EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

fix multiproc executor determine kv cache memory & update Dockerfile

Browse Source
This commit is contained in:
starkwj
2026-04-24 08:31:54 +00:00
parent e4d898b245
commit e17006077a
10 changed files with 145 additions and 153 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
Dockerfile
View File

@@ -21,12 +21,6 @@ ARG PIP_INDEX_URL="https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple"
ARG MOONCAKE_TAG="v0.3.8.post1"
ARG SOC_VERSION="ascend910b1"
# Define environments
ENV DEBIAN_FRONTEND=noninteractive
ENV SOC_VERSION=$SOC_VERSION \
TASK_QUEUE_ENABLE=1 \
OMP_NUM_THREADS=1
WORKDIR /workspace
COPY . /vllm-workspace/vllm-ascend/
@@ -58,6 +52,12 @@ RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /vllm-workspace/vllm
python3 -m pip cache purge
# Install vllm-ascend
ENV DEBIAN_FRONTEND=noninteractive
ENV SOC_VERSION=$SOC_VERSION \
TASK_QUEUE_ENABLE=1 \
OMP_NUM_THREADS=1 \
VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ASCEND_ENABLE_VNPU=1
# Append `libascend_hal.so` path (devlib) to LD_LIBRARY_PATH
# Installing vllm-ascend on x86 can pull upstream triton back in alongside triton-ascend. Remove it immediately after this step.
RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi && \
@@ -86,9 +86,6 @@ RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi
python3 -m pip install "https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/${PTA_WHEEL}" && \
python3 -m pip cache purge
ENV VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ASCEND_ENABLE_VNPU=1
# Install clang-15 (for triton-ascend)
RUN apt-get update -y && \
apt-get -y install clang-15 && \

16
Dockerfile.310p
View File

@@ -20,13 +20,6 @@ FROM quay.io/ascend/cann:8.5.1-310p-ubuntu22.04-py3.11
ARG PIP_INDEX_URL="https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple"
ARG SOC_VERSION="ascend310p1"
# Define environments
ENV DEBIAN_FRONTEND=noninteractive
ENV SOC_VERSION=$SOC_VERSION \
TASK_QUEUE_ENABLE=1 \
OMP_NUM_THREADS=1
RUN apt-get update -y && \
apt-get install -y python3-pip git vim wget net-tools gcc g++ cmake numactl libnuma-dev libjemalloc2 && \
rm -rf /var/cache/apt/* && \
@@ -48,6 +41,12 @@ RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /vllm-workspace/vllm
python3 -m pip cache purge
# Install vllm-ascend
ENV DEBIAN_FRONTEND=noninteractive
ENV SOC_VERSION=$SOC_VERSION \
TASK_QUEUE_ENABLE=1 \
OMP_NUM_THREADS=1 \
VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ASCEND_ENABLE_VNPU=1
# Append `libascend_hal.so` path (devlib) to LD_LIBRARY_PATH
# Installing vllm-ascend on x86 can pull upstream triton back in alongside triton-ascend. Remove it immediately after this step.
RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi && \
@@ -70,9 +69,6 @@ RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi
python3 -m pip install "https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/${PTA_WHEEL}" && \
python3 -m pip cache purge
ENV VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ASCEND_ENABLE_VNPU=1
# Install modelscope (for fast download) and ray (for multinode)
RUN python3 -m pip install modelscope 'ray>=2.47.1,<=2.48.0' 'protobuf>3.20.0' && \
python3 -m pip cache purge

12
Dockerfile.310p.openEuler
View File

@@ -20,10 +20,6 @@ FROM quay.io/ascend/cann:8.5.1-310p-openeuler24.03-py3.11
ARG PIP_INDEX_URL="https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple"
ARG SOC_VERSION="ascend310p1"
ENV SOC_VERSION=$SOC_VERSION \
TASK_QUEUE_ENABLE=1 \
OMP_NUM_THREADS=1
RUN yum update -y && \
yum install -y python3-pip git vim wget net-tools gcc gcc-c++ make cmake numactl numactl-devel jemalloc && \
rm -rf /var/cache/yum
@@ -44,6 +40,11 @@ RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[a
python3 -m pip cache purge
# Install vllm-ascend
ENV SOC_VERSION=$SOC_VERSION \
TASK_QUEUE_ENABLE=1 \
OMP_NUM_THREADS=1 \
VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ASCEND_ENABLE_VNPU=1
# Installing vllm-ascend on x86 can pull upstream triton back in alongside triton-ascend. Remove it immediately after this step.
RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi && \
source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
@@ -66,9 +67,6 @@ RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi
python3 -m pip install "https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/${PTA_WHEEL}" && \
python3 -m pip cache purge
ENV VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ASCEND_ENABLE_VNPU=1
# Install modelscope (for fast download) and ray (for multinode)
RUN python3 -m pip install modelscope 'ray>=2.47.1,<=2.48.0' 'protobuf>3.20.0' && \
python3 -m pip cache purge

14
Dockerfile.a3
View File

@@ -22,11 +22,6 @@ ARG MOONCAKE_TAG=v0.3.8.post1
ARG SOC_VERSION="ascend910_9391"
COPY . /vllm-workspace/vllm-ascend/
# Define environments
ENV DEBIAN_FRONTEND=noninteractive
ENV SOC_VERSION=$SOC_VERSION \
TASK_QUEUE_ENABLE=1 \
OMP_NUM_THREADS=1
RUN pip config set global.index-url ${PIP_INDEX_URL}
@@ -57,6 +52,12 @@ RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /vllm-workspace/vllm
python3 -m pip cache purge
# Install vllm-ascend
ENV DEBIAN_FRONTEND=noninteractive
ENV SOC_VERSION=$SOC_VERSION \
TASK_QUEUE_ENABLE=1 \
OMP_NUM_THREADS=1 \
VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ASCEND_ENABLE_VNPU=1
# Append `libascend_hal.so` path (devlib) to LD_LIBRARY_PATH
# Installing vllm-ascend on x86 can pull upstream triton back in alongside triton-ascend. Remove it immediately after this step.
RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi && \
@@ -85,9 +86,6 @@ RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi
python3 -m pip install "https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/${PTA_WHEEL}" && \
python3 -m pip cache purge
ENV VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ASCEND_ENABLE_VNPU=1
# Install clang-15 (for triton-ascend)
RUN apt-get update -y && \
apt-get -y install clang-15 && \

12
Dockerfile.a3.openEuler
View File

@@ -21,10 +21,6 @@ ARG PIP_INDEX_URL="https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple"
ARG MOONCAKE_TAG="v0.3.8.post1"
ARG SOC_VERSION="ascend910_9391"
ENV SOC_VERSION=$SOC_VERSION \
TASK_QUEUE_ENABLE=1 \
OMP_NUM_THREADS=1
RUN pip config set global.index-url ${PIP_INDEX_URL}
WORKDIR /workspace
@@ -58,6 +54,11 @@ RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[a
python3 -m pip cache purge
# Install vllm-ascend
ENV SOC_VERSION=$SOC_VERSION \
TASK_QUEUE_ENABLE=1 \
OMP_NUM_THREADS=1 \
VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ASCEND_ENABLE_VNPU=1
# Installing vllm-ascend on x86 can pull upstream triton back in alongside triton-ascend. Remove it immediately after this step.
RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi && \
source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
@@ -86,9 +87,6 @@ RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi
python3 -m pip install "https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/${PTA_WHEEL}" && \
python3 -m pip cache purge
ENV VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ASCEND_ENABLE_VNPU=1
# Install clang (for triton-ascend)
RUN yum update -y && \
yum install -y clang && \

12
Dockerfile.openEuler
View File

@@ -21,10 +21,6 @@ ARG PIP_INDEX_URL="https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple"
ARG MOONCAKE_TAG="v0.3.8.post1"
ARG SOC_VERSION="ascend910b1"
ENV SOC_VERSION=$SOC_VERSION \
TASK_QUEUE_ENABLE=1 \
OMP_NUM_THREADS=1
RUN pip config set global.index-url ${PIP_INDEX_URL}
WORKDIR /workspace
@@ -58,6 +54,11 @@ RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[a
python3 -m pip cache purge
# Install vllm-ascend
ENV SOC_VERSION=$SOC_VERSION \
TASK_QUEUE_ENABLE=1 \
OMP_NUM_THREADS=1 \
VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ASCEND_ENABLE_VNPU=1
# Installing vllm-ascend on x86 can pull upstream triton back in alongside triton-ascend. Remove it immediately after this step.
RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi && \
source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
@@ -86,9 +87,6 @@ RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi
python3 -m pip install "https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/${PTA_WHEEL}" && \
python3 -m pip cache purge
ENV VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ASCEND_ENABLE_VNPU=1
# Install clang (for triton-ascend)
RUN yum update -y && \
yum install -y clang && \

1
README.md
View File

@@ -34,5 +34,4 @@ docker build -t $build_image -f ./Dockerfile .
## Limitations
- Restricted by the fact that HCCL cannot be shared, deploying more than one model with multi-GPU (e.g., TP) is not feasible currently.
- The prefix cache will be reset when the LLM is restored, since we just simply discard the KV cache when the LLM is offloaded.

170
vllm_ascend/patch/platform/patch_core.py
View File

@@ -1,20 +1,103 @@
from logging import DEBUG
import os
import signal
import queue
import time
import vllm.envs as envs
from vllm.config import ParallelConfig, VllmConfig
from vllm.logger import logger
from vllm.v1.kv_cache_interface import KVCacheConfig
from vllm.v1.core.kv_cache_utils import (generate_scheduler_kv_cache_config,
get_kv_cache_configs)
from vllm.v1.engine.core import EngineCoreProc, EngineCore
from vllm.tracing import instrument
from vllm.tracing import maybe_init_worker_tracer
from vllm.transformers_utils.config import \
maybe_register_config_serialize_by_value
from vllm.utils.system_utils import decorate_logs, set_process_title
from vllm.v1.engine.core import EngineCoreProc, DPEngineCoreProc, EngineShutdownState
from vllm.v1.engine import EngineCoreRequestType
from vllm.v1.engine.utils import SignalCallback
import vllm_ascend.envs as envs_ascend
def run_engine_core(*args, dp_rank: int = 0, local_dp_rank: int = 0, **kwargs):
"""Launch EngineCore busy loop in background process."""
# Ensure we can serialize transformer config after spawning
maybe_register_config_serialize_by_value()
engine_core: EngineCoreProc | None = None
signal_callback: SignalCallback | None = None
try:
vllm_config: VllmConfig = kwargs["vllm_config"]
parallel_config: ParallelConfig = vllm_config.parallel_config
data_parallel = parallel_config.data_parallel_size > 1 or dp_rank > 0
if data_parallel:
parallel_config.data_parallel_rank_local = local_dp_rank
process_title = f"EngineCore_DP{dp_rank}"
else:
process_title = "EngineCore"
set_process_title(process_title)
maybe_init_worker_tracer("vllm.engine_core", "engine_core", process_title)
decorate_logs()
if data_parallel and vllm_config.kv_transfer_config is not None:
# modify the engine_id and append the local_dp_rank to it to ensure
# that the kv_transfer_config is unique for each DP rank.
vllm_config.kv_transfer_config.engine_id = (
f"{vllm_config.kv_transfer_config.engine_id}_dp{local_dp_rank}"
)
logger.debug(
"Setting kv_transfer_config.engine_id to %s",
vllm_config.kv_transfer_config.engine_id,
)
parallel_config.data_parallel_index = dp_rank
if data_parallel and vllm_config.model_config.is_moe:
# Set data parallel rank for this engine process.
parallel_config.data_parallel_rank = dp_rank
engine_core = DPEngineCoreProc(*args, **kwargs)
else:
# Non-MoE DP ranks are completely independent, so treat like DP=1.
# Note that parallel_config.data_parallel_index will still reflect
# the original DP rank.
parallel_config.data_parallel_size = 1
parallel_config.data_parallel_size_local = 1
parallel_config.data_parallel_rank = 0
engine_core = EngineCoreProc(*args, engine_index=dp_rank, **kwargs)
assert engine_core is not None
def wakeup_engine():
# Wakes up idle engine via input_queue when shutdown is requested
# Not safe in a signal handler - we may interrupt the main thread
# while it is holding the non-reentrant input_queue.mutex
engine_core.input_queue.put_nowait((EngineCoreRequestType.WAKEUP, None))
signal_callback = SignalCallback(wakeup_engine)
def signal_handler(signum, frame):
engine_core.shutdown_state = EngineShutdownState.REQUESTED
signal_callback.trigger()
signal.signal(signal.SIGTERM, signal_handler)
signal.signal(signal.SIGINT, signal_handler)
engine_core.run_busy_loop()
except SystemExit:
logger.debug("EngineCore exiting.")
raise
except Exception as e:
if engine_core is None:
logger.exception("EngineCore failed to start.")
else:
logger.exception("EngineCore encountered a fatal error.")
engine_core._send_engine_dead()
raise e
finally:
signal.signal(signal.SIGTERM, signal.SIG_DFL)
signal.signal(signal.SIGINT, signal.SIG_DFL)
if signal_callback is not None:
signal_callback.stop()
if engine_core is not None:
engine_core.shutdown()
def run_busy_loop(self):
"""Core busy loop of the EngineCore."""
while self._handle_shutdown():
@@ -77,75 +160,8 @@ def _process_input_queue(self):
self._handle_client_request(*req)
@instrument(span_name="Prepare model")
def _initialize_kv_caches(self, vllm_config: VllmConfig) -> KVCacheConfig:
start = time.time()
# Get all kv cache needed by the model
kv_cache_specs = self.model_executor.get_kv_cache_specs()
has_kv_cache = any(kv_cache_spec for kv_cache_spec in kv_cache_specs)
if has_kv_cache:
if envs_ascend.VLLM_ASCEND_ENABLE_VNPU:
# get available memory in idle offload mode
available_gpu_memory = (
self.model_executor.determine_available_memory_vnpu_offload_mode())
self.available_gpu_memory_for_kv_cache = \
available_gpu_memory[0]
elif envs.VLLM_ELASTIC_EP_SCALE_UP_LAUNCH:
# NOTE(yongji): should already be set
# during _eep_scale_up_before_kv_init
assert self.available_gpu_memory_for_kv_cache > 0
available_gpu_memory = [self.available_gpu_memory_for_kv_cache] * len(
kv_cache_specs
)
else:
# Profiles the peak memory usage of the model to determine how
# much memory can be allocated for kv cache.
available_gpu_memory = self.model_executor.determine_available_memory()
self.available_gpu_memory_for_kv_cache = available_gpu_memory[0]
else:
# Attention free models don't need memory for kv cache
available_gpu_memory = [0] * len(kv_cache_specs)
assert len(kv_cache_specs) == len(available_gpu_memory)
# Track max_model_len before KV cache config to detect auto-fit changes
max_model_len_before = vllm_config.model_config.max_model_len
kv_cache_configs = get_kv_cache_configs(
vllm_config, kv_cache_specs, available_gpu_memory
)
# If auto-fit reduced max_model_len, sync the new value to workers.
# This is needed because workers were spawned before memory profiling
# and have the original (larger) max_model_len cached.
max_model_len_after = vllm_config.model_config.max_model_len
if max_model_len_after != max_model_len_before:
self.collective_rpc("update_max_model_len", args=(max_model_len_after,))
scheduler_kv_cache_config = generate_scheduler_kv_cache_config(kv_cache_configs)
vllm_config.cache_config.num_gpu_blocks = scheduler_kv_cache_config.num_blocks
kv_cache_groups = scheduler_kv_cache_config.kv_cache_groups
if kv_cache_groups:
vllm_config.cache_config.block_size = min(
g.kv_cache_spec.block_size for g in kv_cache_groups
)
vllm_config.validate_block_size()
# Initialize kv cache and warmup the execution
self.model_executor.initialize_from_config(kv_cache_configs)
elapsed = time.time() - start
logger.info_once(
"init engine (profile, create kv cache, warmup model) took %.2f seconds",
elapsed,
scope="local",
)
return scheduler_kv_cache_config
# to make multi-proc enginecore get patched
EngineCoreProc.run_engine_core = run_engine_core
EngineCoreProc.run_busy_loop = run_busy_loop
EngineCoreProc._process_input_queue = _process_input_queue
EngineCore._initialize_kv_caches = _initialize_kv_caches

5
vllm_ascend/patch/platform/patch_executor.py
View File

@@ -42,11 +42,6 @@ def reload_vram(self) -> bool:
time.sleep(0.001)
def determine_available_memory_vnpu_offload_mode(self) -> int:
return self.collective_rpc("determine_available_memory_vnpu_offload_mode")
Executor.is_offloaded = is_offloaded
Executor.offload_vram = offload_vram
Executor.reload_vram = reload_vram
Executor.determine_available_memory_vnpu_offload_mode = determine_available_memory_vnpu_offload_mode

41
vllm_ascend/worker/worker.py
View File

@@ -336,6 +336,25 @@ class NPUWorker(WorkerBase):
bytes.
"""
GiB = lambda b: b / GiB_bytes
if envs_ascend.VLLM_ASCEND_ENABLE_VNPU:
allocator = CaMemAllocator.get_instance()
free, total = allocator.get_pool_mem_info()
if self.cache_config.gpu_memory_utilization <= 0.9:
logger.warning(
"GPU memory utilization is set to %.2f. For VNPU mode, it is recommended to set gpu_memory_utilization to a larger value",
self.cache_config.gpu_memory_utilization,
)
available_kv_cache_memory = int(
total * self.cache_config.gpu_memory_utilization - (total - free)
)
available_kv_cache_memory = int(max(available_kv_cache_memory, 0))
self.available_kv_cache_memory_bytes = available_kv_cache_memory
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)
# Execute a forward pass with dummy inputs to profile the memory usage
# of the model.
@@ -363,28 +382,6 @@ class NPUWorker(WorkerBase):
return int(self.available_kv_cache_memory_bytes)
@torch.inference_mode()
def determine_available_memory_vnpu_offload_mode(self) -> int:
GiB = lambda b: b / GiB_bytes
allocator = CaMemAllocator.get_instance()
free, total = allocator.get_pool_mem_info()
if self.cache_config.gpu_memory_utilization <= 0.9:
logger.warning(
"GPU memory utilization is set to %.2f. For VNPU mode, it is recommended to set gpu_memory_utilization to a larger value",
self.cache_config.gpu_memory_utilization,
)
available_kv_cache_memory = int(
total * self.cache_config.gpu_memory_utilization - (total - free)
)
available_kv_cache_memory = int(max(available_kv_cache_memory, 0))
self.available_kv_cache_memory_bytes = available_kv_cache_memory
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 execute_model(
self,
scheduler_output: "SchedulerOutput",