kai/infra-docker-configs
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

perf(unraid): enable Ampere mixed precision inference

Browse Source
This commit is contained in:
Codex
2026-08-15 10:18:23 -07:00
parent cd29046592
commit 593186b325
2 changed files with 10 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
servers/unraid/tail2-pose-server/app/tail2/models.py
+5
View File
@@ -29,6 +29,9 @@ class Models:
def load(self) -> None: def load(self) -> None:
from mmdet.apis import init_detector from mmdet.apis import init_detector
from mmpose.apis import init_model from mmpose.apis import init_model
torch.backends.cudnn.benchmark = True
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
self.detector = init_detector(self.det_config, self.det_checkpoint, device="cuda:0") self.detector = init_detector(self.det_config, self.det_checkpoint, device="cuda:0")
self.detector.test_cfg.nms.iou_threshold = float(os.getenv("NMS_IOU_THRESHOLD", ".60")) self.detector.test_cfg.nms.iou_threshold = float(os.getenv("NMS_IOU_THRESHOLD", ".60"))
self.pose = init_model(self.pose_config, self.pose_checkpoint, device="cuda:0") self.pose = init_model(self.pose_config, self.pose_checkpoint, device="cuda:0")
@@ -41,6 +44,7 @@ class Models:
height, width = frame.shape[:2] height, width = frame.shape[:2]
t0 = time.perf_counter_ns() t0 = time.perf_counter_ns()
init_default_scope("mmdet") init_default_scope("mmdet")
with torch.autocast(device_type="cuda", dtype=torch.float16):
result = inference_detector(self.detector, frame) result = inference_detector(self.detector, frame)
pred = result.pred_instances.cpu().numpy() pred = result.pred_instances.cpu().numpy()
keep = (pred.labels == 0) & (pred.scores >= threshold) keep = (pred.labels == 0) & (pred.scores >= threshold)
@@ -52,6 +56,7 @@ class Models:
det_ms = (time.perf_counter_ns() - t0) / 1e6 det_ms = (time.perf_counter_ns() - t0) / 1e6
t1 = time.perf_counter_ns() t1 = time.perf_counter_ns()
init_default_scope("mmpose") init_default_scope("mmpose")
with torch.autocast(device_type="cuda", dtype=torch.float16):
pose_results = inference_topdown(self.pose, frame, bboxes=boxes) if len(boxes) else [] pose_results = inference_topdown(self.pose, frame, bboxes=boxes) if len(boxes) else []
pose_ms = (time.perf_counter_ns() - t1) / 1e6 pose_ms = (time.perf_counter_ns() - t1) / 1e6
people = [] people = []
servers/unraid/tail2-pose-server/compose.yaml
+3 -3
View File
@@ -84,7 +84,7 @@ services:
NDI_SOURCE_NAME: TAIL 2_1621D2 (OBSBOT) NDI_SOURCE_NAME: TAIL 2_1621D2 (OBSBOT)
NDI_SOURCE_IP: 192.168.50.207 NDI_SOURCE_IP: 192.168.50.207
MODEL_PROFILE: balanced MODEL_PROFILE: balanced
BENCHMARK_PROFILE: rtmdet-tiny_rtmw-m_gpu1 BENCHMARK_PROFILE: rtmdet-tiny_rtmw-m_gpu1_fp16
MAX_PEOPLE: "4" MAX_PEOPLE: "4"
DETECTION_THRESHOLD: "0.35" DETECTION_THRESHOLD: "0.35"
NMS_IOU_THRESHOLD: "0.60" NMS_IOU_THRESHOLD: "0.60"
@@ -115,7 +115,7 @@ services:
restart: "no" restart: "no"
environment: environment:
DETECTOR_NAME: rtmdet-tiny-person DETECTOR_NAME: rtmdet-tiny-person
BENCHMARK_PROFILE: rtmdet-tiny_rtmw-m_gpu1 BENCHMARK_PROFILE: rtmdet-tiny_rtmw-m_gpu1_fp16
entrypoint: ["/bin/bash", "-lc"] entrypoint: ["/bin/bash", "-lc"]
depends_on: depends_on:
tail2-pose-server: tail2-pose-server:
@@ -146,7 +146,7 @@ services:
restart: "no" restart: "no"
environment: environment:
DETECTOR_NAME: rtmdet-tiny-person DETECTOR_NAME: rtmdet-tiny-person
BENCHMARK_PROFILE: rtmdet-tiny_rtmw-m_gpu1 BENCHMARK_PROFILE: rtmdet-tiny_rtmw-m_gpu1_fp16
entrypoint: ["/bin/bash", "-lc"] entrypoint: ["/bin/bash", "-lc"]
depends_on: depends_on:
tail2-pose-server: tail2-pose-server: