Video-to-Video
Diffusers
English
Chinese
video-editing
streaming
quantization
int8
text-stream-stripping
Instructions to use kylinfires/Joy-face_change_optimized with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use kylinfires/Joy-face_change_optimized with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("kylinfires/Joy-face_change_optimized", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
Sync main 91da25b: text K/V host cache (repeat sessions 1.4 s) + audit fixes
Browse files- README.md +2 -17
- deploy/xvideo/serving/joyomni_streaming.py +46 -1
README.md
CHANGED
|
@@ -1,19 +1,3 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
language:
|
| 4 |
-
- en
|
| 5 |
-
- zh
|
| 6 |
-
library_name: diffusers
|
| 7 |
-
pipeline_tag: video-to-video
|
| 8 |
-
base_model: jdopensource/JoyAI-Video-Edit
|
| 9 |
-
tags:
|
| 10 |
-
- video-editing
|
| 11 |
-
- streaming
|
| 12 |
-
- quantization
|
| 13 |
-
- int8
|
| 14 |
-
- text-stream-stripping
|
| 15 |
-
---
|
| 16 |
-
|
| 17 |
# Joy-face_change_optimized
|
| 18 |
|
| 19 |
[JoyAI-Video-Edit](https://github.com/jd-opensource/JoyAI-Video-Edit) 的优化分支(Apache-2.0)。
|
|
@@ -73,11 +57,12 @@ PyTorch 回退与 CUDA 版**逐张量验证一致**:99.997% 元素完全相同
|
|
| 73 |
|---|---:|---:|---:|
|
| 74 |
| 基线 | 25.5 GiB(26124 MiB) | 1.8 s | 0.314 s |
|
| 75 |
| + 文本 K/V + 权重卸载 | 19.5 GiB(19926 MiB) | 6.4 s | 0.280 s |
|
| 76 |
-
| + tight(关 graph) | **15.3 GiB**(15668 MiB) | 10.7 s | 0.314 s |
|
| 77 |
| + tight + `JOYOMNI_KV_DTYPE=int8` | **14.7 GiB** | 10.7 s | 0.316 s |
|
| 78 |
|
| 79 |
> 注意:文本 K/V 缓存**本身不省显存**——启动时 24.8 GiB vs 基线 23.6 GiB(缓存 0.46 + graph 池 0.46),
|
| 80 |
> 必须和 `JOYOMNI_TEXT_OFFLOAD` 一起用。表中数字为"跑完一个会话"后的 nvidia-smi,括号内为原始 MiB。
|
|
|
|
| 81 |
|
| 82 |
- **16 GB 卡可跑**:`JOYOMNI_VRAM_CAP_GB=16` 的硬上限下 17 帧会话跑通(reserved 14.5 GiB,峰值 13.7 GiB)。
|
| 83 |
- dit 每 chunk 0.219 → 0.185 s(graph 稳态,文本分支不再参与)。
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# Joy-face_change_optimized
|
| 2 |
|
| 3 |
[JoyAI-Video-Edit](https://github.com/jd-opensource/JoyAI-Video-Edit) 的优化分支(Apache-2.0)。
|
|
|
|
| 57 |
|---|---:|---:|---:|
|
| 58 |
| 基线 | 25.5 GiB(26124 MiB) | 1.8 s | 0.314 s |
|
| 59 |
| + 文本 K/V + 权重卸载 | 19.5 GiB(19926 MiB) | 6.4 s | 0.280 s |
|
| 60 |
+
| + tight(关 graph) | **15.3 GiB**(15668 MiB) | 10.7 s(首次 prompt)/ **1.4 s**(缓存命中) | 0.314 s |
|
| 61 |
| + tight + `JOYOMNI_KV_DTYPE=int8` | **14.7 GiB** | 10.7 s | 0.316 s |
|
| 62 |
|
| 63 |
> 注意:文本 K/V 缓存**本身不省显存**——启动时 24.8 GiB vs 基线 23.6 GiB(缓存 0.46 + graph 池 0.46),
|
| 64 |
> 必须和 `JOYOMNI_TEXT_OFFLOAD` 一起用。表中数字为"跑完一个会话"后的 nvidia-smi,括号内为原始 MiB。
|
| 65 |
+
> 文本 K/V 抓取的结果按(prompt embeds + 参考图 latent + steps + seed)做主机侧缓存(LRU 2),同一组输入的新会话直接拷回显存:首帧 **11.1 → 1.4 s**,逐帧与首次抓取 **138.13 dB** 完全一致(不同 prompt 各自 8–11 s 首次抓取,LRU 满时淘汰最旧一条)。
|
| 66 |
|
| 67 |
- **16 GB 卡可跑**:`JOYOMNI_VRAM_CAP_GB=16` 的硬上限下 17 帧会话跑通(reserved 14.5 GiB,峰值 13.7 GiB)。
|
| 68 |
- dit 每 chunk 0.219 → 0.185 s(graph 稳态,文本分支不再参与)。
|
deploy/xvideo/serving/joyomni_streaming.py
CHANGED
|
@@ -215,6 +215,11 @@ class JoyOmniRuntime:
|
|
| 215 |
self.ref_kv_cache: dict[str, dict] = {}
|
| 216 |
self.ref_kv_order: list[str] = []
|
| 217 |
self.ref_kv_cache_max = 2
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 218 |
self.output_quality = 60
|
| 219 |
self.lossless_output = False
|
| 220 |
self.graph_runners: dict[tuple, StreamingGraphRunner] = {}
|
|
@@ -693,7 +698,7 @@ class JoyOmniV2VStreamingSession:
|
|
| 693 |
self._prefill_or_reuse_reference_kv()
|
| 694 |
|
| 695 |
if text_kv_env_enabled():
|
| 696 |
-
self._text_kv_by_step = self.
|
| 697 |
|
| 698 |
self._maybe_prepare_graph_runner()
|
| 699 |
self._maybe_offload_text_stream()
|
|
@@ -761,6 +766,46 @@ class JoyOmniV2VStreamingSession:
|
|
| 761 |
flush=True,
|
| 762 |
)
|
| 763 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 764 |
@torch.no_grad()
|
| 765 |
def _capture_text_kv_by_step(self) -> list[dict[int, tuple[torch.Tensor, torch.Tensor]]]:
|
| 766 |
"""Prefill the text stream once per denoise step.
|
|
|
|
| 215 |
self.ref_kv_cache: dict[str, dict] = {}
|
| 216 |
self.ref_kv_order: list[str] = []
|
| 217 |
self.ref_kv_cache_max = 2
|
| 218 |
+
# Host-side text K/V cache, keyed by everything the capture depends on
|
| 219 |
+
# (prompt embeds, reference latent, step count, seed).
|
| 220 |
+
self.text_kv_cache: dict[str, list] = {}
|
| 221 |
+
self.text_kv_order: list[str] = []
|
| 222 |
+
self.text_kv_cache_max = 2
|
| 223 |
self.output_quality = 60
|
| 224 |
self.lossless_output = False
|
| 225 |
self.graph_runners: dict[tuple, StreamingGraphRunner] = {}
|
|
|
|
| 698 |
self._prefill_or_reuse_reference_kv()
|
| 699 |
|
| 700 |
if text_kv_env_enabled():
|
| 701 |
+
self._text_kv_by_step = self._text_kv_from_cache_or_capture()
|
| 702 |
|
| 703 |
self._maybe_prepare_graph_runner()
|
| 704 |
self._maybe_offload_text_stream()
|
|
|
|
| 766 |
flush=True,
|
| 767 |
)
|
| 768 |
|
| 769 |
+
def _text_kv_key(self) -> str:
|
| 770 |
+
"""Hash the exact inputs of the text K/V capture."""
|
| 771 |
+
digest = hashlib.sha1()
|
| 772 |
+
embeds = self.streaming_cond_embeds.detach().to("cpu", torch.float32).contiguous()
|
| 773 |
+
digest.update(embeds.numpy().tobytes())
|
| 774 |
+
if self.ref_image_latent is None:
|
| 775 |
+
digest.update(b"no-reference")
|
| 776 |
+
else:
|
| 777 |
+
reference = self.ref_image_latent.detach().to("cpu", torch.float32).contiguous()
|
| 778 |
+
digest.update(reference.numpy().tobytes())
|
| 779 |
+
digest.update(f"{int(self.settings.num_inference_steps)}|{int(self.settings.seed)}".encode())
|
| 780 |
+
return digest.hexdigest()[:16]
|
| 781 |
+
|
| 782 |
+
def _text_kv_from_cache_or_capture(self) -> list[dict[int, tuple[torch.Tensor, torch.Tensor]]]:
|
| 783 |
+
key = self._text_kv_key()
|
| 784 |
+
cached = self.runtime.text_kv_cache.get(key)
|
| 785 |
+
if cached is not None:
|
| 786 |
+
caches = [
|
| 787 |
+
{layer: tuple(t.to(self.device) for t in pair) for layer, pair in step.items()}
|
| 788 |
+
for step in cached
|
| 789 |
+
]
|
| 790 |
+
print(f"#####[TEXT-KV] reused host cache ({key})", flush=True)
|
| 791 |
+
return caches
|
| 792 |
+
|
| 793 |
+
started = time.perf_counter()
|
| 794 |
+
caches = self._capture_text_kv_by_step()
|
| 795 |
+
self.runtime.text_kv_cache[key] = [
|
| 796 |
+
{layer: tuple(t.to("cpu") for t in pair) for layer, pair in step.items()}
|
| 797 |
+
for step in caches
|
| 798 |
+
]
|
| 799 |
+
self.runtime.text_kv_order.append(key)
|
| 800 |
+
while len(self.runtime.text_kv_order) > self.runtime.text_kv_cache_max:
|
| 801 |
+
self.runtime.text_kv_cache.pop(self.runtime.text_kv_order.pop(0), None)
|
| 802 |
+
print(
|
| 803 |
+
f"#####[TEXT-KV] cached on host ({key}, "
|
| 804 |
+
f"{len(self.runtime.text_kv_order)} entr(ies), {time.perf_counter() - started:.2f}s)",
|
| 805 |
+
flush=True,
|
| 806 |
+
)
|
| 807 |
+
return caches
|
| 808 |
+
|
| 809 |
@torch.no_grad()
|
| 810 |
def _capture_text_kv_by_step(self) -> list[dict[int, tuple[torch.Tensor, torch.Tensor]]]:
|
| 811 |
"""Prefill the text stream once per denoise step.
|