ecc-pytorch-build-resolver

Trace tensor shapes by printing shape, dtype, and device before the failing line, then align layer dimensions. For 'mat1 and mat2 shapes cannot be multiplied', fix the nn.Linear in_features to match the previous layer's output size.

This error means some tensors are on CPU while the model or other tensors are on GPU. Apply .to(device) consistently to the model and all input tensors so every operand in the computation shares one device.

CUDA OOM comes from oversized live tensors, retained autograd graphs, large batches, or fragmentation. Inspect live allocations with torch.cuda.memory_summary in the failing process, release unintended references, and consider gradient checkpointing rather than relying on cache clearing.

No. It works only within the existing local environment and explicitly avoids package installs, driver changes, training launches, and model downloads. Missing prerequisites like drivers or GPUs are reported rather than worked around.

An in-place operation such as x += 1 or in-place ReLU overwrote a tensor needed by autograd. Replace in-place ops with out-of-place equivalents like x = x + 1 so the computation graph stays intact.

Stop and report if the same error persists after three fix attempts, the fix requires fundamental architecture changes, the cause is hardware or driver incompatibility, or OOM persists even at batch_size=1.