Ensure diff_gaussian_rasterization is installed at runtime

Added ensure_diff_gaussian utility to check/install diff_gaussian_rasterization from GitHub if missing. Updated app.py to call this before loading Trellis pipeline. requirements.txt now explicitly includes the package.

README.md

@@ -4,7 +4,7 @@ emoji: 🐝
 colorFrom: yellow
 colorTo: purple
 sdk: gradio
-python_version: 3.10.13
+python_version: 3.12.12
 sdk_version: 6.10.0
 app_file: app.py
 pinned: true

app.py

@@ -23,7 +23,7 @@ import accelerate
 from transformers import AutoTokenizer, DPTImageProcessor, DPTForDepthEstimation
 from trellis.pipelines import TrellisImageTo3DPipeline
 from trellis.representations import Gaussian, MeshExtractResult
-from trellis.utils import render_utils, postprocessing_utils
+#from trellis.utils import render_utils, postprocessing_utils
 from pathlib import Path
 import utils.hex_hura as hex_hura
 from utils.storage import (
@@ -1000,30 +1000,71 @@ def generate_3d_asset_part1(depth_image_source, randomize_seed, seed, input_imag
 
     return depth_img, image_path, output_name, final_seed
 
-@spaces.GPU(duration=150,progress=gr.Progress(track_tqdm=True))
-def generate_3d_asset_part2(depth_img, image_path, output_name, seed, steps, model_resolution, video_resolution, depth_alpha: int, multi_mode: str, req: gr.Request, progress=gr.Progress(track_tqdm=True)):
+@spaces.GPU(duration=150, progress=gr.Progress(track_tqdm=True))
+def generate_3d_asset_part2(
+    depth_img,
+    image_path,
+    output_name,
+    seed,
+    steps,
+    model_resolution,
+    video_resolution,
+    depth_alpha: int,
+    multi_mode: str,
+    req: gr.Request,
+    progress=gr.Progress(track_tqdm=True),
+):
+    from trellis.utils import render_utils
     # Open image using standardized defaults
     image_raw = Image.open(image_path).convert("RGB")
     resized_image = resize_image_with_aspect_ratio(image_raw, model_resolution, model_resolution)
     depth_img = Image.open(depth_img).convert("RGBA")
     resized_depth_image = resize_image_with_aspect_ratio(depth_img, model_resolution, model_resolution)
     resized_depth_image_enhanced = apply_lut(resized_depth_image, "./LUT/Contrast.cube", 100)
-    combined_depth_img = combine_depth_map_with_Image(resized_image, resized_depth_image_enhanced, model_resolution, model_resolution, depth_alpha)
+    combined_depth_img = combine_depth_map_with_Image(
+        resized_image,
+        resized_depth_image_enhanced,
+        model_resolution,
+        model_resolution,
+        depth_alpha,
+    )
 
-    use_multi = (multi_mode != "singleimage")
+    use_multi = multi_mode != "singleimage"
+    processed_images = None
     if use_multi:
         images = [resized_image, combined_depth_img, resized_depth_image]
-        processed_images = [TRELLIS_PIPELINE.preprocess_image(image,  max_resolution=model_resolution, remove_bg=False) for image in images]       
+        processed_images = [
+            TRELLIS_PIPELINE.preprocess_image(image, max_resolution=model_resolution, remove_bg=False)
+            for image in images
+        ]
 
     if TRELLIS_PIPELINE is None:
-        gr.Warning(f"Trellis Pipeline is not initialized: {TRELLIS_PIPELINE.device()}")
-        return [None, None, depth_img]
-    else:
         # Preprocess and run the Trellis pipeline with fixed sampler settings
+        gr.Warning("Trellis Pipeline is not initialized")
+        return [None, None, depth_img]
+
+    try:
+        # Prefer backends that work on Blackwell sm_120
+        os.environ["ATTN_BACKEND"] = "sdpa"
+        os.environ["XFORMERS_DISABLED"] = "1"
         try:
-            TRELLIS_PIPELINE.cuda()
+            from trellis.modules.attention import config as attn_config
+            if hasattr(attn_config, "set_backend"):
+                attn_config.set_backend("sdpa")
+        except Exception:
+            pass
+
+        TRELLIS_PIPELINE.cuda()
+
+        # no_grad (not inference_mode): diff_gaussian_rasterization uses autograd.Function
+        with torch.no_grad():
+            #with torch.autocast("cuda", dtype=torch.bfloat16):
             if not use_multi:
-                processed_image = TRELLIS_PIPELINE.preprocess_image(combined_depth_img, max_resolution=model_resolution, remove_bg = False)
+                processed_image = TRELLIS_PIPELINE.preprocess_image(
+                    combined_depth_img,
+                    max_resolution=model_resolution,
+                    remove_bg=False,
+                )
                 outputs = TRELLIS_PIPELINE.run(
                     processed_image,
                     seed=seed,
@@ -1057,39 +1098,40 @@ def generate_3d_asset_part2(depth_img, image_path, output_name, seed, steps, mod
 
             # Validate the mesh
             mesh = outputs['mesh'][0]
-            meshisdict = isinstance(mesh, dict)
-            if meshisdict:
-                vertices = mesh['vertices']
-                faces = mesh['faces']
-            else:
-                vertices = mesh.vertices
-                faces = mesh.faces
+        meshisdict = isinstance(mesh, dict)
+        if meshisdict:
+            vertices = mesh["vertices"]
+            faces = mesh["faces"]
+        else:
+            vertices = mesh.vertices
+            faces = mesh.faces
 
-            print(f"Mesh vertices: {vertices.shape}, faces: {faces.shape}")
-            if faces.max() >= vertices.shape[0]:
-                raise ValueError(f"Invalid mesh: face index {faces.max()} exceeds vertex count {vertices.shape[0]}")
-        except Exception as e:
-            gr.Warning(f"Error generating 3D asset: {e}")
-            print(f"Error generating 3D asset: {e}")
-            torch.cuda.empty_cache()
-            torch.cuda.ipc_collect()
-            return None,None, depth_img
+        print(f"Mesh vertices: {vertices.shape}, faces: {faces.shape}")
+        if faces.max() >= vertices.shape[0]:
+            raise ValueError(f"Invalid mesh: face index {faces.max()} exceeds vertex count {vertices.shape[0]}")
+    except Exception as e:
+        gr.Warning(f"Error generating 3D asset: {e}")
+        print(f"Error generating 3D asset: {e}")
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
+        return None, None, depth_img
 
         # Ensure data is on GPU and has correct type
-        if not vertices.is_cuda or not faces.is_cuda:
-            raise ValueError("Mesh data must be on GPU")
-        if vertices.dtype != torch.float32 or faces.dtype != torch.int32:
-            if meshisdict:
-                mesh['faces'] = faces.to(torch.int32)
-                mesh['vertices'] = vertices.to(torch.float32)
-            else:
-                mesh.faces = faces.to(torch.int32)
-                mesh.vertices = vertices.to(torch.float32)
+    if not vertices.is_cuda or not faces.is_cuda:
+        raise ValueError("Mesh data must be on GPU")
+    if vertices.dtype != torch.float32 or faces.dtype != torch.int32:
+        if meshisdict:
+            mesh["faces"] = faces.to(torch.int32)
+            mesh["vertices"] = vertices.to(torch.float32)
+        else:
+            mesh.faces = faces.to(torch.int32)
+            mesh.vertices = vertices.to(torch.float32)
 
-        user_dir = os.path.join(constants.TMPDIR, str(req.session_hash))
-        os.makedirs(user_dir, exist_ok=True)
+    user_dir = os.path.join(constants.TMPDIR, str(req.session_hash))
+    os.makedirs(user_dir, exist_ok=True)
 
-        video = render_utils.render_video(outputs['gaussian'][0], resolution=video_resolution, num_frames=64, r=1, fov=45)['color']
+    with torch.no_grad():
+        video = render_utils.render_video(outputs["gaussian"][0], resolution=video_resolution, num_frames=64, r=1, fov=45)["color"]
         try:
             #video_rf = render_utils.render_video(outputs['radiance_field'][0], resolution=video_resolution, num_frames=64, r=1, fov=45)['color']
             video_geo = render_utils.render_video(outputs['mesh'][0], resolution=video_resolution, num_frames=64, r=1, fov=45)['normal']
@@ -1097,21 +1139,19 @@ def generate_3d_asset_part2(depth_img, image_path, output_name, seed, steps, mod
         except Exception as e:
             gr.Info(f"Error rendering video: {e}")
             print(f"Error rendering video: {e}")
-        video_path = os.path.join(user_dir, f'{output_name}.mp4')
-        imageio.mimsave(video_path, video, fps=8)
+    video_path = os.path.join(user_dir, f"{output_name}.mp4")
+    imageio.mimsave(video_path, video, fps=8)
 
         #snapshot_results = render_utils.render_snapshot_depth(outputs['radiance_field'][0], resolution=1280, r=1, fov=80)
         #depth_snapshot = Image.fromarray(snapshot_results['normal'][0]).convert("L")
-        depth_snapshot = depth_img
-
-        state = pack_state(outputs['gaussian'][0], outputs['mesh'][0], output_name)
+    depth_snapshot = depth_img
+    state = pack_state(outputs["gaussian"][0], outputs["mesh"][0], output_name)
 
     if torch.cuda.is_available():
         torch.cuda.empty_cache()
         torch.cuda.ipc_collect()
     return [state, video_path, depth_snapshot]
 
-
 @spaces.GPU(duration=90,progress=gr.Progress(track_tqdm=True))
 def extract_glb(
     state: dict,
@@ -1130,6 +1170,7 @@ def extract_glb(
     Returns:
         str: The path to the extracted GLB file.
     """
+    from trellis.utils import postprocessing_utils
     user_dir = os.path.join(constants.TMPDIR, str(req.session_hash))
     gs, mesh, name = unpack_state(state)
     glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False)

pre-requirements.txt

@@ -3,5 +3,5 @@ pip>=26.2
 # Official PyTorch + torchvision for CUDA 12.8 (sm_120 / ZeroGPU)
 --extra-index-url https://download.pytorch.org/whl/cu128
 
-torch==2.9.1
-torchvision==0.24.1
+torch==2.10.0
+torchvision==0.25.0

requirements.txt

@@ -14,7 +14,7 @@ sentencepiece
 --extra-index-url https://download.pytorch.org/whl/cu128
 
 # xformers for torch 2.10+ (official cu128 wheels)
-xformers>=0.0.34
+xformers>=0.0.33.post2
 
 # Other dependencies
 Haishoku
@@ -61,7 +61,12 @@ pyvista==0.44.2
 pymeshfix==0.17.0
 igraph==0.11.8
 git+https://github.com/EasternJournalist/utils3d.git@3913c65d81e05e47b9f367250cf8c0f7462a0900
+
+# spconv + cumm (Python 3.12 has cu128 wheels)
+--extra-index-url https://ratharog.github.io/cumm-spconv/
+cumm-cu128==0.9.1
 spconv-cu128==2.4.1
+
 #gradio_litmodel3d==0.0.1
 #linux only
 #https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.3/flash_attn-2.7.3+cu12torch2.6cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
@@ -70,19 +75,34 @@ spconv-cu128==2.4.1
 
 # -------------------------------------------------
 # CUDA extensions for torch 2.9.1 + cu128
+# CUDA extensions for torch 2.10.0 + cu128 + Python 3.12
 # -------------------------------------------------
-flash-attn @ https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.0/flash_attn-2.8.3+cu128torch2.9-cp310-cp310-linux_x86_64.whl
+# flash-attn @ https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.0/flash_attn-2.8.3+cu128torch2.9-cp310-cp310-linux_x86_64.whl
+# flash-attn @ https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.0/flash_attn-2.8.3+cu128torch2.9-cp310-cp310-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl
+# flash-attn @ https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.0/flash_attn-2.8.3+cu128torch2.9-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl
+# https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.0/flash_attn-2.8.3+cu128torch2.9-cp312-cp312-linux_x86_64.whl
+# https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.0/flash_attn-2.8.3+cu128torch2.10-cp312-cp312-linux_x86_64.whl
+https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.0/flash_attn-2.8.3+cu128torch2.10-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl
+# temporarily disabled – ABI mismatch with current torch 2.9.1+cu128
 
 ninja
 
-# nvdiffrast – ready wheel for 2.9.1 + cu128 (Python 3.10 Linux)
---find-links https://pozzettiandrea.github.io/nvdiffrast-full-wheels/cu128-torch291/
-nvdiffrast
+# nvdiffrast – direct wheel (GitHub Pages index is down)
+#https://github.com/PozzettiAndrea/cuda-wheels/releases/download/nvdiffrast-latest/nvdiffrast-0.4.0%2Bcu128torch2.9-cp310-cp310-manylinux_2_34_x86_64.manylinux_2_35_x86_64.whl
+#https://github.com/PozzettiAndrea/cuda-wheels/releases/download/nvdiffrast-latest/nvdiffrast-0.4.0%2Bcu128torch2.9-cp312-cp312-manylinux_2_34_x86_64.manylinux_2_35_x86_64.whl
+https://github.com/PozzettiAndrea/cuda-wheels/releases/download/nvdiffrast-latest/nvdiffrast-0.4.0%2Bcu128torch2.10-cp312-cp312-manylinux_2_34_x86_64.manylinux_2_35_x86_64.whl
 
 # diff_gaussian_rasterization – no reliable 2.9.1 wheel, build from source
-git+https://github.com/graphdeco-inria/diff-gaussian-rasterization.git
+#git+https://github.com/graphdeco-inria/diff-gaussian-rasterization.git  ###############run inside app.py to build from source in misc.py ##############
+https://huggingface.co/spaces/Surn/HexaGrid/resolve/main/wheels/diff_gaussian_rasterization-0.0.0-cp312-cp312-linux_x86_64.whl?download=true
+# (do not put the git URL here if you use ensure_diff_gaussian())
 
 #Windows only
 #https://huggingface.co/spaces/Surn/HexaGrid/resolve/main/wheels/flash_attn-2.7.4.post1-cp312-cp312-win_amd64.whl?download=true
 #https://huggingface.co/spaces/Surn/HexaGrid/resolve/main/wheels/diff_gaussian_rasterization-0.0.0-cp312-cp312-win_amd64.whl?download=true
 #https://huggingface.co/spaces/Surn/HexaGrid/resolve/main/wheels/nvdiffrast-0.3.3-py3-none-any.whl
+
+#https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.28/flash_attn-2.8.3+cu128torch2.9-cp312-cp312-win_amd64.whl
+#https://github.com/PozzettiAndrea/cuda-wheels/releases/download/nvdiffrast-latest/nvdiffrast-0.4.0%2Bcu128torch2.9-cp312-cp312-win_amd64.whl
+
+#https://github.com/PozzettiAndrea/cuda-wheels/releases/download/nvdiffrast-latest/nvdiffrast-0.4.0%2Bcu128torch2.10-cp312-cp312-win_amd64.whl

utils/misc.py

@@ -4,6 +4,7 @@ import sys
 import logging
 import os
 import subprocess
+import sys
 from utils.constants import TMPDIR, MAX_SEED
 import numpy as np
 
@@ -206,4 +207,4 @@ def number_to_letter(n: int, upper_case: bool = True):
     while abs(n) > 0:
         n, remainder = divmod(abs(n) - 1, 26)
         result = chr(a_char + remainder) + result
-    return result
+    return result

wheels/diff_gaussian_rasterization-0.0.0-cp312-cp312-linux_x86_64.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7acb876ee0e0150cec13f3f9388d76410da40a2719a69c3c574343ef4ebba8d5
+size 4159087