Update app.py

app.py

@@ -438,6 +438,10 @@ _PLACEMENT_ALLOWANCE, _MARGIN = 8, 1.15
 # The ZeroGPU per-call ceiling. A booking above it is refused with `ZeroGPU illegal duration` once the request is
 # already in flight, so `generate` checks it up front and says which knob to turn instead.
 _MAX_BOOKING = int(os.environ.get("H3_MAX_BOOKING", "1500"))
+# Free-tier testing mode: forces the main Space's booking to exactly this many seconds regardless of the actual
+# request. Paired with the conditioner Space's own fixed 8s booking (both xlarge), for a combined 148s against
+# the shared 150s free-tier ceiling.
+MAXIMIZE_GPU_DURATION = int(os.environ.get("H3_MAXIMIZE_GPU_DURATION", "140"))
 
 
 def get_duration(
@@ -453,9 +457,14 @@ def get_duration(
     sharpen,
     multiplier,
     seed,
+    lora_strengths,
+    maximize_gpu,
     *a,
     **k,
 ):
+    if maximize_gpu:
+        return MAXIMIZE_GPU_DURATION
+
     height, width, num_frames, steps = int(height), int(width), int(num_frames), int(steps)
     multiplier = max(1, int(multiplier))
     latent_frames = (num_frames - LATENTS_PER_CHUNK) // FRAMES_PER_CHUNK * LATENTS_PER_CHUNK + 2
@@ -475,7 +484,7 @@ def get_duration(
     return max(60, int((denoise + decode + post) * _MARGIN) + _PLACEMENT_ALLOWANCE)
 
 
-@spaces.GPU(duration=140, size=GPU_SIZE)
+@spaces.GPU(duration=get_duration, size=GPU_SIZE)
 def _generate(
     prompt_embeds,
     text_token_tags,
@@ -490,6 +499,7 @@ def _generate(
     multiplier,
     seed,
     lora_strengths,
+    maximize_gpu,
 ):
     """The only thing on GPU time: the denoise loop, the two decoders and the workflow's post chain.
     The mp4 is muxed here rather than in the caller: a `@spaces.GPU` return crosses a process boundary by pickling,
@@ -591,6 +601,7 @@ def generate(
     lora_e_strength=DEFAULT_LORA_E_STRENGTH,
     lora_f_strength=DEFAULT_LORA_F_STRENGTH,
     lora_g_strength=DEFAULT_LORA_G_STRENGTH,
+    maximize_gpu=False,
     progress=gr.Progress(track_tqdm=True),
 ):
     """One request through the PlagueKind graph. Every parameter but the prompt carries the default its UI
@@ -649,6 +660,7 @@ def generate(
         multiplier,
         int(seed),
         lora_strengths,
+        bool(maximize_gpu),
     )
     # The same call `spaces` will book the worker with, so the report can show the fit against the measurement.
     booked_seconds = get_duration(*call)
@@ -775,62 +787,6 @@ with gr.Blocks(title="PlagueKind · MiniMax-H3") as demo:
                     value=DEFAULT_STEPS,
                     info="PlagueKind: 15-20 on the linear_quadratic grid.",
                 )
-                lora_1_strength = gr.Slider(
-                    label="Distilled LoRA",
-                    minimum=0.0,
-                    maximum=2.0,
-                    step=0.05,
-                    value=DEFAULT_LORA_1_STRENGTH,
-                )
-                lora_a_strength = gr.Slider(
-                    label="Anthro Enhancer LoRA",
-                    minimum=0.0,
-                    maximum=2.0,
-                    step=0.05,
-                    value=DEFAULT_LORA_A_STRENGTH,
-                )
-                lora_b_strength = gr.Slider(
-                    label="Reasoning Enhancer LoRA",
-                    minimum=0.0,
-                    maximum=2.0,
-                    step=0.05,
-                    value=DEFAULT_LORA_B_STRENGTH,
-                )
-                lora_c_strength = gr.Slider(
-                   label="HM-AIO LoRA",
-                   minimum=0.0,
-                   maximum=2.0,
-                   step=0.05,
-                   value=DEFAULT_LORA_C_STRENGTH,
-                )
-                lora_d_strength = gr.Slider(
-                   label="Anthro Realism LoRA",
-                   minimum=0.0,
-                   maximum=2.0,
-                   step=0.05,
-                   value=DEFAULT_LORA_D_STRENGTH,
-                )
-                lora_e_strength = gr.Slider(
-                   label="SB LoRA",
-                   minimum=0.0,
-                   maximum=2.0,
-                   step=0.05,
-                   value=DEFAULT_LORA_E_STRENGTH,
-                )
-                lora_f_strength = gr.Slider(
-                   label="Moaxx LoRA",
-                   minimum=0.0,
-                   maximum=2.0,
-                   step=0.05,
-                   value=DEFAULT_LORA_F_STRENGTH,
-                )
-                lora_g_strength = gr.Slider(
-                   label="Fluid Enhancer LoRA",
-                   minimum=0.0,
-                   maximum=2.0,
-                   step=0.05,
-                   value=DEFAULT_LORA_G_STRENGTH,
-                )
                 schedule = gr.Dropdown(
                     label="Sigma schedule",
                     choices=list(SCHEDULES),
@@ -857,10 +813,72 @@ with gr.Blocks(title="PlagueKind · MiniMax-H3") as demo:
                     value=False,
                     info="Rewrite the prompt on the conditioner Space first, MiniMax's Context-IR style.",
                 )
+                maximize_gpu = gr.Checkbox(
+                    label="Maximize Free Tier ZeroGPU (150 seconds)",
+                    value=False,
+                    info="Forces this request to book exactly 140s (plus 8s on the conditioner) for debugging purposes; does not prevent timeouts.",
+                )
 
         with gr.Column():
             video = gr.Video(label="Video + soundtrack")
             report = gr.Markdown()
+            lora_1_strength = gr.Slider(
+                label="Distilled LoRA",
+                minimum=0.0,
+                maximum=2.0,
+                step=0.05,
+                value=DEFAULT_LORA_1_STRENGTH,
+            )
+            with gr.Accordion("Custom LoRAs", open=False):
+                lora_a_strength = gr.Slider(
+                    label="Anthro Enhancer LoRA",
+                    minimum=0.0,
+                    maximum=2.0,
+                    step=0.05,
+                    value=DEFAULT_LORA_A_STRENGTH,
+                )
+                lora_b_strength = gr.Slider(
+                    label="Reasoning Enhancer LoRA",
+                    minimum=0.0,
+                    maximum=2.0,
+                    step=0.05,
+                    value=DEFAULT_LORA_B_STRENGTH,
+                )
+                lora_c_strength = gr.Slider(
+                    label="HM-AIO LoRA",
+                    minimum=0.0,
+                    maximum=2.0,
+                    step=0.05,
+                    value=DEFAULT_LORA_C_STRENGTH,
+                )
+                lora_d_strength = gr.Slider(
+                    label="Anthro Realism LoRA",
+                    minimum=0.0,
+                    maximum=2.0,
+                    step=0.05,
+                    value=DEFAULT_LORA_D_STRENGTH,
+                )
+                lora_e_strength = gr.Slider(
+                    label="SB LoRA",
+                    minimum=0.0,
+                    maximum=2.0,
+                    step=0.05,
+                    value=DEFAULT_LORA_E_STRENGTH,
+                )
+                lora_f_strength = gr.Slider(
+                    label="Moaxx LoRA",
+                    minimum=0.0,
+                    maximum=2.0,
+                    step=0.05,
+                    value=DEFAULT_LORA_F_STRENGTH,
+                )
+                lora_g_strength = gr.Slider(
+                    label="Fluid Enhancer LoRA",
+                    minimum=0.0,
+                    maximum=2.0,
+                    step=0.05,
+                    value=DEFAULT_LORA_G_STRENGTH,
+                )
 
     first_frame.upload(_fit_keyframe, [first_frame, canvas], [first_frame, canvas])
     last_frame.upload(_fit_keyframe, [last_frame, canvas], [last_frame, canvas])
@@ -885,6 +903,7 @@ with gr.Blocks(title="PlagueKind · MiniMax-H3") as demo:
         lora_e_strength,
         lora_f_strength,
         lora_g_strength,
+        maximize_gpu,
     ]
 
     run.click(generate, controls, [video, report], api_name="generate")