t5 encoder fp8 seems to be working

Dockerfile

@@ -53,6 +53,14 @@ RUN python3.11 -m pip install --no-cache-dir \
     "git+https://github.com/ModelTC/LightX2V.git" || \
     echo "LightX2V install failed — config.video.enabled must stay false until fixed"
 #
+# sgl-kernel (fp8 T5 encoder acceleration). The PyPI wheel lacks SM120
+# (Blackwell) CUTLASS kernels; use SGLang's cu128 wheel index instead.
+# Our wan22.py patches fp8_scaled_mm → torch._scaled_mm at runtime for
+# Blackwell GPUs, but the sgl_kernel package itself must still be present.
+RUN python3.11 -m pip install --no-cache-dir --no-deps \
+    "sgl-kernel @ https://github.com/sgl-project/whl/releases/download/v0.3.14.post1/sgl_kernel-0.3.14.post1%2Bcu128-cp310-abi3-manylinux2014_x86_64.whl" || \
+    echo "sgl-kernel install failed — fp8 T5 will fall back to bf16"
+#
 # MuseTalk (audio-driven lip-sync) — same story.
 RUN python3.11 -m pip install --no-cache-dir \
     "git+https://github.com/TMElyralab/MuseTalk.git" || \