Spaces:
Running on Zero
Running on Zero
triple-fallback artifact export + triton cache
Browse filesAfter compile, tries in order:
1. AOTInductor aoti_compile_and_package → .pt2 (best)
2. torch.compiler.save_cache_artifacts → mega-cache .bin (PyTorch 2.7+)
3. In-memory only (fallback)
Also: TRITON_CACHE_DIR set for kernel persistence
Download button finds all artifact types
app.py
CHANGED
|
@@ -52,9 +52,10 @@ try:
|
|
| 52 |
del _torch
|
| 53 |
except ImportError:
|
| 54 |
pass
|
| 55 |
-
# Persist
|
| 56 |
_inductor_cache = os.path.join(os.path.expanduser("~"), ".cache", "corridorkey", "inductor")
|
| 57 |
os.environ.setdefault("TORCHINDUCTOR_CACHE_DIR", _inductor_cache)
|
|
|
|
| 58 |
|
| 59 |
# Workaround: Gradio cache_examples bug with None outputs.
|
| 60 |
_original_read_from_flag = gr.components.Component.read_from_flag
|
|
@@ -380,17 +381,29 @@ def _load_greenformer(img_size):
|
|
| 380 |
time.time() - t_compile, e)
|
| 381 |
torch.cuda.empty_cache()
|
| 382 |
|
| 383 |
-
# Export compiled
|
| 384 |
pt2_path = f"/tmp/corridorkey_{img_size}.pt2"
|
|
|
|
|
|
|
|
|
|
| 385 |
try:
|
| 386 |
-
unwrapped = model._orig_mod if hasattr(model, '_orig_mod') else model
|
| 387 |
-
dummy_export = torch.zeros(1, 4, img_size, img_size, dtype=torch.float16, device="cuda")
|
| 388 |
exported = torch.export.export(unwrapped, (dummy_export,))
|
| 389 |
-
torch.
|
| 390 |
-
logger.info("
|
| 391 |
pt2_path, os.path.getsize(pt2_path) / 1024 / 1024)
|
| 392 |
except Exception as e:
|
| 393 |
-
logger.info("
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 394 |
|
| 395 |
logger.info("GreenFormer loaded on CUDA (img_size=%d)", img_size)
|
| 396 |
return model
|
|
@@ -1027,7 +1040,7 @@ with gr.Blocks(title="CorridorKey") as demo:
|
|
| 1027 |
pt2_btn = gr.Button("Check for .pt2", size="sm")
|
| 1028 |
def _get_pt2():
|
| 1029 |
import glob
|
| 1030 |
-
files = glob.glob("/tmp/corridorkey_*.
|
| 1031 |
return files if files else None
|
| 1032 |
pt2_btn.click(fn=_get_pt2, outputs=[pt2_download])
|
| 1033 |
|
|
|
|
| 52 |
del _torch
|
| 53 |
except ImportError:
|
| 54 |
pass
|
| 55 |
+
# Persist compilation caches across ZeroGPU sessions
|
| 56 |
_inductor_cache = os.path.join(os.path.expanduser("~"), ".cache", "corridorkey", "inductor")
|
| 57 |
os.environ.setdefault("TORCHINDUCTOR_CACHE_DIR", _inductor_cache)
|
| 58 |
+
os.environ.setdefault("TRITON_CACHE_DIR", os.path.join(os.path.expanduser("~"), ".cache", "corridorkey", "triton"))
|
| 59 |
|
| 60 |
# Workaround: Gradio cache_examples bug with None outputs.
|
| 61 |
_original_read_from_flag = gr.components.Component.read_from_flag
|
|
|
|
| 381 |
time.time() - t_compile, e)
|
| 382 |
torch.cuda.empty_cache()
|
| 383 |
|
| 384 |
+
# Export compiled artifacts for future pre-compiled loading
|
| 385 |
pt2_path = f"/tmp/corridorkey_{img_size}.pt2"
|
| 386 |
+
unwrapped = model._orig_mod if hasattr(model, '_orig_mod') else model
|
| 387 |
+
dummy_export = torch.zeros(1, 4, img_size, img_size, dtype=torch.float16, device="cuda")
|
| 388 |
+
# Method 1: AOTInductor .pt2 (best — standalone compiled artifact)
|
| 389 |
try:
|
|
|
|
|
|
|
| 390 |
exported = torch.export.export(unwrapped, (dummy_export,))
|
| 391 |
+
torch._inductor.aoti_compile_and_package(exported, package_path=pt2_path)
|
| 392 |
+
logger.info("AOTInductor .pt2 exported: %s (%.1f MB)",
|
| 393 |
pt2_path, os.path.getsize(pt2_path) / 1024 / 1024)
|
| 394 |
except Exception as e:
|
| 395 |
+
logger.info("AOTInductor export failed (%s), trying mega-cache...", type(e).__name__)
|
| 396 |
+
# Method 2: Portable mega-cache (PyTorch 2.7+)
|
| 397 |
+
try:
|
| 398 |
+
artifact_bytes, cache_info = torch.compiler.save_cache_artifacts()
|
| 399 |
+
cache_path = f"/tmp/corridorkey_{img_size}_cache.bin"
|
| 400 |
+
with open(cache_path, "wb") as f:
|
| 401 |
+
f.write(artifact_bytes)
|
| 402 |
+
logger.info("Mega-cache saved: %s (%.1f MB, %s)",
|
| 403 |
+
cache_path, len(artifact_bytes) / 1024 / 1024, cache_info)
|
| 404 |
+
except Exception as e2:
|
| 405 |
+
logger.info("Mega-cache also failed (%s) — triton kernels cached in memory only",
|
| 406 |
+
type(e2).__name__)
|
| 407 |
|
| 408 |
logger.info("GreenFormer loaded on CUDA (img_size=%d)", img_size)
|
| 409 |
return model
|
|
|
|
| 1040 |
pt2_btn = gr.Button("Check for .pt2", size="sm")
|
| 1041 |
def _get_pt2():
|
| 1042 |
import glob
|
| 1043 |
+
files = glob.glob("/tmp/corridorkey_*.*")
|
| 1044 |
return files if files else None
|
| 1045 |
pt2_btn.click(fn=_get_pt2, outputs=[pt2_download])
|
| 1046 |
|