[pull] main from NVIDIA:main#1314

Merged

pull[bot] merged 11 commits into

LarryXFly:mainfrom

May 29, 2026

Commits on May 29, 2026

[https://nvbugs/6189416][fix] Add a Blackwell-specific reference entry (extra_acc_spec=sm100_fp8, accuracy=46. (#14484)
tensorrt-cicd
authored
[#14619][perf] AutoDeploy: tune Llama-3.1-8B-Instruct-FP8 TP=2/4 config and handle CG max bs when it is unset in the yaml (#14622)
MrGeva
authored
[#13561][feat] AutoDeploy: enable MLIR elementwise fusion and trtllm_gen MoE on Nano NVFP4 (#14554)
MrGeva
authored
[https://nvbugs/6221841][fix] Detect via the raw config_dict whether the user actually set a top-level rope_th (#14624)
tensorrt-cicd
authored
[None][perf] Replace Parakeet audio encoder with native trtllm layers (#14474)
aswinvisva
authored
[https://nvbugs/6194552][fix] stabilize Triton Mamba softplus (#14652)
hnover-nv
authored
[TRTLLM-13043][chroe] add VisualGen context to AGENTS.md (#14732)
zhenhuaw-me
authored
[None][feature] Add thinking token budget control (#14665)
tijyojwad
authored
[TRTLLM-13050][test] Remove two-model eagle3 spec-decoding tests (#14735)
QiJune
authored
[TRTLLM-12653][feat] LTX-2 Ulysses cross-attention for v2a with audio padding (#14044)
luyiyun1021
authored
[None][refactor] Flatten thop.attention sequence kwargs + rename rotary_embedding_* to rope_* (#14569)
yuxianq
authored