@@ -267,13 +267,6 @@ def run_inference_on_file(
267267 if "memory_factor" not in ds_info :
268268 missing_mem_keys .append ("memory_factor" )
269269 memory_factor = float (ds_info .get ("memory_factor" , 160 ))
270- if missing_mem_keys :
271- _suggest_memory_estimation_script (
272- idx ,
273- model_path ,
274- f"Memory parameter(s) { missing_mem_keys } not set in the model's dataset.json; falling back to defaults. { memory_base = } , { memory_factor = } " ,
275- logger = logger ,
276- )
277270
278271 use_folds_arg = tuple (folds ) if len (folds ) != 5 else None
279272 # Include every setting that changes the loaded predictor so a cache hit is always equivalent
@@ -362,11 +355,43 @@ def run_inference_on_file(
362355 if padd != 0 :
363356 p = (padd , padd )
364357 input_nii = [i .apply_pad ([p , p , p ], mode = "reflect" ) for i in input_nii ]
358+
359+ def _defaults_fit_gpu (memory : float , gpu : int | None , safety_factor : float = 2.0 ) -> bool :
360+ """Whether the fallback ``memory_base`` reservation is a small fraction of GPU memory.
361+
362+ Returns ``True`` when ``memory_base * safety_factor <= total_gpu_memory_mb`` —
363+ i.e. the fallback reservation leaves ample headroom for tile scheduling and
364+ the missing ``memory_base``/``memory_factor`` defaults are safe to use
365+ without warning. Returns ``False`` on CUDA-unavailable systems or when GPU
366+ memory can't be probed, so the load-time warning still fires on
367+ memory-constrained setups (CPU-only, small GPUs, or driver errors).
368+ """
369+ try :
370+ import torch
371+
372+ if not torch .cuda .is_available ():
373+ return False
374+ device = torch .device (f"cuda:{ gpu } " ) if gpu is not None else torch .device ("cuda:0" )
375+ _ , total_bytes = torch .cuda .mem_get_info (device )
376+ total_mb = total_bytes / (1024 ** 2 )
377+ except Exception : # noqa: BLE001
378+ return False
379+ return memory * safety_factor <= total_mb
380+
381+ num_classes = int (nnunet .label_manager .num_segmentation_heads )
382+ est_full_mb = estimate_peak_ram_mb (input_nii [0 ].shape , num_classes , len (input_nii ))
383+ # Only warn if memory requirement is non-trivial.
384+ if missing_mem_keys and not _defaults_fit_gpu (est_full_mb , gpu ):
385+ _suggest_memory_estimation_script (
386+ idx ,
387+ model_path ,
388+ f"Memory parameter(s) { missing_mem_keys } not set in the model's dataset.json; falling back to defaults. { memory_base = } , { memory_factor = } " ,
389+ logger = logger ,
390+ )
391+
365392 if _cpu_chunks is None or _cpu_chunks <= 1 :
366- num_classes = int (nnunet .label_manager .num_segmentation_heads )
367393 total_ram_mb = _get_total_ram_mb ()
368394 target_ram_mb = total_ram_mb * 0.5
369- est_full_mb = estimate_peak_ram_mb (input_nii [0 ].shape , num_classes , len (input_nii ))
370395 if est_full_mb > target_ram_mb :
371396 shape = input_nii [0 ].shape
372397 split_axis = int (np .argmax (shape ))
0 commit comments