tensorRT: encoder/decoder fp8 tiers + DiT precision rename (fp16mixed→fp16); retire w8_bf16 & medium bf16 - #92

Merged

Cortexelus merged 6 commits into

mainfrom

decoder-encoder-quant-tiers

Aug 2, 2026

Commits on Aug 2, 2026

tensorRT: add quantized encoder+decoder tiers (fp8 / w8_bf16, wide profile, pad node)

Cortexelus
and
claude
committed
4411f0a
tensorRT/quantize: add bf16 baselines as rows in all 4 tier tables + fp8-vs-fp8_fast

Cortexelus
and
claude
committed
b2dc1ad
tensorRT: wire decoder/encoder quant tiers into CLI, gradio, install.sh

Cortexelus
and
claude
committed
5b8860f
tensorRT: AOT SWA plugin for sm_120 tier builds + restore missing _arch.py

Cortexelus
and
claude
committed
0952f19
tensorRT: retire the w8_bf16 decoder/encoder tier (redundant with bf16)

Cortexelus
and
claude
committed
8f48ee0
tensorRT: rename DiT/T5Gemma precision fp16mixed → fp16, retire medium bf16 tier

Cortexelus
and
claude
committed
c7e8800