Migration Guide: TensorRT Backend Removed#
Note
Breaking change. The TensorRT engine backend has been removed. PyTorch is now the sole execution backend for TensorRT LLM (AutoDeploy, built on the PyTorch backend, remains available).
What changed#
Removed |
Replacement / new behavior |
|---|---|
|
Raises |
|
Use |
|
Use |
|
No engine-build step — HuggingFace checkpoints load directly |
Per-model |
Not needed — no checkpoint conversion |
|
Omit, or pass |
|
Dropped (no longer installed) |
How to migrate#
Python API — PyTorch is the default, so drop the backend argument:
from tensorrt_llm import LLM
llm = LLM(model="<hf_model>")
Command line — --backend pytorch is the default, so no engine is needed:
trtllm-serve <hf_model>
There is no separate checkpoint-conversion or engine-build step.
Examples#
PyTorch usage lives with the LLM API examples (for example
quickstart_advanced.py, quickstart_multimodal.py) and the model-specific
deployment guides. The per-model convert_checkpoint.py /
trtllm-build example directories have been removed.
Retained reference docs#
The legacy tree under docs/source/legacy/ is kept for cross-reference only.
Pages that are purely about the TensorRT engine mechanism carry a caution banner;
the rest document concepts still relevant to the PyTorch backend.
Features & concepts: attention, KV cache management / reuse, speculative decoding, expert parallelism, disaggregated serving, low-precision allreduce, executor API.
Performance & tuning: analysis, benchmarking, tuning guide.
Architecture: adding a model, checkpoint format, conversion workflow, weights loader.
Reference: precision, memory, multimodal support matrix, troubleshooting, key features, PyTorch overview.
Dev environment: build a Docker image, develop on Runpod.