Qwen3.5 paged attention bypasses the shared AttnBlock seam

Row: BACKEND-ROCM

Gap

The numerical-characterization review for #2773 found that Qwen3.5 paged attention does not enter dense_attn::AttnBlock. The repository requires that shared seam for decode. This issue owns reconciliation of that existing route; #2773 owns characterization of the route that actually executes.

At local revision 41b6aa2560a576c28fa15f2a3ff4505305f9b850, src/vllm/model_executor/models/qwen3_5.cpp::RunDenseLayerPaged calls FullAttnBlockPaged at line 7758. FullAttnBlockPaged starts at line 5726 and calls dense_attn::WriteKvCache at line 5909. The file has no call to dense_attn::AttnBlock.

This is a source finding, not a new runtime failure. The custom route predates this characterization: git log -S'FullAttnBlockPaged' identifies the paged-forward additions f1ae0183f and 13fd1ea7f. Characterization probes must name this executing path rather than assert the mandated seam is already reached.

Scope and gates

Prepare a committed spec before implementation. Determine whether the shared seam can represent Qwen3.5 paged attention; extend it where necessary, then route the model through it without changing the pinned vLLM behavior or dtype defaults. Preserve the quantized arms.

Require a failing test through ModelRegistry::Forward, focused correctness gates, and a fresh review that deletes the production call site and proves the gate fails. Read the complete matching active-pin vLLM chain and run the identical workload before acceptance.

The active vLLM pin is e126687a9a828d513c01a07cd69f025f27d63280. Its local source objects are available. A runnable matching ROCm oracle remains PENDING in #2773. No performance result or shared-seam completion is claimed here.

Ownership

BACKEND-ROCM owns the gap. The spec .agents/specs/rocm-qwen35-08b-cpu-gfx1100-numerics.md will name this issue under ## Owed and identify the existing route explicitly. Wiring is separate from #2773's instrumentation change and needs its own reviewed spec. The exception records existing debt; it does not waive the shared-seam requirement.

Start with src/vllm/model_executor/models/qwen3_5.cpp, especially RunDenseLayerPaged and FullAttnBlockPaged, then read the active-pin vLLM chain at e126687a9a828d513c01a07cd69f025f27d63280. Prepare the required spec and a failing ModelRegistry::Forward test before implementation. Done means the paged route uses dense_attn::AttnBlock without changing pinned behavior or dtype defaults, preserves quantized arms, and passes the focused correctness gates.