Merge main into ja: apply Experimental-Turbo changes to Japanese space

app.py

@@ -4,6 +4,9 @@ import time
 import threading
 import traceback
 
+# cudaMallocAsync bypasses NVML memory queries that fail on MIG GPU instances
+os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "backend:cudaMallocAsync")
+
 import gradio as gr
 import numpy as np
 import spaces
@@ -34,6 +37,33 @@ print("torch.__version__ =", torch.__version__, flush=True)
 print("Using device:", device, flush=True)
 print(f"CUDA device_count={torch.cuda.device_count()}, is_available={torch.cuda.is_available()}", flush=True)
 
+
+def _log_env():
+    import importlib.metadata as _meta
+    if torch.cuda.is_available():
+        p = torch.cuda.get_device_properties(0)
+        print(f"[env] GPU: {p.name}, VRAM={p.total_memory/1024**3:.1f}GB, cap={p.major}.{p.minor}", flush=True)
+    print(f"[env] CUDA (torch build): {torch.version.cuda}", flush=True)
+    print(f"[env] cuDNN: {torch.backends.cudnn.version()}", flush=True)
+    for pkg in ["spaces", "diffusers", "transformers", "gradio", "accelerate", "peft", "torchvision"]:
+        try:
+            print(f"[env] {pkg}=={_meta.version(pkg)}", flush=True)
+        except Exception as e:
+            print(f"[env] {pkg}==? ({e})", flush=True)
+    try:
+        mem = {}
+        with open("/proc/meminfo") as f:
+            for line in f:
+                k, v = line.split(":", 1)
+                mem[k.strip()] = v.strip()
+        total_gb = int(mem["MemTotal"].split()[0]) / 1024**2
+        avail_gb = int(mem["MemAvailable"].split()[0]) / 1024**2
+        print(f"[env] RAM: {total_gb:.0f}GB total, {avail_gb:.0f}GB available", flush=True)
+    except Exception as e:
+        print(f"[env] RAM: unavailable ({e})", flush=True)
+
+_log_env()
+
 # TF32 matmul: ~10-15% free speedup on Ampere/Hopper (bfloat16 accumulation paths benefit too)
 torch.backends.cuda.matmul.allow_tf32 = True
 torch.backends.cudnn.allow_tf32 = True
@@ -70,7 +100,7 @@ _hb = _start_heartbeat("transformer")
 _transformer = QwenImageTransformer2DModel.from_pretrained(
     "prithivMLmods/Qwen-Image-Edit-Rapid-AIO-V23",
     torch_dtype=dtype,
-    device_map="cuda",
+    device_map="cpu",
 )
 _hb.set()
 print(f"[startup] transformer loaded in {time.perf_counter()-_t0_load:.1f}s", flush=True)
@@ -82,7 +112,7 @@ pipe = QwenImageEditPlusPipeline.from_pretrained(
     "FireRedTeam/FireRed-Image-Edit-1.1",
     transformer=_transformer,
     torch_dtype=dtype,
-).to(device)
+)
 _hb.set()
 print(f"[startup] pipeline loaded in {time.perf_counter()-_t1_load:.1f}s", flush=True)
 
@@ -372,11 +402,15 @@ def _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height, m
         print(f"[infer] GPU: {p.name}, total={p.total_memory/1024**3:.1f}GB, cap={p.major}.{p.minor}")
         torch.cuda.reset_peak_memory_stats()
 
+    # Sequential offload: only the component currently doing work (text_encoder,
+    # then transformer, then vae — see model_cpu_offload_seq) sits on GPU at a
+    # time, instead of the full ~37GB pipeline resident simultaneously. Needed
+    # to fit on smaller MIG slices (e.g. the 47GB 2g.48gb partition) without OOM.
+    if getattr(pipe.transformer, "_hf_hook", None) is None:
+        pipe.enable_model_cpu_offload(device=device)
+        print(f"[infer] enabled sequential cpu offload on {device}")
     print(f"[infer] {_gpu_mem_str(_cuda_ok)} — t={time.perf_counter()-t0:.1f}s")
 
-    torch.cuda.empty_cache()
-    print(f"[infer] cache cleared — {_gpu_mem_str(_cuda_ok)}")
-
     print(f"[infer] {len(pil_images)} image(s) pre-decoded, output={width}x{height}, seed={seed}")
 
     generator = torch.Generator(device=device).manual_seed(seed)
@@ -418,6 +452,9 @@ def _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height, m
         timer.print_timings()
         raise
     finally:
+        # No manual pipe.to("cpu") — that fights the offload hooks' own device
+        # bookkeeping. Hooks already return each component to CPU after its
+        # forward; empty_cache() just reclaims the freed GPU blocks.
         gc.collect()
         torch.cuda.empty_cache()
         print(f"[infer] ===== END t={time.perf_counter()-t0:.1f}s =====")
@@ -432,7 +469,7 @@ with gr.Blocks() as demo:
     guidance_scale = gr.Slider(minimum=1.0, maximum=10.0, step=0.1, value=1.0, elem_id="gradio-guidance", elem_classes="hidden-input", container=False)
     steps = gr.Slider(minimum=1, maximum=50, step=1, value=4, elem_id="gradio-steps", elem_classes="hidden-input", container=False)
     mode = gr.Textbox(value="fast", elem_id="gradio-mode", elem_classes="hidden-input", container=False)
-    gpu_duration = gr.Slider(minimum=10, maximum=120, step=5, value=20, elem_id="gradio-gpu-duration", elem_classes="hidden-input", container=False)
+    gpu_duration = gr.Slider(minimum=10, maximum=120, step=5, value=30, elem_id="gradio-gpu-duration", elem_classes="hidden-input", container=False)
     result = gr.Image(elem_id="gradio-result", elem_classes="hidden-input", container=False, format="png")
 
     example_idx = gr.Textbox(value="", elem_id="example-idx-input", elem_classes="hidden-input", container=False)

static/mode_toggle.js

@@ -1,6 +1,6 @@
 () => {
 window.__selectedMode = 'fast';
-var MODE_GPU_DURATION = { fast: 20, high_detail: 30 };
+var MODE_GPU_DURATION = { fast: 30, high_detail: 60 };
 window.__setMode = function(m) {
     window.__selectedMode = m;
     var fast = document.getElementById('mode-btn-fast');

templates/app.html

@@ -120,8 +120,8 @@
                 <div class="settings-group-body">
                     <div class="slider-row">
                         <label>GPU時間（秒）</label>
-                        <input type="range" id="custom-gpu-duration" min="10" max="120" step="5" value="20">
-                        <span class="slider-val" id="custom-gpu-duration-val">20</span>
+                        <input type="range" id="custom-gpu-duration" min="10" max="120" step="5" value="30">
+                        <span class="slider-val" id="custom-gpu-duration-val">30</span>
                     </div>
                     <div class="slider-row">
                         <label>シード</label>