GitLab

Enable 720p model inference on low-spec GPUs/CPUs and accelerate T5/CLIP quantized models with vLLM operators

* add step & cfg distillation wan model

* bug fixed

* reconstruct quantization and fix memory leak bug.

* Support lazy load inference.

* reconstruct quantization

* Fix hunyuan bugs

* deleted tmp file

---------
Co-authored-by: root <root@pt-c0b333b3a1834e81a0d4d5f412c6ffa1-worker-0.pt-c0b333b3a1834e81a0d4d5f412c6ffa1.ns-devsft-3460edd0.svc.cluster.local>
Co-authored-by: gushiqiao <gushqiaio@sensetime.com>
Co-authored-by: gushiqiao <gushiqiao@sensetime.com>

* fix offload extra latency in the first step by pre-allocating pinned memory

* pre-commit

---------
Co-authored-by: “de1star” <“843414674@qq.com”>

* fix offload

* fix offload

---------
Co-authored-by: “de1star” <“843414674@qq.com”>

* function hunyuan_t2v_tea, hunyuan_t2v_taylorseer, modify the fresh_threshold of taylorseer

* hunyuan i2v,t2v + tea,tay; wan i2v,t2v + tea function, add log files

* 删除了TeaCace Scheduler的多余属性

* 删除了多余目录

* 修复了TeaCaching部分的bug,目前t2v, i2v feature caching均可跑通

* Update attn_weight.py

---------
Co-authored-by: Yang Yong(雍洋) <yongyang1030@163.com>

* Support load advance ptq model.

* Update run_wan_i2v_advanced_ptq.sh

---------
Co-authored-by: gushiqiao <gushiqiao@sensetime.com>
Co-authored-by: Yang Yong(雍洋) <yongyang1030@163.com>

* [feature]: add Wan Sparge infer

* Update scripts/run_wan_t2v_sparge.sh
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* [minor]: fix typo and use config style

* [minor]: remove breakpoint

* [feature]: add all attn class

* [minor]: remove args

* [minor]: remove shared weights

---------
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* [bugs fixed] fixed bugs for cpu offload.

* [rename] rename causal_model -> causvid_model

* [feature] add prompt enhancer

* [feature] add prompt enhancer

* [rename] rename causal_model -> causvid_model

* Support hunyuan offload and teacache.

* Fix

* Fix

---------
Co-authored-by: gushiqiao <gushiqiao@sensetime.com>

Co-authored-by: gushiqiao <gushiqiao@sensetime.com>

* [minor]: optimize dockerfile for fewer layer

* [feature]: add pre-commit lint, update readme for contribution guidance

* [minor]: fix run shell privileges

* [auto]: first lint without rule F, fix rule E

* [minor]: fix docker file error

Co-authored-by: gushiqiao <gushiqiao@sensetime.com>

* 修改了main.py, t5的model, wan的model、三个weights文件和三个infer文件, 并且在common的conv3d算子中注册新算子

* 修改了Conv3dWeightForceBF16算子，更新了wan的pre_weights中对此算子的使用

* 修复了import中的bug

* 修复了WanPreWeights, WanTransformerWeights没有self.config的bug

* 修复了WanPreWeights, WanTransformerWeights没有self.config的bug

* 修复了config的bug，目前在使用cpu_offload的时候，vae阶段有tensor不在同一device的bug

* 修复了vae阶段迁移的bug

* 修复了scale在mean和inv_std迁移后仍需重新赋值的bug