Update: zh/train-optimize-sd-intel.md (huggingface#1303)

* update soc3-zn * Update _blog.yml Try to resolve conflicts * Update: proofreading zh/ethics-soc-3.md * add how-to-generate cn version Signed-off-by: Yao, Matrix <matrix.yao@intel.com> * unity game in hf space translation completed * Update: punctuations of how-to-generate.md * hf-bitsandbytes-integration cn done Signed-off-by: Yao, Matrix <matrix.yao@intel.com> * Proofread hf-bitsandbytes-integration.md * Proofread: red-teaming.md * Update: add red-teaming to zh/_blog.yml * Update _blog.yml * Update: add red-teaming to zh/_blog.yml Fix: red-teaming title in zh/_blog.yml * Fix: red-teaming PPLM translation * deep-learning-with-proteins cn done Signed-off-by: Yao, Matrix <matrix.yao@intel.com> * Add: stackllama.md * if blog translation completed * Update unity-in-spaces.md Add a link for AI game * Update if.md Fix “普罗大众” to “普惠大众” * deep-learning-with-proteins cn done Signed-off-by: Yao, Matrix <matrix.yao@intel.com> * add starcoder cn Signed-off-by: Yao, Matrix <matrix.yao@intel.com> Update: formatting and punctuations of starcoder.md * add starcoder cn Signed-off-by: Yao, Matrix <matrix.yao@intel.com> * Update: proofreading zh/unity-in-spaces.md * fix(annotated-diffusion.md): fix image shape desc in PIL and Tensor (huggingface#1080) modifiy the comment after ToTensor with the correct image shape CHW * Add text-to-video blog (huggingface#1058) Adds an overview of text-to-video generative models, task specific challenges, datasets, and more. Co-authored-by: Omar Sanseviero <osanseviero@gmail.com> Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Fix broken link in text-to-video.md (huggingface#1083) * Update: proofreading zh/unity-in-spaces.md Fix: incorrect _blog.yml format * Update: proofreading zh/deep-learning-with-proteins.md * update ethics-diffusers-cn (#6) * update ethics-diffusers * update ethics-diffusers --------- Co-authored-by: Zhongdong Yang <zhongdong_y@outlook.com> * Update: proofreading zh/ethics-diffusers.md * 1. introducing-csearch done (#11) 2. text-to-video done Signed-off-by: Yao, Matrix <matrix.yao@intel.com> * Update: proofread zh/text-to-video.md * Update: proofreading zh/introducing-csearch.md * generative-ai-models-on-intel-cpu cn done (#13) Signed-off-by: Yao, Matrix <matrix.yao@intel.com> Update: proofread zh/generative-ai-models-on-intel-cpu.md Signed-off-by: Yang, Zhongdong <zhongdong_y@outlook.com> * add starchat-alpha zh translation (#10) * Preparing blogpost annoucing `safetensors` security audit + official support. (huggingface#1096) * Preparing blogpost annoucing `safetensors` security audit + official support. * Taking into account comments + Grammarly. * Update safetensors-official.md * Apply suggestions from code review Co-authored-by: Omar Sanseviero <osanseviero@gmail.com> * Update safetensors-official.md * Apply suggestions from code review Co-authored-by: Luc Georges <McPatate@users.noreply.github.com> * Apply suggestions from code review Co-authored-by: Luc Georges <McPatate@users.noreply.github.com> * Apply suggestions from code review * Update safetensors-official.md Co-authored-by: Luc Georges <McPatate@users.noreply.github.com> * Apply suggestions from code review * Adding thumbnail. * Include changes from Stella. * Update safetensors-official.md * Update with Stella's comments. * Remove problematic sentence. * Rename + some rephrasing. * Apply suggestions from code review Co-authored-by: DeltaPenrose <128761972+DeltaPenrose@users.noreply.github.com> * Apply suggestions from code review Co-authored-by: DeltaPenrose <128761972+DeltaPenrose@users.noreply.github.com> * Update safetensors-security-audit.md Co-authored-by: DeltaPenrose <128761972+DeltaPenrose@users.noreply.github.com> * Last fixes. * Apply suggestions from code review Co-authored-by: DeltaPenrose <128761972+DeltaPenrose@users.noreply.github.com> --------- Co-authored-by: Omar Sanseviero <osanseviero@gmail.com> Co-authored-by: Luc Georges <McPatate@users.noreply.github.com> Co-authored-by: DeltaPenrose <128761972+DeltaPenrose@users.noreply.github.com> * Hotfixing safetensors. (huggingface#1131) * Removing the checklist formatting is busted. (huggingface#1132) * Update safetensors-security-audit.md (huggingface#1134) * [time series transformers] update dataloader API (huggingface#1135) * update dataloader API * revert comment * add back Cached transform * New post: Hugging Face and IBM (huggingface#1130) * Initial version * Minor fixes * Update huggingface-and-ibm.md Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update huggingface-and-ibm.md Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Resize image * Update blog index --------- Co-authored-by: Julien Simon <julsimon@huggingface.co> Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Show authors of safetensors blog post (huggingface#1137) Update: proofread zh/starchat-alpha.md * add megatron-training & assisted-generation (#8) * add megatron-training * add megatron-training * add megatron-training * add megatron-training * add assisted-generation * add assisted-generation * add assisted-generation * Update: proofreading zh/assisted-generation * Update: proofread zh/megatron-training.md * rwkv model blog translation completed (#12) * rwkv model blog translation completed * add 3 additional parts in the blog tail * Update: proofread zh/rwkv.md * Fix: missing subtitle/notes for image references. * encoder-decoder cn done (#14) Signed-off-by: Yao, Matrix <matrix.yao@intel.com> Co-authored-by: Zhongdong Yang <zhongdong_y@outlook.com> * Update: proofread zh/encoder-decoder.md * constrained-beam-search cn done (#15) Signed-off-by: Yao, Matrix <matrix.yao@intel.com> Update: proofread zh/constrained-beam-search.md * Update: zh/unity-api.md + zh/unity-asr.md * unity ai speech recognition blog translation completed * add (GameObject) to attach its Chinese translation * finish unity-api translation * add unity series entry to zh/_blog.yml * Update: proofread zh/unity-{api,asr}.md * Update zh/falcon.md Signed-off-by: Yao, Matrix <matrix.yao@intel.com> Update: zh/falcon.md * instruction-tuning-sd cn done (#21) Signed-off-by: Yao, Matrix <matrix.yao@intel.com> * Update: zh/instruction-tuning-sd.md * fine-tune-whisper cn done (#23) Signed-off-by: Yao, Matrix <matrix.yao@intel.com> * Update: zh/fine-tune-whisper.md * add mms_adapters and policy (#22) Update: zh/policy-ntia-rfc.md * Update: refine zh/mms_adapters.md Update: remove incompleted file * Update: zh/llm-leaderboard.md, zh/autoformer.md * add llm-leaderboard CN translation * add CN translation for autoformer * Update: proofreading zh/autoformer.md * BridgeTower blog post (huggingface#1118) * Update BridgeTower blog post (huggingface#1277) * LLM Eval: minor typos and nits (huggingface#1263) * Fix anchor link to custom pipeline section. (huggingface#485) * Update: zh/llm-leaderboard.md, zh/autoformer.md * add llm-leaderboard CN translation * add CN translation for autoformer Update: proofreading zh/autoformer.md Update: proofreading zh/llm-leaderboard.md * Update: proofreading zh/ethics-soc-4.md * Update "How to deploy LLM" blog post to use `huggingface_hub` in example (huggingface#1290) * Use InferenceClient from huggingface_hub * Update inference-endpoints-llm.md Co-authored-by: Pedro Cuenca <pedro@huggingface.co> --------- Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update BridgeTower blog post (huggingface#1295) * Removed duplicate numbering (huggingface#1171) * Update: zh/evaluating-mmlu-leaderboard.md Signed-off-by: Yao, Matrix <matrix.yao@intel.com> Co-authored-by: Zhongdong Yang <zhongdong_y@outlook.com> Update: proofreading zh/evaluating-mmlu-leaderboard.md * Translate train-optimize-sd-intel.md to zh (#16) * Translate "stackllama" into Chinese * Create train-optimize-sd-intel.md Add new Update: zh/train-optimize-sd-intel.md --------- Signed-off-by: Yao, Matrix <matrix.yao@intel.com> Signed-off-by: Yang, Zhongdong <zhongdong_y@outlook.com> Co-authored-by: innovation64 <liyang19991126@126.com> Co-authored-by: Yao, Matrix <matrix.yao@intel.com> Co-authored-by: SuSung-boy <872414318@qq.com> Co-authored-by: Luke Cheng <2258420+chenglu@users.noreply.github.com> Co-authored-by: yaoqih <40328311+yaoqih@users.noreply.github.com> Co-authored-by: Shiliang Chen <36809537+csl122@users.noreply.github.com> Co-authored-by: Alara Dirik <8944735+alaradirik@users.noreply.github.com> Co-authored-by: Omar Sanseviero <osanseviero@gmail.com> Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> Co-authored-by: 李洋 <45715979+innovation64@users.noreply.github.com> Co-authored-by: Yao Matrix <yaoweifeng0301@126.com> Co-authored-by: Hoi2022 <120370631+Hoi2022@users.noreply.github.com> Co-authored-by: Nicolas Patry <patry.nicolas@protonmail.com> Co-authored-by: Luc Georges <McPatate@users.noreply.github.com> Co-authored-by: DeltaPenrose <128761972+DeltaPenrose@users.noreply.github.com> Co-authored-by: Victor Muštar <victor.mustar@gmail.com> Co-authored-by: Kashif Rasul <kashif.rasul@gmail.com> Co-authored-by: Julien Simon <3436143+juliensimon@users.noreply.github.com> Co-authored-by: Julien Simon <julsimon@huggingface.co> Co-authored-by: Pedro Cuenca <pedro@huggingface.co> Co-authored-by: gxy-gxy <57594446+gxy-gxy@users.noreply.github.com> Co-authored-by: regisss <15324346+regisss@users.noreply.github.com> Co-authored-by: Lucain <lucainp@gmail.com> Co-authored-by: Eswar Divi <76403422+EswarDivi@users.noreply.github.com> Co-authored-by: Qi Zhang <82949744+Vermillion-de@users.noreply.github.com>
huggingface-cn · Jul 12, 2023 · 7667c29 · 7667c29
1 parent 0dd3d70
commit 7667c29
Show file tree

Hide file tree

Showing 2 changed files with 131 additions and 9 deletions.
diff --git a/zh/_blog.yml b/zh/_blog.yml
@@ -571,6 +571,18 @@
     - community
     - research
 
+- local: generative-ai-models-on-intel-cpu
+  title: "越小越好：Q8-Chat，在英特尔至强 CPU 上体验高效的生成式 AI"
+  thumbnail: /blog/assets/143_q8chat/thumbnail.png
+  author: andyll7772
+  date: May 16, 2023
+  tags:
+    - llm
+    - nlp
+    - inference
+    - intel
+    - quantization
+
 - local: instruction-tuning-sd
   title: "使用 InstructPix2Pix 对 Stable Diffusion 进行指令微调"
   author: sayakpaul
@@ -583,17 +595,17 @@
     - research
     - guide 
 
-- local: generative-ai-models-on-intel-cpu
-  title: "越小越好：Q8-Chat，在英特尔至强 CPU 上体验高效的生成式 AI"
-  thumbnail: /blog/assets/143_q8chat/thumbnail.png
-  author: andyll7772
-  date: May 16, 2023
+- local: train-optimize-sd-intel
+  title: "基于 NNCF 和 🤗 Optimum 面向 Intel CPU 对 Stable Diffusion 优化"
+  author: AlexKoff88
+  thumbnail: /blog/assets/train_optimize_sd_intel/thumbnail.png
+  date: May 25, 2023
   tags:
-    - llm
-    - nlp
-    - inference
+    - diffusers
+    - cpu
     - intel
-    - quantization
+    - guide
+      - quantization
 
 - local: unity-asr
   title: "如何在 Unity 游戏中集成 AI 语音识别？"

diff --git a/zh/train-optimize-sd-intel.md b/zh/train-optimize-sd-intel.md
@@ -0,0 +1,110 @@
+---
+title: 基于 NNCF 和 🤗 Optimum 面向 Intel CPU 对 Stable Diffusion 优化
+thumbnail: /blog/assets/train_optimize_sd_intel/thumbnail.png
+authors:
+- user: AlexKoff88
+  guest: true
+- user: MrOpenVINO
+  guest: true
+- user: helenai
+  guest: true
+- user: sayakpaul
+- user: echarlaix
+translators:
+- user: vermillion
+- user: zhongdongy
+  proofreader: true
+---
+
+# 基于 NNCF 和 🤗 Optimum 面向 Intel CPU 对 Stable Diffusion 优化
+
+基于隐空间的扩散模型 (Latent Diffusion Model)，是解决文本到图片生成问题上的颠覆者。Stable Diffusion 是最著名的一例，广泛应用在商业和工业。Stable Diffusion 的想法简单且有效: 从噪声向量开始，多次去噪，以使之在隐空间里逼近图片的表示。
+
+但是，这样的方法不可避免地增加了推理时长，使客户端的体验大打折扣。众所周知，一个好的 GPU 总能有帮助，确实如此，但其损耗大大增加了。就推理而言，在 2023 年上半年 (H1’23)，一个好 CPU 实例 ([r6i.2xlarge](https://aws.amazon.com/ec2/instance-types/r6g/)，8 vCPUs ，64 GB 内存) 价格是 0.504 $/h，同时，类似地，一个好 GPU 实例 ([g4dn.2xlarge](https://aws.amazon.com/ec2/instance-types/g4/)，NVIDIA T4，16 GB 内存) 价格是 0.75 $/h ，是前者的近 1.5 倍。
+
+这就使图像生成的服务变得昂贵，无论持有者还是用户。该问题在面向用户端部署就更突出了: 可能没有 GPU 能用！这让 Stable Diffusion 的部署变成了棘手的问题。
+
+在过去五年中，[OpenVINO](https://docs.openvino.ai/) 集成了许多高性能推理的特性。 其一开始为计算机视觉模型设计，现今仍在许多模型的推理性能上取得最佳表现，包括 [Stable Diffusion](https://huggingface.co/blog/zh/stable-diffusion-inference-intel)。然而，对资源有限型的应用，优化 Stable Diffusion 远不止运行时的。这也是 OpenVINO [NNCF](https://github.com/openvinotoolkit/nncf)(Neural Network Compression Framework) 发挥作用的地方。
+
+在本博客中，我们将理清优化 Stable Diffusion 模型的问题，并提出对资源有限的硬件 (比如 CPU) 减负的流程。尤其是和 PyTorch 相比，我们速度提高了 5.1 倍，内存减少了 4 倍。
+
+## Stable Diffusion 的优化
+
+在 Stable Diffusion 的 [管线](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/overview) 中，UNet 的运行是最计算昂贵的。因此，对模型的推理速度，针对 UNet 的优化能带来足够的效益。
+
+然而事实表明，传统的模型优化方法如 8-bit 的后训练量化，对此不奏效。主要原因有两点: 其一，面向像素预测的模型，比如语义分割、超分辨率等，是模型优化上最复杂的，因为任务复杂，参数和结构的改变会导致无数种变数; 其二，模型的参数不是很冗余，因为其压缩了其数以千万计的 [数据集](https://laion.ai/blog/laion-5b/) 中的信息。这也是研究者不得不用更复杂的量化方法来保证模型优化后的精度。举例而言，高通 (Qualcomm) 用分层知识蒸馏 (layer-wise Knowledge Distillation) 方法 ([AdaRound](https://arxiv.org/abs/2004.10568)) 来 [量化](https://www.qualcomm.com/news/onq/2023/02/worlds-first-on-device-demonstration-of-stable-diffusion-on-android) Stable Diffusion。这意味着，无论如何，模型量化后的微调是必要的。既然如此，为何不用 量化感知的训练 ([Quantization-Aware Trainning, QAT](https://arxiv.org/abs/1712.05877))，其对原模型的微调和参数量化是同时进行的？因此，我们在本工作中，用 token 合并 ([Token Merging](https://arxiv.org/abs/2210.09461)) 方法结合 [NNCF](https://github.com/openvinotoolkit/nncf), [OpenVINO](https://www.intel.com/content/www/us/en/developer/tools/openvino-toolkit/overview.html) 和 [Diffusers](https://github.com/huggingface/diffusers) 实践了该想法。
+
+## 优化流程
+
+我们通常从训练后的模型开始优化。在此，我们从宝可梦数据集 ([Pokemons dataset](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions)，包含图片和对应的文本描述) 上微调的 [模型](https://huggingface.co/svjack/Stable-Diffusion-Pokemon-en)。
+
+我们对 Stable Diffusion 用 Diffusers 中的 [图片 - 文本微调之例](https://huggingface.co/docs/diffusers/training/text2image)，结合 NNCF 中的 QAT (参见训练的 [脚本](https://github.com/huggingface/optimum-intel/tree/main/examples/openvino/stable-diffusion))。我们同时改变了损失函数，以同时实现从源模型到部署模型的知识蒸馏。该方法与通常的知识蒸馏不同，后者是把源模型蒸馏到小些的模型。我们的方法主要将知识整理作为附加的方法，帮助提高最后优化的模型的精度。我们也用指数移动平均方法 (Exponential Moving Average, EMA) 让我们训练过程更稳定。我们仅对模型做 4096 次迭代。
+
+基于一些技巧，比如梯度检查 (gradient checkpointing) 和 [保持 EMA 模型](https://github.com/huggingface/optimum-intel/blob/bbbe7ff0e81938802dbc1d234c3dcdf58ef56984/examples/openvino/stable-diffusion/train_text_to_image_qat.py#L941) 在内存 (RAM) 而不是虚拟内存 (VRAM) 中。整个优化过程能用一张 GPU 在一天内完成。
+
+## 量化感知的训练之外呢 ?
+
+量化模型本身就能带来模型消耗、加载、内存、推理速度上的显著提高。但量化模型蛮大的优势在能和其他模型优化方法一起，达到加速的增益效果。
+
+最近，Facebook Research 针对视觉 Transformer 模型，提出了一个 [Token Merging](https://arxiv.org/abs/2210.09461) 方法。该方法的本质是用现有的方法 (取平均、取最大值等) 把冗余的 token 和重要的 token 融合。这在 self-attention 块之前完成，后者是 Transformer 模型最消耗算力的部分。因此，减小 token 的跨度能减少 self-attention 块消耗的时间。该方法也已被 Stable Diffusion 模型 [采用](https://arxiv.org/pdf/2303.17604.pdf)，并在面向 GPU 的高分辨率优化上有可观的表现。
+
+我们改进了 Token Merging 方法，以便用 OpenVINO，并在注意力 UNet 模型上采用 8-bit 量化。这包含了上述含知识蒸馏等的所有技术。对量化而言，其需要微调，以保证数值精度。我们也从 [宝可梦数据集](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions) 上训练的 [模型](https://huggingface.co/svjack/Stable-Diffusion-Pokemon-en) 开始优化和微调。下图体现了总体的优化工作流程。
+
+![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/train-optimize-sd-intel/overview.png)
+
+结果的模型在有限资源的硬件上是高度有效的，如客户机或边缘 CPU。如上文所述，把 Token Merging 方法和量化方法叠加能带来额外的推理增益。
+
+<div class="flex flex-row">
+<div class="grid grid-cols-2 gap-4">
+<figure>
+<img class="max-w-full rounded-xl border-2 border-solid border-gray-600" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/train-optimize-sd-intel/image_torch.png" alt="Image 1" />
+<figcaption class="mt-2 text-center text-sm text-gray-500">PyTorch FP32，推理时长：230.5 秒，内存消耗：3.44 GB</figcaption>
+</figure>
+<figure>
+<img class="max-w-full rounded-xl border-2 border-solid border-gray-600" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/train-optimize-sd-intel/image_fp32.png" alt="Image 2" />
+<figcaption class="mt-2 text-center text-sm text-gray-500">OpenVINO FP32，推理时长：120 秒 (<b>1.9 倍</b>)，内存消耗：3.44 GB</figcaption>
+</figure>
+<figure>
+<img class="max-w-full rounded-xl border-2 border-solid border-gray-600" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/train-optimize-sd-intel/image_quantized.png" alt="Image 3" />
+<figcaption class="mt-2 text-center text-sm text-gray-500">OpenVINO 8-bit，推理市场：59 秒 (<b>3.9 倍</b>)，内存消耗：0.86 GB(<b>0.25 倍</b>)</figcaption>
+</figure>
+<figure>
+<img class="max-w-full rounded-xl border-2 border-solid border-gray-600" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/train-optimize-sd-intel/image_tome_quantized.png" alt="Image 4" />
+<figcaption class="mt-2 text-center text-sm text-gray-500">ToMe + OpenVINO 8-bit, 推理速度：44.6 秒 (<b>5.1 倍</b>)，内存消耗：0.86 GB (<b>0.25 倍</b>)</figcaption>
+</figure>
+</div>
+</div>
+
+用不同模型优化方法的图片生成的结果 [展示](https://huggingface.co/spaces/AlexKoff88/stable_diffusion)。输入提示词为 “cartoon bird”，随机种子为 42。模型用 OpenVINO 2022.3，来自 [Hugging Face Space](https://huggingface.co/docs/hub/spaces-overview)，用“CPU 升级”的实例: 第三代 Intel® Xeon® Scalable Processors，和 Intel® 深度学习加速技术。
+
+## 结果
+
+我们用优化模型不完整的流程以得到两种模型: 基于 8-bit 量化的和基于 Token Merging 量化的，并和 PyTorch 作为基准比较。我们也把基准先转化成 vanilla OpenVINO (FP32) 的模型，以用以分析性比较。
+
+上面的结果图展示了图像生成和部分模型的特性。如你所见，仅转化成 OpenVINO 就带来大的推理速度提高 ( **1.9** 倍)。用基于 8-bit 的量化加速和 PyTorch 相比带来了 **3.9** 倍的推理速度。量化的另外一个重要提高在于内存消耗减少，0.25 倍之于 PyTorch，同时也提高了加载速度。在量化之上应用 Token Merging (ToME) (**融合比为 0.4**) 带来了 **5.1 倍** 的提速，同时把模型内存消耗保持在原水平上。我们不提供输出结果上的质量改变，但如你所见，结果还是有质量的。
+
+下面我们展示将最终优化结果部署在 Intel CPU 上代码。
+
+```python
+from optimum.intel.openvino import OVStableDiffusionPipeline
+
+# Load and compile the pipeline for performance.
+name = "OpenVINO/stable-diffusion-pokemons-tome-quantized-aggressive"
+pipe = OVStableDiffusionPipeline.from_pretrained(name, compile=False)
+pipe.reshape(batch_size=1, height=512, width=512, num_images_per_prompt=1)
+pipe.compile()
+
+# Generate an image.
+prompt = "a drawing of a green pokemon with red eyes"
+output = pipe(prompt, num_inference_steps=50, output_type="pil").images[0]
+output.save("image.png")
+
+```
+
+在 Hugging Face [Optimum Intel](https://huggingface.co/docs/optimum/main/en/intel/index) 库中你可以找到训练和量化 [代码](https://github.com/huggingface/optimum-intel/tree/main/examples/openvino/stable-diffusion)。比较优化过的和原模型的 notebook 代码在 [这里](https://github.com/huggingface/optimum-intel/blob/main/notebooks/openvino/stable_diffusion_optimization.ipynb)。你可以在 Hugging Face Hub 上找到 [OpenVINO](https://huggingface.co/OpenVINO) 下的 [许多模型](https://huggingface.co/models?library=openvino&sort=downloads)。另外，我们在 Hugging Face Spaces 上建了一个 [demo](https://huggingface.co/spaces/AlexKoff88/stable_diffusion)，以运行 [带第三代 Intel Xeon Scalable 的 r6id.2xlarge 实例]。
+
+## 一般的 Stable Diffusion 模型呢？
+
+正如我们在宝可梦图像生成任务中展现的一样，仅用小量的训练资源，对 Stable Diffusion 管线实现高层次的优化是可能的。同时，众所周知，训练一般的 Stable Diffusion 模型是一个 [昂贵的任务](https://www.mosaicml.com/blog/training-stable-diffusion-from-scratch-part-2)。但是，有充足的资金和硬件资源，用上述方法优化一般的模型生成高分辨率的模型是可能的。我们唯一的警告是关于 Token Merging 方法，其会减弱模型容忍性。这里衡量标准是，训练数据越复杂，优化模型时的融合比就该越小。
+
+如果你乐于读本博客，那你可能对另外一篇 [博客](https://huggingface.co/blog/zh/stable-diffusion-inference-intel) 感兴趣，它讨论了在第四代 Intel Xeon CPU 上其他互补的 Stable Diffusion 模型优化方法。