[Blog] Add "Full-Stack Performance Optimization of AR+DiT in SGL-Diffusion"

[ping1jing2]

diffusion blog first version

[BBuf]

A few documentation and reproducibility issues to address before publishing.

[BBuf]

A few additional comments on the figures.

[ping1jing2]

i addressed all the comments above, but i think the below items need to be double check:

1. PR #30683 is still open and #31320 is still a draft. Please mark these sections experimental and pin commit SHAs, or publish after they merge.----added a note below. file location

   To resolve these issues, we contributed three progressively staged PRs that evolve the system from a monolith to a fully decoupled, heterogeneous distributed architecture:

   Note: PR #31320 has not been merged yet. The commit SHA we adopt here is 4e520bd

2. Please qualify optimal with the tested workload/configuration; the appendix later uses 2-NPU SP denoisers, which contradicts this absolute claim.-----added test result below. file location

   DiT denoising is optimal at batch=1 on a single NPU for both latency and throughput.

   Batch size	AR (s)	Denoising, 30 step (s)	Denoising, 30 steps (s)
   1	20.4	0.407	12.2
   2	21.3 (+4.4%)	0.854 (+110%)	25.6 (+110%)
   4	22.8 (+12%)	1.98 (+386%)	59.6 (+389%)
   8	25.9 (+27%)	3.73 (+816%)	112.2 (+820%)
   16	29.4 (+44%)	7.24 (+1679%)	217.3 (+1681%)
   32	33.2 (+63%)	14.0 (+3339%)	420.6 (+3348%)