Add optional Turbo validation toggle

app.py

@@ -91,7 +91,7 @@ def start_training(
     dataset_rows, lora_name, instance_prompt, validation_prompt, rank, lora_alpha, max_train_steps,
     learning_rate, lr_scheduler, resolution, repeats, train_batch_size, gradient_accumulation_steps,
     seed, optimizer, use_8bit_adam, cache_latents, gradient_checkpointing, offload, quantization,
-    lora_layers, validation_epochs, hub_model_id, flavor, timeout,
+    lora_layers, run_validation, validation_epochs, hub_model_id, flavor, timeout,
     profile: gr.OAuthProfile | None = None, oauth_token: gr.OAuthToken | None = None,
 ):
     if oauth_token is None or profile is None:
@@ -113,7 +113,8 @@ def start_training(
         "optimizer": optimizer, "use_8bit_adam": bool(use_8bit_adam),
         "cache_latents": bool(cache_latents), "gradient_checkpointing": bool(gradient_checkpointing),
         "offload": bool(offload), "quantization": quantization, "lora_layers": lora_layers,
-        "validation_epochs": validation_epochs, "hf_token": oauth_token.token,
+        "run_validation": bool(run_validation), "validation_epochs": validation_epochs,
+        "hf_token": oauth_token.token,
     }
     try:
         res = jobs.submit(params, image_paths, captions, flavor=flavor, timeout=timeout)
@@ -213,6 +214,12 @@ with gr.Blocks(title="Krea 2 LoRA Trainer") as demo:
                         train_batch_size = gr.Number(label="Batch size", value=1, precision=0)
                         gradient_accumulation_steps = gr.Number(label="Grad accumulation", value=1, precision=0)
                 with gr.Accordion("Validation", open=False):
+                    run_validation = gr.Checkbox(
+                        label="Run validation on Turbo", value=True,
+                        info="Generate preview images on Krea 2 Turbo during/after training. "
+                             "Turn off to train faster (and to skip it while the diffusers PR's "
+                             "validation path is still being finalized upstream).",
+                    )
                     validation_prompt = gr.Textbox(
                         label="Validation prompt", placeholder="(defaults to your trigger)",
                         info="Generated on Turbo every N epochs to preview progress.",
@@ -259,8 +266,8 @@ with gr.Blocks(title="Krea 2 LoRA Trainer") as demo:
         inputs=[dataset_state, lora_name, instance_prompt, validation_prompt, rank, lora_alpha,
                 max_train_steps, learning_rate, lr_scheduler, resolution, repeats, train_batch_size,
                 gradient_accumulation_steps, seed, optimizer, use_8bit_adam, cache_latents,
-                gradient_checkpointing, offload, quantization, lora_layers, validation_epochs,
-                hub_model_id, flavor, timeout],
+                gradient_checkpointing, offload, quantization, lora_layers, run_validation,
+                validation_epochs, hub_model_id, flavor, timeout],
         outputs=[status, joblink, job_id],
     )
     refresh_btn.click(refresh, inputs=[job_id], outputs=[mon_status, mon_logs])

jobs.py

@@ -94,12 +94,17 @@ def build_train_args(params: dict, hub_model_id: str) -> list[str]:
         "--max_train_steps", str(int(params["max_train_steps"])),
         "--optimizer", str(params["optimizer"]),
         "--seed", str(int(params["seed"])),
-        "--validation_prompt", val_prompt,
-        "--validation_epochs", str(int(params["validation_epochs"])),
-        "--num_validation_images", "2",
         "--push_to_hub",
         "--hub_model_id", hub_model_id,
     ]
+    if params.get("run_validation", True):
+        args += [
+            "--validation_prompt", val_prompt,
+            "--validation_epochs", str(int(params["validation_epochs"])),
+            "--num_validation_images", "2",
+        ]
+    else:
+        args += ["--skip_final_inference"]
     if params.get("lora_layers"):
         args += ["--lora_layers", str(params["lora_layers"]).strip()]
     if params.get("gradient_checkpointing", True):