moe-l2

English | 中文

MoE expert offload for low-VRAM GPUs — run 100B+ MoE models (DeepSeek, Qwen, Mixtral) on 10-11 GB cards. A transparent, OpenAI-compatible proxy that predicts which experts your prompt needs and preloads them into a shared-memory LRU cache, so you can run 16 GB+ MoE models on 10-11 GB GPUs with up to ~60% VRAM savings.

⭐ Found this useful? Give us a star — it helps others discover the project. ★ Star on GitHub

💬 Measured results & quick start — full benchmark table, memory numbers, and install guide: Discussion #2

Real-world benchmark

Speed = RTX 4090 measured (2026-08-19, bins-v0.6.0 per-slot-lock build, full-chain moe-l2 start --gpu): DS-V2-Lite 139-154 t/s (single-turn 141, long-context 139), Qwen3.6-A3B 25.5-44.2 t/s (mixed-domain). 2080 Ti full-chain (v0.6.0): Qwen 16.6-28.6 t/s, DS-V2-Lite 86-94 t/s. 5090 (SM120a, v0.6.0): Qwen 28-52.5 t/s, DS 141-151 t/s. Qwen3-235B-A22B: ~3.9 t/s steady (2026-08-11). See models-benchmark.md.

Without moe-l2, a 10-11 GB card cannot load these models at all — it OOMs immediately. With moe-l2, a 32B MoE fits in ~9.3 GB VRAM (on-demand pin experts on Qwen3.6-A3B, GPU compute). DeepSeek-V4-Flash (157B params / 85 GB file, 256 experts, top-6) runs on a 10-11 GB card at 8.3-9.1 GB VRAM — with selective pin (v4_top100.map), RSS 26.8 GB (from 84.4 GB whole-pin, −68%); on-demand fallback RSS 17.5 GB (VRAM 16.5-16.7 GB, measured 2026-08-10). ⚠️ No valid speed data — upstream llama.cpp bug (#25582): deepseek4 MoE expert layers on CUDA produce garbled output (reproduces on pure vanilla llama-server, no moe-l2 involved); UD-IQ2_M/Q4_K_XL both affected. Waiting for upstream fix. Full report: deepseek-v4-flash-verify-20260805.md · Qwen3-235B-A22B (235B params / 85.7 GB file, 128 experts, top-8) runs on a 24 GB card at ~3.9 t/s steady — selective pin (top-60/layer, 98.5% coverage): 24 GB VRAM + 55 GB RAM, RSS 80.8 → 54.7 GB (−33%), measured 2026-08-11. Full report: qwen3-235b-a22b-q2k-benchmark.md · All measured models: models-benchmark.md

Visual demo (RTX 4090, 2026-08-16)

Qwen3.6-35B-A3B (32B MoE) — standard vs moe-l2	DeepSeek-V2-Lite (16B MoE) — 8 GB card vs 24 GB card

Summary: ~57-60% less VRAM — a 10-11 GB card runs what used to need 24 GB (RTX 4090 measured 2026-08-19, bins-v0.6.0: DS 139-154 t/s @ ~10.1 GB / Qwen 25.5-44.2 t/s @ ~6.7-7.2 GB):

Live capture (2026-08-16, bins-v0.5.0 C-scheme): Qwen3.6-35B-A3B generating 2,344 tokens with VRAM at ~5.4 GB (~32 t/s) — watch the VRAM curve stay flat the whole run:

examples/demo-assets/demo-vram-animation.mp4 (45 s, 1280×720) · raw telemetry: examples/demo-assets/rec_data.csv · full generated text: examples/demo-assets/rec_full.txt

Benchmarked on RTX 4090 (2026-08-19, v0.6.0)

We benchmarked Qwen3.6-A3B (32B MoE) and DeepSeek-V2-Lite (16B MoE, 64 experts) on RTX 4090 with selective pin + per-domain table switch + A3 cache (bins-v0.6.0 per-slot-lock build, 2026-08-19): experts stay in CPU RAM (zero VRAM), the router map pre-selects per-domain hot experts and the scheduler copies only the activated experts to GPU each step, hot experts are cached in VRAM. DS-V2-Lite 139-154 t/s (~10.1 GB VRAM, ~6.4-6.7 GB RSS), Qwen3.6-A3B 25.5-44.2 t/s (mixed-domain; ~6.7-7.2 GB VRAM, ~8.9-11.3 GB RSS). Full reports: qwen3.6-a3b-iq2m-benchmark.md · deepseek-v2-lite-q2k-benchmark.md · models-benchmark.md

Selective pin — 低内存模式（V4 RSS 对比，2026-08-10 实测）

Measured on RTX 4090 (2026-08-10, bins-v0.4.0): whole-pin 84 GB → selective pin 26.8 GB (router-map top-K) → on-demand 17.5 GB. RSS −68% (V4 memory data; speed N/A — upstream deepseek4 CUDA bug #25582). Also: speed vs RSS scatter.

Selective pin（低内存模式，v0.4.0 起；当前主路径为 v0.6.0 按领域换表） — a router map (top-K experts per layer, e.g. v4_top100.map 43 layers) pre-pins the hot experts as host-pinned; experts outside the map fall back to on-demand pin. No env vars needed for whole-pin default; pass --router-map <file> or --router-top-k N to moe-l2 start --gpu:

moe-l2 start --model model.gguf --gpu --router-map v4_top100.map

Multi-architecture binaries (bins-v0.6.0, 2026-08-19)

One binary for all NVIDIA consumer GPUs — GTX 1080 (sm_61) through RTX 50-series (sm_120a). Built with CUDA 12.8; no per-GPU compilation needed. moe-l2 download-bins fetches it automatically. bins-v0.6.0 includes the selective pin (router-map driven) + GPU cache prefill + on-demand pin main path + expert-page eviction v3.1 (MOE_L2_LRU_MAX_EXPERTS=N) + layered pin (MOE_L2_PIN_LAYERS) + A3 cache 32768 slots + cuda-libs (no libnccl — not needed for single-GPU) + per-domain router table switch (POST /moe-set-domain) + per-slot lock optimization (concurrent correctness restored — v0.5.0 was lock-free and DS concurrent output was garbage) + soft_resize / retain-hot-experts (v2, default single-table; MOE_L2_POOL_SIZE=3 opt-in) + VRAM-adaptive main table top-k (≤12G=top-75 / >12G=top-100) + proxy concurrency fixes (throttled domain switch + background retrain — DS long 500 gone) + flywheel table persistence (A-fix, no more domain loss on restart) + DEFAULT_TOP_K 100.

* All rows = bins-v0.6.0 full-chain (2026-08-19, moe-l2 start --gpu, per-domain table switch + A3 cache, default single-table). 4090: Qwen 25.5-44.2 t/s (mixed-domain) / DS 139-154 t/s (single-turn 141, long-context 139), VRAM 6.7-7.2/10.1 GB, RSS 8.9-11.3 / 6.4-6.7 GB; 2080 Ti: Qwen 16.6-28.6 / DS 86-94 t/s, VRAM 5.3/10.0 GB; 5090: Qwen 28-52.5 (peak 52.5) / DS 141-151 t/s, VRAM 5.6/10.2 GB. 3080 Ti row remains v3.1 multi-arch (bins-v0.3.0). All outputs verified clean (no garbage) on all three cards, 2026-08-19.

Concurrent requests — shared cache, no speed loss (2026-08-12)

4 parallel slots share one A3 expert cache / selective-pin table — verified on 2080 Ti and 4090 with Qwen3.6-35B-A3B, DS-V2-Lite and DeepSeek-V4-Flash (256 experts, spread routing):

⚠️ V4 行速度 N/A — 上游 deepseek4 CUDA 专家路径 bug（#25582）。

Concurrent throughput = 2.3-2.5× a single session; cross-domain vs same-domain is only -5-7% — no per-domain cache pools needed. VRAM grows only by the per-slot KV cache (+2.9 GB for 4 slots), RAM stays flat (+0.2 GB). One AI PC can serve multiple users at once. Full report: concurrent-cache-sharing-20260812.md

Quick start

One-line install (Linux x86_64 + NVIDIA GPU):

curl -fsSL https://raw.githubusercontent.com/yalun753/moe-l2/main/scripts/install.sh | bash

The installer checks your GPU/driver/Python, installs moe-l2 from PyPI, downloads the pre-built CUDA binaries, optionally downloads a demo model (Qwen3.6-35B-A3B, ~11.5 GB, resumable), then runs a self-check.

Manual install:

pip install moe-l2                   # keyword-only predictor (zero extra deps)
pip install moe-l2[predictor]        # hybrid: keyword + semantic embedding
moe-l2 download-bins                 # pre-built CUDA llama-server (on-demand pin patched)
moe-l2 model download --model qwen3.6-35b   # optional demo model (~11.5 GB)
moe-l2 start --model model.gguf --gpu

Useful commands:

moe-l2 doctor                        # environment self-check (GPU/CUDA/Python/disk)
moe-l2 model list                    # list downloadable models
moe-l2 model download --model <name> # download model (resumable, via hf-mirror)

Your tools (curl, Open WebUI, LangChain) connect to localhost:11435 — no client changes needed.

How it works

MoE models have many "experts" but only activate a few per token. moe-l2 predicts your prompt's domain (codegen, math, chinese_tech, etc.) and preloads the relevant experts into an mmap'd LRU cache before they're needed.

user → moe-l2 proxy (localhost:11435)
    ├── predict domain (keyword → TF-IDF → semantic)
    ├── on-demand pin experts (lazy mmap + register, zero VRAM)
    ├── hot experts cached in VRAM (A3 LRU)
    └── forward to llama-server (localhost:11436, CUDA GPU)
        └── GPU reads pinned experts via PCIe DMA; cold pages evicted

Usage

1. L2 proxy with GPU on-demand pin (recommended)

Start the transparent proxy with the bundled on-demand pin llama-server:

moe-l2 start --model /models/DeepSeek-V2-Lite.Q4_K_M.gguf --gpu

The proxy exposes OpenAI-compatible endpoints — all your tools work through it (curl, open-webui, langchain):

# streaming
curl http://localhost:11435/v1/chat/completions -d '{
  "model":"qwen3:4b",
  "messages":[{"role":"user","content":"write a Python script"}],
  "stream":true
}'

# blocking
curl http://localhost:11435/v1/chat/completions -d '{
  "model":"qwen3:4b",
  "messages":[{"role":"user","content":"hello"}],
  "stream":false
}'

2. Monitor cache stats

moe-l2 stats --port 11435

Example output:

moe-l2 cache stats
  requests:     47
  hits:         42     (89.4%)
  misses:        5
  slots_used:  32/48  (66.7%)
  memory:     456 MB  (68.3% of 668 MB)

3. Use as a library

from moe_l2 import predict, predict_hybrid, domain_to_expert_ids
from moe_l2.cache import L2Cache

# Predict domain (zero-dependency mode)
domain = predict("print hello world")  # → "codegen"

# Or use the hybrid semantic predictor
domain = predict_hybrid("how do I sort a list?")  # → "codegen"

# Preload experts
cache = L2Cache(model_path="model.gguf", l2_size="4GB")
cache.preload(domain_to_expert_ids[domain])

Architecture

┌────────────────────────────────────────────────────────────┐
│                       HTTP client                          │
│     curl / open-webui / langchain / any OpenAI client      │
└──────────┬─────────────────────────────────────────────────┘
           │ POST /api/chat
           ▼
┌────────────────────────────────────────────────────────────┐
│               moe-l2 Proxy (port 11435)                     │
│                                                            │
│  ┌─────────────────────────────────────────────────────┐   │
│  │  Domain Predictor                                    │   │
│  │  - Keyword mode: zero deps, instant classification   │   │
│  │  - Hybrid mode: +sentence-transformers for context   │   │
│  └─────────────────────┬───────────────────────────────┘   │
│                        │ domain                             │
│  ┌─────────────────────▼───────────────────────────────┐   │
│  │  L2 Cache (mmap'd shared memory)                    │   │
│  │  - LRU eviction policy                              │   │
│  │  - Async preload: next-prediction prefetch          │   │
│  │  - Thread-safe concurrent access                    │   │
│  │  - Zero-copy mmap from SSD → RAM                    │   │
│  └─────────────────────┬───────────────────────────────┘   │
│                        │ forward request                    │
└────────────────────────┼───────────────────────────────────┘
                         ▼
┌────────────────────────────────────────────────────────────┐
│         llama-server (port 11436, CUDA GPU)                │
│         on-demand pin experts: lazy mmap, zero VRAM       │
│         GPU reads pinned experts via PCIe DMA             │
│         hot experts cached in VRAM (A3 LRU, 2048 slots)   │
│         cold expert pages evicted (v3.1, RSS capped)      │
└────────────────────────────────────────────────────────────┘

CLI reference

Options:

• --model auto: scan /opt/data/models/*.gguf
• --l2-size 4GB / --l2-size 512MB: target cache size (proxy-only mode)
• --port 11435 (default)
• --gpu: enable GPU mode (requires CUDA + NVIDIA GPU; spawns bundled on-demand pin llama-server on 11436)

GPU binaries: Not tracked in git (bundled as llama_bins.tar.gz, ~1.6 GB multi-architecture on the bins-v0.6.0 release — sm_61/75/86/89/120a, one binary for all NVIDIA consumer GPUs, ships cuda-libs). Fetched at runtime via moe-l2 download-bins. When you pip install moe-l2, binaries are included. For git-clone users, run moe-l2 download-bins to fetch them from GitHub Release.

Platform requirements

• Linux x86_64 only — pre-built binaries target Linux AMD64 (CUDA .so + llama-server)
• macOS, Windows, and ARM Linux are not supported
• NVMe SSD strongly recommended
• NVIDIA GPU required for --gpu mode

More data

The speed tradeoff is intentional and small: expert weights live in CPU RAM (lazy mmap, zero VRAM) and are pinned on first touch — the GPU reads them directly via PCIe DMA, hot experts are cached in VRAM (A3 LRU), and cold pages are evicted to keep RSS capped. On the 2026-08-19 bins-v0.6.0 build, DS-V2-Lite reaches 139-154 t/s gen at ~10.1 GB VRAM — faster than full-GPU at ~43% of the VRAM.

Expert offload & cache fast path (llama.cpp)

Beyond the proxy layer, moe-l2 ships llama.cpp patches that compile expert handling directly into the CUDA backend — no proxy needed. Two mechanisms:

1. Selective pin expert GPU fast path (2026-08-10, current main path). Expert tensors live in CPU RAM via lazy mmap (zero VRAM). A router map (top-K experts per layer) pre-pins the hot experts as host-pinned (cudaHostRegister), so the GPU reads them directly via PCIe DMA; experts outside the map fall back to on-demand pin. Hot experts are cached in VRAM (A3 LRU, 32768 slots max) and cold pages are evicted (v3.1) to keep RSS capped. Measured (2026-08-19, bins-v0.6.0): DS 139-154 / Qwen 25.5-44.2 t/s (4090, mixed-domain). V4: ⚠️ speed N/A — upstream deepseek4 CUDA expert bug (#25582).

2. A3 LRU expert cache (historical, --expert-cache). An LRU cache that keeps recent experts on GPU. In the old --cpu-moe CPU-compute architecture it cut VRAM from 6.6 GB → 1.2 GB (5.64×) at 8.2 t/s. In the current on-demand pin architecture the cache is hooked into the scheduler copy layer (GGML_CUDA_EXPERT_CACHE) and only pays off for small, frequently-hit experts (see below).

When the cache helps (and when it doesn't)

The sched-cache only pays off when experts are small and frequently hit. Verified on RTX 4090 (host-buffer, 2026-08-02):

Key findings (2026-08-02, cache hooked into the scheduler input-copy layer):

• The cache sits in copy_experts: on hit it does a D2D copy (no PCIe round-trip), on miss it falls back to the pinned-host CPU→GPU path and writes back. It only intercepts single-expert groups.
• Benefit = expert size × hit rate. DS (1.55 MB, top-6) wins big; Qwen (~1 MB) pays for itself at best; Mixtral (252 MB, top-2) never hits enough to pay for its VRAM slots.
• Recommended: GGML_CUDA_EXPERT_CACHE=0.25 for DS-class models (16 slots/layer cover all hot experts, VRAM unchanged). Leave it off for Qwen/Mixtral.

Run the demo yourself: bash examples/demo_a3_compression.sh (edit paths first).

Related work

AirLLM (lyogavin/airllm, ~29k stars)

AirLLM is a general-purpose layer-offload scheme for very large models. Its scheduling granularity is the full Transformer layer: during inference only one layer's weights stay in VRAM, everything else is swapped to/from disk — giving an extreme low-VRAM floor (4GB GPU runs 70B). But it has three weaknesses: ① every generated token requires reading/writing a full layer to/from disk, so IO cost is huge and interactive speed is very low; ② no MoE-specific routing prediction or expert hot cache (per-expert streaming only started in 2026-07, with Kimi K3), so repeated prompts keep triggering heavy disk reads; ③ built on native Hugging Face Transformers, with no OpenAI-compatible serving interface out of the box, making it awkward to wire into Open WebUI, LangChain, etc.

Which to choose: pick moe-l2 if you run MoE models (DeepSeek/Qwen) locally for chat, have an 8–12GB older NVIDIA card, want an OpenAI API for tooling, or use multi-shard giant GGUFs. Pick AirLLM if you need dense (non-MoE) models, use Windows/macOS/AMD or CPU-only environments (moe-l2 currently requires Linux + NVIDIA), only do one-shot batch generation, or must stay with native HF weights.

Testing

Automated CI runs on every push (GitHub Actions, Python 3.10–3.13): ruff lint, pytest with coverage (fail below 50%), and package build. Status badge:

• 113 tests covering the Python scheduler core: domain predictor (keyword boundaries, fallback), L2 cache (LRU eviction, pinning, domain switching), GGUF weight reader (synthetic models), transparent proxy (live fake-backend HTTP, blocking + SSE), CLI helpers and the training data flywheel.
• Coverage: 72–88% on the core modules (cache 88%, proxy 78%, gguf_reader 73%, predictor 72%), ~55% total.
• Run locally:

  uv sync --group dev
  uv run pytest tests/
  uv run ruff check moe_l2/ tests/
  

The C++ side (llama.cpp on-demand-pin / expert-cache patches) is GPU-bound and is verified by the end-to-end benchmark reports in references/ — see models-benchmark.md.

Project status

• ✅ Domain predictor (keyword + optional semantic)
• ✅ L2 cache (mmap LRU, thread-safe, async preload)
• ✅ Transparent proxy (HTTP/SSE forwarding)
• ✅ CLI with auto model detection, GPU mode, and collect (routing data → expert map)
• ✅ Selective pin + GPU prefill (2026-08-10, v0.4.0, current main path): router-map-driven top-K pin → V4 RSS 84.4 → 26.8 GB (on-demand fallback 17.5 GB; ⚠️ V4 speed N/A — upstream deepseek4 CUDA bug #25582); DS 139-154 / Qwen 25.5-44.2 t/s on 4090 (2026-08-19, bins-v0.6.0); GPU cache prefill lifts cold-start round1 10.7 → 19.7 t/s (+84%). (Prior milestones: host-buffer fast path 08-02 → on-demand pin 08-07 → selective pin 08-10.)
• ✅ bins-v0.6.0 (2026-08-19): cache slot cap 16384→32768; per-slot lock optimization (concurrent correctness restored — v0.5.0 lock-free DS concurrent was garbage; same-machine A/B 4090: Qwen 31.76 vs 32.42 ≈parity, DS 140.43 vs 113.15 +24%); retain-hot-experts v2 (default single-table, MOE_L2_POOL_SIZE=3 opt-in); VRAM-adaptive main top-k (≤12G=75 / >12G=100); proxy concurrency fixes (DS long 500 gone). Three-card regression (2080Ti/4090/5090) all outputs clean.
• ✅ Expert cache boundary verified on Mixtral 8x7B / RTX 4090 (2026-08-02, sched-cache): cache benefit = expert size × hit rate — DS-V2-Lite (1.55 MB, top-6) gets Prompt +211% / Gen +5% at cache=0.25; Qwen (~1 MB) and Mixtral (252 MB, top-2) get no gain. Recommended: cache=0.25 for DS-class, off otherwise.
• ✅ DeepSeek-V4-Flash (157B MoE) verified (2026-08-05): 85 GB 3-shard GGUF runs on 2080 Ti (11 GB) — VRAM 8.3-9.1 GB, RSS capped by expert-page eviction v3.1 (fixed-expert-count LRU, MOE_L2_LRU_MAX_EXPERTS), multi-shard GGUF parsing fix shipped. Full report
• ✅ PyPI package (moe-l2)

License

Apache 2.0. See LICENSE for details.