Locate the allocation
Connect a growing tensor shape to the line that materializes it.
HEADROOM
GPU REPAIR LAB v0.1
BUILT FOR THE MOMENT YOUR GPU SAYS NO
Make it fit.
Prove it works.
Headroom finds the memory bottleneck, proposes a bounded repair, and tests whether the outputs you actually need still survive.
01 / THE WORKBENCH
One memory pattern. Different promises.
Choose a built-in workload and set its constraints. No code uploads, account, or personal data needed.
01
Choose a workload
Same allocation pattern. Different output contract.
Set the limits
Fixed workloads only · local runner · no user code execution
02AWAITING RUN
Trace the pressure
Source-linked evidence, then an independent decision.
nearest.pyPYTORCH
01def nearest(query, reference):02 delta = query[:, None, :] - reference[None, :, :]03 distance = torch.sqrt(torch.sum(delta * delta, dim=-1))04 return torch.argmin(distance, dim=1)
FULL PAIRWISE ALLOCATION · O(N × M) MEMORY
MEMORY PROFILERUN TO MEASURE
Your measured result will appear here.
06 / THE METHOD
Nothing passes on a good story alone.
Bound the repair
Use a small, reviewable strategy set. Model output is never executed as code.
Challenge the output
Independent cases check shape, values, ties, memory, and latency.
Keep the evidence
Download the settings, measurements, decision, and regression assertion.