Codex commited on
Commit
eb9efc0
·
1 Parent(s): e662caa

Keep ZeroGPU calls under proxy token lifetime

Browse files
MainProc/src/group_pipeline.py CHANGED
@@ -113,7 +113,7 @@ def _parse_view_batch(
113
 
114
  max_batch_views = max(1, _env_int(
115
  "LENTA_GROUP_FULL_OCR_BATCH_VIEWS" if force_full_variants else "LENTA_GROUP_OCR_BATCH_VIEWS",
116
- 2 if force_full_variants else 4,
117
  ))
118
  if len(views) > max_batch_views:
119
  for start in range(0, len(views), max_batch_views):
 
113
 
114
  max_batch_views = max(1, _env_int(
115
  "LENTA_GROUP_FULL_OCR_BATCH_VIEWS" if force_full_variants else "LENTA_GROUP_OCR_BATCH_VIEWS",
116
+ 1,
117
  ))
118
  if len(views) > max_batch_views:
119
  for start in range(0, len(views), max_batch_views):
MainProc/src/ocr.py CHANGED
@@ -65,11 +65,11 @@ def _env_flag(name: str, default: bool = False) -> bool:
65
 
66
  def _gpu_duration(base_name: str, default_base: float, count: int = 1, per_item: float = 2.0) -> int:
67
  seconds = _env_float(base_name, default_base) + max(1, count) * per_item
68
- return max(5, min(120, int(seconds + 0.999)))
69
 
70
 
71
  def ocr_one_gpu_duration(*_args, **_kwargs) -> int:
72
- return _gpu_duration("LENTA_OCR_ONE_GPU_SECONDS_BASE", 12.0, 1, 0.0)
73
 
74
 
75
  def extract_text_many_gpu_duration(*args, **kwargs) -> int:
@@ -86,11 +86,11 @@ def extract_text_many_gpu_duration(*args, **kwargs) -> int:
86
  has_full = bool(force_full_variants)
87
 
88
  if has_full:
89
- per_item = _env_float("LENTA_OCR_FULL_GPU_SECONDS_PER_IMAGE", 28.0)
90
- return _gpu_duration("LENTA_OCR_FULL_GPU_SECONDS_BASE", 35.0, count, per_item)
91
 
92
- per_item = _env_float("LENTA_OCR_GPU_SECONDS_PER_IMAGE", 18.0)
93
- return _gpu_duration("LENTA_OCR_GPU_SECONDS_BASE", 30.0, count, per_item)
94
 
95
 
96
  def bbox_to_list(bbox) -> list | None:
@@ -518,7 +518,7 @@ def _read_profile_variants_many(
518
  variant_meta: list[tuple[int, str]] = []
519
  for image_index in sorted(unresolved):
520
  if use_full_variants:
521
- variants = build_ocr_variants_for_profile(images[image_index], profile)
522
  else:
523
  variants = build_ocr_variant_stage_for_profile(images[image_index], profile, stage_index)
524
  for variant in variants:
@@ -564,6 +564,15 @@ def _read_profile_variants_many(
564
  return selected_results
565
 
566
 
 
 
 
 
 
 
 
 
 
567
  def _resolve_ocr_image_path(
568
  image_name: str,
569
  *,
 
65
 
66
  def _gpu_duration(base_name: str, default_base: float, count: int = 1, per_item: float = 2.0) -> int:
67
  seconds = _env_float(base_name, default_base) + max(1, count) * per_item
68
+ return max(5, min(55, int(seconds + 0.999)))
69
 
70
 
71
  def ocr_one_gpu_duration(*_args, **_kwargs) -> int:
72
+ return _gpu_duration("LENTA_OCR_ONE_GPU_SECONDS_BASE", 45.0, 1, 0.0)
73
 
74
 
75
  def extract_text_many_gpu_duration(*args, **kwargs) -> int:
 
86
  has_full = bool(force_full_variants)
87
 
88
  if has_full:
89
+ per_item = _env_float("LENTA_OCR_FULL_GPU_SECONDS_PER_IMAGE", 25.0)
90
+ return _gpu_duration("LENTA_OCR_FULL_GPU_SECONDS_BASE", 25.0, count, per_item)
91
 
92
+ per_item = _env_float("LENTA_OCR_GPU_SECONDS_PER_IMAGE", 20.0)
93
+ return _gpu_duration("LENTA_OCR_GPU_SECONDS_BASE", 20.0, count, per_item)
94
 
95
 
96
  def bbox_to_list(bbox) -> list | None:
 
518
  variant_meta: list[tuple[int, str]] = []
519
  for image_index in sorted(unresolved):
520
  if use_full_variants:
521
+ variants = _build_full_fallback_variants(images[image_index], profile)
522
  else:
523
  variants = build_ocr_variant_stage_for_profile(images[image_index], profile, stage_index)
524
  for variant in variants:
 
564
  return selected_results
565
 
566
 
567
+ def _build_full_fallback_variants(image, profile: str) -> list[dict]:
568
+ """Return a compact set of heavy OCR variants that can fit one ZeroGPU lease."""
569
+ max_variants = max(1, _env_int("LENTA_OCR_FULL_VARIANTS_PER_IMAGE", 2))
570
+ variants: list[dict] = []
571
+ for stage_index in (2, 1, 0):
572
+ variants.extend(build_ocr_variant_stage_for_profile(image, profile, stage_index))
573
+ return variants[:max_variants]
574
+
575
+
576
  def _resolve_ocr_image_path(
577
  image_name: str,
578
  *,
MainProc/src/ocr_preprocess_profiles.py CHANGED
@@ -80,8 +80,8 @@ def build_ocr_variant_stages_for_profile(image, profile: str = "default") -> lis
80
  def ocr_variant_stage_count(profile: str = "default") -> int:
81
  """Return how many staged OCR variant groups a profile has."""
82
  if profile == "red_promo":
83
- return 3
84
- return 3
85
 
86
 
87
  def build_ocr_variant_stage_for_profile(image, profile: str = "default", stage_index: int = 0) -> list[dict]:
 
80
  def ocr_variant_stage_count(profile: str = "default") -> int:
81
  """Return how many staged OCR variant groups a profile has."""
82
  if profile == "red_promo":
83
+ return 2
84
+ return 2
85
 
86
 
87
  def build_ocr_variant_stage_for_profile(image, profile: str = "default", stage_index: int = 0) -> list[dict]:
app.py CHANGED
@@ -134,9 +134,9 @@ def ensure_detector_device(model):
134
 
135
  def detector_gpu_duration(frames, conf):
136
  count = len(frames) if frames is not None else 1
137
- seconds = env_float("LENTA_DETECT_GPU_SECONDS_BASE", 12.0)
138
- seconds += count * env_float("LENTA_DETECT_GPU_SECONDS_PER_FRAME", 3.0)
139
- return max(15, min(120, int(math.ceil(seconds))))
140
 
141
 
142
  def serialize_detector_results(results):
@@ -172,7 +172,7 @@ def predict_detector_batch(frames, conf):
172
  predict_kwargs = {
173
  "conf": conf,
174
  "verbose": False,
175
- "batch": max(1, min(len(frames), env_int("LENTA_DETECT_BATCH_SIZE", 2))),
176
  "device": device,
177
  }
178
  if env_flag("LENTA_DETECT_HALF", False) and device == "cuda":
@@ -449,7 +449,7 @@ def extract_grouped_input(
449
  tracks = {}
450
  prev_processed_gray = None
451
  total_saved = 0
452
- detector_batch_size = max(1, env_int("LENTA_DETECT_BATCH_SIZE", 2))
453
 
454
  def detections_from_result(frame, result):
455
  frame_height, frame_width = frame.shape[:2]
 
134
 
135
  def detector_gpu_duration(frames, conf):
136
  count = len(frames) if frames is not None else 1
137
+ seconds = env_float("LENTA_DETECT_GPU_SECONDS_BASE", 18.0)
138
+ seconds += count * env_float("LENTA_DETECT_GPU_SECONDS_PER_FRAME", 8.0)
139
+ return max(20, min(55, int(math.ceil(seconds))))
140
 
141
 
142
  def serialize_detector_results(results):
 
172
  predict_kwargs = {
173
  "conf": conf,
174
  "verbose": False,
175
+ "batch": max(1, min(len(frames), env_int("LENTA_DETECT_BATCH_SIZE", 1))),
176
  "device": device,
177
  }
178
  if env_flag("LENTA_DETECT_HALF", False) and device == "cuda":
 
449
  tracks = {}
450
  prev_processed_gray = None
451
  total_saved = 0
452
+ detector_batch_size = max(1, env_int("LENTA_DETECT_BATCH_SIZE", 1))
453
 
454
  def detections_from_result(frame, result):
455
  frame_height, frame_width = frame.shape[:2]