feat(gemma4): feature-complete backend with DFlash + MTP + sparse-FA decode (supersedes PR #175 skeleton)

[dusterbloom]

Supersedes #171, #176, #179, #184, #185 — those were split against pre-#175 main. After #175 merged howard0su's gemma4/ skeleton into main, the split chain was overtaken by upstream work. This PR re-cuts against current main as a single cohesive replacement: howard0su's 341-LoC AR-only Gemma4Backend skeleton → our feature-complete impl (DFlash + MTP + sparse-FA decode + per-layer embedding + MoE expert metadata + asymmetric KV).

Validation (RTX 3090, Gemma4-26B-A4B-it Q4_K_M, apples-to-apples binary built from this branch's libdflash27b.a)

Cell	Config	Prefill (tok/s)	Decode (tok/s)	TTFT	MTP accept	VRAM
1	MoE MTP γ=1 + TQ3+sparse-FA @ 16k (8k-tok prompt)	2137	43.86	46 ms	0.56	18.54 GB
2	MoE MTP γ=1 + TQ3+sparse-FA @ 64k (50k-tok prompt)	1544	20.86	99 ms	0.52	19.03 GB

Byte-identical to upstream llama-cli AR Q4 @ temp=0 greedy on this same model (Q4_0 KV is lossless; our forward implements the same Gemma4 architecture; identical argmax).

Slop-fix parity: cell 1 on pre-slop binary (May-13 feature/gemma4-support tip) = 42.85 tok/s decode vs. 43.86 tok/s on this branch's post-slop binary — within run-to-run noise. Confirms S1–S10 cleanups are zero-perf-cost.

Scope by layer

Core (target forward + AR decode + loader)

• gemma4/gemma4_loader.cpp — Per-layer-embedding (PLE) table, MoE expert metadata, asymmetric KV override for MTP donor layers
• gemma4/gemma4_graph.cpp — iSWA target forward in pure ggml, per-layer-embedding decode path
• gemma4/gemma4_backend.cpp — decode_autoregressive, prefill, snapshot save/restore
• gemma4/gemma4_daemon.{h,cpp} — protocol wrapper, mirrors run_laguna_daemon
• gemma4/gemma4_internal.h — 17-line passthrough to ../internal.h (struct defs sit alongside other archs)
• Replaces howard's Gemma4Weights/Gemma4Cache names with our existing GemmaTargetWeights/GemmaTargetCache family (canonical across all archs in this tree)

TQ3 KV (TurboQuant)

• Ring-buffer + SWA boundary correctness for long contexts
• Asymmetric KV override: when MTP loads, donor layer K/V forced to Q8 (TQ3 unreadable by MTP cross-attn)

Sparse-FA decode dispatch (commit 2 — refactor: rename use_pflash → use_sparse_fa)

• use_sparse_fa flag + F5 v2 gate at gemma4_graph.cpp can_sparse_fa predicate
• Decode-only (n_tokens == 1) because CUDA dispatcher at deps/llama.cpp/ggml/src/ggml-cuda/fattn.cu:572-576 has no sparse-FA kernel for head_dim=512 + mask + S>1 (BEST_FATTN_KERNEL_NONE abort)
• Not pFlash — see naming clarification below

DFlash speculative decode

• gemma4_dflash_graph.cpp (top-level — separate model, the drafter)
• gemma4_backend.cpp::decode_dflash (425 LoC inner loop, DDTree verify)
• target_feat capture in target graph (hidden states fed to draft KV prefill)
• Draft KV prefill helper (run_draft_kv_prefill)

MTP γ=1 (Multi-Token Prediction)

• gemma4_mtp_graph.cpp (top-level — separate head, the MTP cross-attn)
• gemma4_backend.cpp::decode_mtp
• mtp_h_prev capture (post-output-norm hidden, fed to MTP cross-attn)
• Asymmetric KV override activation

Naming clarification (commit 2)

The previous internal naming used use_pflash / pflash_alpha / s_pflash_tq3 / DFLASH_PFLASH_TQ3 to refer to the ggml_flash_attn_sparse op. This collides with the PFlash product (Python speculative-prefill compression module at pflash/, which calls pflash_register_ggml_kernel() to register its BSA forward as the implementation of ggml_flash_attn_sparse).

Commit 2 renames the Gemma4-scoped flag to use_sparse_fa etc. PFlash product files (pflash_ggml_adapter.{cpp,h}, pflash/ dir, README/RESULTS prose) are untouched — those are correctly named.

A real pFlash (= prefill flash, a D=512 sparse-FA-with-mask CUDA kernel) is future work, separate from this PR, blocked on a D=512 kernel that doesn't exist yet in upstream ggml-cuda.

Out of scope (future PRs)

• MTP γ=2 — needed for >1× chain speculative decode (1.5-2.5 days estimated)
• Real pFlash D=512 sparse-FA-with-mask kernel — would widen the F5 v2 gate to allow prefill dispatch (20-32 senior-CUDA-days estimated per a private kernel-author scoping)
• Bench harness (test_gemma4_dflash.cpp) — currently lives only on feature/gemma4-support branch; will be brought into dflash/test/gemma4/ as a follow-up

Commits

1. b0b4b7e feat(gemma4): feature-complete backend with DFlash + MTP + sparse-FA decode (supersedes PR #175 skeleton) — 22 files, +7887/-1081
2. 9b311d0 refactor(gemma4): rename use_pflash → use_sparse_fa (clarify ggml_flash_attn_sparse is not the PFlash product) — 8 files, +46/-46

Closes

#171 (target loader), #176 (target graph), #179 (sparse-FA integration), #184 (backend), #185 (daemon routing) — all superseded by this PR.