PoC: per-layer split or fused-uniform QKV topology

[malaiwah]

Purpose

Proof of concept for #297: allow one checkpoint to choose QKV topology per full-attention block.

• split: existing q/k/v EXL3 objects, independent K/codebook/scale choices, preserving mixed precision and the serving consumer's three-launch route.
• fused_uniform: concatenate the BF16 q/k/v source weights in frozen q,k,v order, quantize that matrix jointly at one common K/codebook/scale, and serialize one qkv payload for a one-launch consumer.

This is the practical hybrid requested by the downstream allocator: retain split widths where q/v sensitivity matters; use uniform fusion only where a new fused direct marginal shows no decision-relevant fidelity loss. A heterogeneous-width one-launch tuple remains future kernel/format work.

Implementation

Adds versioned exl3_qkv_topology/1 metadata and a deterministic converter plan:

• complete split-default rows with per-layer fused overrides;
• direct BF16 q/k/v loading and exact q,k,v-order concatenation;
• one joint quantization call retaining the input calibration qmap;
• mixed split/fused Attention construction and runtime routing;
• config, safetensors/index and tensor-storage logical reconstruction metadata;
• fail-closed common-K/codebook/scale, unknown/duplicate declaration, missing/duplicate payload and component checks;
• fused TP explicitly rejected by the PoC rather than silently mis-sharded.

No load-time re-encoding and no duplicate split+fused payload for one block. Split remains the default.

Verification

The companion source verifier pins the patch/base/patched identities and mixed-schema invariants. Focused tests were executed in the immutable CUDA 13.2/r34 environment after compiling the pinned source extension:

6 passed, 1 warning in 1.89s

Coverage includes mixed maps/runtime construction, BF16 concatenation/split bit identity, common-K/codebook/scale enforcement, unknown/duplicate declarations, and missing/duplicate payloads.

A consumer PoC is open at local-inference-lab/vllm#454. The ~88 us split versus ~36 us fused figure remains modeled; this PR enables the real block/end-logit/graph/TG experiment and makes no performance claim.

AI assistance was used to implement and review the PoC; the patch identity, extension build and focused tests were independently executed as described.

[malaiwah]

Adversarial review follow-up is now in 3777614a5af305d8e2db4bab87f940983e9c71e0.

• No-plan conversion now short-circuits topology handling and preserves existing behavior, including ordinary 3inst conversions.
• Opted-in topology handles Qwen3.5/3.8 interleaved Q/G with the real doubled q projection width ([12288,1024,1024] for the target).
• Schema rows/completeness are scoped to the primary text decoder; vision and MTP stay on their established payloads.
• The producer's opted-in schema domain now matches the consumer: K3..K8 and mcg/mul1, with lexical full-path ordering.
• Added realistic interleaved, side-model exclusion, no-plan, and domain-boundary tests.

Focused result on the pinned GG test image: 21 passed.

[malaiwah]

Actual layer-3 producer pilot found a remaining gated-Qwen execution bug: the fused path serialized the correct 14,336-wide QG/K/V payload but Attention.project_qkv still split only 6,144 Q values. 0f92b6f4 now splits the complete 12,288-wide interleaved Q/G component, deinterleaves it through the established fast/fallback paths, and returns the 24×256 query plus 6,144 gate values. Added a forward regression over the real 24/4/256 geometry. Focused producer tests: 22 passed; combined producer+GDN source tests: 26 passed.

[malaiwah]

The next actual pilot reached the extension fast path and exposed one more concrete contract: torch.split returns a non-contiguous Q/G view, while ext.deinterleave_qg requires contiguous input. dce9cfa materializes only that fused Q/G slice before the existing deinterleave kernel and adds a half-precision fast-path regression that asserts contiguity. Producer tests: 23 passed; combined producer+GDN tests: 27 passed.