update

python_server/model/Inpaint/flux_fill/loader.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import os
 from pathlib import Path
 from typing import Any, Callable
 
@@ -61,16 +62,32 @@ def make_flux_fill_predictor(
         local_files_only=local_files_only,
     )
 
+    # 显存：默认顺序 CPU offload（各子模块用时再上 GPU，峰值远低于 pipe.to(cuda)）。
+    # 整卡加载易占满 32GB 再在推理时 OOM。环境变量 FLUX_FILL_OFFLOAD=none 可改回整卡（需有余量）。
+    # 勿用 enable_model_cpu_offload（与 transformers 5.x CLIP 的 _hf_hook 问题相关）。
+    used_sequential = False
     if device == "cuda":
+        offload = os.environ.get("FLUX_FILL_OFFLOAD", "sequential").strip().lower()
         try:
             pipe.enable_vae_tiling()
         except Exception:
             pass
+        try:
+            pipe.enable_vae_slicing()
+        except Exception:
+            pass
         try:
             pipe.enable_attention_slicing()
         except Exception:
             pass
-        pipe.to("cuda")
+        if offload not in ("none", "gpu", "0", "false", "no"):
+            try:
+                pipe.enable_sequential_cpu_offload(gpu_id=0)
+                used_sequential = True
+            except Exception:
+                used_sequential = False
+        if not used_sequential:
+            pipe.to("cuda")
     else:
         pipe.to(device)
 
@@ -99,6 +116,13 @@ def make_flux_fill_predictor(
         run_w = _align_spatial_dim_for_flux_fill(run_w)
         run_h = _align_spatial_dim_for_flux_fill(run_h)
 
+        # 推理分辨率硬顶（像素域），防止极大 crop + 高 max_side 时注意力/VAE 爆显存；输出仍会缩回 orig。
+        run_cap = int(os.environ.get("FLUX_FILL_RUN_MAX_SIDE", "4096"))
+        if run_cap >= 256 and max(run_w, run_h) > run_cap:
+            s = run_cap / float(max(run_w, run_h))
+            run_w = _align_spatial_dim_for_flux_fill(int(round(run_w * s)))
+            run_h = _align_spatial_dim_for_flux_fill(int(round(run_h * s)))
+
         if (run_w, run_h) != (orig_w, orig_h):
             image_run = image.resize((run_w, run_h), resample=Image.BICUBIC)
             mask_run = mask.resize((run_w, run_h), resample=Image.NEAREST)
@@ -116,22 +140,35 @@ def make_flux_fill_predictor(
 
         if device == "cuda":
             try:
+                import gc
+
+                gc.collect()
                 torch.cuda.empty_cache()
             except Exception:
                 pass
 
-        out = pipe(
-            prompt=p_clip,
-            prompt_2=p_t5,
-            image=image_run,
-            mask_image=mask_run,
-            height=run_h,
-            width=run_w,
-            strength=st,
-            guidance_scale=float(guidance_scale),
-            num_inference_steps=int(num_inference_steps),
-            max_sequence_length=int(max_sequence_length),
-        ).images[0]
+        try:
+            out = pipe(
+                prompt=p_clip,
+                prompt_2=p_t5,
+                image=image_run,
+                mask_image=mask_run,
+                height=run_h,
+                width=run_w,
+                strength=st,
+                guidance_scale=float(guidance_scale),
+                num_inference_steps=int(num_inference_steps),
+                max_sequence_length=int(max_sequence_length),
+            ).images[0]
+        finally:
+            if device == "cuda":
+                try:
+                    import gc
+
+                    gc.collect()
+                    torch.cuda.empty_cache()
+                except Exception:
+                    pass
 
         out = out.convert("RGB")
         if out.size != (orig_w, orig_h):