Genesis vLLM Patches — Complete Reference

This file is the curated narrative reference for every Genesis runtime patch. For each patch you get: ID, title, what it does, status (ON / opt-in / deprecated), env flag to toggle, upstream PR (if backported), and credit.

📊 Full machine-readable table — auto-generated from dispatcher/registry.py — at PATCHES_AUTO.md. Regenerate via python3 scripts/generate_patches_md.py. CI gate: --check mode (audit 2026-05-11).

🛡️ Release policy — which patch-proof mode gates a public release — is documented in RELEASE_POLICY.md. The current public gate is require-static; two stricter ratchets (require-bench, require-baseline) are available for operators preparing a hardened deploy.

Current state (v12.1.0, 2026-06-01)

Total PATCH_REGISTRY entries: 329 — range covers P1–P109 legacy + PN8–PN522 modern (through the July 2026 PN517–PN522 wave) + G4_01–G4_85 Gemma 4 / MoE family + sub-patches (P5b, P7b, P15B, P18b, P38B, P39a, P61c, P67b, P67c, P79d, PN26b, PN40-classifier) + library/diagnostic (P51, P102, P103) + the standalone SNDR_WORKSPACE_001 entry. P56/P57 archived 2026-05-05; PN71 burned by abandoned prealloc_v2 experiment; PN78 tombstoned; P39a flipped to community per the strict-AND audit.

Library modules (not in PATCH_REGISTRY but available for patches to use): sndr.cache.eviction_policies — LRU/2Q/ARC eviction policy classes (37 unit tests, scan-pollution verified). Awaiting full backport of vllm#40270 with proper BlockPool integration before being added as a registry patch.

Metric	Count
Total PATCH_REGISTRY entries	329
Tier=community (Apache 2.0, sndr)	324 (all entries)
Tier=engine (commercial, sndr_engine)	0 (PN72 reclassified to community 2026-05-08; sndr_engine namespace reserved but empty)
Default-on at boot	56
Lifecycle=experimental	221
Lifecycle=legacy (pre-dispatcher)	28
Lifecycle=retired	60
Lifecycle=research	3
Lifecycle=stable	13 (G4_01, G4_02, G4_03, G4_04, G4_05 [retired], G4_09, G4_11, G4_12, G4_13, G4_16, G4_23, G4_25, PN33, PN35 — ratchet active with `stable_kind` declared; G4_14 retired 2026-07-05; see CONTRIBUTING.md § Promoting a patch to lifecycle=stable)
Lifecycle=coordinator	4 (env-flag-only, no real binding)
Implementation status=full	267
Implementation status=marker_only / placeholder / partial / retired	37 (22 + 2 + 7 + 6)
Apply-loop coverage (apply_module set)	305 / 329 = 92.6%
Spec-only (intentional, allow-listed)	17 (P1, P17, P18b, P20, P23, P29, P32, P51, P102, PN60, PN63, PN64, PN256, PN261, G4_70, G4_70B, G4_70C)

Engine tier (the strict-AND boundary)

A patch is tier="engine" only when ALL four conditions hold (Sander's strict AND rule, audit 2026-05-08):

NOT present on the public github repo.
NO external author credit in title/credit text.
NO PR link / PR number in title/credit text.
NO upstream_pr / related_upstream_prs field.

After the strict-AND + github-presence audit (2026-05-08):

PN72 was the only candidate but reclassified to community since its real algorithm lives in sndr/engines/vllm/kernels_legacy/ngram_frequency_filter.py (subsystem boundary). sndr_engine namespace remains reserved but empty; engine_available() returns False.

Everything previously parked in engine (P67/P67b/P67c, PN21..PN24, PN26, PN29, PN38, PN40, PN57, P82, PN16, PN65, plus 30+ legacy P*) was reclassified to community and ships in sndr/engines/vllm/patches/<family>/ under Apache 2.0.

Tombstones (deprecated, no-op)

PN78 — post-warmup empty_cache() wrap. Deprecated 2026-05-07 same day it was created: apply() returns "skipped" unconditionally. Investigation found vllm pin already calls torch.accelerator.empty_cache() twice in gpu_model_runner.py:capture_model() (lines 6213 + 6244), so the wrap would be redundant. Architectural lesson: monkey-patches in apply_all don't reach spawn'd workers — only source-level edits do.

Recent additions (June–July 2026 waves)

PN520 (experimental, GENESIS_ENABLE_PN520_QWEN3_5_LOAD_WEIGHTS) — restore the imperative Qwen3.5/3.6 GDN weight loader (revert of vllm#47058). Battle-validated 2026-07-04 on dev748: 96 in_proj_ba shards routed through the restored loader; the INT4-27B degeneration (coherent boot, garbage output) is CURED — see the fleet-sweep table in BENCHMARKS.md.
PN521 (experimental, GENESIS_ENABLE_PN521_TQ_RAW_TAIL_VERIFY) — TurboQuant raw-bf16-tail spec-verify: fixes the INT4 non-pow2-GQA (Qwen3.6-27B AutoRound) TQ k8v4 × MTP token-repetition collapse. Required for MTP on the 27B TQ preset (root-caused 2026-07-02).
PN521_SPLIT_K (experimental, GENESIS_ENABLE_PN521_SPLIT_K) — split-K (Flash-Decoding) occupancy fix for the PN521 raw-tail verify — ~11× faster verify kernel single-stream.
PN522 (experimental, GENESIS_ENABLE_PN522_TQ_RAW_TAIL_WARMUP) — PN521 raw-tail kernel pre-capture warmup — removes the first-request JIT latency spike on the 27B spec-verify path.
PN517 (experimental, GENESIS_ENABLE_PN517_INIT_SNAPSHOT_BEFORE_NCCL, vllm#45517) — init MemorySnapshot before NCCL: asymmetric TP+PP OOM guard + startup_free observability.
PN518 (experimental, GENESIS_ENABLE_PN518_INC_HYBRID_FP8_DETECT, vllm#46322) — INCConfig hybrid INT4+FP8 AutoRound silent-skip trap-closer (detect-and-WARN; latent, default-OFF).
PN519 (experimental, GENESIS_ENABLE_PN519_SWA_TILE_BASE, vllm#46087) — SWA/chunked KV-tile loop starts exactly at first_allowed_key (tile_base fix in triton_unified_attention consumers).
PN400 (retired — merged upstream, vllm#45656) — is_sym qzeros guard for symmetric AutoRound/GPTQ Marlin MoE (fixes the vllm#43409 regression).
PN401 (experimental, GENESIS_ENABLE_PN401_TQ_PREFILL_CONTINUATION_GUARD, vllm#46461) — TurboQuant prefill continuation guard against silently dropping cached prefix K/V on co-batched continuations. Enabled in the 27B/35B TQ model YAMLs.
PN402 (experimental, GENESIS_ENABLE_PN402_SANITIZE_INVALID_DRAFT_TOKENS, vllm#46574) — sanitize invalid (−1 / over-vocab) MTP draft token ids before batch prep (prevents a single bad draft from IMA-crashing the runner).
PN392 (retired, dev491-era) — qwen3_coder streaming tool-call within-call coalescing regression fix; PN394 (retired — merged as vllm#46047) — qwen3 partial-param < truncation fix.
PN351 (experimental, GENESIS_ENABLE_PN351, vllm#43257) — Triton unified_attention head_dim≥512 tune (pre-existing patch, listed here for its July re-anchor). Its launch-variant anchor was re-derived dual-variant during the dev714→dev748 preflight; the 2026-07-04 fleet-sweep exercise of the lane ran on dev714 (stale image_digest, post-release audit), and the 2026-07-05 re-run battle-validated it on dev748: the applied variant (pn351_kernel_launch_warps_stages_mmprefix, the dev748 anchor variant with the upstream MM_PREFIX_CLAMP_SW insert) was read back from the LIVE dev748 container's triton_unified_attention.py, and the head_dim=512 Gemma4-31B kvauto-chat lane served chat + 7/7 tool-calls with window accept 0.744 at K=3 (TPOT 9.42 ms, noisy CV — within CV of dev714, no gain claim; see BENCHMARKS.md).

Recent additions (v11.0 sprint — historical, 2026-05)

PN95 (2026-05-09, Path C v7.73.x foundation) — Genesis-original tier-aware KV cache + vision sub-tier + Mamba SSM exclusion. Solves club-3090 issue #58 long-context+vision OOM on hybrid-GDN models (which all upstream CPU offload paths — vLLM --cpu-offload-gb, SimpleCPUOffloadConnector, LMCache, SGLang HiCache — crash on because Mamba state lives outside the KV pool). Two anchors: single_type_kv_cache_manager.py::cache_blocks notify_admit, and block_pool.py::get_cached_block notify_touch. Runtime singleton TierManager (sndr/cache/tier_manager.py) owns per-tier EvictionPolicy + CPU-pool slab + vision-first demote ordering + Mamba-exclusion filter + spec-decode hot ring. Schema: cache_config.tiers: [CacheTier(...), ...] with exclude_mamba_ssm: true (mandatory on hybrid GDN). Default OFF; opt in via GENESIS_ENABLE_PN95_TIER_AWARE_CACHE=1 + declared tiers in YAML. Days 5-7 (vision-token MM tagging + Mamba runtime classifier walk + live 27B Lorbus bench) deferred to live integration.
PN16_V6 (2026-05-09, Sprint 4) — Genesis-original streaming <think> token-budget enforcer. Per-request stateful truncator (sndr/engines/vllm/middleware/think_streaming_truncator.py) wraps OpenAIServingChat.chat_completion_stream_generator via class-rebind. When request has tools attached AND GENESIS_PN16_MAX_THINKING_STREAM_TOKENS > 0, counts delta.reasoning_content chunks; once budget exceeded, drops subsequent reasoning chunks and emits one-shot [Genesis] truncation note as delta.content. tool_calls + plain content always pass. Cache-safe (no chat_template mutation) + compute-neutral (model still generates internally). Stacks on V8: V8 nudges, V6 enforces. Default OFF; opt in via GENESIS_ENABLE_PN16_V6_STREAMING_TRUNCATOR=1.
PN90 (2026-05-09, Wave 3.1) — vllm#40269 backport: probabilistic draft rejection. Captures softmax(logits) in _greedy_sample of MTP/Eagle/DFlash drafters, stacks per K-step into [total_draft_tokens, vocab] 2D layout, feeds to rejection_sampler via gpu_model_runner.py site (replaces literal `None,
draft_probs). Rejection rule becomes min(1, target/draft)`
instead of argmax-or-bonus. Expected +0.5-2% acceptance on MTP K=3. Atomic via MultiFilePatchTransaction (proposer + runner). Default OFF; opt in via SNDR_ENABLE_PN90_PROBABILISTIC_DRAFT=1.
PN82 (2026-05-08, PR38 Day 1) — vllm#41873 backport zeroing padded is_prefilling rows so Mamba/hybrid backends don't read stale True from condense() rotation. Hybrid-only via applies_to, default OFF.
PN55v2 (2026-05-08, PR38 Day 2) — same patch_id as PN55, body upgraded from list-only to recursive iterator covering Tensor / list / tuple / Mapping / None / non-tensor sentinels. Unified backport of vllm#41602 + vllm#41896.
PN80 (2026-05-07) — vllm#41845 backport of LoRA tensorizer device kwarg (single-line fix not yet in nightly image).
PN79 (2026-05-07) — In-place SSM state for GDN chunk prefill (vllm#41824 backport). Most complex patch in the registry: 17 anchors atomic across 3 files via MultiFilePatchTransaction. A/B null-impact on single-shot; multi-turn validation pending. lifecycle="experimental".

Archived dead-ends

P56 / P57 archived 2026-05-05 — TQ spec-decode dead-ends superseded by P65 (CG downgrade). Files live in Genesis_internal_docs/_archive/dead_patches/p56_p57_tq_specdec_deadends/.
PN71 — abandoned prealloc_v2 experiment, never landed.

Native upstream features (no Genesis backport needed)

The following capabilities are native to the current pin (0.23.1rc1.dev748+g2dfaae752 — SSOT: sndr/pins.yaml) and require no Genesis patch — Genesis would only add a stub. Documented here so users know where to enable them. NOTE: the per-feature caveats in this section were last re-verified 2026-06-11 against the dev259-era pin — re-verify a specific flag path against dev748 before relying on it.

CPU KV cache offloading (vllm#37160 lineage, merged upstream) — canonical knob in the current pin is --kv-offloading-size <GiB> (CacheConfig.kv_offloading_size), which routes to the native OffloadingConnector by default (kv_offloading_backend: native; VLLM_USE_SIMPLE_KV_OFFLOAD=1 selects the legacy SimpleCPUOffloadConnector instead). The old explicit --kv-transfer-config '{"kv_connector": "SimpleCPUOffloadConnector", ...}' form still works but is no longer the documented path.

Caveats for Genesis stack (re-verified 2026-06-11 against the pin):
- Hybrid GDN models (Qwen3.6-A3B family) ARE now schedulable by the native offload path — the earlier "Mamba state lives outside the KV cache" caveat is stale. In the current pin Mamba/GDN state is a MambaSpec KV-cache group; the OffloadingConnector scheduler handles it explicitly (single-state group, see kv_connector/v1/offloading/scheduler.py), and GDN conv-state sub-projection transfer helpers exist (kv_connector/v1/ssm_conv_transfer_utils.py, used by the NIXL worker; handles GDN [Q, K, V] conv layout).
- Hybrid models require --enable-prefix-caching to align per-group block sizes (vllm/v1/kv_offload/base.py asserts gpu_block_size % hash_block_size == 0). Our 27B PROD profile runs prefix caching OFF — offload stays blocked there until that constraint is lifted or the profile changes.
- MTP spec-decode + native offload was broken in 0.22.x-era pins: the OffloadingConnectorScheduler wrongly included EAGLE/MTP draft-attention groups in load/store scheduling → out-of-bounds GPU block index → cuMemcpyBatchAsync segfault. The upstream fix vllm#44784 MERGED 2026-06-16 and is native (and evolved) in dev748: is_eagle_group is a first-class KVCacheGroup field set by the engine core (v1/core/kv_cache_utils.py) and consumed by kv_connector/v1/offloading/scheduler.py, plus a scheduler-level volatile-trailing-block exclusion that removes the OOB source at its root. The Genesis vendor of the fix (PN383) is retired; retirement re-verified by code on the pristine dev748 tree 2026-07-05. MTP + native offload has NOT yet been exercised on our lanes — validate on a throwaway container before enabling it on the MTP PROD configs (K=5 on 35B / K=4 on 27B as of 2026-07).
- Frees GPU VRAM worth ~tens of K context at the cost of CPU↔GPU PCIe traffic on miss; payoff is prefix-reuse on agent loops (second-pass TTFT restore instead of full prefill).
Other connectors in the pin: LMCacheConnector, LMCacheMPConnector, MooncakeConnector, NixlConnector, OffloadingConnector, MultiConnector, FlexKVConnector, DecodeBenchConnector. See vllm docs for matrix.

Source of truth: sndr.dispatcher.PATCH_REGISTRY (rich metadata; lifecycle-tracked; schema-validated) + sndr.dispatcher.spec.iter_patch_specs() (canonical apply_module resolver; the new spec-driven loop).
Legacy parking lot: sndr.apply._per_patch_dispatch — the 4542-line @register_patch registry kept as fallback during the PR38 Day 6-8 migration to registry-driven apply (opt-in via SNDR_APPLY_VIA_SPECS=1).
All patches default OFF unless explicitly noted. Production launch script enables a curated set via env flags.
Credits: every backport names its upstream author + PR. Genesis-original patches are explicitly labelled. See CREDITS.md.
Status legend:
- default ON — patch self-activates when its config gate passes
- opt-in — requires SNDR_ENABLE_<patch>=1 env var (the historical GENESIS_ENABLE_<patch>=1 form is an accepted, deprecated alias and still works — the two resolve identically; see CONFIGURATION.md)
- deprecated / retired — superseded by another patch; kept for archeology
- library — utility module loaded by other patches, no direct env flag

How to enable / disable a patch

# Enable an opt-in patch (`:nightly` = the current pin; the explicit-hash
# tag vllm/vllm-openai:nightly-2dfaae752
# is the unambiguous reference — see sndr/pins.yaml current_image):
docker run -e GENESIS_ENABLE_P82=1 -e GENESIS_P82_THRESHOLD_SINGLE=0.3 ... vllm/vllm-openai:nightly

# Disable a default-ON patch (rare — usually for A/B testing):
docker run -e GENESIS_ENABLE_P67_TQ_MULTI_QUERY_KERNEL=0 ... vllm/vllm-openai:nightly

# See full list of env vars + defaults:
# → CONFIGURATION.md

Where the code lives

Integrations (text-patcher hooks + runtime hooks): sndr/engines/vllm/patches/<family>/<patch_id>_*.py — organized by subsystem family (Phase 6 reorg, ~15 subdirs: attention/, spec_decode/, scheduler/, kv_cache/, memory/, kernels/, compile_safety/, loader/, lora/, middleware/, moe/, multimodal/, observability/, quantization/, reasoning/, serving/, tool_parsing/, worker/). Resolution is layout-agnostic via compat/categories.module_for(patch_id).
Kernels (Triton / CUDA): sndr/engines/vllm/kernels/
Dispatcher metadata (P56+): sndr/dispatcher/registry.py:PATCH_REGISTRY
Registration: sndr/apply/_per_patch_dispatch.py:@register_patch
Per-patch CHANGELOG entries: CHANGELOG.md (root, search by patch ID — audit 2026-05-11)

How a patch is decided to apply or skip

Every patch goes through the same dispatcher decision tree at boot. The tree below visualises the order of checks — if any check fails, the patch is SKIPPED with a logged reason; if all pass, it APPLIES.

flowchart TD
    START[boot: apply_all iterates PATCH_REGISTRY] --> ENV{env_flag set<br/>to truthy?}
    ENV -->|no| SKIP_OPTIN[SKIP: opt-in not engaged]
    ENV -->|yes, OR default_on=True| LIFE{lifecycle in<br/>SAFETY_GATED_STATES?<br/>retired/deprecated}
    LIFE -->|yes| SKIP_RETIRED[SKIP: retired/deprecated<br/>--allow-retired to override]
    LIFE -->|no| MATCH{applies_to gates<br/>match runtime?<br/>model_class / is_hybrid / GPU}
    MATCH -->|no| SKIP_MISMATCH[SKIP: applies_to mismatch]
    MATCH -->|yes| REQS{requires_patches<br/>all also applied?}
    REQS -->|no| SKIP_DEP[SKIP: missing dependency]
    REQS -->|yes| CONFLICT{conflicts_with<br/>any active patch?}
    CONFLICT -->|yes| SKIP_CONFLICT[SKIP: conflicts]
    CONFLICT -->|no| WIRING{wiring import<br/>+ apply succeeds?}
    WIRING -->|exception| FAIL[FAILED: see logs]
    WIRING -->|drift marker found| SKIP_UPSTREAM[SKIP: upstream merged]
    WIRING -->|anchor not found| SKIP_DRIFT[SKIP: anchor drift]
    WIRING -->|marker already present| IDEM[APPLIED: idempotent]
    WIRING -->|all clean| APPLY[APPLIED: text-patch written]

    style APPLY fill:#90EE90,stroke:#2d6a2d
    style IDEM fill:#90EE90,stroke:#2d6a2d
    style FAIL fill:#FFB6B6,stroke:#a02020
    style SKIP_OPTIN fill:#FFE4B5,stroke:#a05a00
    style SKIP_RETIRED fill:#D3D3D3,stroke:#555
    style SKIP_MISMATCH fill:#FFE4B5,stroke:#a05a00
    style SKIP_DEP fill:#FFE4B5,stroke:#a05a00
    style SKIP_CONFLICT fill:#FFE4B5,stroke:#a05a00
    style SKIP_UPSTREAM fill:#D3D3D3,stroke:#555
    style SKIP_DRIFT fill:#FFE4B5,stroke:#a05a00

Reading guide:

🟢 green = patch active (production effect)
🟡 amber = predictable skip (operator decision or runtime mismatch)
⚪ grey = retired/upstream-merged (no longer needed)
🔴 red = unexpected failure (read logs!)

Boot summary (printed by apply_all) shows the verdict for each patch with the reason. Use sndr patches diff-upstream to see which patches are at risk of upstream-merge drift (the old genesis lifecycle-audit verb no longer exists).

Patches by category

TurboQuant integration

ID	Title	Status	Env Flag	Upstream	Credit
P4	TurboQuant hybrid model support	opt-in	—	—	Genesis (see source / CHANGELOG)
P98	TQ WorkspaceManager revert (vllm#40941 perf hotfix)	opt-in	`GENESIS_ENABLE_P98`	#40941	Reverts upstream PR #40941 WorkspaceManager indirection in turboquant_attn — required for TQ k8v4 on hybrid GDN models (else AssertionError on workspace lock)
P101	TQ continuation 64-token slicing (vllm#41123 SELECTIVE)	opt-in	`GENESIS_ENABLE_P101`	#41123	Selective backport of vllm#41123 TQ on hybrid models — takes the `_CONTINUATION_DECODE_SLICE` improvement only
PN14	TQ decode IOOB safe_page_idx clamp	opt-in	`GENESIS_ENABLE_PN14_TQ_DECODE_OOB_CLAMP`	#40074	devarakondasrikanth @adobe (vllm#40074, OPEN) — fixes upstream issue: TQ decode page-index out-of-bounds clamp
PN353A	TurboQuant MetadataBuilder workspace reserve	opt-in	`GENESIS_ENABLE_PN353A`	#44053	Genesis backport of OPEN vllm#44053 (Bot1822, supersedes #40798). Closes long-context `AssertionError: workspace locked, cannot grow` on TQ + chunked-prefill by reserving the decode-scratch workspace in the MetadataBuilder. PN399 dependency. Enabled in the Qwen YAMLs on the dev148 pin.
PN399	Consolidated single-owner TQ decode-scratch fixed-buffer (IMA-safe FULL cudagraph)	opt-in	`GENESIS_ENABLE_PN399_TQ_DECODE_SCRATCH_IMA`	#46067	Genesis consolidated backport+improvement of OPEN vllm#46067 (Bugfix/TurboQuant — fix CUDA illegal memory access in FULL cudagraph). Single-owner fixed decode-scratch buffer; removes the dead PN118 / PN353A decode reservations. Promoted to PROD on dev148. requires PN353A.

Memory / buffers

ID	Title	Status	Env Flag	Upstream	Credit
P6	TurboQuant-aware attention page size	opt-in	—	—	Genesis (see source / CHANGELOG)
P5	KV cache page size unification	opt-in	—	—	Genesis (see source / CHANGELOG)
P22	TurboQuant shared dequant prealloc	opt-in	—	—	Genesis (see source / CHANGELOG)
P26	TurboQuant prefill output prealloc	opt-in	—	—	Genesis (see source / CHANGELOG)
P44	TQ mixed-batch attn_out pool	opt-in	—	—	Genesis (see source / CHANGELOG)
P46	GDN gating buffer pool	opt-in	—	—	Genesis (see source / CHANGELOG)
P39a	FLA chunk_scaled_dot_kkt persistent A pool	opt-in	—	—	Genesis (see source / CHANGELOG)
P38	TQ _continuation_prefill persistent workspace	opt-in	—	—	Genesis (see source / CHANGELOG)
P36	TurboQuant shared decode buffers	opt-in	—	—	Genesis (see source / CHANGELOG)
P32/P33	TurboQuant cu_2 + synth_seq_lens preallocs	opt-in	—	—	Genesis (see source / CHANGELOG)
P28	GDN core_attn_out prealloc	opt-in	—	—	Genesis (see source / CHANGELOG)
P14	block_table tail zero-fill	opt-in	—	—	Genesis (see source / CHANGELOG)
P20	TurboQuant continuation-prefill FP16 rotate	opt-in	—	—	Genesis (see source / CHANGELOG)
PN12	FFN intermediate scratch pool — Cliff 1 fix on TQ3 path	opt-in	`GENESIS_ENABLE_PN12_FFN_INTERMEDIATE_POOL`	#34207	Genesis-original 2026-04-29 — Cliff 1 fix on TQ3 path. Closes 138 MiB OOM at 192K context
PN17	FA2 softmax_lse runtime clamp (Cliff 1 mechanism A, Issue #11)	opt-in	`GENESIS_ENABLE_PN17_FA2_LSE_CLAMP`	—	Genesis-original 2026-04-30 in response to noonghunna's Genesis Issue #11 cross-rig diagnosis (RTX 3090). Clamps `max_seqlen_k` from `attn_metadata.max_seq_len` (= max_model_len during cudagraph capture) to actual chunk max from `seqused_k.max()` at runtime. Closes Cliff 1 FA2 softmax_lse mechanism; widens long-text-no-vision safe envelope from ~150K to ~205K. Cudagraph-safe (capture-path falls back to original). Reference: Dao-AILab/flash-attention#1011.
PN19	Scoped max_split_size_mb during model load (vllm#41268)	opt-in	`GENESIS_ENABLE_PN19_SCOPED_MAX_SPLIT`	#41268	Backport of vllm#41268 (MatthewBonanni, OPEN). PyTorch 2.10+ introduced load-time allocator fragmentation; mitigates by temporarily setting `max_split_size_mb=20` (PyTorch minimum) during `Worker.load_model`, restores prior on exit. Cudagraph-safe (load-time only). Self-detects torch < 2.11 lacking `_accelerator_setAllocatorSettings` and falls through unchanged. Estimated 200-500 MiB headroom on H100; unverified on Ampere — operator should measure via nvidia-smi peak before relying on it.
PN25	SiluAndMul forward_native opaque-op pool — Cliff 1 mech B fix	opt-in	`GENESIS_ENABLE_PN25_SILU_INDUCTOR_SAFE`	—	Genesis-original 2026-05-01 — closes Cliff 1 mech B (PN12 inductor inline bypass) on TQ3 + spec-decode + TP=1 configs. Registers `genesis::silu_and_mul_pooled` torch op as Inductor-opaque, ensuring FFN intermediate buffer pool fires even when Inductor inlines `forward_native`. Worker-spawn safe (issue #16 fix 2026-05-02).
PN31	FA varlen persistent out buffer (issue #15, sister to P38)	opt-in	`GENESIS_ENABLE_PN31_FA_VARLEN_PERSISTENT_OUT`	—	Genesis-original 2026-05-02 — closes #15 OOM at flash_attn_varlen_func on budget-constrained single-GPU. Per-shape persistent `out` buffer eliminates per-call malloc pressure inside FA C extension. Memory cost: ~16-64 MiB per shape × layer. NULL impact on 2×A5000 PROD (we have headroom); designed for 1×3090/1×4090 community users.

Kernel performance

ID	Title	Status	Env Flag	Upstream	Credit
P3	TurboQuant BF16->FP8 cast (Ampere fix)	opt-in	—	—	Genesis (see source / CHANGELOG)
P23	Marlin FP32_REDUCE env override	opt-in	—	—	Genesis (see source / CHANGELOG)
P31	MoE router fp32 softmax	opt-in	—	—	Genesis (see source / CHANGELOG)
P81	fp8 block-scaled MM low-M decode tuning (vllm#40925)	opt-in	`GENESIS_ENABLE_P81_FP8_BLOCK_SCALED_M_LE_8`	#40925	Backport of vllm#40925 (tonyliu312, OPEN). Specializes w8a8_triton_block_scaled_mm default config for M<=8 (single-reque
P87	Marlin W4A16/W8A16 sub-tile output dim pad-on-load (vllm#40361)	opt-in	`GENESIS_ENABLE_P87`	#40361	Backport of vllm#40361 (OPEN). MarlinLinearKernel requires per-rank out_features divisible by GPTQ_MARLIN_MIN_THREAD_N=64. Sub-tile shards (Qwen3.5 GatedDeltaNet at TP>=2, Intel/Qwen3.6-35B-A3B-int4-AutoRound n=32 shard) fail can_implement and force a slow non-Marlin fallback. P87 zero-pads weights/scales/qzeros at load. Class-rebind on MarlinLinearKernel.{can_implement, process_weights_after_loading, apply_weights}. NOTE 2026-04-28: v764 boots with P87+P93 enabled trigger torch.dynamo "marked as skipped" error during cudagraph capture; deep dive deferred. Use at your own risk.
P91	AutoRound row-parallel group cdiv + start-idx fix (vllm#39460)	opt-in	`GENESIS_ENABLE_P91`	#39460	Backport of vllm#39460 non-MoE portion (CLOSED but valid). Fixes silent dequant corruption when AutoRound INT4/INT8 checkpoints have row-parallel layers whose input_size_per_partition is not divisible by group_size at TP>=2. Text-patch on gptq_marlin.py + parameter.py. Hypothesized cause of Lorbus INT4 < INT8 perf gap.
P93	AllSpark bypass for INT8 W8A16 group_size=-1 (force Marlin path)	archived	—	—	Genesis-original 2026-04-28. Archived 2026-05-06: P93 was removed from PATCH_REGISTRY after v764c boot crash blocked measurement. Conceptual approach (prepend "AllSparkLinearKernel" to VLLM_DISABLED_KERNELS to force Marlin fallback for INT8 W8A16 group_size=-1) is documented here for historical reference. If revived, must be rewritten as a text-patch and re-registered.
P72	profile_run M cap (unblocks `--max-num-batched-tokens > 4096` on MoE)	opt-in	`GENESIS_ENABLE_P72_PROFILE_RUN_CAP`	—	Genesis-original (Dynamo fake-tensor mismatch workaround for moe_align_block_size symbolic shape)
P37	MoE intermediate cache pool (opt-in)	opt-in	—	—	Genesis (see source / CHANGELOG)
P17/P18	Marlin MoE per-SM tuning	opt-in	—	—	Genesis (see source / CHANGELOG)
P1/P2	FP8 kernel dispatcher	opt-in	—	—	Genesis (see source / CHANGELOG)

Spec-decode

ID	Title	Status	Env Flag	Upstream	Credit
P62	structured-output spec-decode timing fix	opt-in	`GENESIS_ENABLE_P62_STRUCT_OUT_SPEC_TIMING`	#36138	sfbemerk (vllm#36138), cicirori (vllm#34650)
P60b	GDN+ngram Triton kernel offset	opt-in	`GENESIS_ENABLE_P60B_TRITON_KERNEL`	#40738	tdoublep (vllm#40738)
P60	GDN+ngram state recovery	opt-in	`GENESIS_ENABLE_P60_GDN_NGRAM_FIX`	#40738	tdoublep (vllm#40738), bhaktatejas922 (#39273)
P63	MTP/Eagle drafter GDN state recovery	deprecated	`GENESIS_ENABLE_P63_MTP_GDN_STATE_RECOVERY`	—	Genesis-original (hypothesis disproven 2026-04-25)
P64	qwen3coder MTP streaming early-return fix	opt-in	`GENESIS_ENABLE_P64_QWEN3CODER_MTP_STREAMING`	#39598	kotori-yan (vllm#39598)
P65	TurboQuant spec-decode cudagraph downgrade	opt-in	`GENESIS_ENABLE_P65_TURBOQUANT_SPEC_CG_DOWNGRADE`	—	Genesis-original (root cause for noonghunna #40880)
P66	cudagraph_capture_sizes spec-decode divisibility filter	opt-in	`GENESIS_ENABLE_P66_CUDAGRAPH_SIZE_FILTER`	#23679	Genesis-original (mirrors fhl2000 vllm#23679 closed)
P70	Auto-strict-ngram (force prompt_lookup_min>=8)	opt-in	`GENESIS_ENABLE_P70_AUTO_STRICT_NGRAM`	—	Genesis-original (vllm#40875 enforcement)
P67	TurboQuant multi-query kernel for spec-decode K+1	opt-in	`GENESIS_ENABLE_P67_TQ_MULTI_QUERY_KERNEL`	—	Genesis-original (proper fix for noonghunna #40880; replaces P65 workaround)
P71	Block-verify rejection sampler (vllm#40819 + gemini bug	opt-in	`GENESIS_ENABLE_P71_BLOCK_VERIFY`	#40819	Backport of vllm#40819 (Z. Golpayegani draft) + Sun et al. arXiv 2403.10444 + 2 critical fixes from gemini-code-assist r
P77	Adaptive ngram K controller (EMA + hysteresis + auto-di	opt-in	`GENESIS_ENABLE_P77_ADAPTIVE_NGRAM_K`	—	Genesis-original (port of SGLang adaptive_spec_params.py EMA+hysteresis Apache-2.0 + Nightjar arXiv 2512.22420 auto-disa
P79b	Async × spec-decode proposer-sync backport (vllm#40610)	opt-in	`GENESIS_ENABLE_P79B_ASYNC_PROPOSER_SYNC`	#40610	Backport of vllm#40610 (OPEN draft, tracked from #40608). Re-records prepare_inputs_event AFTER spec-decode proposer GPU
P82	SGLang threshold_single OR-clause acceptance (BIASED —	opt-in	`GENESIS_ENABLE_P82`	—	SGLang team (sgl-project/sglang) speculative_sampling.cuh — port of the threshold_single OR-clause that breaks the struc
P83	MTP keep-last-cached-block (vllm#38182 downstream symptom)	opt-in (research)	`GENESIS_ENABLE_P83`	—	Genesis (root-cause analysis on vllm#38182 by uOnePiece + @Angazenn). Empirically DISPROVEN as actual cause for our workload; kept as research artifact.
P84	hash_block_size override (vllm#38182 actual root cause for hybrid)	opt-in (research)	`GENESIS_ENABLE_P84` + `GENESIS_P84_HASH_BLOCK_SIZE=16`	—	Genesis-original 2026-04-27. scheduler.py:234 + engine/core.py:209 dual-site override; lets prefix-cache hash at finer granularity than the LCM-padded hybrid block_size.
P85	Hybrid fine-shadow prefix cache (MambaManager fix)	opt-in (research)	`GENESIS_ENABLE_P85`	—	Genesis-original 2026-04-27. MambaManager scale-factor shadow-hash entries + eviction-safety verify. Requires P84 to give fine hashes. NOTE: bisect 2026-04-27 confirmed v756 sustained-load crash NOT caused by P85 (reproduced without it).
P86	ngram batch_propose O(N*K) → O(N+K) direct-fill (vllm#40876)	opt-in	`GENESIS_ENABLE_P86`	#40876	Backport of vllm#40876 (aaronagent). Replaces O(N*K) `i in valid_ngram_requests` membership scan with O(N+K) direct-fill loop. Algorithmic improvement, no behavioral change.
P94	Spec-decode prepare_next_token_ids_padded zero-alloc (vllm#41043)	opt-in	`GENESIS_ENABLE_P94`	#41043	Backport of vllm#41043 (wangluochao902, OPEN). Removes GPU→CPU `.tolist()` sync + list-comprehension + `np.array(...)` allocation in spec-decode hot path. PR author measured P99 TPOT -9.3% on Llama-3.1-8B + Eagle3 TP=4. Applies to all spec methods (Eagle, MTP, ngram, draft model). For our MTP single-stream (estimate written at the K=3 era; PROD now K=5): expected +2-4% wall TPS + tighter CV.
P57	TQ spec-decode capture-safe buffers	deprecated	`GENESIS_ENABLE_P57_SPEC_DECODE_CAPTURE_SAFE`	—	noonghunna (#40831), gdn_attn.py reference
P56	TQ spec-decode safe-path guard	deprecated	`GENESIS_ENABLE_P56_SPEC_DECODE_GUARD`	—	noonghunna (#40807, #40831)
PN9	Independent drafter attention backend (vllm#39930)	self-retired	—	#39930	MatthewBonanni (vllm#39930, MERGED). Self-retires when upstream merged — use `--speculative-config.attention_backend` instead. Removed from start scripts 2026-05-02.
PN30	DS conv state layout + spec-decode AL>1 fix (issue #17)	opt-in	`GENESIS_ENABLE_PN30_DS_LAYOUT_SPEC_DECODE`	—	Genesis-original 2026-05-02 (closes #17 noonghunna LCB v6 50/50 crash). Two-file text-patch: replaces `NotImplementedError` raise in `mamba_utils.py:get_conv_copy_spec` with `.contiguous()` copy + module-level temp-tensor list; adds stream sync + cleanup in `do_mamba_copy_block`. Cost ~10-50us per batch when path active.

Structured-output / Qwen3 parser

ID	Title	Status	Env Flag	Upstream	Credit
P15	Qwen3 None/null tool arg parser	opt-in	—	—	Genesis (see source / CHANGELOG)
P12	Qwen3 <tool_call> implicit reasoning end	opt-in	—	—	Genesis (see source / CHANGELOG)
P29	tool parser IndexError guard	opt-in	—	—	Genesis (see source / CHANGELOG)
P61c	Qwen3Coder deferred-commit until `<function=` header	opt-in	`GENESIS_ENABLE_P61C_QWEN3CODER_DEFERRED_COMMIT`	—	Closes club-3090#72 (troymroberts 2026-05-06). Defers `is_tool_call_started=True` until `<function=` confirms within 64-char slack — fixes 30-120s SSE silence on prose containing literal `<tool_call>`.
P61b	Qwen3 streaming partial-tag overlap guard	opt-in	`GENESIS_ENABLE_P61B_STREAMING_OVERLAP`	#40783	ExtReMLapin (vllm#40783)
P61	Qwen3 multi-tool first-occurrence	opt-in	`GENESIS_ENABLE_P61_QWEN3_MULTI_TOOL`	#40783	ExtReMLapin (vllm#40783)
P68/P69	long-context tool-call adherence	opt-in	`GENESIS_ENABLE_P68_AUTO_FORCE_TOOL`	—	Genesis-original (long-ctx tool adherence mitigation)
PN70	Tool schema subset filter (companion to P68 v7.72.1)	opt-in	`GENESIS_ENABLE_PN70_TOOL_SCHEMA_FILTER`	—	Genesis-original (closes club-3090#57 option-3)
PN72	Frequency-based ngram draft post-filter (llama.cpp `draft_min_sample_size`-style; rejects spurious 2-token-suffix matches)	opt-in	`GENESIS_ENABLE_PN72_FREQUENCY_NGRAM_DRAFTER`	—	Genesis-original 2026-05-06 (composes with P70; tunables `GENESIS_PN72_MIN_OBSERVATIONS=4`, `GENESIS_PN72_FREQUENCY_WINDOW=1024`)
PN77	FP8 lm_head compression (BF16→FP8 e4m3 + per-channel scale; ~606 MiB/rank on 27B Qwen3.6, ~243 MiB/rank on 35B). Phase E.5 redesign: subclass `Genesis_FP8_LMHead_EmbeddingMethod` swap via single text-patch on `process_weights_after_loading` walker; uses `replace_parameter` (preserves `weight_loader`); hardware-tier dispatch — Marlin on Ampere, `torch._scaled_mm` on Ada/Hopper/Blackwell, cast-back fallback otherwise.	opt-in	`GENESIS_ENABLE_PN77_FP8_LM_HEAD`	—	Genesis-original 2026-05-07 (Phase E.5 architectural redesign after E.2-3 weight_loader-orphan blocker; cosine_sim ≥0.999 quality gate validated; auto-retire on vllm#41000 lm_head_quantized merge)
PN78	[DEPRECATED 2026-05-07] post-warmup empty_cache wrap — upstream `gpu_model_runner.py:6213/6244` already calls it; PN78 would be redundant 3rd call. Also monkey-patch wouldn't reach spawn'd workers. Retained for documentation only.	deprecated	`GENESIS_ENABLE_PN78_POST_WARMUP_CACHE_RELEASE` (ignored)	—	Genesis-original 2026-05-07 (deprecated same day after source-code investigation)
PN80	[ADDED 2026-05-07] LoRA tensorizer device kwarg (vllm#41845 backport). Single-line fix: passes `device=device` to TensorDeserializer in lora/lora_model.py. Without it, deserializer first goes to host RAM then transfers to GPU, causing OOM on memory-constrained rigs. With kwarg, streams directly to GPU. Genesis 35B/27B PROD does not use LoRA — patch ready for community deployments. Default OFF. Upstream merged in vllm#41845 (2026-05-07); native since the 2026-05 pin `0.21.1rc0+g626fa9bba5` (and every pin after, incl. current dev748) — Genesis backport retained for operators on older nightlies.	experimental	`GENESIS_ENABLE_PN80_LORA_TENSORIZER_DEVICE`	#41845	Backport of Or Ozeri @ IBM (vllm#41845, MERGED 2026-05-07)
PN79	[FULL IMPL 2026-05-07 + LIVE A/B VALIDATED] In-place SSM state for GDN chunk prefill (vllm#41824 backport). Eliminates per-decode-step gather/scatter copies of `initial_state`/`final_state` by passing `ssm_state_indices` directly to Triton kernel. Author claims 4.5–36 GiB cumulative fp32 traffic eliminated per multi-turn session. ORTHOGONAL TO PN59 (which fixes prefill peak; empirical 2026-05-06 — PN59 streaming path almost never fires under chunked-prefill, so PN79 is the actual decode-side win). 17 anchors atomic across 3 files via MultiFilePatchTransaction (Sub-1 chunk.py: 7, Sub-2 chunk_delta_h.py: 7, Sub-3 gdn_linear_attn.py: 3). 27B Lorbus INT4+TQ k8v4 A/B 2026-05-07: TPS 105.3 (PN79=1) vs 104.2 (baseline) — within noise; VRAM identical; tool 10/10 match. Single-shot win unproven, multi-turn evidence pending (Cliff 2 reproducer + memory traffic profiler on roadmap). 58 PN79 tests + 2260 full Genesis suite pass. conflicts_with [PN59, PN54]. Default OFF.	experimental	`GENESIS_ENABLE_PN79_INPLACE_SSM_STATE`	#41824	Backport of Kermit-C (vllm#41824, OPEN)
P59	Qwen3 reasoning embedded tool_call recovery	opt-in	`GENESIS_ENABLE_P59_QWEN3_TOOL_RECOVERY`	#39055	ZenoAFfectionate (vllm#39055)
P40	TurboQuant GQA-grouped decode stage1 (opt-in)	opt-in	—	—	Genesis (see source / CHANGELOG)
P24	fused_moe num_warps/num_stages overlay	opt-in	—	—	Genesis (see source / CHANGELOG)
P18b	TurboQuant decode stage1 tune	opt-in	—	—	Genesis (see source / CHANGELOG)
PN394	Qwen3 partial-param value `<` truncation fix	retired	`GENESIS_ENABLE_PN394_QWEN3_PARTIAL_PARAM_LT_FIX`	#46047	Backport of vllm#46047 (MERGED 2026-06-18; IN the deployed pin since dev301, so present in current dev748). Fixed the qwen3 tool-call parser truncating a partial parameter value at a literal `<`. Ran ON at dev148 where the merge was absent; retired 2026-07-02 — the upstream fix ships in the engine and the patch self-skips.

Hybrid / GDN / Mamba

ID	Title	Status	Env Flag	Upstream	Credit
P34	Mamba zero-collapse deadlock guard	opt-in	—	—	Genesis (see source / CHANGELOG)
P7b	GDN dual-stream via torch.library.custom_op (opt-in)	opt-in	—	—	Genesis (see source / CHANGELOG)
P7	GDN dual-stream in_proj parallelism	opt-in	—	—	Genesis (see source / CHANGELOG)
PN11	GDN a/b contiguity in fix_query_key_value_ordering (vllm#41142)	opt-in	`GENESIS_ENABLE_PN11_GDN_AB_CONTIGUOUS`	#41142	Yeuvoir (vllm#41142, OPEN). Fixes upstream issue #41112: in `GatedDeltaNet.fix_query_key_value_ordering` the `a` and `b` slices need explicit `.contiguous()` calls

Cudagraph

ID	Title	Status	Env Flag	Upstream	Credit
P78	TurboQuant .tolist() capture-guard (adapted from noongh	opt-in	`GENESIS_ENABLE_P78_TOLIST_CAPTURE_GUARD`	—	Adapted from noonghunna's patch_tolist_cudagraph.py (Apache-2.0, github.com/noonghunna/qwen36-27b-single-3090). Surgical
P67b	TurboQuant spec-verify forward() routing (FULL CG enabl	opt-in	—	—	Genesis (see source / CHANGELOG)
P95	Marlin TP cudagraph cap on Ampere (vllm#40385)	opt-in	`GENESIS_ENABLE_P95`	#40385	Backport of vllm#40385 (OPEN as of 2026-04-28). Defensive cap of `max_cudagraph_capture_sizes` to avoid OOM on TP>=2 with Marlin kernels
P100	FlashInfer FULL CUDA graph for spec-decode (vllm#41127)	opt-in	`GENESIS_ENABLE_P100`	#41127	Backport of vllm#41127 (open 2026-04-28) — enables FlashInfer FULL cudagraph mode for spec-decode. Battle-validated 2026-07-04 on dev748: its 6 anchors were re-anchored dual-variant during the dev714→dev748 preflight, then runtime-validated on the 27B fp8kv lane (FlashInfer FULL-CG spec-decode: coherent generation, 0 errors — fleet sweep, `BENCHMARKS.md`)
PN13	CUDAGraphWrapper gc.collect/empty_cache lambda arity (vllm#41235)	opt-in	`GENESIS_ENABLE_PN13_CUDA_GRAPH_LAMBDA_ARITY`	#41235	roikoren755 (vllm#41235, OPEN). Fixes worker-death TypeError in CUDAGraphWrapper

Scheduler / chunked-prefill

ID	Title	Status	Env Flag	Upstream	Credit
P74	Auto chunk-clamp via long_prefill_token_threshold (P72	opt-in	`GENESIS_ENABLE_P74_CHUNK_CLAMP`	—	Genesis-original (zero-VRAM-cost prealloc-overflow safety net for P72-unblocked batched_tokens>4096)
P58	async-scheduler -1 placeholder fix	opt-in	`GENESIS_ENABLE_P58_ASYNC_PLACEHOLDER_FIX`	#40768	z1ying (vllm#40768)

Other

ID	Title	Status	Env Flag	Upstream	Credit
P8	KV hybrid reporting (per-token capacity)	opt-in	—	—	Genesis (see source / CHANGELOG)
P27	Qwen3 BEFORE-THINK fallback	opt-in	—	—	Genesis (see source / CHANGELOG)
P5b	KV page-size pad-smaller-to-max (env-opt-in)	opt-in	—	—	Genesis (see source / CHANGELOG)
P79c	Stale spec_token_ids cleanup for unscheduled requests (	opt-in	`GENESIS_ENABLE_P79C_STALE_SPEC_TOKEN_CLEANUP`	#37629	Backport of vllm#37629 (OPEN, fixes #36906). Cleanup pass after main scheduling loop clears spec_token_ids for unschedul
P79d	Preempt async-discard backport (vllm#38624)	opt-in	`GENESIS_ENABLE_P79D_PREEMPT_ASYNC_DISCARD`	#38624	Backport of vllm#38624 (CodersAcademy006, OPEN). Adds discard_latest_async_tokens=True + num_output_placeholders=0 to _preempt_request() — fixes silent token duplication after preemption-resume on async + EAGLE/MTP/ngram_gpu paths. Idempotent.
P75	Auto-enable Suffix Decoding (vllm#25784 Arctic Inferenc	opt-in	`GENESIS_ENABLE_P75_SUFFIX_DECODING`	#25784	Backport-enabler of vllm#25784 (Arctic Inference Suffix Decoding) — operator convenience: auto-swap method=ngram→suffix
P51	TQ-active runtime layer-level guard	library	—	—	Pre-dispatcher library patch in kernels/dequant_buffer.py — runtime layer-level TQ-active detection (skips TQ preallocs on layers where TQ is not active). No env toggle, defensive runtime check. Companion to model_detect's config-level TQ check.
P102	Unified spec-decode metadata + disagreement tracker (TRT-LLM style)	opt-in	`GENESIS_ENABLE_P102`	—	Genesis-original (Sander 2026-04-29). First-class spec_meta module in spec_meta.py — diagnostic-only opt-in observability layer for spec-decode predicate disagreement (e.g. should_dispatch_p67 divergence between proposer and verify paths).
PN60	Quant arg vs config.json validator (preflight DX)	default-on	`GENESIS_ENABLE_PN60`	—	Genesis-original 2026-05-05 (apnar club-3090#51 finding). Doctor extension cross-checks operator's `--quantization` CLI arg against the model's `config.json:quantization_config.quant_method` BEFORE vLLM loads. Emits one-line remediation hint instead of a 30-line pydantic ValidationError.
PN61	qwen3_vl loader KeyError → text-only auto-fallback	opt-in	`GENESIS_ENABLE_PN61`	—	Genesis-original 2026-05-05 (apnar club-3090#51 NVFP4 finding). Catches `KeyError: 'blocks.0.attn.proj.weight'` in `qwen3_vl.load_weights` when an NVFP4 quant strips the ViT tower; emits WARN + auto-sets `language_model_only=True` instead of crashing. Same defensive pattern as P29 IndexError guard.
PN62	Text-only ViT scratch skip (memory savings, 3-5 GiB on 27B-NVFP4)	opt-in	`GENESIS_ENABLE_PN62`	—	Genesis-original 2026-05-05 (apnar club-3090#51 highest-impact gap); Wave 6 real hook 2026-05-09. When `mm_limits_all_zero AND --language-model-only`, PN62 wraps `GPUModelRunner.profile_run` and flips `MultiModalConfig.skip_mm_profiling=True` BEFORE the original runs, so vllm dev93's native short-circuit at the encoder-profiling branch (`gpu_model_runner.py:5879`) fires. Saves ~3-5 GiB ViT scratch on qwen3_vl + NVFP4 single-card boot. Sister to PN35 (text-only inputs_embeds skip).
PN63	fp8_e5m2 advisory for consumer Blackwell (gpu_profile recommendation)	default-on	`GENESIS_ENABLE_PN63`	—	Genesis-original 2026-05-05 (apnar club-3090#51 empirical). Adds advisory entry to `gpu_profile.PATCH_RECOMMENDATIONS` recommending `--kv-cache-dtype fp8_e5m2` over `fp8_e4m3` on consumer Blackwell (sm 12.0) until vLLM e4m3 codepath matures. Suggest-only; operator passes via CLI.
PN64	Marlin MoE per-SM tuning placeholder for SM 12.0	opt-in	`GENESIS_ENABLE_PN64`	—	Genesis-original 2026-05-05 (apnar club-3090#51 — boot log shows `[Genesis] skipped: P17/P18 Marlin MoE per-SM tuning — no tuning entry for SM (12, 0)`). PN64 adds a placeholder copying SM (9, 0) Hopper config until empirical sweep data lands from sm_120. Author-blocked: needs real 5090 sweep — solicit from apnar/jhsmith409.
PN65	Genesis structured API access log middleware (operator UX)	opt-in	`GENESIS_ENABLE_PN65`	—	Genesis-original 2026-05-05 (Sander request 'по апи лог невзрачный надо тоже проработать'). Replaces uvicorn's bare `INFO: 127.0.0.1:45116 - "GET /v1/models" 401` with `[Genesis-API] 200 POST /v1/chat/completions 34ms prompt=46t completion=400t tools=1 client=127.0.0.1`. Suppresses /health polling by default (GENESIS_PN65_LOG_HEALTH=1 to include). Status-aware level (2xx INFO / 4xx WARN / 5xx ERROR + exception type).
PN66	Multiturn `</think>` leak fix in DelegatingParser (vllm#41696 backport)	opt-in	`GENESIS_ENABLE_PN66`	#41696	Backport of vllm#41696 (panpan0000, OPEN as of 2026-05-05). Removes the buggy `prompt_reasoning_checked` short-circuit in `vllm.parser.abstract_parser.DelegatingParser.parse_delta` that walked the FULL prompt for `</think>` and prematurely set `reasoning_ended=True` from a previous turn's `</think>`. Defensive backport for multi-turn DSML/Hermes/Qwen3 chat clients sending full history.
PN67	thinking_token_budget inverted-bool fix (vllm#41674 backport, 1-line)	opt-in	`GENESIS_ENABLE_PN67`	#41674	Backport of vllm#41674 (JasonKeyiL, OPEN as of 2026-05-04). Single-token fix in `vllm/v1/worker/gpu_input_batch.py:894` — removes `not` from `or not thinking_budget_tracks_reqs`. Bug: thinking_token_budget silently ignored for any request without penalty parameters. NULL on Genesis PROD; defensive for users who experiment. Trivial backport, zero risk.
PN122	Sprint 2.6 v2 — CUDA graph dispatch trace wire-in (formerly `SPRINT26_CG_DISPATCH_TRACE`; renamed 2026-05-14 to conform to canonical `P[N]?\d+` patch ID convention)	opt-in	`GENESIS_ENABLE_PN122_CG_DISPATCH_TRACE` (legacy `GENESIS_ENABLE_SPRINT26_CG_DISPATCH_TRACE` accepted for one release)	—	Genesis-original (Sandermage). Wires the Sprint 2.6 v1 cudagraph dispatch instrumentation into the live decision path so dispatch divergences surface in boot logs instead of post-mortem diffs. Diagnostic-only; off by default. Family: observability.
PN96	Persistent Marlin MoE workspace (Wave 9 dev209 perf-restore)	default-on	`GENESIS_ENABLE_PN96`	—	Genesis-original 2026-05-12 (Wave 9 dev209 35B regression RCA). Upstream `experts/marlin_moe.py::MarlinExperts.apply` calls `fused_marlin_moe(...)` without passing `workspace=`, so `_fused_marlin_moe` allocates fresh via `marlin_make_workspace_new(device, 4)` per call. PN96 wraps MarlinExperts.apply to cache a persistent workspace per-instance + monkey-patches fused_marlin_moe to honor a thread-local default when workspace=None. Target: recover part of -2.82% TPS / +2.86% TPOT 35B A3B-FP8 regression. NO-OP on non-Marlin paths (27B hybrid GDN+Mamba INT4). Auto-skips on dev93-era layout. Family: moe.

Category descriptions

TurboQuant integration

Genesis uses Beidi Chen's TurboQuant k8v4 KV cache (FP8 K + 4-bit V) on Ampere. These patches add Ampere-specific casts, page-size unification, hybrid model support, and decode/prefill workspace management for the TurboQuant attention path.

Memory / buffers

Pre-allocated singleton buffers shared across all 36 attention layers via GenesisPreallocBuffer.get_or_create(). Eliminates per-call torch.empty allocations on the hot path. Toggleable via GENESIS_BUFFER_MODE=shared|per_layer.

Kernel performance

Triton / CUDA kernel optimizations: Marlin FP8 reduce, MoE warps/stages tuning, fp8 block-scaled MM low-M decode (P81), and the flagship P67 multi-query kernel for spec-decode K+1 verify (Genesis-original, +32% TPS).

Spec-decode

Speculative decoding fixes and acceptance heuristics: GDN+ngram state recovery (P60/P60b backport), block-verify rejection sampler (P71, Sun 2024), MTP cudagraph fixes, async-scheduler placeholder, and P82 SGLang threshold_single OR-clause (production: +12% TPS at threshold=0.3).

Structured-output / Qwen3 parser

Qwen3 thinking + tool-call output handling: reasoning-end timing (P62), multi-tool first-occurrence (P61), streaming overlap (P61b), embedded tool_call recovery (P59), MTP streaming early-return (P64), long-context tool-format reminder (P68/P69).

Hybrid / GDN / Mamba

Patches for hybrid attention models (GDN, Mamba): dual-stream parallelism (P7/P7b/P28), gating buffers (P46), zero-collapse deadlock guard (P34), FLA chunk_scaled_dot_kkt persistent A pool (P39a).

Cudagraph

Cudagraph capture safety: spec-decode divisibility filter (P66), TurboQuant CG downgrade (P65), .tolist() capture-guard (P78).

Scheduler / chunked-prefill

Scheduler-level fixes for long-context decode + MoE: profile_run M cap (P72, unblocks --max-num-batched-tokens > 4096), auto chunk-clamp via long_prefill_token_threshold (P74).

Other

Misc fixes: KV cache page size unification (P5/P5b), block_table tail zero-fill (P14), Marlin FP32_REDUCE env override (P23).

Community-issue fixes (v7.65, 2026-05-02)

Patches addressing reported community bugs. Mix of default-OFF (cross-rig validation pending) and default-ON (root-cause correctness fixes).

ID	Title	Issue	Default	Env flag
P15B	FA varlen `max_seqlen_k` clamp on TQ path	#15	OFF	`GENESIS_ENABLE_P15B_FA_VARLEN_CLAMP`
P38B	P38 compile-safe in-source hook (aot_compile-safe)	#14	OFF	`GENESIS_ENABLE_P38B_COMPILE_SAFE`
PN30	DS conv state layout + spec-decode AL>1	#17	OFF	`GENESIS_ENABLE_PN30_DS_LAYOUT_SPEC_DECODE`
PN31	FA varlen persistent `out` buffer (sister to P38)	#15	OFF	`GENESIS_ENABLE_PN31_FA_VARLEN_PERSISTENT_OUT`
PN32	GDN chunked-prefill (Cliff 2 single-24GB-GPU OOM)	noonghunna	OFF	`GENESIS_ENABLE_PN32_GDN_CHUNKED_PREFILL`
PN33	Spec-decode warmup K-aware (vllm#37521 extended to MTP/ngram)	ampersandru/noonghunna	ON	`GENESIS_ENABLE_PN33_SPEC_DECODE_WARMUP_K` (disable: `GENESIS_DISABLE_PN33_SPEC_DECODE_WARMUP_K=1`)
PN34	WorkspaceManager runtime lock relaxation (PN33 companion for runtime decode)	noonghunna	OFF	`GENESIS_ENABLE_PN34_WORKSPACE_LOCK_RELAX`
PN35	Skip inputs_embeds buffer for text-only models (vllm#35975 backport, ~64 MiB GPU + 64 MiB CPU per buffer × 2 sites)	AjAnubolu/noonghunna/club-3090#32	ON	`GENESIS_ENABLE_PN35_INPUTS_EMBEDS_OPTIONAL` (disable: `GENESIS_DISABLE_PN35_INPUTS_EMBEDS_OPTIONAL=1`)
PN38	DFlash quantized drafter sub-kernels (vllm#40425 backport, 4 sub-patches)	infatoshi/Sander	OFF	`GENESIS_ENABLE_PN38_DFLASH_QUANT_DRAFTER`
PN40	Spec-decode omnibus (sub-A DFlash K-norm + sub-B persistent buffer pool + sub-C adaptive K controller + sub-D workload classifier + sentinel)	Sander	OFF	`GENESIS_ENABLE_PN40_DFLASH_OMNIBUS` (per-sub: `GENESIS_PN40_ENABLE_SUB_{A,B,C,D}=0` to disable)
PN40-classifier	PN40 sub-D workload classifier middleware (chat_completion serving hook — companion of PN40 sub-D)	Sander	OFF	`GENESIS_ENABLE_PN40_DFLASH_OMNIBUS` (toggled jointly with PN40 master)
PN50	GDN proj fusion (SGLang#21019 — Qwen3.5/3.6 contiguous-projection Triton kernel; +7.4% TPS claimed)	Yuan Luo (SGLang)/Sander	OFF	`GENESIS_ENABLE_PN50_GDN_FUSED_PROJ`
PN51	Qwen3 streaming `enable_thinking=false` content routing (vllm#40816 backport — fixes OpenAI-compatible streaming clients)	keehawkes/Sander	OFF	`GENESIS_ENABLE_PN51_QWEN3_STREAMING_THINKING_DISABLED`
PN52	prompt_logprobs eviction fix during chunked prefill (vllm#41411 backport — multi-file)	Joachim Studnia (Mistral)/Sander	OFF	`GENESIS_ENABLE_PN52_PROMPT_LOGPROBS_EVICTION`
PN54	GDN contiguous-call deduplication (P0.7 Cliff 2b OOM mitigation)	Sander/MLX-LM#1077 inspiration	OFF	`GENESIS_ENABLE_PN54_GDN_CONTIGUOUS_DEDUP`
PN55	wake_up crash fix on hybrid (Mamba/DeltaNet) — vllm#41602 backport	kevglynn/Sander	OFF	`GENESIS_ENABLE_PN55_WAKE_UP_HYBRID_KV`
PN56	Qwen3Coder XML parse fallback — vllm#41466 backport	ToastyTheBot/Sander	OFF	`GENESIS_ENABLE_PN56_QWEN3CODER_XML_FALLBACK`
PN57	TurboQuant centroids disk-persistent cache — vllm#41418-inspired	TheTom inspiration / Sander	OFF	`GENESIS_ENABLE_PN57_TQ_CENTROIDS_DISK_CACHE`
P107	MTP truncation detector at reasoning→tool_call boundary — vllm#41467	ToastyTheBot/Sander	OFF	`GENESIS_ENABLE_P107_MTP_TRUNCATION_DETECTOR`
PN58	Spec-decode reasoning boundary validation — narrower alt to P62 (vllm#40962, mutex with P62)	ToastyTheBot/Sander	OFF	`GENESIS_ENABLE_PN58_SPEC_REASONING_BOUNDARY` (+ `VLLM_SPEC_REASONING_BOUNDARY_VALIDATION=1`)
PN59	Streaming-GDN orchestrator (Variant D Phase 2) — true Cliff 2b OOM fix (window-iterative, eliminates 805 MiB peak)	Sander/noonghunna validation	OFF	`GENESIS_ENABLE_PN59_STREAMING_GDN` (+ `GENESIS_VARIANT_D_WINDOW_NT=4`)

v7.68 cross-rig fixes from noonghunna (2026-05-02)

After v7.66 cross-rig validation by noonghunna and an independent CLI cross-check on 1× 3090 + 2× 3090, three real bugs in v7.66 were diagnosed and fixed in v7.68:

Bug	Root cause	Fix in v7.68
PN30 layout corruption	`.contiguous()` produced compact buffer (10240×5) but destination strided full state_len (10240×6) → memcpy corrupted DS row strides on every offset>0 copy	New part3 patches `collect_mamba_copy_meta` to build dst-shaped temp via `state[dest].clone()` + tail copy. Old part1 path becomes fail-closed RuntimeError.
PN25 v7.66 TP=1 spawn fail	`Library("genesis", "FRAGMENT")` constructed inside Dynamo trace context on TP=1 spawn config → `instantiate_user_defined_class_object` crash	Text-patch `activation.py` to register at module-import time (BEFORE any trace context). `forward_native` body reads only the cached module global. P7b extended with same pattern preventively.
PN33 partial close	Boot-time warmup K-aware fix closes profile_run path but runtime decode `_decode_attention` workspace_lock still fires on rare paths	New PN34 — companion patch relaxing the strict runtime AssertionError to WARN+grow. Default OFF; engage when PN33 alone doesn't close.

Diagnosis credit: noonghunna + independent CLI cross-check (club-3090 commit 9af1a52, a62ad78, 2b5ab4d, all on 2026-05-02).

PN33 is default-ON because it's a root-cause correctness fix, not experimental. The vanilla _dummy_sampler_run warms the rejection sampler with K=1 dummy draft tokens regardless of real num_speculative_tokens. This causes (a) under-counted KV-cache profile → mid-stream OOM via propose_draft_token_ids (ampersandru on club-3090#16), and (b) TurboQuant WorkspaceManager lock fails when real K-token spec-decode tries to grow workspace beyond warmup-reserved size (noonghunna on club-3090 disc #19). Both share one root cause; PN33 fixes the root, both symptoms disappear. Genesis extends the upstream PR (vllm-project/vllm#37521 by itailang) beyond its use_eagle() gate to cover all spec-decode methods (EAGLE + MTP + ngram + draft-model) since the rejection sampler memory footprint scales with K regardless of method. Disable only if K-sized warmup itself OOMs on a tight rig.

Reasoning + middleware

PN16 — Lazy-reasoner request hook (per-request enable_thinking). Middleware variant; configured via GENESIS_ENABLE_PN16_LAZY_REASONER.

DFlash spec-decode (PN21–PN24)

DFlash combine_hidden_states + SWA + aux-layer indexing fixes for spec-decode + DFlash hybrid:

PN21 — DFlash SWA support partial backport (vllm#40898). Env: GENESIS_ENABLE_PN21_DFLASH_SWA.
PN22 — Local argmax for TP draft (vllm#39419). Env: GENESIS_ENABLE_PN22_LOCAL_ARGMAX_TP.
PN23 — DFlash combine_hidden_states dtype cast (vllm#40334). Env: GENESIS_ENABLE_PN23_DFLASH_DTYPE_FIX.
PN24 — DFlash aux layer +1 indexing fix (vllm#40727). Env: GENESIS_ENABLE_PN24_DFLASH_AUX_LAYER_FIX.

TurboQuant unified pack (PN26 / PN26b)

PN26 — TQ unified perf pack (centroids prebake + sparse-V scaffold). Genesis-original.
PN26b — Sparse-V tile-skip kernel (BLASST λ=a/L for SM86; first NVIDIA Ampere implementation). Genesis-original Triton kernel.

MoE / merge_attn_states / chunk_o numerics

PN27 — Revert MoERunnerInterface PluggableLayer (vllm#41440 backport).
PN28 — merge_attn_states NaN guard (vllm#39148 backport).
PN29 — GDN chunk_o scale-fold (vllm#41446 pattern (c) backport). Default OFF; empirical −0.4% on 27B (Welch p=0.008) — kept opt-in for future Blackwell.

Compatibility matrix

Every opt-in patch listed here is default_on=False. The PROD launch scripts already set the right combination for each model — this section is what to consult when customising. The default-on subset ships in the V1 / V2 presets and is summarised in MODELS.md + CONFIGS_AUTO.md.

Conventions:

🟢 SAFE — empirically verified neutral or positive on the listed model. No tool-call regression, no stability issue.
🟡 OPT-IN-ONLY — patch is applied but its effect depends on workload (prefix-caching, GQA shape, online quant, ...).
🔴 DO-NOT-ENABLE — empirically broken on the listed model (regression confirmed in this repo's commit history).
⚠ DEPRECATED — superseded by another patch; kept for git history but should NOT be enabled.

How patches interact

The dispatcher's validate_registry() call (runs at every boot) checks for requires_patches and conflicts_with declarations and refuses to boot on a violation. As of HEAD there are 0 declared conflicts between patches — they compose additively. The "conflicts" below are EMPIRICAL (observed at runtime), not schema-declared. Within a category, patches are typically alternatives (enable one, not all). Across categories, patches are typically additive.

spec_decode — currently in PROD on 35B-A3B-FP8

Patch	Status	Effect	Notes
P58 Async-scheduler -1 placeholder fix	🟢 SAFE	spec-decode + cudagraph workloads no longer loop or IMA	Backport vllm#40768
P60 GDN+ngram state recovery (Phase 1)	🟢 SAFE	tool-call recovery on hybrid models	Backport vllm#40738
P60b GDN+ngram Triton kernel offset (Phase 2)	🟢 SAFE	composes with P60	Backport vllm#40738
P66 cudagraph_capture_sizes divisibility filter	🟢 SAFE	boot 2–4× faster, less peak GPU memory	Genesis-original
P67 TurboQuant multi-query kernel (spec K+1)	🟢 SAFE on 35B (GQA=8 = pow-2); 🔴 27B (GQA=6 non-pow-2) — guard auto-skips	self-protected by power-of-2 guard	Genesis-original
P67b TQ spec-verify forward() routing	🟢 SAFE on 35B; auto-skips on 27B	routes K+1 batches through P67 before `_prefill_attention`	Genesis-original
P82 SGLang threshold_single OR-clause acceptance	🟢 SAFE on 35B; `GENESIS_P82_THRESHOLD_SINGLE=0.3` recommended	+5–10% acceptance on conservative draft scenarios	Backport SGLang

spec_decode — opt-in only on certain workloads

Patch	When to enable	Why off by default
P59 Qwen3 reasoning embedded tool_call recovery	OpenAI-compatible streaming clients	Overlaps with P62; dual-apply on same code is redundant.
P65 TQ spec-decode CG downgrade	Never on Ampere (superseded by P67b)	Forces PIECEWISE which costs ~5% TPS.
P70 Auto-strict-ngram (`prompt_lookup_min>=8`)	ngram workloads only	We use MTP, not ngram → no-op.
P71 Block-verify rejection sampler	Hopper / Blackwell experiments	Sampling change; needs retraining.
P75 Auto-enable Suffix Decoding	Arctic Inference workloads	Different acceptance heuristic.
P77 Adaptive ngram K controller	ngram only	We're on MTP.
P83 + P85 MTP keep-last-cached-block + hybrid fine-shadow prefix cache	If `--enable-prefix-caching` ON	PROD doesn't use prefix-caching (P83+P84+P85 stack regressed −29% in our 4-arm A/B).
PN8 MTP/draft online-quant propagation	RETIRED 2026-07-05 — do not enable	Anchor gone since dev672 (benign-skips); vllm#40849 still OPEN, feature NOT native. The historical "−1066 MiB on 35B FP8" applied on ≤dev424-era pins only.

spec_decode — deprecated

Patch	Why
P56 TQ spec-decode safe-path guard	Superseded by P65 (which itself is superseded by P67b).
P57 TQ spec-decode capture-safe buffers	Research artefact; never reproduced positive effect.
P61 Qwen3 multi-tool first-occurrence	Superseded by P12 v2.

structured_output

Patch	Status	When	Effect
P61b Streaming partial-tag overlap guard	🟢 SAFE on PROD	streaming tool-call	stops `<tool_call` partial fragment closing prematurely
P62 Spec-decode reasoning-end timing fix	🟢 SAFE on PROD	spec-decode + grammar	reasoning-aware grammar acceptance + spec-token validation
P64 Qwen3coder MTP streaming early-return fix	🟢 SAFE on PROD	streaming + MTP	removes early `return` that drops parameters when MTP bundles last param + `</function>`
P68 Auto force `tool_choice=required` for long-context	🟢 SAFE on PROD	long-context (≥50K chars)	threshold via `GENESIS_P68_P69_LONG_CTX_THRESHOLD_CHARS`
P69 Long-context tool-format reminder injection	🟢 SAFE on PROD	long-context	companion to P68

perf_hotfix

Patch	Status	Effect	Empirical
P98 TQ WorkspaceManager revert	🟡 OPT-IN-ONLY	expected +15–25% TPS recovery on Ampere small-batch single-stream	tested 35B PROD 2026-04-30: +1.0% — within ±1% noise
P99 WorkspaceManager.get_simultaneous memoization	🟢 SAFE on 35B PROD	helper for P98	already in PROD; baked into baseline
P100 FlashInfer FULL CUDA graph for spec-decode	🟢 SAFE on 35B PROD	UNIFORM_BATCH cudagraph for K+1 spec-verify	NO-OP on TQ backend; only fires on FlashInfer
P101 TQ continuation 64-token slicing	🟢 SAFE on 35B PROD	+3–12% TPS on long-context	already in PROD baseline

kv_cache

Patch	Status	When	Notes
P84 hash_block_size override (vllm#38182)	🔴 DO NOT ENABLE on 27B PROD	prefix-caching ON + `GENESIS_P84_HASH_BLOCK_SIZE` set	−29% TPS regression in 4-arm A/B
P85 Hybrid fine-shadow prefix cache	🔴 DO NOT ENABLE on 27B PROD	companion to P84	same regression

kernel / quantization

Patch	Status	When	Notes
P81 fp8 block-scaled MM low-M decode tuning	🟢 SAFE on 35B PROD	FP8 block-scaled checkpoints	Backport vllm#40925
P87 Marlin W4A16/W8A16 sub-tile output dim pad-on-load	🟢 SAFE on 27B PROD	INT4 AutoRound checkpoints	Backport vllm#40361
P91 AutoRound row-parallel group cdiv + start-idx fix	🟢 SAFE on 27B PROD	row-parallel AutoRound INT4	Backport vllm#39460
PN14 TQ decode IOOB safe_page_idx clamp	🟡 OPT-IN-ONLY	defensive on TQ k8v4	tested 35B PROD: +0.1% (noise)

compile_safety

Patch	Status	Effect	Empirical
P72 profile_run M cap	🟢 SAFE on PROD	unblocks `--max-num-batched-tokens > 4096` on MoE	already PROD
P74 Auto chunk-clamp via long_prefill_token_threshold	🟢 SAFE on PROD	prevents prealloc buffer overflow when batched=8192	already PROD
P78 TurboQuant `.tolist()` capture-guard	🟡 OPT-IN-ONLY (Site B only)	falls back to `flash_attn_varlen_func` during cudagraph capture	Site A NOT applied — substituting captured-time constant produces garbage output

cudagraph_safety / model_correctness / memory_savings

Patch	Status	When	Empirical
PN11 GDN a/b contiguity (vllm#41142)	🟢 SAFE on 35B PROD	hybrid GDN — Quentin Machu fix	defensive; community-validated
PN12 FFN intermediate scratch pool (Cliff 1 fix)	🟡 OPT-IN-ONLY	INT4 AutoRound 27B at long-ctx	tested 35B PROD: +0.16% (noise)
PN13 CUDAGraphWrapper lambda-arity fix (vllm#41235)	🟡 OPT-IN-ONLY	Blackwell GB200 nightly	tested 35B PROD: +0.7% (noise) on Ampere
PN17 FA2 softmax_lse runtime clamp	🟡 OPT-IN-ONLY	FA2 + spec-decode	recent; not yet PROD-tested

Empirical compatibility (HISTORICAL — 2× RTX A5000, 2026-04-30 baseline; PROD has since moved to MTP K=5 on 35B / K=4 on 27B and pin dev748 — treat rows as the last full matrix sweep, not current state)

Patch	35B-A3B-FP8	27B-Lorbus + fp8_e5m2	27B-Lorbus + TQ k8v4	Notes
P58	🟢 PROD	🟢 PROD	🟢 PROD	always safe
P60 + P60b	🟢 PROD	🟢 PROD	🟢 PROD	hybrid GDN — always
P66	🟢 PROD	🟢 PROD	🟢 PROD	always safe
P67 + P67b	🟢 fires (GQA=8)	🟡 auto-skip (GQA=6)	🟡 auto-skip (GQA=6)	pow-2 guard
PN521 + PN521_SPLIT_K + PN522 (added 2026-07)	n/a	n/a	🟢 required for MTP	raw-bf16-tail spec-verify closes the non-pow2-GQA (GQA=6) K+1 verify gap that made TQ×MTP collapse into token repetition on the 27B; split-K ~11× verify speedup + warmup removes the first-request JIT spike
P67/P67b under FULL cudagraph	🟢 PROD	n/a	🔴 `_prefill_attention` `.tolist()` crash + repetition spam	use PIECEWISE on 27B+TQ
P82 (threshold=0.3)	🟢 PROD	🟡 OFF	🟡 OFF	workload-dependent acceptance
P98	🟡 +1% noise	not yet tested	required for boot on hybrid+TQ
P99 + P101	🟢 PROD	🟡 OFF	n/a	already optimised in baseline
P83 + P84 + P85	n/a	n/a	n/a	−29% regression with prefix-cache ON
PN8	🔴 RETIRED 2026-07-05 (anchor gone ≥dev672; was −1066 MiB on ≤dev424 pins)	🟡 no-op (offline INT4)	🟡 no-op (offline INT4)	online-quant only; flags removed from composes/YAMLs
PN11	🟢 PROD	🟡 OFF	🟡 OFF	hybrid defensive
PN12	🟡 +0.16% noise	not yet tested	not yet tested	Cliff 1 fix
PN13	🟡 +0.7% noise	not yet tested	not yet tested	Blackwell-targeted
PN14	🟡 +0.1% noise	not yet tested	not yet tested	TQ defensive

Numbers in "noise" parentheses are from tools/genesis_bench_suite.py --mode standard --runs 25 --max-tokens 1024 on 2× RTX A5000. CV is 5–6%, so anything under ±2% is sampling noise.

Patch plan resolver — `--policy compat|safe|minimal`

Genesis ships 325 registered patches (see the auto-derived count in the header table); a typical preset enables ~50–80 of them. Two operator questions follow naturally:

Which patches actually need to be on for this deployment?
Which ones could I safely drop without losing the speedups I care about?

patch_plan is the answer. It's a read-only resolver that reads each preset's genesis_env and the structured patches_attribution metadata, applies a policy filter, and produces:

the included env-flag set (what the runtime should export),
the excluded set (what the policy dropped and why),
the passthrough set (GENESIS_* parameter keys that are NOT patch toggles and therefore always survive),
a warnings list (conflicts between included patches, candidate_when predicate mismatches against your rig).

The resolver does NOT change runtime apply decisions — the dispatcher's should_apply() remains the canonical runtime gate. patch_plan operates upstream of the runtime: it decides what gets exported in the first place.

Policy modes

Policy	What it drops	When to use
`compat` (default)	Nothing. Every truthy toggle survives. The excluded set only contains operator-disabled (`value=0`) toggles.	Legacy / pre-attribution operators. Byte-identical to old behaviour.
`safe`	Toggles whose attribution declares `role: no_op`. Conservative — only kills patches explicitly documented as inactive on this preset.	Recommended starting point once attribution coverage is meaningful.
`minimal`	Above + `role: suspected_regression` + `role: unknown` (no attribution at all). Aggressive — drops everything that isn't explicitly documented as needed.	Advanced operators who curate attribution and want a lean stack.

Parameter keys always pass through. GENESIS_PN95_CONFIG_KEY, GENESIS_BUFFER_MODE, GENESIS_PROFILE_RUN_CAP_M, and similar configuration values are not patch toggles — they're how the patches behave once they fire. Filtering them would silently break the patches that depend on them. The resolver detects toggles by the GENESIS_ENABLE_ / GENESIS_DISABLE_ prefix; every other GENESIS_* key passes through every policy unchanged.

CLI surface

sndr patches plan --preset prod-qwen3.6-35b-balanced
sndr patches plan --preset prod-qwen3.6-35b-balanced --policy compat --explain
sndr patches plan --preset prod-qwen3.6-35b-balanced --policy minimal --json

# A/B before flipping
sndr compose plan-diff prod-qwen3.6-35b-balanced --from compat --to minimal

# Render a compose file with the filtered env
sndr compose render prod-qwen3.6-35b-balanced --policy minimal -o /etc/sndr/compose.yml

# Launch with the filter applied
sndr launch prod-qwen3.6-35b-balanced --policy minimal --dry-run

# Diagnose against the same policy used at launch
sndr model-config diagnose prod-qwen3.6-35b-balanced --policy minimal

# Capture the plan in a report bundle
sndr report bundle --preset prod-qwen3.6-35b-balanced --scope quality

Authoring attribution

patches_attribution lives inside the model YAML, keyed by registry patch ID (e.g. PN204, not the env-flag name):

patches_attribution:
  PN204:
    role: optional_perf
    bench_evidence: "dev371 35B conc=8: 689 TPS / TTFT 237 ms"
    note: |
      Enabled in qwen3.6-35b-multiconc profile via patches_delta.enable.
      Hopper SM 9.0+ gives +5-10% TPOT per upstream PR.
    candidate_when:
      max_num_seqs_gte: 4

Role taxonomy

Role	Semantics	Required fields
`load_bearing`	Removal causes measurable regression / boot break.	`note`
`defensive`	Cheap safety guard; default-include.	—
`optional_perf`	Conditional perf knob; needs evidence to justify.	`bench_evidence`
`suspected_regression`	Bench showed regression; documented but kept off.	`note`
`no_op`	Included for cross-config consistency but inactive on this preset.	—
`unknown`	Not yet classified (defaults when no attribution exists).	—

`candidate_when` predicates

candidate_when is an optional dict of operator-authored predicates. When the predicate doesn't match the current cfg, the resolver emits an advisory warning (it does not exclude the patch).

Supported predicate keys:

candidate_when:
  max_num_seqs_gte: 4               # cfg.max_num_seqs >= 4
  max_num_seqs_lte: 8
  max_model_len_gte: 100000
  n_gpus_eq: 2
  tool_call_parser: ["qwen3_coder", "qwen3_xml"]    # list membership
  reasoning_parser: ["qwen3"]
  kv_cache_dtype: ["fp8_e5m2", "turboquant_3bit_nc"]
  quantization: ["auto_round"]

Unknown predicate keys produce a warning and don't fail closed — forward-compat for predicates the resolver hasn't been updated to recognise.

Profile-level overrides

A ProfileDef.patches_delta.attribution map can override or extend the model's attribution per patch ID:

# profile/long-ctx-qwen3.6-27b.yaml
patches_delta:
  enable:
    GENESIS_ENABLE_PN204_DUAL_STREAM_INPROJ: '1'
  attribution:
    PN204:
      role: load_bearing
      note: |
        Long-ctx profile keeps PN204 mandatory because OOM risk
        compounds with context length on the multi-conc path.

Per-key full replacement (not field merge): the profile entry is the entry that lands in the composed ModelConfig.patches_attribution.

Audit gates

make audit-patch-attribution     # AT-1 key matches registry; AT-2 role-presence consistency
make audit-patch-plan-resolves   # every V2 preset resolves cleanly under all 3 policies

Coverage ratchet (optional CI gate):

python3 scripts/audit_patch_attribution.py --min-coverage 30

Migration

The policy filter is opt-in everywhere. If no --policy flag is passed:

sndr patches plan runs the dispatcher simulator (legacy);
sndr compose render produces byte-identical output to pre-2026-05-16;
sndr launch exports the full genesis_env matrix;
sndr model-config diagnose compares against the raw matrix.

The resolver still runs in the background under sndr patches plan without --policy to surface advisory warnings (conflicts + candidate_when mismatches). Those warnings are always visible — they don't change behaviour, just visibility.

Reference

Resolver code: sndr/model_configs/patch_plan.py
Schema: PatchAttribution in sndr/model_configs/schema.py, PatchesDelta.attribution in schema_v2.py.
Audit gates: scripts/audit_patch_attribution.py, scripts/audit_patch_plan_resolves.py.

Adding a new patch

Pick a free ID. Run grep -E '^@register_patch' sndr/apply/_per_patch_dispatch.py | head and grep -E '"P[0-9]+' sndr/dispatcher/registry.py to confirm the next available number. Don't reuse retired IDs (P56/P57/P63 are deprecated but kept).
Write integration module: sndr/engines/vllm/patches/<family>/<patch_id>_<name>.py. Use p71_block_verify.py or p82_sglang_acceptance_threshold.py as templates. The family should match a compat/categories.py bucket — _build_module_index will rglob the new file in automatically.
Register in dispatcher (P56+): add an entry to PATCH_REGISTRY in sndr/dispatcher/registry.py.
Hook in apply dispatch: add @register_patch(...) + apply_patch_<id>_* function in sndr/apply/_per_patch_dispatch.py.
Document in CHANGELOG: add a vX.YZ entry to CHANGELOG.md explaining the WHY, empirical data, and ship/reject decision.
Validate:
- Static: python3 -c 'import ast; ast.parse(open("sndr/engines/vllm/patches/<family>/<patch_id>_*.py").read())'
- Container: docker compose down && docker compose up -d (NOT stop/start — see CONFIGURATION.md Container R/W layer note)
- Empirical: blue/green A/B with tools/genesis_bench_suite.py (--compare A.json B.json for the Welch t-test) + the quality-gate probes under tools/quality_gate/. SHIP gate: quality battery clean + a significant TPS win (or whatever the patch targets). (The old genesis_quality_harness.py / genesis_bench_v3.py tools no longer ship.)
Credit upstream in the patch docstring + CREDITS.md if backporting from someone else's PR / project.

FilesExpand file tree

PATCHES.md

Latest commit

History

PATCHES.md

File metadata and controls

Genesis vLLM Patches — Complete Reference

Current state (v12.1.0, 2026-06-01)

Engine tier (the strict-AND boundary)

Tombstones (deprecated, no-op)

Recent additions (June–July 2026 waves)

Recent additions (v11.0 sprint — historical, 2026-05)

draft_probs). Rejection rule becomes min(1, target/draft)`

Archived dead-ends

Native upstream features (no Genesis backport needed)

How to enable / disable a patch

Where the code lives

How a patch is decided to apply or skip

Patches by category

TurboQuant integration

Memory / buffers

Kernel performance

Spec-decode

Structured-output / Qwen3 parser

Hybrid / GDN / Mamba

Cudagraph

Scheduler / chunked-prefill

Other

Category descriptions

TurboQuant integration

Memory / buffers

Kernel performance

Spec-decode

Structured-output / Qwen3 parser

Hybrid / GDN / Mamba

Cudagraph

Scheduler / chunked-prefill

Other

Community-issue fixes (v7.65, 2026-05-02)

v7.68 cross-rig fixes from noonghunna (2026-05-02)

Reasoning + middleware

DFlash spec-decode (PN21–PN24)

TurboQuant unified pack (PN26 / PN26b)

MoE / merge_attn_states / chunk_o numerics

Compatibility matrix

How patches interact

spec_decode — currently in PROD on 35B-A3B-FP8

spec_decode — opt-in only on certain workloads

spec_decode — deprecated

structured_output

perf_hotfix

kv_cache

kernel / quantization

compile_safety

cudagraph_safety / model_correctness / memory_savings

Empirical compatibility (HISTORICAL — 2× RTX A5000, 2026-04-30 baseline; PROD has since moved to MTP K=5 on 35B / K=4 on 27B and pin dev748 — treat rows as the last full matrix sweep, not current state)

Patch plan resolver — --policy compat|safe|minimal

Policy modes

CLI surface

Authoring attribution

Role taxonomy

candidate_when predicates

Profile-level overrides

Audit gates

Migration

Reference

Adding a new patch

draft_probs`). Rejection rule becomes` min(1, target/draft)`

Patch plan resolver — `--policy compat|safe|minimal`

`candidate_when` predicates