[perNyfelt]

Summary

Implements Phase 7.3b WP5: Qwen3-VL vision tower (req/plans/phase7-3b-plan.md), per the design recorded in req/plans/phase7-3b-reference-findings.md §11 (committed on this branch).

Lands the vision tower + mergers — the last model-side piece before Qwen3VlModel composition in WP6:

1. New jmlx-core components (se.alipsa.jmlx.nn): PatchMerger (the 2×2 spatial merge), TwoAxisRope/TwoAxisRotation (interleaved two-axis rotary preparation) and TwoAxisRotaryAttention (the vision block's qkv-projection attention), each with its own strict core tests.
2. Qwen3VlVisionTower (jmlx-models): patch embed + 3-D interpolated position embeddings (the torch 2.6.0 CPU linspace_kernel ported exactly), N Qwen3VlVisionBlocks, DeepStack level merges + the main merger, returning Qwen3VlVisionFeatures (merged rows + per-level DeepStack rows) hoisted into the tower's scope. Each image runs in a private child scope that closes on return (the SmolVlmModel per-request pattern). A test-only capture sink feeds the golden.
3. Strict HF intermediate-feature golden: tools/hf-reference/qwen3vl_vision.py (manual, reviewed, locked 4.57.6 venv — never Gradle-invoked) generates block-level captures (norm/attn/mlp outputs, DeepStack + merged rows, position-lookup corner indices) for 4 probe images, plus a tiny synthetic vision checkpoint (HF default init) and the images, all SHA-256-pinned in tools/hf-reference/provenance.json and verified by verifyHfReferenceGoldens.
4. Qwen3VlVisionTowerGoldenTest (tagged full-float32): every case max diff ≤ 3.1e-6 against the 1e-4 / 7-significant-digit tolerance; 3/3 executed.
5. Qwen3VlVisionPeakTest (ordinary test): real-2B tower geometry at reduced depth 3, one synthetic 1024×1024 image (grid (1, 64, 64), 4096 patch rows).

Root cause found by the golden

The first golden run diverged at fc1 of block 0: Java applied norm2 to the block input. The pinned 4.57.6 Qwen3VLVisionBlock.forward is two sequential statements —

hidden_states = hidden_states + self.attn(self.norm1(hidden_states), ...)
hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))

— so line 274's norm2(hidden_states) receives x + attn(norm1(x)), the updated residual. A forward-pre-hook on the actual model's blk.norm2 confirmed its input digest equals the attention-added residual, not the block input. Fixed in Qwen3VlVisionBlock.forward; the findings §11.1 verbatim transcription was always correct. (The norm1/norm2 identical-digest anomaly seen during debugging is not a bug: the tiny checkpoint uses HF default init, so every LayerNorm is identity.) The attention path itself verified within the golden's 1e-4 tolerance (measured max diff 3.1e-6) before the fc1 divergence point.

Exit evidence — measured per-image peak (findings §11.4)

Method per findings §11.4 (all statics materialized before the settled baseline so the delta excludes weight materialization):

• Sustained per-image evaluated peak: 16,777,216 B — exactly 16 MiB: the merged 1024 × 2048 float32 rows plus one DeepStack level, nothing else retained after the per-image scope closes.
• Transient attention workspace: 335,544,320 B = exactly 320 MiB (0.3125 GiB) over baseline (the bracket runs after every per-request scope closes, so it carries no retained-row offset). MLX's fused scaled-dot-product kernel tiles the score computation — the full 16-head 4096×4096 float32 score matrix (1,073,741,824 B = 1 GiB) is never resident.
• Environment: M5 Max, ordinary test task (TF32 default mode). Both figures are far below the plan §4 5 GiB planning bound.

CI

The "Assert required native suites executed" step gains Qwen3VlVisionTowerGoldenTest (jmlx-models list + float32_suites, so it is required in :jmlx-models:float32GoldenTest) and Qwen3VlVisionPeakTest (jmlx-models list, ordinary test).

Validation (all local, green)

• :jmlx-core:check (432 test / 30 float32, 0 failures; the single skip is the pre-existing MoeMlpBenchmarkTest), :jmlx-models:check (383 test / 50 float32, 0 failures), root verifyHfReferenceGoldens, spotlessApply clean, git diff --check clean.
• Result XMLs (unfiltered reruns): Qwen3VlVisionTowerGoldenTest 3/3 in float32GoldenTest, Qwen3VlVisionPeakTest 1/1 in test; the CI-suite assertion logic passes against them.
• No golden/fixture regeneration: Java was made to match the goldens; committed digests verified unchanged by verifyHfReferenceGoldens.
• No native pins, no generated bindings, no Gradle task changes.

Docs

• Findings §11.1–§11.4: verbatim 4.57.6 block source, golden schema/provenance, measurement method and the measured peak (replacing PEAK-TBD).
• AGENTS.md: WP5 paragraph in "Phase 7.3b additions".

Explicitly deferred

• Qwen3VlModel composition, loader dispatch and the DeepStack consumption of VisualFeatureInjection → WP6 (plan §7); generation/lifecycle/cache policy → WP7.