TensorRT-LLM 0.18.0 listed no new features: it deprecated Windows, updated dependencies (TensorRT 10.9, CUDA 12.8.1, ModelOpt 0.25) and noted an SBSA known issue.
Think of a phone app update. Sometimes the test versions get exciting new features, but the official update ships without them. Its notes only say that support for an old phone system will end, some parts inside were updated, and one bug is known. The TensorRT-LLM 0.18.0 release was like that. Its notes say the features from the test versions are not included. They list three things: Windows support is being dropped, the libraries it depends on were updated, and one setup has a known problem. Features such as FP4 on Blackwell GPUs or speculative decoding exist in the library, but this release did not add them. So when you ask what a release shipped, read its notes, not the feature list of the whole product.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
TensorRT-LLM (written "TensorRT LLM" by NVIDIA) is NVIDIA's open-source LLM inference library built on TensorRT. Along with vLLM and SGLang, it is a main choice for serving large language models. A release note answers a narrow question: what changed in that version. It is easy to answer instead with what the product can do in general.
The v0.18.0 release (2025-04-02) shows the difference. Its notes say that features from the 0.18.0.dev pre-releases are not included in this release. What it lists is a Windows deprecation, dependency updates and one known issue. Options that name real TensorRT-LLM features (FP4 on Blackwell, in-flight batching, speculative decoding) are wrong for this release because the notes do not list them as new.
The library has since moved to a stable 1.x line.
What the 0.18.0 release notes contain
The v0.18.0 notes have four parts, and none of them is a feature list.
Key features. The notes state: "Features that were previously available in the 0.18.0.dev pre-releases are not included in this release." Work that appeared in the dev builds did not ship in the stable 0.18.0.
Breaking change. Windows platform support is deprecated as of v0.18.0, and all Windows-related code is to be removed in future releases.
Known issue. The PyTorch workflow on SBSA (Arm server) platforms is incompatible with bare-metal environments such as Ubuntu 24.04. The notes recommend the PyTorch NGC container on those platforms.
Infrastructure. The base Docker image moves to the NVIDIA PyTorch 25.03 image and the backend image to Triton Server 25.03. Dependencies move to TensorRT 10.9, CUDA 12.8.1 and NVIDIA ModelOpt 0.25 (Linux only).
Point releases v0.18.1 (2025-04-09) and v0.18.2 (2025-04-16) followed. So options D, E and F, which restate these items, are the correct picks.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- The v0.18.0 notes (2025-04-02) state that features from the 0.18.0.dev pre-releases are not included, and list Windows deprecation as the one breaking change.
- Point releases v0.18.1 (2025-04-09) and v0.18.2 (2025-04-16) followed 0.18.0.
What an interviewer would ask next. Try answering before peeking at the approach.
QWhat is NVFP4 and how does it differ from vanilla FP4?
NVFP4 stores E2M1 4-bit elements (magnitudes up to 6) with one FP8 E4M3 scale per block of 16 elements, plus one FP32 per-tensor scale. MXFP4, the OCP format, uses the same E2M1 elements with one power-of-two E8M0 scale per 32-element block. The smaller blocks and finer scales of NVFP4 track local value ranges more closely. Both are accelerated on Blackwell fifth-generation Tensor Cores. Hopper has no FP4 tensor cores.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Crediting 0.18.0 with headline features (FP4 kernels, chunked prefill, Medusa and EAGLE) because they exist in TensorRT-LLM. The 0.18.0 notes list no new features, only a Windows deprecation, dependency updates and a known issue.
60 second bullets to scan on the way to the call.
What the v0.18.0 release notes list and what they say about dev pre-release features
The one breaking change in 0.18.0
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.