revert batched-CFG attempt: FA3 has no sm_120 build; drop kernels dep (broke diffusers import)

app.py

@@ -74,17 +74,6 @@ def _read_example(name: str) -> str:
     return (EXAMPLES_DIR / name).read_text(encoding="utf-8").strip()
 
 
-# Warm the FA3 kernel cache (download only; the binary is loaded lazily inside
-# the GPU worker on first use). batch_cfg has a per-call sequential fallback.
-HAS_FA3 = True
-try:
-    from huggingface_hub import snapshot_download as _sd
-
-    _sd("kernels-community/flash-attn3", allow_patterns=["build/torch211-cxx11-cu130*", "*.json", "*.md"])
-except Exception as _fa3_exc:  # noqa: BLE001
-    print(f"[startup] FA3 kernel prefetch failed (lazy load will retry): {_fa3_exc}", flush=True)
-print(f"[startup] batched CFG (FA3) enabled: {HAS_FA3}", flush=True)
-
 print(f"[startup] downloading {MODEL_ID} (refiner excluded) ...", flush=True)
 t0 = time.perf_counter()
 model_dir = snapshot_download(
@@ -290,14 +279,9 @@ def _video_gpu_duration(caption, height, width, num_frames, steps, *args, **kwar
 
 
 def _run_pipe(target, **call_kwargs):
-    """Run with batched CFG (one packed forward per step) when FA3 is present;
-    fall back to sequential CFG if the packed path fails on this GPU."""
-    if HAS_FA3:
-        try:
-            return target(**call_kwargs, batch_cfg=True), "batched"
-        except Exception as exc:  # noqa: BLE001
-            print(f"[fallback] batch_cfg failed ({type(exc).__name__}: {str(exc)[:150]}), "
-                  "retrying sequential CFG", flush=True)
+    # Sequential CFG. Batched CFG requires FA3 varlen attention, and the
+    # kernels-community FA3 binaries only target sm_80/sm_90a — ZeroGPU's
+    # RTX PRO 6000 is sm_120, so the packed path cannot run here.
     return target(**call_kwargs, batch_cfg=False), "sequential"
 
 

lingbot_video/transformer_lingbot_video.py

@@ -15,10 +15,7 @@ from diffusers.models.modeling_outputs import Transformer2DModelOutput
 from diffusers.models.modeling_utils import ModelMixin
 
 try:
-    # Space-local shim: lazily loads the kernels-community FA3 build. Named
-    # fa3_shim because a metadata-less top-level flash_attn_interface module
-    # breaks diffusers' import-time flash-attn probing.
-    from fa3_shim import flash_attn_varlen_func as flash_attn_varlen_func_v3
+    from flash_attn_interface import flash_attn_varlen_func as flash_attn_varlen_func_v3
 except Exception:  # pragma: no cover - optional CUDA kernel.
     flash_attn_varlen_func_v3 = None
 

requirements.txt

@@ -10,4 +10,3 @@ numpy>=1.26
 pillow>=10.4.0
 scipy>=1.11
 json_repair>=0.30
-kernels>=0.6