viable/strict/1790060510: flex_gemm: accept 1-D captured tensors (#196183)
- PyTorch: 1328 events in the last 90 days
- PyTorch: 1313th Release in the last 90 days
- Previous: earlier the same day · viable/strict/1790054880: [ROCm] Skip conv_transpose1d complex32 reference test on MI200 (#198025)
What happened
A closure that captures a 1-D tensor, the natural nn.Parameter shape for a norm weight [N] or a router bias [E] , was rejected with the capture-profile error even when the body wrote w[None, :] , because the profile was checked on the captured tensor rather than on how the body broadcasts it. Running FlexGEMM over DeepSeek-V3's router hit this immediately, and the only workaround was hoisting w2 = w[None, :] outside…
Summary assembled by rule from the sources below