apextoaster
/
onnx-web
1
0
Fork 0
Code Actions

feat(api): add server setting for CUDA memory limit (#211)

This commit is contained in:
Sean Sube 2023-03-18 13:40:15 -05:00
parent af326a784f
commit aec540a524
Signed by: ssube
GPG Key ID: 3EED7B957D362AF1
3 changed files with 29 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
api/onnx_web/server/context.py
View File

@ -7,6 +7,10 @@ from .model_cache import ModelCache
logger = getLogger(__name__)
DEFAULT_CACHE_LIMIT = 2
DEFAULT_JOB_LIMIT = 10
DEFAULT_IMAGE_FORMAT = "png"
class ServerContext:
def __init__(
@ -19,12 +23,14 @@ class ServerContext:
any_platform: bool = True,
block_platforms: Optional[List[str]] = None,
default_platform: Optional[str] = None,
image_format: str = "png",
cache_limit: Optional[int] = 1,
image_format: str = DEFAULT_IMAGE_FORMAT,
cache_limit: int = DEFAULT_CACHE_LIMIT,
cache_path: Optional[str] = None,
show_progress: bool = True,
optimizations: Optional[List[str]] = None,
extra_models: Optional[List[str]] = None,
job_limit: int = DEFAULT_JOB_LIMIT,
memory_limit: Optional[int] = None,
) -> None:
self.bundle_path = bundle_path
self.model_path = model_path
@ -35,15 +41,22 @@ class ServerContext:
self.block_platforms = block_platforms or []
self.default_platform = default_platform
self.image_format = image_format
self.cache = ModelCache(cache_limit)
self.cache_limit = cache_limit
self.cache_path = cache_path or path.join(model_path, ".cache")
self.show_progress = show_progress
self.optimizations = optimizations or []
self.extra_models = extra_models or []
self.job_limit = job_limit
self.memory_limit = memory_limit
self.cache = ModelCache(self.cache_limit)
@classmethod
def from_environ(cls):
memory_limit = environ.get("ONNX_WEB_MEMORY_LIMIT", None)
if memory_limit is not None:
memory_limit = int(memory_limit)
return cls(
bundle_path=environ.get(
"ONNX_WEB_BUNDLE_PATH", path.join("..", "gui", "out")
@ -57,8 +70,10 @@ class ServerContext:
block_platforms=environ.get("ONNX_WEB_BLOCK_PLATFORMS", "").split(","),
default_platform=environ.get("ONNX_WEB_DEFAULT_PLATFORM", None),
image_format=environ.get("ONNX_WEB_IMAGE_FORMAT", "png"),
cache_limit=int(environ.get("ONNX_WEB_CACHE_MODELS", 2)),
cache_limit=int(environ.get("ONNX_WEB_CACHE_MODELS", DEFAULT_CACHE_LIMIT)),
show_progress=get_boolean(environ, "ONNX_WEB_SHOW_PROGRESS", True),
optimizations=environ.get("ONNX_WEB_OPTIMIZATIONS", "").split(","),
extra_models=environ.get("ONNX_WEB_EXTRA_MODELS", "").split(","),
job_limit=int(environ.get("ONNX_WEB_JOB_LIMIT", DEFAULT_JOB_LIMIT)),
memory_limit=memory_limit,
)

12
api/onnx_web/server/load.py
View File

@ -268,13 +268,19 @@ def load_platforms(context: ServerContext) -> None:
):
if potential == "cuda":
for i in range(torch.cuda.device_count()):
options = {
"device_id": i,
}
if context.memory_limit is not None:
options["arena_extend_strategy"] = "kSameAsRequested"
options["gpu_mem_limit"] = context.memory_limit
available_platforms.append(
DeviceParams(
potential,
platform_providers[potential],
{
"device_id": i,
},
options,
context.optimizations,
)
)

3
api/onnx_web/worker/pool.py
View File

@ -41,13 +41,12 @@ class DevicePoolExecutor:
self,
server: ServerContext,
devices: List[DeviceParams],
max_jobs_per_worker: int = 10,
max_pending_per_worker: int = 100,
join_timeout: float = 1.0,
):
self.server = server
self.devices = devices
self.max_jobs_per_worker = max_jobs_per_worker
self.max_jobs_per_worker = server.job_limit
self.max_pending_per_worker = max_pending_per_worker
self.join_timeout = join_timeout