Merge pull request #1150 from bghira/feature/clip-evaluation

add documentation updates

OPTIONS.md

@@ -203,7 +203,7 @@ A lot of settings are instead set through the [dataloader config](/documentation
 - **What**: Output image resolution, measured in pixels, or, formatted as: `widthxheight`, as in `1024x1024`. Multiple resolutions can be defined, separated by commas.
 - **Why**: All images generated during validation will be this resolution. Useful if the model is being trained with a different resolution.
 
-### `--validation_model_evaluator`
+### `--evaluation_type`
 
 - **What**: Enable CLIP evaluation of generated images during validations.
 - **Why**: CLIP scores calculate the distance of the generated image features to the provided validation prompt. This can give an idea of whether prompt adherence is improving, though it requires a large number of validation prompts to have any meaningful value.
@@ -472,8 +472,8 @@ usage: train.py [-h] [--snr_gamma SNR_GAMMA] [--use_soft_min_snr]
                 [--model_card_note MODEL_CARD_NOTE]
                 [--model_card_safe_for_work] [--logging_dir LOGGING_DIR]
                 [--benchmark_base_model] [--disable_benchmark]
-                [--validation_model_evaluator {clip,none}]
-                [--pretrained_validation_model_name_or_path PRETRAINED_VALIDATION_MODEL_NAME_OR_PATH]
+                [--evaluation_type {clip,none}]
+                [--pretrained_evaluation_model_name_or_path pretrained_evaluation_model_name_or_path]
                 [--validation_on_startup] [--validation_seed_source {gpu,cpu}]
                 [--validation_torch_compile]
                 [--validation_torch_compile_mode {max-autotune,reduce-overhead,default}]
@@ -1243,12 +1243,12 @@ options:
   --disable_benchmark   By default, the model will be benchmarked on the first
                         batch of the first epoch. This can be disabled with
                         this option.
-  --validation_model_evaluator {clip,none}
+  --evaluation_type {clip,none}
                         Validations must be enabled for model evaluation to
                         function. The default is to use no evaluator, and
                         'clip' will use a CLIP model to evaluate the resulting
                         model's performance during validations.
-  --pretrained_validation_model_name_or_path PRETRAINED_VALIDATION_MODEL_NAME_OR_PATH
+  --pretrained_evaluation_model_name_or_path pretrained_evaluation_model_name_or_path
                         Optionally provide a custom model to use for ViT
                         evaluations. The default is currently clip-vit-large-
                         patch14-336, allowing for lower patch sizes (greater

documentation/quickstart/FLUX.md

@@ -413,6 +413,7 @@ Currently, the lowest VRAM utilisation (9090M) can be attained with:
 - Batch size: 1, zero gradient accumulation steps
 - DeepSpeed: disabled / unconfigured
 - PyTorch: 2.6 Nightly (Sept 29th build)
+- Using `--quantize_via=cpu` to avoid outOfMemory error during startup on <=16G cards.
 
 Speed was approximately 1.4 iterations per second on a 4090.
 

helpers/configuration/cmd_args.py

@@ -1331,7 +1331,7 @@ def get_argument_parser():
         ),
     )
     parser.add_argument(
-        "--validation_model_evaluator",
+        "--evaluation_type",
         type=str,
         default=None,
         choices=["clip", "none"],
@@ -1341,7 +1341,7 @@ def get_argument_parser():
         )
     )
     parser.add_argument(
-        "--pretrained_validation_model_name_or_path",
+        "--pretrained_evaluation_model_name_or_path",
         type=str,
         default="openai/clip-vit-large-patch14-336",
         help=(

helpers/training/evaluation.py

@@ -25,9 +25,9 @@ def from_config(args):
         """Instantiate a ModelEvaluator from the training config, if set to do so."""
         if not StateTracker.get_accelerator().is_main_process:
             return None
-        if args.validation_model_evaluator is not None and args.validation_model_evaluator.lower() != "" and args.validation_model_evaluator.lower() != "none":
-            model_evaluator = model_evaluator_map[args.validation_model_evaluator]
-            return globals()[model_evaluator](args.pretrained_validation_model_name_or_path)
+        if args.evaluation_type is not None and args.evaluation_type.lower() != "" and args.evaluation_type.lower() != "none":
+            model_evaluator = model_evaluator_map[args.evaluation_type]
+            return globals()[model_evaluator](args.pretrained_evaluation_model_name_or_path)
 
         return None