xslandCursor
e7b62f2b2e
问题：接口2 与接口3/5 乱序调用时耗时抖动（最差 15~22s）。两个根因： 1. GPU 24G 常驻 21.4G，Flux-2(3.9G) 无法完全驻留显存，每次采样动态换页， 速度随空闲显存波动（2s~8s）； 2. ComfyUI 单队列 FIFO，接口2 排在接口3/5 批量任务后面。 改动： - hairline/comfyui.py: run() 新增 front 参数，/prompt 带 "front": true 插队到队列最前； redraw.py 透传；service.py 接口2 三处调用（女重绘 + 男有/无遮罩）传 front=True， 接口3/5 仍走普通队列。 - add_hair.json / 0716add-hair-api.json: 节点61 CLIPLoader device default→cpu。 qwen CLIP(4G) 不再占显存（文本条件缓存常年命中），ComfyUI 显存 8.8G→4.5G， Flux-2 完全驻留，采样稳定 ~3-5s。代价：换 prompt 后首次请求 CPU 编码 ~11s（一次性）。 - 提示词全局统一为「填充遮罩区域的头发，皮肤加一点磨皮，再加一点美颜」： app.py 4处默认值、service.py _REDRAW_PROMPT、redraw.py _DEFAULT_PROMPT、 4个工作流节点60内置文案、测试页(test_interface2/3/7/12/12_final)、local_test。 任何两个不同 prompt 交替提交都会打爆 CLIP 编码缓存（--cache-classic 只存最近一次）， 之前测试页旧文案与服务端不一致导致交替测试每次 +11s。 - app.py: 接口7 /api/v1/hair/grow-v2 下线（业务弃用；add_hair2.json 的 Klein-9b 会把常驻 Klein-4b 挤出显存）。保留 stub 返回 1007 明确报错，避免裸 404。 实测（1024 档）：接口2女 8.5~10s、接口2男 ~5s、接口3 ~7-10s，交替混跑无尖刺。 Co-authored-by: Cursor <cursoragent@cursor.com>
70 lines
2.6 KiB
Python
70 lines
2.6 KiB
Python
#!/usr/bin/env python3
|
|
"""Steps sweep + resolution test on the working 9B fp8-fast config."""
|
|
import io, time
|
|
import requests
|
|
import numpy as np
|
|
from PIL import Image, ImageFilter
|
|
import app as A
|
|
|
|
COMFY = "http://127.0.0.1:8188"
|
|
MODEL = "flux2.0/flux-2-klein-9b-fp8.safetensors"
|
|
DTYPE = "fp8_e4m3fn_fast"
|
|
|
|
|
|
def upload(scale=1.0):
|
|
image = Image.open("用来重绘.jpg").convert("RGB")
|
|
mask_img = Image.open("用来重绘.png").convert("RGBA")
|
|
if scale != 1.0:
|
|
w, h = image.size
|
|
w, h = int(w * scale) // 8 * 8, int(h * scale) // 8 * 8
|
|
image = image.resize((w, h), Image.LANCZOS)
|
|
mask_data = np.max(np.array(mask_img), axis=2)
|
|
m = Image.fromarray(mask_data, mode="L")
|
|
if m.size != image.size:
|
|
m = m.resize(image.size, Image.LANCZOS)
|
|
m = m.filter(ImageFilter.GaussianBlur(radius=4))
|
|
alpha = Image.eval(m, lambda x: 255 - x)
|
|
r, g, b = image.split()
|
|
rgba = Image.merge("RGBA", (r, g, b, alpha))
|
|
buf = io.BytesIO(); rgba.save(buf, format="PNG"); buf.seek(0)
|
|
up = requests.post(f"{COMFY}/upload/image",
|
|
files={"image": ("hair_input.png", buf, "image/png")}).json()
|
|
return up["name"], image.size
|
|
|
|
|
|
def run(fname, steps):
|
|
wf = A.build_workflow(fname, "填充遮罩区域的头发，皮肤加一点磨皮，再加一点美颜")
|
|
wf["16"]["inputs"]["unet_name"] = MODEL
|
|
wf["16"]["inputs"]["weight_dtype"] = DTYPE
|
|
wf["1"]["inputs"]["steps"] = steps
|
|
pid = requests.post(f"{COMFY}/prompt", json={"prompt": wf}).json()["prompt_id"]
|
|
dl = time.time() + 120
|
|
while time.time() < dl:
|
|
time.sleep(0.1)
|
|
h = requests.get(f"{COMFY}/history/{pid}").json()
|
|
if pid in h and "17" in h[pid].get("outputs", {}):
|
|
ts = {m[0]: m[1].get("timestamp") for m in h[pid]["status"]["messages"]}
|
|
return (ts["execution_success"] - ts["execution_start"]) / 1000.0
|
|
return float("nan")
|
|
|
|
|
|
print("=== 步数扫描 (9B fp8-fast, 原分辨率 1024x775) ===", flush=True)
|
|
fname, sz = upload(1.0)
|
|
run(fname, 6) # warmup
|
|
res = {}
|
|
for s in [2, 3, 4, 6, 8]:
|
|
t = min(run(fname, s), run(fname, s))
|
|
res[s] = t
|
|
print(f" steps={s}: {t:.2f}s", flush=True)
|
|
# derive per-step cost & fixed overhead via two points
|
|
per = (res[8] - res[2]) / 6
|
|
fixed = res[2] - per * 2
|
|
print(f" -> 每步 ~{per:.3f}s, 固定开销(VAE/编码/colormatch/加载) ~{fixed:.2f}s", flush=True)
|
|
|
|
print("\n=== 分辨率影响 (steps=4) ===", flush=True)
|
|
for scale in [1.0, 0.75, 0.6]:
|
|
fn, s2 = upload(scale)
|
|
run(fn, 4) # warmup
|
|
t = min(run(fn, 4), run(fn, 4))
|
|
print(f" {s2[0]}x{s2[1]} ({s2[0]*s2[1]/1e6:.2f}MP): {t:.2f}s", flush=True)
|