jax-ml/jax · error · NotImplementedError
The SVD algorithm parameter is not implemented on TPU.
Error message
The SVD algorithm parameter is not implemented on TPU.
What it means
The TPU implementation of SVD only supports the DEFAULT algorithm; explicitly requesting a different SvdAlgorithm (e.g. QR or JACOBI_ANNIHILATOR, which select CPU/GPU algorithms) raises NotImplementedError in the TPU frontend wrapper.
Solutions
- Omit the algorithm argument (DEFAULT works on TPU) or pass SvdAlgorithm.DEFAULT.
- Conditionally select the algorithm per backend: only set non-DEFAULT on CPU/GPU.
- If a specific algorithm is required, run that op with jax.default_device(jax.devices('cpu')[0]) or on GPU.
Example fix
# before u, s, vt = jnp.linalg.svd(a, algorithm=SvdAlgorithm.QR) # after u, s, vt = jnp.linalg.svd(a) # DEFAULT (POLAR) on TPU
Defensive patterns
Strategy: fallback
Validate before calling
import jax
if jax.default_backend() == 'tpu':
algorithm = None # DEFAULT only on TPU Try / catch
try:
u, s, vt = jnp.linalg.svd(a, algorithm=algorithm)
except NotImplementedError:
u, s, vt = jnp.linalg.svd(a) # DEFAULT algorithm fallback Prevention
- Backend-gate any non-DEFAULT algorithm choice.
- Keep algorithm as a config knob defaulted to None.
When it happens
Trigger: Calling jax.numpy.linalg.svd(a, algorithm=lax_linalg.SvdAlgorithm.QR) (or any non-DEFAULT value) on a TPU backend, via jax._src.tpu.linalg.svd._svd_tpu.
Common situations: Tuning algorithm for CPU/GPU performance or stability then running the same code on TPU; cross-backend pipelines promoted to TPU.
Related errors
- Only the POLAR (which is also DEFAULT on TPU) SVD algorithm…
- Batching over dynamic grid values is not supported yet.
- Cannot split a Pallas key. Use fold_in instead to generate…
- dma_start not implemented in LoJAX yet.
- Failed to find assignment for logical_axis_index
AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27).
Data as JSON: /api/errors/bfbb39761785e713.
Report an issue: GitHub.
Appendix: source
Thrown at jax/_src/tpu/linalg/svd.py:260
body_f = lambda args: (
jnp.array(True),
jnp.full_like(u_out, np.nan),
jnp.full_like(s_out, np.nan),
jnp.full_like(v_out, np.nan),
)
_, u_out, s_out, v_out = lax.while_loop(
cond_f, body_f, (is_finite, u_out, s_out, v_out)
)
if is_flip:
return (v_out, s_out, u_out.T.conj())
return (u_out, s_out, v_out.T.conj())
def _svd_tpu(a, *, full_matrices, compute_uv, subset_by_index, algorithm=None):
if algorithm is not None and algorithm != lax_linalg.SvdAlgorithm.DEFAULT:
raise NotImplementedError(
"The SVD algorithm parameter is not implemented on TPU.")
batch_dims = a.shape[:-2]
fn = functools.partial(
svd,
full_matrices=full_matrices,
compute_uv=compute_uv,
subset_by_index=subset_by_index,
)
for _ in range(len(batch_dims)):
fn = api.vmap(fn)
if compute_uv:
u, s, vh = fn(a)
return [s, u, vh]
else:
s = fn(a)
return [s]View on GitHub (pinned to 1e1c6a8fc0)