Skip to content

0.6.6: follow-ups from the Model K softmax self-check review (#254) #261

Open

Description

@phantoslayer21

Follow-ups from the review rounds of #254 (Model K softmax weight rounding in one mixed-precision FMA, with a create-time self-check). None of them changes the output of a frame that passes the self-check. They are tracked here for 0.6.6.

1. Shape coverage beyond the sampled sizes

The self-check runs each selectable softmax stage at the window offsets the frames use, but only at sampled output sizes (for example, the m stage only at one large base). A defect that depends on the map size, such as a small m map or partial windows at an edge, could pass the check. Add shape-dependent negative controls, and cover boundary and partial-window shapes, or verify the compiled data flow independently of the sampled sizes.

2. Preparation that silently skips a port

The create-time preparation step returns success for a manifest entry whose variant does not match or whose root signature is unsupported, while the launch path later refuses the same entry. Such an entry, even with a corrupt shader, then passes create and fails on the first frame. Make preparation fail on these cases, and test them.

3. Stale embedded manifest and shader pairs

The manifest and each shader fall back to the embedded copies independently. A runtime built from a reused output folder could embed an old, matching pair, which then passes the integrity check when the loose files are absent. The fixed-weight stages are not in the output-hash self-check. Bind the embedded artifacts to the intended build's sources (the new per-entry source digests are a start), and refuse stale ones before embedding.

4. Self-check digest inputs and reference provenance

The digest that binds the reference header to its sources does not yet include the stage table and layout helpers, the FP16 conversion used to make the synthetic inputs, or the separately compiled probe. The header generator also stamps the current digest onto a reference build without checking that build's own provenance. Include those inputs, and verify the reference build before stamping.

5. Create time

Native create with the self-check takes about 1.7 s at 1440p Quality, and the game-path executor creates the runtime twice. Reduce the duplicate check (for example, reuse a verified result safely within one process), and measure end-to-end creation on supported devices.

Activity

  1. added this to the 0.6.6 milestone on Nov 16, 2026
  2. phantoslayer21 commented on Nov 16, 2026

    @phantoslayer21
    OwnerAuthor

    One more follow-up, from the fix for Model K's direct-route legacy-input frames:

    6. Nullable kernel arguments

    A struct argument that is a null buffer pointer is packed as arena offset 0, so a kernel that reads or writes that field touches the start of the arena. The fix for the Model K direct-route legacy input adds a narrow create-time assert and a launch-time refusal for that one case. A general rule would be safer: each kernel's argument structure declares which pointer fields may be null (fields the kernel never reads in that variant), and packing refuses a null pointer in every other field, for every model.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions