This document is relevant for: Trn2, Trn3

Feature and model compatibility matrix#

Use this page to check feature support before configuring a deployment. For configuration details, see the features guide.

Feature/model matrix#

Feature

GPT-OSS

Qwen3-VL

Continuous batching

✅

✅

Segmented prefill

✅

✅

Prefix caching (APC)

✅

✅

Speculative decoding (EAGLE3)

✅

❌

FP8 weight quantization (static)

❌

❌

MXFP8 weight quantization

❌

✅

MXFP4 weight quantization

✅ ¹

❌

KV cache FP8

✅

❌

Multimodal (image input)

❌

✅

Disaggregated inference (1P1D / xPyD)

✅

❌

Structured outputs / tool calling

✅

✅

On-device sampling

✅

✅

Tensor parallelism

✅

✅

Data parallelism

✅

✅

Expert parallelism

✅

N/A

Vision encoder parallelism

N/A

✅

¹ Trn3 only.

Legend#

  • ✅ Supported and tested

  • ❌ Not supported

  • N/A Not applicable to this architecture