Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio video: keep Wan2.1-T2V-1.3B in fp16 on fp16-only GPUs (#15147)

Daniel Han47b4c3e57240
diff --git a/studio/backend/core/inference/video_families.py b/studio/backend/core/inference/video_families.py
index ebde2871f28..d95209174bb 100644
--- a/studio/backend/core/inference/video_families.py
+++ b/studio/backend/core/inference/video_families.py
@@ -864,6 +864,8 @@ _FAMILIES: tuple[VideoFamily, ...] = (
resolution_presets = ((832, 480), (480, 832)),
bf16_components_gb = (2.8, 11.4, 0.5),
vae_force_fp32 = True,
+ # Finite in float16 like TI2V-5B (ComfyUI WAN21_T2V runs fp16): T4 keeps the DiT fp16, not the fp32 promotion.
+ fp16_guard = "native",
# UMT5-XXL pre-cast fp8, one file for all four Wan bases (diffusion_te_prequant._TE_EQUIVALENT_BASES), the
# int8 ConvRot one beside it (TE_INT8_CONVROT_FILES). fp8 by default, as ComfyUI's Wan templates load
# umt5_xxl_fp8_e4m3fn_scaled (workflow templates video_wan2_2_5B_ti2v.json, video_wan2_2_14B_t2v.json,
diff --git a/studio/backend/tests/test_diffusion_fp16_guard.py b/studio/backend/tests/test_diffusion_fp16_guard.py
index 23b31893165..c37deec10ed 100644
--- a/studio/backend/tests/test_diffusion_fp16_guard.py
+++ b/studio/backend/tests/test_diffusion_fp16_guard.py
@@ -56,7 +56,7 @@ def test_video_families_declare_only_native():
def test_video_capability_flags():
- for name in ("wan2.2-ti2v-5b", "hunyuanvideo-1.5", "hunyuanvideo-1.5-720p"):
+ for name in ("wan2.2-ti2v-5b", "wan2.1-t2v-1.3b", "hunyuanvideo-1.5", "hunyuanvideo-1.5-720p"):
assert guard.fp16_promotes_to_fp32(_video(name)) is False, name
# unmeasured families, and A14B (fp16 drifts from fp32 as far as bf16 does), keep the blanket promotion
for name in ("ltx-2", "minimax-h3", "wan2.2-t2v-a14b"):