Upload folder using huggingface_hub

app.py

@@ -25,12 +25,11 @@
 
 # Must precede the torch import: the allocator reads this at initialization.
 #
-# Without it the VAE decode dies with
-# "NVML_SUCCESS == r INTERNAL ASSERT FAILED at CUDACachingAllocator.cpp" once
-# denoising finishes and the decoder asks for its large contiguous buffers.
-# Same failure, same fix as wan-2-2-5b-video: ZeroGPU hands out a partitioned
-# GPU, and expandable segments keep the allocator off the NVML query path that
-# breaks there.
+# Fragmentation insurance, NOT the fix for the NVML assert this Space used to
+# die on. That was measured: the ZeroGPU worker printed
+# 'expandable_segments:True' on the runs that still crashed, so the setting was
+# applied the whole time and made no difference. What actually decides it is
+# WORK_AREA below. Kept because it costs nothing and reduces fragmentation.
 import os
 
 os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
@@ -96,9 +95,16 @@ NEG_EMB = _fetch("neg_emb.pt", Path("./neg_emb.pt"))
 # scaled UP to it and large inputs down, because the model was only trained at
 # high resolution.
 #
-# Overridable so the value can be tuned against the ZeroGPU VAE-decode ceiling
-# without a code change: the decoder's peak allocation scales with this, and
-# that allocation is what trips the caching allocator's NVML assert.
+# THIS is what decides whether the VAE decode survives on ZeroGPU. The
+# decoder's peak contiguous allocation scales with it, and that allocation is
+# what trips "NVML_SUCCESS == r INTERNAL ASSERT FAILED" in the caching
+# allocator. Measured: 2560x1440 (the upstream value) fails, 1920x1080 passes
+# in 14.7s. Left overridable so the ceiling can be probed without a code
+# change, but do not raise the default without re-testing a real run.
+#
+# The cost is honest to state: the model was trained at high resolution, so
+# restoring at 2MP rather than 3.7MP gives up some of its headroom on very
+# large outputs. It still upscales ~4.9x from a 360x240 input.
 WORK_AREA = int(os.environ.get("SEEDVR2_WORK_AREA", 1920 * 1080))
 
 # Single-process "distributed" context. The model code routes every device

configs_3b/main.yaml

@@ -53,14 +53,13 @@ vae:
   memory_limit:
     # CHANGED FROM UPSTREAM (Upsampler): 0.5 -> 0.15 GiB.
     #
-    # This is the VAE's own decode-splitting threshold: a conv whose input would
-    # exceed it is split and run in pieces. At the upstream 0.5 the decoder still
-    # asks for buffers large enough to send ZeroGPU's caching allocator down the
-    # NVML path that dies with "NVML_SUCCESS == r INTERNAL ASSERT FAILED", which
-    # is the same class of failure wan-2-2-5b-video solved by keeping the decode
-    # tiled and sliced. Smaller splits mean more convs but no single large
-    # contiguous allocation. Output is unaffected: the splitting carries the
-    # overlap cache between pieces, so it is exact rather than approximate.
+    # The VAE's own decode-splitting threshold: a conv whose input would exceed
+    # it is split and run in pieces. Lowering this did NOT by itself clear the
+    # NVML assert this Space used to die on -- it moved the failure from the
+    # conv to the torch.cat that reassembles the splits. WORK_AREA in app.py is
+    # what actually decides that. Kept at 0.15 because smaller allocations are
+    # still the right default here and the splitting is exact: it carries its
+    # overlap cache between pieces, so output is unchanged.
     conv_max_mem: 0.15
     norm_max_mem: 0.15
   checkpoint: ./ckpts/ema_vae.pth

upsampler_theme.py

@@ -1,54 +0,0 @@
-"""Shared Upsampler look-and-feel for every Upsampler/* HF Space.
-
-`create_and_push.py` uploads this file alongside each Space's app.py, so every
-Space imports the exact same theme, CSS, header, and footer. Keep this the
-single source of truth (v3 recipe): Soft indigo/purple theme, gradient primary
-button, Gradio's own footer hidden, 1000px max width, minimal header/footer.
-"""
-
-import gradio as gr
-
-UPSAMPLER_THEME = gr.themes.Soft(
-    primary_hue=gr.themes.colors.indigo,
-    secondary_hue=gr.themes.colors.purple,
-    neutral_hue=gr.themes.colors.slate,
-    font=[gr.themes.GoogleFont("Inter"), "system-ui", "sans-serif"],
-).set(
-    button_primary_background_fill="linear-gradient(90deg, #6366f1 0%, #a855f7 100%)",
-    button_primary_background_fill_hover="linear-gradient(90deg, #4f46e5 0%, #9333ea 100%)",
-    button_primary_text_color="#ffffff",
-    button_primary_border_color="*primary_500",
-)
-
-# Hide Gradio's built-in footer and keep the app narrow and centered.
-UPSAMPLER_CSS = """
-footer { display: none !important; }
-.gradio-container { max-width: 1000px !important; margin: 0 auto !important; }
-#usp-header h1 { font-size: 1.7rem; font-weight: 700; margin: 0 0 .25rem; }
-#usp-header p { opacity: .6; margin: 0; }
-#usp-footer { opacity: .5; font-size: .85rem; margin-top: 1.25rem; }
-#usp-footer a { text-decoration: none; }
-"""
-
-
-def header_html(title: str, subtitle: str) -> str:
-    return f"""<div id="usp-header">
-  <h1>{title}</h1>
-  <p>{subtitle}</p>
-</div>"""
-
-
-_LINK_STYLE = "color:#8b7cf6;font-weight:600;text-decoration:none"
-
-
-def footer_html(description: str, tool_url: str, tool_anchor: str) -> str:
-    """SEO footer (v4 recipe): a short paragraph describing what the model
-    does (unique per Space, keyword-bearing) plus the Upsampler attribution
-    with a deep link to the matching /free-* tool on upsampler.com. No model
-    credit/license line (the README frontmatter carries the license). Spaces
-    target model-name queries; the site pages keep the intent queries
-    ("free X no signup"), so the two never compete."""
-    return f"""<div id="usp-footer">
-  <p style="margin:0 0 10px">{description}</p>
-  <p style="margin:0">Maintained by <a href="https://upsampler.com" target="_blank" rel="noopener" style="{_LINK_STYLE}">Upsampler</a>. Check out the <a href="{tool_url}" target="_blank" rel="noopener" style="{_LINK_STYLE}">{tool_anchor}</a>, no sign-up required.</p>
-</div>"""