Issues

A pooling layer constructed with default arguments leaves its spatial extents unresolved

#919

· khatchad opened

on Sep 12, 2026

A comprehension in a method's default value has no call-graph invoke

#918

· khatchad opened

on Sep 12, 2026

A list comprehension's `if` filters do not reach the IR, and the windowed-batcher rule sizes a filtered comprehension as unfiltered

#917

· khatchad opened

on Sep 12, 2026

An `Embedding` layer's output carries a sequence extent its input does not have

#916

· khatchad opened

on Sep 12, 2026

A parameter of a `tf.function` with an `input_signature` is typed with the caller's concrete shape where the body sees the signature's

#915

· khatchad opened

on Sep 12, 2026

Generators map an empty (bottom) input shape set to top through the `ret.isEmpty() ? null : ret` idiom

#913

· khatchad opened

on Sep 11, 2026

A `tf.slice` bounded by a runtime tensor emits `Unresolved` where TensorFlow's static shape is `None` (`Dynamic`)

#912

· khatchad opened

on Sep 11, 2026

A `scalar_literal + X` concatenation is typed as a rank-1 sequence when `X` is an evidence-free tensor whose chain the provenance walk cannot root at a modeled library

#911

· khatchad opened

on Sep 11, 2026

`NpPermutation` resolves its array argument through an empty points-to set, losing shape and dtype

#910

· khatchad opened

on Sep 10, 2026

Dimension arithmetic resolves values, not terms, so `n + (k - n)` cannot cancel

#909

· khatchad opened

on Sep 10, 2026

A resolved extent does not distinguish "constant at every analysed call site" from "invariant"

#908

· khatchad opened

on Sep 10, 2026

Parameters whose every axis is fixed by a literal tuple in the only driver are emitted with no rank

#907

· khatchad opened

on Sep 9, 2026

Navigation Menu

Uh oh!

All issues

A pooling layer constructed with default arguments leaves its spatial extents unresolved

A comprehension in a method's default value has no call-graph invoke

A list comprehension's `if` filters do not reach the IR, and the windowed-batcher rule sizes a filtered comprehension as unfiltered

An `Embedding` layer's output carries a sequence extent its input does not have

A parameter of a `tf.function` with an `input_signature` is typed with the caller's concrete shape where the body sees the signature's

Generators map an empty (bottom) input shape set to top through the `ret.isEmpty() ? null : ret` idiom

A `tf.slice` bounded by a runtime tensor emits `Unresolved` where TensorFlow's static shape is `None` (`Dynamic`)

A `scalar_literal + X` concatenation is typed as a rank-1 sequence when `X` is an evidence-free tensor whose chain the provenance walk cannot root at a modeled library

`NpPermutation` resolves its array argument through an empty points-to set, losing shape and dtype

Dimension arithmetic resolves values, not terms, so `n + (k - n)` cannot cancel

A resolved extent does not distinguish "constant at every analysed call site" from "invariant"

Parameters whose every axis is fixed by a literal tuple in the only driver are emitted with no rank

Uh oh!

Search results