FireRed-Experimental-Turbo

Running on Zero

someone-in-the-world Claude Sonnet 4.6 commited on Jul 20

Commit

2f7a843

1 Parent(s): acdbece

Raise VAE tile threshold to HIGH_DETAIL.max_dim (2048px) to skip tiling

Tiling only activates when image size exceeds the threshold; at 768px the
threshold was always triggered for high_detail outputs (up to 2048px),
adding tile-split and blend overhead. FAST mode (<=768px) is unaffected.
This is an investigation step to measure peak memory and decode speed
with tiling disabled on high_detail — guided by issue #6.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Files changed (1) hide show

app.py +1 -1

app.py CHANGED Viewed

@@ -191,7 +191,7 @@ pipe = QwenImageEditPlusPipeline.from_pretrained(
     torch_dtype=dtype,
 )
 _hb.set()
-pipe.vae.enable_tiling(tile_sample_min_height=Mode.FAST.max_dim, tile_sample_min_width=Mode.FAST.max_dim)
 print(f"[startup] pipeline loaded in {time.perf_counter()-_t1_load:.1f}s", flush=True)
 print("[startup] setting cuDNN SDPA attention processor...", flush=True)

     torch_dtype=dtype,
 )
 _hb.set()
+pipe.vae.enable_tiling(tile_sample_min_height=Mode.HIGH_DETAIL.max_dim, tile_sample_min_width=Mode.HIGH_DETAIL.max_dim)
 print(f"[startup] pipeline loaded in {time.perf_counter()-_t1_load:.1f}s", flush=True)
 print("[startup] setting cuDNN SDPA attention processor...", flush=True)