()
context = allocator.use_memory_pool(tag="kv_cache")
else:
from contextlib import nullcontext
context = nullcontext() # type: ignore
with context:
self.model_runner.initialize_kv_cache(kv_cache_config)
# Restrict to mamba and full attn hybrid models (e.g. Qwen3.x).
#
# When eagle3 is enabled with num_speculative_tokens>1, mamba blocks may be reallocated to full blocks if
# the target and draft models share the same kv cache tensor (e.g. unaligned full attn layers with
# different num_kv_heads and head_size). In addition, for performance reasons, the current mtp/eagle path
# does not update seq_lens_cpu with num_rejected_tokens for step>1, since it would require d2h sync. As a
# result, seq_lens_cpu can become stale and some blocks will be unintentionally used.
#
# If an uncleared mamba block is later reused, the stale state combined with the incorrect seq_lens_cpu may
# lead to NaNs and reduced acceptance rate.
if (
kv_cache_config.needs_kv_cache_zeroing
and hasattr(self.model_runner, "_init_kv_zero_meta")
and self.vllm_config is not None
and self.vllm_config.speculative_config is not None
and self.vllm_config.speculative_config.method == "eagle3"
and self.vllm_config.speculative_config.num_speculative_tokens > 1
):
self.model_runner._init_kv_zero_meta()
def profile(self, is_start: bool = True, profile_prefix: str | None = None):
# Check if profiling is enabled (RFC #6954 - align with upstream vLLM)
if self.profiler_config is None or self.profiler_config.profiler is None:
raise RuntimeError(
"Profiling is not enabled. Please set --profiler-config to enable "
"profiling. Example: "
"'--profiler-config.profiler=torch --profiler-config.torch_profiler_dir"
"=YOUR_DIR_PATH_TO_DUMP_TRACE'"
)
if is_start:
from vllm.distributed.utils import get_worker_rank_suffix
rank_suffix = get_worker_rank_suffix(global_rank=self.rank)
trace_name = f"{profile_prefix}_{rank_suffix}" if profile_prefix else rank_suffix
if self.profiler is None:
self.profiler = TorchNPUProfilerWrapper(self.profiler_config, trace_name)
logger.debug("Starting torch profiler with trace name: %s", trace_name)
self.profiler.start() # type: ignore[attr-defined]
else:
# Profiler already initialized. Restart profiling but keep
# the original trace name from the first initialization.
self.profiler.start()
else:
if self.profiler is None:
logger.warning("Profiler was not started, nothing to stop.")
return
self.profiler.stop()
def add_lora(self, lora_request: LoRARequest) -> bool:
return self.model_runner.add_lora(lora_request)
def remove_lora(self, lora_id: int) -> bool:
return self.model_runner.remove_lora(lora_id)
def list_loras(self) -> set[int]:
return self.model_runner.list_loras()
def pin_lora(self, lora_id: int) -> bool:
return self.model_runner.pin_lora(lora_id)
def reset_encoder_cache(self) -> None:
self.model_runner.reset_encoder_cache()
def execute_dummy_batch(self) -> None:
self.log_memory_stats()
self.model_runner._dummy_run(num_tokens=self.model_runner.decode_token_per_req, uniform_decode=True)
def _init_worker_distributed_environment(self) -> None:
"""Initialize the distributed environment."""
init_batch_invariance()
init_distributed_environment(
self.parallel_config.world_size, self.rank, self.distributed_init_method, self.local_rank, "hccl"
)
ensure_model_parallel_initialized(
self.parallel_config.tensor_parallel_size,
self.parallel_config.pipeline_parallel_size,
self.parallel_config.prefill_context_parallel_size,
self.parallel_config.decode_context_parallel_size,
)
init_ascend_model_parallel(self.parallel_config)
ensure_ec_transfer_initialized(self.vllm_config)
def get_supported_pooling_tasks(self):
return self.model_runner.get_supported_pooling_tasks()
def get_supported_tasks(self) -> "tuple[SupportedTask, ...]":
return self.model_runner.get_supported_tasks()
def take_draft_token_ids(self) -> DraftTokenIds | None:
return self.model_runner.take_draft_token_ids()
def check_health(self) -> None:
import subprocess
logger.debug("check_health starting for rank %s...", self.local_rank)
try:
result = subprocess.run(
["npu-smi", "info", "-i", str(self.local_rank), "-t", "health"],
capture_output=True,
text=True,
timeout=10,
)
if result.returncode == 0:
parse_text_output(result.stdout)
logger.debug("check_health success for rank %s.", self.local_rank)
else:
logger.warning("query NPU card %s fail: %s", self.local_rank, result.stderr)
except subprocess.TimeoutExpired:
logger.warning("query NPU card %s timeout.", self.local_rank)
except FileNotFoundError:
logger.warning("npu-smi tool not found.")
except Exception as e:
logger.error("query NPU card %s fail: %s", self.local_rank, e)
return
def parse_text_output(output) -> None:
lines = output.strip().split("\n")
for i, line in enumerate(lines):
line = line.strip()
if "Health" in line:
if line.split(":")[-1].strip() != "OK":
raise RuntimeError("NPU card health status is not OK")
return