minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[diffusion] FLUX.2 bit-exact residual-gate fast path (H200 klein-4B 50-step denoise -1.2%) (#33823)

Browse Source 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Xiaoyu Zhang
2026-08-06 22:54:40 +08:00
committed by GitHub
co-authored by Claude Fable 5
parent dd98c9572a
commit 591cfb0881
2 changed files with 51 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/kernels/ops/diffusion/test_residual_gate_add.py
+4 -1
View File
@@ -18,10 +18,13 @@ CASES = [
((1, 512, 4096), (1, 512, 4096)),
((1, 17, 65), (1, 1, 65)),
((1, 17, 65), (1, 17, 65)),
# FLUX.1 1024^2 shapes: dual-stream image/text and single-stream joint.
# FLUX.1 / FLUX.2-klein 1024^2 shapes (D=3072): dual-stream image/text
# and single-stream/joint concat; gates are [1, 1, D] modulation rows.
((1, 4096, 3072), (1, 1, 3072)),
((1, 512, 3072), (1, 1, 3072)),
((1, 4608, 3072), (1, 1, 3072)),
# FLUX.2-dev (D=6144) joint sequence.
((1, 4608, 6144), (1, 1, 6144)),
]