DSV4: Runtime gap inventory vs local main and PR #29

[kentwait]

Question

Given the fork's mainline (Gemma-only, binary full-attention mask, INT4/INT8 affine weights, GQA KV rings, 128-expert layers, top-8 softmax router) and unmerged upstream PR drumih#29 (Qwen 3.6: ModelFamily plumbing, gated-DeltaNet linear attention, partial RoPE, attention output gate, 256-expert top-8 layers), what does DeepSeek-V4 additionally require?

Deliver a gap list mapping each missing piece → runtime area (Metal kernels / KV cache / expert IO and slots / repack and format / tokenizer / product surface). Include: MLA latent KV cache layout, DSA sparse attention, per-layer KV compression, mxfp4 expert dequant (if needed), the noaux_tc router, the SwiGLU clamp, o_groups, YaRN, 256-expert ~13 MB-class expert blobs vs the 16-slot LFU cache, and the fork/upstream state of this checkout.

Read this repo's Sources/ and docs/, and PR drumih#29's code at /Users/kent/repos/kentwait/turbo-fieldfare-upstream (branch qwen36-support). Do not modify sources.

[kentwait]

Closed as resolved.

[kentwait]

Resolved.
Gap inventory complete (see dsv4-runtime-gaps.md): fork main has no Qwen code (no ModelFamily); PR drumih#29 is the plumbing template and must land/rebase first. 15 gaps; the hard ones: (1) MLA latent KV cache + sparse-attention kernel + compressor/indexer (new kernel families), (2) FP8/I8 or mxfp4 quantization (new dequant paths + repack format), (3) ~13.36 MB expert blobs vs the 16-slot cache → ~9 GB slot reservation if all layers open (memory-contract design decision, now D3). Reusable from drumih#29: family plumbing, partial RoPE, 256-expert repack, XML tool parser. Verdict: large but bounded and tractable; no hardware feature is missing.
Assets: branch research/dsv4-feasibility — findings file scratch/research/dsv4-runtime-gaps.md.