feat(gemma4): full 12B forward from one token through window-edge pfill

[FeathBow]

Description

Closes #878.

The full text-tower forward, prefill form: scaled embedding, all 48 layers dispatched by the parsed layer map, the final norm, the tied LM head, and the final-logit softcap — every prompt at or below the 1024-token window, where causal attention IS sliding attention. Geometry and the end constants come from the weights' own config inside the forward — one source of truth for every buffer size handed to CUDA — and token ids are bounds-checked against the embedding before upload, since the embedding kernel validates neither its output width nor the row it reads.

The end constants are the ones a from-the-paper implementation gets wrong: the embedding multiplier is the bf16 rounding of sqrt(hidden_size) (exactly 62.0 at hidden 3840 — the reference casts the scale to the weight dtype before multiplying), derived in one place and asserted against the fixture's recorded value before use; the LM head is the embedding matrix itself; the softcap rides the typed config, validated positive at parse.

The oracle compares log-probabilities at the reference's top-64 ids per position (host log-softmax over the f32 upcast, f64 accumulation for the 262k-way sum), with per-case gates declared from measurement and the window-edge gate distributional — the reasoning and the HF-self-comparison baseline are in the issue.

Test Env

Single GPU (sm_89, x86_64), CUDA 12.9, against the pinned 12B checkpoint.

Verification

• single (1 token): max |dlogprob| 0.162 over the top-64, top-1 exact — under the 0.3 gate with top-1 required exact.
• short (9 tokens): max |dlogprob| 0.638, top-1 exact at all nine positions — under the 0.9 gate with top-1 required exact. Both short cases sit at the bf16 rounding floor.
• edge (1024 tokens, exactly the window): top-1 agreement 821/1024 with max |dlogprob| 16.19, misses spread across all four quarters (229/190/204/198, first miss at position 27) — against the distributional gate: global top-1 >= 0.75, chaos ceiling 20, and per-quarter top-1 >= 0.65 bounding localized damage — a whole-quarter failure cannot pass on the strength of the others. The baseline making those numbers meaningful: two HF attention backends — sdpa vs eager, same weights, same sequence — agree at 802/1024 with max 16.4, per-quarter 224/186/199/193 (worst quarter 72.7%); our forward sits at that noise floor.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update

[FeathBow]

@codex review please

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Swish!

Reviewed commit: b2fff6f08b

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".