INFERENCE RESEARCH

Inference engineering, in practice.

We track the inference stack, explain what is changing, and investigate how systems behave under real workloads.

Understand the stack

Follow the technology. We curate serving engines, kernels, compilers, and infrastructure by where they fit and the problems they address.

Read beyond the headline. Our technical briefs connect developments to practical inference questions, with primary sources, useful context, and the limits of each claim.

Learn through experiments

Measure the trade-offs. We design experiments around latency, throughput, memory, cost, and quality. Methods, configurations, and limitations belong with every result; proposed work stays clearly labeled.

Build with others. Bring a bottleneck, a workload, or an open question. We want to build a community around reproducible experiments and shared findings.

01 / DEVELOPMENTS

Daily Brief

What changed, why it matters, and the primary source.

3 updates
UPDATE

vLLM expands encoder-disaggregated serving.

Metadata-only requests, remote embeddings, and shared encoder outputs extend the encoder / prefill / decode path.

Our readingFor media-heavy workloads, profile encoder capacity and embedding transfer separately from token generation.

Primary source · vLLM ↗Source published 11 Nov 2026

02 / OUR WORK

Lab Notes

Experiments and benchmarks, with the method and evidence.

What a result includes
LAB NOTESComing soon

Experiments are coming soon.

No experiments are published yet. Lab Notes will appear here once we have run the work ourselves, with the method, setup and measured results.

TURINGFOLD / RESEARCH