administrator/frigate
1
0
Fork 0
mirror of https://github.com/blakeblackshear/frigate.git synced 2026-05-07 22:15:28 +03:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

fix: explicitly set enable_mem_pattern for fixed-size models

Browse Source 
Previously relied on ORT's implicit default (True) for fixed-size models,
only flipping the flag in the variable-length branch. Set it explicitly
in both branches to be robust against ORT default changes.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
felalex 2026-05-02 23:51:39 -07:00
parent 1717f21f69
commit cf8638f260
1 changed files with 6 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
frigate/detectors/detection_runners.py
View File

@ -32,7 +32,7 @@ def get_ort_session_options(
is_complex_model: Whether the model needs basic optimization to avoid graph fusion issues. is_complex_model: Whether the model needs basic optimization to avoid graph fusion issues.
variable_length_inputs: Whether the model receives variable-length inputs (e.g. text variable_length_inputs: Whether the model receives variable-length inputs (e.g. text
embeddings). When True, disables memory-pattern caching, which otherwise builds embeddings). When True, disables memory-pattern caching, which otherwise builds
a plan per unique input shape and holds onto mmap regions indefinitely — a major a plan per unique input shape and holds onto mmap regions indefinitely - a major
source of RSS growth in the embeddings_manager process. source of RSS growth in the embeddings_manager process.
Returns: Returns:
@ -40,15 +40,18 @@ def get_ort_session_options(
""" """
sess_options = ort.SessionOptions() sess_options = ort.SessionOptions()
# Disable the CPU BFC arena for all sessions. With the arena enabled ORT pools # Disable the CPU BFC arena for all sessions. With the arena enabled ORT pools
# host-side staging buffers for GPU↔CPU transfers and never releases them back to # host-side staging buffers for GPU -> CPU transfers and never releases them back to
# the OS, causing RSS to grow without bound in long-running embedding processes. # the OS, causing RSS to grow without bound in long-running embedding processes.
sess_options.enable_cpu_mem_arena = False sess_options.enable_cpu_mem_arena = False
if variable_length_inputs: if variable_length_inputs:
# Disable per-shape memory-layout plan caching for models with variable-length # Disable per-shape memory-layout plan caching for models with variable-length
# inputs (Jina CLIP text, PaddleOCR). Each unique sequence length creates a # inputs (Jina CLIP text, PaddleOCR). Each unique sequence length creates a
# new mmap-backed plan that is never freed, leading to unbounded anon-mmap growth. # new mmap-backed plan that is never freed, leading to unbounded anon-mmap growth.
# Fixed-size models (YOLO at 640×640) should keep this enabled for buffer aliasing.
sess_options.enable_mem_pattern = False sess_options.enable_mem_pattern = False
else:
# Fixed-size models (like YOLO ) keep mem_pattern on for buffer aliasing.
# Set explicitly to be robust against ORT default changes.
sess_options.enable_mem_pattern = True
if is_complex_model: if is_complex_model:
sess_options.graph_optimization_level = ( sess_options.graph_optimization_level = (
ort.GraphOptimizationLevel.ORT_ENABLE_BASIC ort.GraphOptimizationLevel.ORT_ENABLE_BASIC