Skip to main content

Diffusion and time-series tasks

These tasks use the same bundle and runtime-root mechanics as text generation, but expose different typed request contracts.

Image generation​

python -m tensorrt_model_connect build black-forest-labs/FLUX.1-schnell \
  -o flux.bundle \
  --precision bf16 \
  --image-height 1024 \
  --image-width 1024

trtmc generate-image flux.bundle \
  --runtime-root /opt/trtmc/lib \
  --prompt "A brass robot reading beside a window" \
  --output robot.png \
  --height 1024 \
  --width 1024 \
  --num-steps 4 \
  --seed 1234

The bundle dimensions are build inputs; sampling steps, guidance, seed, and output path are request inputs. Use generate-image-batch with newline-delimited prompts and a matching seed list when the family implements batch generation.

Video generation​

python -m tensorrt_model_connect build Wan-AI/Wan2.1-T2V-1.3B \
  -o wan.bundle \
  --precision bf16 \
  --image-height 480 \
  --image-width 832 \
  --video-num-frames 81

trtmc generate-video wan.bundle \
  --runtime-root /opt/trtmc/lib \
  --prompt "Ocean waves under moonlight" \
  --output waves.mp4 \
  --height 480 \
  --width 832 \
  --num-steps 20 \
  --seed 1234

Family-specific acceleration policies are encoded in family code rather than an open-ended --set registry.

Time-series forecasting​

The forecast CLI consumes raw float32 values:

python -m tensorrt_model_connect build amazon/chronos-bolt-tiny \
  -o chronos.bundle \
  --precision fp32

trtmc forecast chronos.bundle \
  --runtime-root /opt/trtmc/lib \
  --input history.f32 \
  --frequency H

Optional --mask input uses the family contract. Neural-operator families use solve --branch FILE and optional --trunk FILE instead of comma-separated values on the command line.

For every task, confirm shape, dtype, file format, supported dimensions, and numerical thresholds in the owning family's tests and manifest.