edge-infer

Running multimodal models on edge CPUs — a mini-PC or a gateway box with tens of gigabytes of RAM and no GPU — instead of on servers.

Built as a fork of vLLM-Omni, because it is the only engine with a multi-stage multimodal pipeline, continuous batching and an OpenAI-compatible server; the alternatives were evaluated and the reasoning is in docs/engines/.

The work turned out to be less about kernels than about three things nobody profiles: how much memory is resident while the model loads, how much is resident that nothing ever reads, and whether the number you just measured came from the code path you thought it did. The largest single win here — 843 MB — was a setting that had been marked "no effect, do not retry" because the code path under test discarded it.

This repository holds the engine changes, the measurement harness, the raw results, the baselines they are compared against, and the written analyses. Upstream's own README is preserved at docs/README.vllm-omni.md.

What's here

area	where
4-bit CPU quantization	vllm_omni/model_executor/layers/quantization/cpu_int4.py — W4A16 on PyTorch's tinygemm kernel, prefill reading the kernel's own weight layout instead of a duplicate
Memory	RoPE tables sized to the servable context; weights repacked in blocks; a KV budget that is actually honoured on CPU
CPU platform + vLLM patches	vllm_omni/platforms/cpu/, vllm_omni/patch.py — vLLM is monkey-patched at import; its source is unmodified and no vLLM is redistributed. See docs/edge/environment.md
Device routing	vllm_omni/edge/weight_path.py picks the weight format per device class, from measurements rather than assumption
Device placement	vllm_omni/edge/placement.py decides which parameters leave the GPU when the checkpoint is bigger than the VRAM — cold ones first, because embed_tokens and lm_head are the same size and the opposite decision
Measurement harness	benchmarks/edge_harness/ — the grid runner, benchmarks, memory ledger, fidelity metric and agent suite every number here came from
Raw results	benchmarks/edge_harness/results/ — 100+ artifacts backing the published figures, each with provenance
Baselines	baselines/ — llama.cpp runs and the comparison scripts
Engine analyses	docs/engines/ — llama.cpp, ExecuTorch, MNN, ncnn, vLLM, vLLM-Omni
Study documents	docs/analysis/ — the full reports these summaries distil

Headline results

Spark-X2.5-1.7B, Xeon Platinum 8480C, 24 pinned node-local cores, W4A16 g128. Every figure is a median of at least three interleaved passes against a control arm, on a shared host where end-to-end throughput swings ~30% run to run.

Memory — each measured as its own A/B with the alternative reachable by a flag:

lever	saving	cost
KV budget 1 GiB → 192 MiB	843 MB	fewer cached tokens
RoPE table capped to max_model_len	618 MB	none
One packed weight layout instead of two	615 MB	~1.6× prefill
Blocked weight repack	489 MB of load-time peak	none

These overlap and must not be added — see docs/edge/results.md.

Speed, same machine:

path	decode	prefill (pp512)
W4A8 (AMX int4_scaled_mm_cpu)	74.4 tok/s	—
W4A16 (tinygemm)	71.6 tok/s	2190 tok/s
cpu_gemm_wna16 fallback	32.3 tok/s	—
llama.cpp Q4_K_M (reference)	90.0 tok/s	1121 tok/s

Decode is llama.cpp's; prefill is ours by ~2×. Neither engine is bandwidth-bound — both sit at 33–39% of the measured 222 GB/s ceiling — so the decode gap is the execution model, not the kernels. See docs/edge/comparisons.md.

Why the numbers here are worth reading

This project has retracted five conclusions, every one of them a measurement error rather than a coding error. The harness exists because of them:

• a "3.1 GB saving" that was VmRSS counting the mmap'd checkpoint;
• W4A8 timings taken on a checkpoint whose output head was silently left in bf16;
• a memory benchmark reading /proc/self/status while the model ran in a worker process;
• "KV sizing saves only 188 MB" — measured through a code path that discarded the setting, so both arms ran the same cache;
• a 1.9× prefill cost quoted from the one pass of a grid whose other passes had crashed.

docs/edge/methodology.md is the resulting discipline, and it is the most transferable thing in this repository.

Quick start

# CPU build of vLLM, then this fork
pip install vllm==0.28.0+cpu
VLLM_TARGET_DEVICE=cpu pip install -e .

# a right-sized edge deployment
VLLM_CPU_KVCACHE_SPACE=1 python -c "
from vllm import LLM, SamplingParams
llm = LLM(model='<your-cpu_int4-checkpoint>', quantization='cpu_int4',
          dtype='bfloat16', max_model_len=2048, max_num_seqs=1,
          kv_cache_memory_bytes=192 << 20)   # the single largest memory lever
print(llm.generate(['Hello'], SamplingParams(max_tokens=32))[0].outputs[0].text)
"

Documentation

docs/edge/results.md	every measured figure, with spread and baseline
docs/edge/methodology.md	how to measure on a shared machine without fooling yourself — the five retractions and the rules they left
docs/edge/cpu-4bit.md	the two 4-bit paths, the packed layout as measured, building a checkpoint
docs/edge/comparisons.md	against llama.cpp and the other engines considered
docs/edge/environment.md	pins, machine, and why there is no vLLM patch series
docs/edge/local-text-mode.md	the single-stage local text engine: device capabilities, memory admission, explicit refusals, cancellation
docs/edge/external-stage-backends.md	placing a stage on the Radeon 890M or the AMD NPU: out-of-process workers, proven placement, the shared-RAM ledger
docs/edge/harness.md	running a comparison
docs/edge/laptop-heterogeneous.md	Qwen3.8-27B and Spark-X2.5-4B on a Ryzen AI laptop — what the CPU, NPU and GPU are each worth when decode is bandwidth-bound
ROADMAP.md	what is next, with expected gains and how each would be verified — and what was rejected

Licence

Apache-2.0, inherited from vLLM-Omni. See LICENSE and NOTICE.