Spaces:
Runtime error
Runtime error
Commit History
-
Fix RoIAlign/RoIPool output_size to always be a tuple
The downstream code accesses output_size[0] and output_size[1], so normalize int inputs to (int, int) tuples. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>d868632 -
Implement full CPU forward for MultiScaleDeformableAttention
Replace NotImplementedError stub with the actual pure-Python CPU implementation using F.grid_sample bilinear interpolation, matching the upstream mmcv multi_scale_deformable_attn_pytorch logic. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>0b92003 -
Import MultiScaleDeformableAttention in ops __init__ for registry
The shim's __init__.py must import the multi_scale_deform_attn module so that @ATTENTION.register_module() is executed at import time, making it available in the attention registry. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>4e247c3 -
Register MultiScaleDeformableAttention in mmcv ATTENTION registry
The stub class needs to be registered with @ATTENTION.register_module() so that BaseTransformerLayer can find it when building the model. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>dac41e5 -
Add sigmoid_focal_loss, deform_conv2d, get_onnxruntime_op_path to shim
These are imported by mmdet during module loading and must be available even if not used at inference time. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>875a4a2 -
Add missing mmcv.ops stubs: RoIPool, DeformConv2d, etc.
The shim was missing RoIPool (needed by mmdet.apis.inference) and other ops imported at module load time. Add stubs for all ops referenced by mmdet to prevent ImportError at startup. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>168a22c
-
Fix Dockerfile: extract mmcv.ops shim to separate Python file
Move inline Python code from Dockerfile RUN command into setup_mmcv_shim.py to fix Docker parse errors. Use mmcv lite instead of mmcv-full to avoid C++ compilation. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>2d204ac