update

2026-05-15 00:46:37 +08:00
parent b4c7697e8a
commit f64e624a39
8 changed files with 190 additions and 100 deletions
@@ -51,7 +51,7 @@ class InpaintRequest(ImageInput):
    strength: float = Field(0.8, ge=0.0, le=1.0)
    negative_prompt: Optional[str] = Field("", description="负向 prompt")
    mask_b64: Optional[str] = Field(None, description="mask PNG base64（可选）")
-    max_side: int = Field(1024, ge=128, le=2048)
+    max_side: int = Field(1024, ge=128, le=8192, description="推理时长边上限；不超过则按输入全分辨率推理后再还原")
    guidance_scale: Optional[float] = Field(None, ge=0.0, le=200.0, description="扩散 CFG；不传则各后端默认")
    num_inference_steps: Optional[int] = Field(None, ge=1, le=150, description="采样步数；不传则各后端默认")

@@ -63,13 +63,14 @@ def make_flux_fill_predictor(

    if device == "cuda":
        try:
-            pipe.enable_model_cpu_offload()
+            pipe.enable_vae_tiling()
        except Exception:
-            try:
-                pipe.enable_vae_tiling()
-            except Exception:
-                pass
-            pipe.to("cuda")
+            pass
+        try:
+            pipe.enable_attention_slicing()
+        except Exception:
+            pass
+        pipe.to("cuda")
    else:
        pipe.to(device)

@@ -78,10 +78,8 @@ def _enable_memory_opts(pipe, device: str) -> None:
            pipe.enable_vae_tiling()
        except Exception:
            pass
-        try:
-            pipe.enable_model_cpu_offload()
-        except Exception:
-            pass
+        # 不使用 enable_model_cpu_offload：与 transformers 5.x 下 CLIPTextModel 等子模块的 accelerate hook
+        # 组合易触发 `'_hf_hook'` AttributeError；整管已 .to(cuda) 时以 slicing/tiling 省显存即可。


 def _align_size(orig_w: int, orig_h: int, max_side: int) -> tuple[int, int]:
@@ -123,8 +121,7 @@ def _make_sdxl_inpaint_predictor(
        use_safetensors=True,
    ).to(device)

-    # 省显存设置（尽量不改变输出语义）
-    # 注意：CPU offload 会明显变慢，但能显著降低显存占用。
+    # 省显存：attention slicing / VAE slicing tiling（不再 enable_model_cpu_offload，见 _enable_memory_opts 注释）
    if device == "cuda":
        try:
            pipe.enable_attention_slicing()
@@ -138,10 +135,6 @@ def _make_sdxl_inpaint_predictor(
            pipe.enable_vae_tiling()
        except Exception:
            pass
-        try:
-            pipe.enable_model_cpu_offload()
-        except Exception:
-            pass

    def _predict(
        image: Image.Image,
@@ -244,10 +237,7 @@ def _make_controlnet_predictor(_: UnifiedInpaintConfig):
            pipe.enable_vae_tiling()
        except Exception:
            pass
-        try:
-            pipe.enable_model_cpu_offload()
-        except Exception:
-            pipe.to(device)
+        pipe.to(device)
    else:
        pipe.to(device)

@@ -361,9 +351,22 @@ def _make_lama_predictor(cfg: UnifiedInpaintConfig):
        max_side: int = 1024,
        **_kwargs,
    ) -> Image.Image:
-        # big-lama：仅 (RGB, mask)；无文本条件。
-        _ = (strength, guidance_scale, num_inference_steps, max_side)
-        return lama(image, mask)
+        # big-lama：仅 (RGB, mask)；无文本条件。与 SDXL/FLUX 一致：可按 max_side 降采样推理后再还原到输入分辨率。
+        _ = (strength, guidance_scale, num_inference_steps)
+        image = image.convert("RGB")
+        mask = mask.convert("L")
+        orig_w, orig_h = image.size
+        run_w, run_h = _align_size(orig_w, orig_h, max_side=max_side)
+        if (run_w, run_h) != (orig_w, orig_h):
+            image_run = image.resize((run_w, run_h), Image.BICUBIC)
+            mask_run = mask.resize((run_w, run_h), Image.NEAREST)
+        else:
+            image_run = image
+            mask_run = mask
+        out = lama(image_run, mask_run)
+        if out.size != (orig_w, orig_h):
+            out = out.resize((orig_w, orig_h), resample=Image.BICUBIC)
+        return out

    return _predict