New model architectures at low cost: one quant dispatch table and shared attention/FFN blocks #5315

Scope: Making the next architecture cheap: quant-type dispatch consolidated from about 30 files into one table checked against ggml.h ids, the shared attention/FFN blocks adopted, contract-first model families (qwen3_moe), the qwen35moe forward, and tests for the GDN layer loader.

Why it matters: Each new architecture today re-implements dispatch and blocks; Any Model means adding one is a contract and a small module.

Done when: One quant dispatch table is the only source of truth (gate against ggml.h ids), and qwen3_moe and qwen35moe are added through the shared blocks with a parity receipt each.

Epic: #4001. Folded under triage rule 7b (2026-10-14): each row below was its own issue, closed as not planned and kept here as a sub-ticket. Reopen an issue when its work starts or its PR opens.

Sub-tickets

• ARCH: quant-type dispatch is duplicated across ~30 files with no single source of truth — propose a dedicated consolidation release #3418 ARCH: quant-type dispatch is duplicated across ~30 files with no single source of truth — propose a dedicated consolidation release (milestone 0.74.0); branch la-74/3418
  • accept: the outcome in the title, with a receipt
• ARCH: shared attention/FFN building blocks exist but 0% adopted — new architectures still cost ~1,600 lines each #3422 ARCH: shared attention/FFN building blocks exist but 0% adopted — new architectures still cost ~1,600 lines each (milestone 0.74.0); branch feat/3422-pp-arch-adoption
  • accept: exit criterion is measurable: a new dense or MoE architecture (the next real test case
• PP-QUANT-001 M2: reconciliation gate — GgmlType vs the vendored ggml.h id list at a pinned sha; exhaustive matches, no _ => #3431 PP-QUANT-001 M2: reconciliation gate — GgmlType vs the vendored ggml.h id list at a pinned sha; exhaustive matches, no _ => (milestone 0.74.0); branch la-74/3431
  • accept: the outcome in the title, with a receipt
• qwen35moe (Qwen3.5-35B-A3B, hybrid Gated DeltaNet + MoE): CPU and CUDA forward #4801 qwen35moe (Qwen3.5-35B-A3B, hybrid Gated DeltaNet + MoE): CPU and CUDA forward (milestone backlog)
  • accept: Done when:
• E7: qwen3_moe model-family contract first (contract → code → parity) #5056 E7: qwen3_moe model-family contract first (contract → code → parity) (milestone 0.74.0)
  • accept: Done when: pv validate is green on the family contract; ArchConstraints for qwen3moe is generated from it; CPU parity vs llama.cpp is measured on more than
• coverage: test the Qwen3.5 GDN layer loader (extract_linear_layer_generic) #5214 coverage: test the Qwen3.5 GDN layer loader (extract_linear_layer_generic) (milestone 0.71.1)
  • accept: the outcome in the title, with a receipt