Guard every assumption that changes indexing or semantics

For a two-dimensional strided view with element strides s0 and s1, the storage element offset of index [i,j] is offset0+i·s0+j·s1. Contiguous row-major shape [2,3] usually has strides [3,1], but the transpose of a contiguous [3,2] tensor has the same [2,3] shape and strides [1,2]. A kernel that silently substitutes contiguous indexing changes which elements it reads. Shape-only testing can miss this if every benchmark input is created with the same layout constructor.

Use deliberately distinctive contents to expose the mismatch. Storage [0,1,2,3,4,5] viewed with shape [2,3], strides [1,2], and zero storage offset gives rows [0,2,4] and [1,3,5]. At index [0,1], the correct value is two while assumed contiguous indexing reads one. Both are finite, ordinary numbers, so a crash detector will not catch this bug. A layout guard or a genuinely stride-general implementation is necessary; a convenient allocation pattern is not a substitute.

A dtype assumption can select different instructions, accumulation precision, and byte offsets. A device assumption identifies where data resides and which generated code can access it. A static scalar option may select an activation, axis, or training branch. If compilation treats that option as fixed, later changes need an appropriate dispatch condition or a dynamic representation. Do not guard an entire weight tensor's identity merely because that was easy unless identity is the actual required condition; changing weight contents and replacing weight objects are different events.

PyTorch 2.9's documented guards and recompilation behavior illustrate these distinctions. A failed condition can cause a new variant or eager execution under the runtime's policy, with existing valid variants still potentially reusable. Cache limits and names of internal configuration settings are version-specific, so this lesson does not prescribe a universal limit. Audit the generated assumptions for the chosen release. Then test that each deliberately violated assumption routes safely, rather than accepting that warm calls on the initial example succeed.

Build paired fixtures that hold contents and shape stable while changing layout, or hold shape and layout stable while changing a static option. Expected behavior should identify whether the call reuses a compatible artifact, selects another artifact, recaptures, or falls back. A guard check must precede any operation that relies on it. Testing only that the final numerical result is sometimes close can conceal an invalid dispatch whose wrong address happens to contain a similar value.

The cache lab helps separate artifact diversity from correctness of artifact guards. Reset and set both C=4,s=4 to inspect four repeating identifiers with twelve initial hits; Reset and set both C=4,s=1 to inspect sixteen identifiers with zero initial hits. Predict before Play or Step. Time-zero already includes sixteen requests and playback caps at 4,096. Identifiers can stand for distinct fictional variants, but this fully associative LRU stream does not evaluate shape, layout, device, or value conditions and cannot validate any real dispatch policy.