feat(amd): prepare gfx950 SiTU prefill

[panditsa]

Summary

• add a SiTU-v2 epilogue option to the gfx950 package-prefill stage
• preserve logical W13 K metadata while retaining the padded W2 aliases required by stage 2
• keep both aliases as zero-copy views over the preshuffled weights

Performance scope

This is a foundation-only stack entry. At this boundary no production caller passes situ_linear_beta, so DO_SITU remains false and there is no standalone end-to-end performance claim. #1138 adds the A8W4 core and #1139 registers/selects it.

Tests

• gfx950 MXFP4 regression suite: 4 passed, 6 skipped
• metadata check confirms W13 exposes logical K, W2 retains physical padded K, and both aliases preserve the original data pointers
• downstream selector and end-to-end coverage is in (WIP) feat(moe): select gfx950 TP A8W4 SiTU #1139

Stack

• depends on perf(kimi3): tune small-batch latent projection #1135

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 58b5342883

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Keep non-tile-aligned W13 shapes off this kernel

When the logical packed K is not a multiple of 128 (for example, GPT-OSS hidden size 2880 gives original_k_pk=1440), shortening this alias makes the package eligibility check in fused/moe.py pass. However, prefill_stage1.py uses cdiv(K, 256) and performs unmasked 128-byte activation and eight-scale loads for the final tile, while _quantize_mxfp4_activation allocates only the logical 1440 packed bytes and 90 scales. Batches of at least nine tokens can therefore read past the quantized activation buffers; either pad/mask the activation tail or keep these shapes on the fallback path.

Useful? React with 👍 / 👎.