cleanup: VRAM-layer audit — dead modules, staging gate, one reserve floor, honest budget logs

[kekzl]

What

Verified findings from the 2026-07-07 VRAM management audit — full write-up with per-finding evidence, the deliberate-design list, and measurement-gated follow-ups in docs/audit/vram_audit_2026_07_07.md. Net: −757 LOC, one real allocation-gate fix, the rest is dead scaffolding and diagnostic honesty.

Dead code removed

• GDNState module: init() never called, Engine::gdn_state_ never assigned (always null, every read guarded); GDN models run recurrent state through SSMState. Also removes the dead InferenceState::gdn_state/gdn_seq_id fields and the never-read gdn_layer_map_.
• DeviceAllocator + PinnedAllocator: zero runtime callers (unit tests only); the async default pool is tuned directly and prefill pinned staging uses raw cudaHostAlloc.
• MemoryManager facade collapsed: its remaining job was holding a VRAMAllocator — Engine now holds it directly; the single compute_budget caller uses the free function.
• Stranded plan_tier_of() wrapper (comment-only references), Phase-4b constexpr actually_free = true bisect leftover with dead branches.

Real fix

• moe_.raw_staging_buf gated on any_host_packed_experts: its only consumers are the !packed.on_device branches of the legacy MoE forward, but it was allocated whenever packed experts existed — dead weight (1 expert raw, MiB-range) on every all-on-device GGUF-MoE load. Deliberately NOT using the LRU predicate with its gpt-oss exemption — gpt-oss' transiently host-resident experts keep the staging fallback.

Consolidation / honesty

• nvfp4_beneficial() hoisted to core/qtype.h as single policy source (the budget re-derived it as a local lambda — the G1 duplication class, in sync today, drift waiting to happen).
• New vram_reserve_floor() helper replaces the 6 true max(total/10, 256 MiB) derivations; normalizes one copy that had drifted to a 1 MiB floor. Semantically distinct reserves (kv-floor clamps, additive safeties) intentionally untouched.
• Budget log no longer reports a phantom fp8= reservation when FP8 prefill is resolved off (the sm_120 default): fp8_cache_bytes had zero consumers in that state. Strategy line now reads NVFP4_DECODE (fp8 prefill off). KV sizing math unchanged.
• Removed the vacuous !plan.failed guard on the fix: capture-poisoned engine wedge + planner-driven KV pool sizing #875 GGUF reserve branch (the budget-side plan runs unconstrained → failed unreachable there) and documented the deliberate two-plan design at both sites.
• MemAccount per-pool notes wired for weight-cache build totals + executor workspaces; the header now promises exactly what is wired.

Perf

Perf-neutral by design — no hot-path changes, no baseline refresh needed.

• verify-fast: tests PASS, smoke PASS, pp512 +5.3% / pp4096 +73% vs baseline (above, FA2).
• decode tg128 read −20.7% with the depressed-host signature (mem 405 MHz median during bench; script auto-degraded to WARN per Q8_0 GGUF perf drift since 2026-05-29 baseline: tg128 259->238 (-8%), pp512 8260->7150 (-13%) #526).
• graphs on/off gate re-measured on warmed clocks: 2.58× (threshold 1.3×) — the scripted 1.26× run was the same cold-clock artifact.
• MoE smokes: GGUF Q4_K_M 30B logs MoE staging buffer: skipped (all packed experts on device), output coherent; Coder-30B-FP4 292 tok/s decode, VramBudgetReserve 2/2 green.

🤖 Generated with Claude Code