Issues

`jnp.linalg.svd` gradients return `nan` for intermediate overflow

bug

#40185

· shadow-orange41 opened

on Aug 25, 2026

[Mosaic GPU] Support collective tcgen05.commit arrivals in multidimensional clusters

enhancement

#40165

· saladpalad opened

on Aug 24, 2026

XLA - CPU thunk runtime: `jax.grad` through `lax.scan` is 60x slower when the carry holds many scalar leaves

#40150

· LegrandNico opened

on Aug 24, 2026

Repeated XLA compiles of jnp.argsort on the CUDA backend crashes

bug

#40142

· quantum-ml-geek opened

on Aug 22, 2026

`eigh` forward- and reverse-mode eigenvector gradients return all-`NaN` for eigenvalues above `2^p` (mantissa-precision cliff in the JVP rule's div-by-zero guard)

bug

#40141

· shadow-orange41 opened

on Aug 22, 2026

splash attention: `dsinks` gradient loses precision by rounding before the reduction

bug

#40128

· ds-hwang opened

on Aug 21, 2026

XLA:CPU: a small dynamic-update-slice inside a loop rewrites the whole buffer each iteration (0.11.1 regression)

#40101

· mmcky opened

on Aug 20, 2026

Mosaic GPU `broadcast_in_dim` unnecessarily restricts ranks > 2

bug

#40081

· samanklesaria opened

on Aug 18, 2026

Support out-of-bounds predication for the `cp.async` for GMEM→SMEM (pre-Hopper)

enhancement

#40048

· samanklesaria opened

on Aug 17, 2026

[RFC] Unsafe opaque PJRT C-API handles on jax.Array for external PJRT-C-API interop

#40038

· lokic233 opened

on Aug 17, 2026

jnp.remainder gives a zero result the sign of the dividend, not the divisor

#40028

· krrishapatel opened

on Aug 17, 2026

Allow dicts with mixed-type keys by comparing by (type name, value)

enhancement

#40026

· carlosgmartin opened

on Aug 16, 2026

Navigation Menu

Uh oh!

All issues

`jnp.linalg.svd` gradients return `nan` for intermediate overflow

[Mosaic GPU] Support collective tcgen05.commit arrivals in multidimensional clusters

XLA - CPU thunk runtime: `jax.grad` through `lax.scan` is 60x slower when the carry holds many scalar leaves

Repeated XLA compiles of jnp.argsort on the CUDA backend crashes

`eigh` forward- and reverse-mode eigenvector gradients return all-`NaN` for eigenvalues above `2^p` (mantissa-precision cliff in the JVP rule's div-by-zero guard)

splash attention: `dsinks` gradient loses precision by rounding before the reduction

XLA:CPU: a small dynamic-update-slice inside a loop rewrites the whole buffer each iteration (0.11.1 regression)

Mosaic GPU `broadcast_in_dim` unnecessarily restricts ranks > 2

Support out-of-bounds predication for the `cp.async` for GMEM→SMEM (pre-Hopper)

[RFC] Unsafe opaque PJRT C-API handles on jax.Array for external PJRT-C-API interop

jnp.remainder gives a zero result the sign of the dividend, not the divisor

Allow dicts with mixed-type keys by comparing by (type name, value)

Uh oh!

Search results