[Diffusion] Improve qwen image edit performace to align with LightX2V

[BBuf]

main

sglang generate --model-path Qwen/Qwen-Image-Edit-2511 --prompt "Change the person to a standing position, bending over to hold the dog's front paws."  --image-path "/home/lmsys/bbuf/LightX2V/examples/qwen_image/1.png"

100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 40/40 [00:30<00:00,  1.29it/s]
[12-25 07:11:20] [DenoisingStage] average time per step: 0.7722 seconds
[12-25 07:11:20] [DenoisingStage] finished in 30.8956 seconds
[12-25 07:11:20] [DecodingStage] started...
[12-25 07:11:20] [DecodingStage] finished in 0.5390 seconds
[12-25 07:11:20] Output saved to outputs/Change_the_person_to_a_standing_position_bending_over_to_hold_the_dog_s_front_paws._20251225-071047_08c591d4.jpg
[12-25 07:11:20] Pixel data generated successfully in 33.13 seconds

pr

sglang generate --model-path Qwen/Qwen-Image-Edit-2511 --prompt "Change the person to a standing position, bending over to hold the dog's front paws."  --image-path "/home/lmsys/bbuf/LightX2V/examples/qwen_image/1.png"

[12-25 07:00:34] [DenoisingStage] started...
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 40/40 [00:25<00:00,  1.58it/s]
[12-25 07:00:59] [DenoisingStage] average time per step: 0.6327 seconds
[12-25 07:00:59] [DenoisingStage] finished in 25.3114 seconds
[12-25 07:00:59] [DecodingStage] started...
[12-25 07:01:00] [DecodingStage] finished in 0.5667 seconds

Motivation

use upstream fa3 , not sgl-kernel fa3 : 1.7ms->1.2ms

main:

pr:

flashinfer rope: 241us->82us

main

pr:

revert pack qkv to avoid unaligned cat kernel: 154.5us->70us

main:

pr:

fuse_scale_shift_gate_select01_kernel: 148us->66us。

main:

pr:

one-step one-lyaer result

• LightX2V(5.1ms)

• sglang main(6.1ms)

• pr(5.19ms)

Each layer of every step is very close now.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[BBuf]

/tag-and-rerun-ci

[sunxxuns]

this PR failed AMD CI's cc @merrymercy @Ying1123 should not be merged.
As flashinfer now only supports Nvidia hardware, should not become a chokepoint for other hardwares. Thanks.
PS: fix in #16287