drop the sageattention wheel: this GPU class runs cuDNN (8/20 bench); runtime falls back gracefully

README.md

@@ -23,8 +23,9 @@ DiT + streaming VAE pipeline, served on ZeroGPU.
   module scope under ZeroGPU's CUDA-emulation layer.
 - Each session leases a real GPU in a `@spaces.GPU` fork; the main process is a
   thin WebSocket byte-pipe (H.264 both ways, MJPEG fallback).
-- Custom CUDA kernels (`joyomni_ops`, `sageattention`) ship as prebuilt cp310 /
-  torch-2.9.1 wheels in `wheels/` — matching ZeroGPU's supported stack.
+- The custom CUDA kernels (`joyomni_ops`: FP8 GEMM + fused norm/rope) ship as a
+  prebuilt cp310 / torch-2.9.1 wheel in `wheels/` — matching ZeroGPU's supported
+  stack. Attention runs on plain cuDNN SDPA (fastest on this GPU class).
 
 ## Notes
 
@@ -33,5 +34,5 @@ DiT + streaming VAE pipeline, served on ZeroGPU.
   not the first user session.
 - `torch.compile` is inert in ZeroGPU forks; each fork instead pays a short warmup
   (CUDA-graph capture, KV fill), so the first 2-3 chunks are slower.
-- A session runs up to ~150s of GPU time. Free ZeroGPU quota is limited — this is a
-  small-scale demo.
+- A session runs up to ~120s of GPU time (shorter retries when the queue is busy).
+  Free ZeroGPU quota is limited — this is a small-scale demo.

app.py

@@ -10,7 +10,7 @@ to ZeroGPU's fork model, without modifying the code it vendors:
     in-process.
 
 Hardware: ZeroGPU size="xlarge" = RTX Pro 6000 Blackwell (sm_120, 96GB);
-the sm_120 joyomni_ops + SageAttention wheels in wheels/ match it.
+the sm_120 joyomni_ops wheel in wheels/ matches it.
 """
 from __future__ import annotations
 
@@ -41,9 +41,10 @@ os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
 os.environ.setdefault("GRADIO_SSR_MODE", "false")
 
 # Kernel path: FP8 + CUDA graph on plain cuDNN SDPA. Sage int8 stays OFF — the Space GPU
-# is an RTX PRO 6000 Blackwell (probe-verified), where cuDNN beats sage at every serving
-# shape (2026-08-20 bench; sage only wins on GeForce's half-rate fp32-accum). The wheel
-# still ships, so JOYOMNI_SAGE_ATTN=1 in Space settings can re-enable it for an A/B.
+# is an RTX PRO 6000 Blackwell (probe-verified), where cuDNN beats sage across the 480p
+# serving shapes (2026-08-20 bench; sage only wins on GeForce's half-rate fp32-accum).
+# The sage wheel is no longer shipped (recover from git history if a GeForce tier ever
+# appears); JOYOMNI_SAGE_ATTN=1 without it just logs a fallback to SDPA (dit.py).
 os.environ.setdefault("JOYOMNI_FP8_IMG", "1")
 os.environ.setdefault("JOYOMNI_FP8_TXT", "1")
 os.environ.setdefault("JOYOMNI_CUDA_GRAPH", "1")
@@ -86,7 +87,7 @@ CKPT_ROOT = Path(os.environ.get("JOYOMNI_CKPT_ROOT", str(_DATA / "checkpoints"))
 SESSION_DURATION = int(os.environ.get("JOYOMNI_SPACES_DURATION", "120"))
 
 
-# Prebuilt sm_120 wheels: pip runs before repo files exist on Spaces — install here.
+# Prebuilt sm_120 wheel: pip runs before repo files exist on Spaces — install here.
 def _ensure_local_wheels() -> None:
     import importlib.util
     import subprocess
@@ -94,7 +95,6 @@ def _ensure_local_wheels() -> None:
     wheel_dir = DEPLOY_ROOT / "wheels"
     specs = [
         ("joyomni_ops", "joyomni_ops-0.1.0-cp310-cp310-linux_x86_64.whl"),
-        ("sageattention", "sageattention-2.2.0-cp310-cp310-linux_x86_64.whl"),
     ]
     for mod, whl in specs:
         if importlib.util.find_spec(mod) is not None:

wheels/sageattention-2.2.0-cp310-cp310-linux_x86_64.whl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:983e1f21123b81916f110077e19982f97c69e0de4fb0bdf7ef8501baf0c6f9eb
-size 14974158