What happened
A change to Tencent-Hunyuan’s Hy-Nomas README adds an “Inference with SGLang” section. It describes SGLang Diffusion serving tencent/Hy-Nomas for offline video jobs as well as realtime sessions that receive segments one at a time. On first start, the system downloads the distilled DiT, PixelUMM decoder, and text encoder, then assembles and caches a local pipeline. The README says the same server can accept realtime WebSocket sessions and return frames as they are decoded.
The instructions distinguish the current development setup from a stable release. Hy-Nomas support is not yet in a stable SGLang release, so the README recommends the nightly lmsysorg/sglang:dev Docker image, which is rebuilt daily from main. It says to switch to lmsysorg/sglang:latest after a release includes Hy-Nomas. Builders can alternatively install SGLang from source using Python 3.12 and the python[diffusion] extra. The Docker example mounts the local cache directory so downloaded weights and the assembled pipeline persist across containers.
The documented server command runs on one GPU; a second example uses two GPUs with sequence parallelism enabled. For offline generation, the sample client posts to /v1/videos with a JSON prompt, two segments of 120 frames each, and a total of 240 frames at 1280 by 704 resolution and 24 frames per second. It specifies eight inference steps, a guidance scale of 1.0, and seed 42, then polls the job until it completes or fails and downloads the resulting MP4. These are example values, not a general performance guarantee.
For realtime use, the README points to /v1/realtime_video/generate. It says each segment is generated as it arrives and frames stream back during decoding. The SGLang guide linked from the README covers the realtime protocol, camera control, layerwise offload, and KV-cache budget. The README also lists implementation optimizations: Ulysses sequence parallelism and tensor parallelism for the DiT, PixelUMM decoding windows spread across GPUs, layerwise host-memory offload, FlashAttention-3 for segment attention on Hopper GPUs (with FlexAttention as fallback on other GPUs), and fused elementwise kernels described as bitwise identical to the unfused path.
What the archive adds
This is a new fact on a known story, not a new model release. SGLang merges Hy-Nomas support for offline and real-time streaming video generation covered the serving integration; this repository change adds Tencent’s official usage instructions. Our earlier coverage of SGLang-Omni concerned a different Tencent model and serving support, so it does not establish anything further about this guide.
Trend signal
This supports generative media becomes application infrastructure: the guide describes both queued video jobs and a realtime interface that accepts successive segments. That makes the integration more concrete for application builders, although documentation alone does not show that the path is reliable at production scale or broadly adopted.
Why now
The source establishes a sequence: SGLang support was available, and Tencent’s model repository now documents how to use it. It does not explain why Tencent published the instructions at this point, so no broader timing rationale can be inferred.
Connections to other events
The earlier SGLang Hy-Nomas support note is the direct connection: it covered the serving capability; this README change documents how to invoke it. The guide also sits alongside Tencent’s separate Hy-SimulMT serving coverage, but that is a different model and does not demonstrate shared implementation details.
Why it matters
For a team evaluating Hy-Nomas, an official repository recipe reduces the gap between knowing a runtime supports a model and attempting a local deployment. The distinction between offline jobs and realtime segment streaming matters architecturally: a batch workflow can submit a complete request and retrieve a finished file, while a realtime workflow must manage incoming segments and consume output while decoding continues.
The examples also expose practical choices before a team invests in integration work: whether it can use a nightly image, whether it has one or two GPUs available, how it will retain model downloads between containers, and whether its request shape fits the documented segment-based interface. The listed optimizations suggest where the serving path addresses compute and memory constraints, but they do not quantify end-to-end speed, cost, or hardware requirements. Teams should treat the values as a reproducible starting point, then benchmark their own prompts, hardware, and latency targets rather than assuming the sample configuration predicts production behavior.
Limits and uncertainty
The README is an implementation guide, not an evaluation report. It provides no measured throughput, latency, memory footprint, output-quality comparison, or cost figures. Its sample request and optimization descriptions do not establish how performance changes with different video lengths, prompts, GPU configurations, or concurrent sessions. Hy-Nomas support is explicitly not yet in a stable SGLang release; the guide points readers to a nightly image until a release includes it, without naming a release date. The source also does not establish production reliability, service-level guarantees, or the support status of every listed optimization on every GPU.
Skill to build now
If Hy-Nomas is relevant to your work, use the documented two-segment, 240-frame request as a baseline experiment. Record the SGLang image or source revision, GPU count, resolution, frame rate, inference steps, completion time, and whether output frames arrive incrementally. This will reveal whether the offline or realtime interface fits your application, while keeping the test tied to the configuration actually documented.
Reader takeaway
Tencent’s README now gives builders a concrete SGLang path for Hy-Nomas, including an offline request example and a realtime endpoint. It lowers setup friction, but support is still tied to a nightly build, and the guide provides no performance evidence. Treat it as an integration recipe to test, not proof of production readiness.
Open question
When will Hy-Nomas support appear in a stable SGLang release, and what measured performance will that release document?