anodyine/vllm-omni
1
0 Fork 0
Code Issues 4 Pull Requests 2 Actions Packages Projects Releases Wiki Activity

[BugFix][Nightly CI] Adjust Qwen-Image accuracy thresholds for stable FA-deterministic mode (#5963)

Browse Source 
Signed-off-by: NumberWan <wantszkin2003@gmail.com>
This commit is contained in:
NumberWan authored and GitHub committed 2026-08-15 09:38:30 +00:00
1 parent e57574ae29
commit f39b8684eb
1 file changed
+7 -7
Download Patch File Download Diff File Expand all files Collapse all files
tests/e2e/accuracy/test_qwen_image.py
+7 -7
View File
@@ -29,14 +29,14 @@ HEIGHT = 512
NUM_INFERENCE_STEPS = 20
TRUE_CFG_SCALE = 4.0
SEED = 42
SSIM_THRESHOLD = 0.97
# The FA3-hub-matched comparison clears 0.97 with margin. The SDPA-matched
# fallback (hub kernel unavailable for the image's torch) landed at 0.9687 in
# build 2954 — the attention math matches but the surrounding stacks (vLLM
# compiled path vs eager diffusers) still differ slightly. Keep the strict
# gate for the FA3 path and a slightly looser one for the fallback only.
SSIM_THRESHOLD = 0.94
# With --fa-deterministic, H100 nightlies stably land near SSIM 0.958 / PSNR 27.8
# vs Diffusers (see #5734 / #5963), so the primary gates are 0.94 / 27.0.
# The SDPA-matched fallback (hub kernel unavailable for the image's torch)
# landed at SSIM 0.9687 in build 2954 — keep a slightly tighter SSIM gate for
# that path only.
SSIM_THRESHOLD_SDPA_FALLBACK = 0.96
PSNR_THRESHOLD = 30.0
PSNR_THRESHOLD = 27.0
MODEL_2512_ID = "Qwen/Qwen-Image-2512"
MODEL_2512_ENV_VAR = "QWEN_IMAGE_2512_MODEL"