fix(trt_llm): warn when python_version is ignored

[blockgroot]

🚀 What

Adds a warning when python_version is set on a TRT-LLM model, since it currently has no effect there.

Fixes #2168

Problem: for any TRT-LLM model (decoder/Briton, encoder/BEI, encoder-bert, or v2), _fill_trt_llm_versions (truss/contexts/docker_build_setup.py:34-84) forces config.base_image + a fixed python_executable_path — /usr/local/briton/venv/bin/python for decoder, /usr/bin/python3 for the others. Once config.base_image is set this way, serving_image_builder.py:970-975 skips the python_version-derived image tag entirely, and server.Dockerfile.jinja:32-34 symlinks python to that fixed path. So python_version is accepted, validated, and stored — but has zero effect on the running container, with no indication to the user. Reproduced directly against TrussConfig:

$ uv run python -c "
from truss.base.truss_config import TrussConfig
from truss.base.constants import TRTLLM_PYTHON_EXECUTABLE
cfg = TrussConfig.from_dict({
    'model_name': 'repro', 'python_version': 'py39',
    'model_metadata': {'tags': ['openai-compatible']},
    'resources': {'accelerator': 'A100', 'use_gpu': True},
    'trt_llm': {'build': {'max_seq_len': 1000, 'max_batch_size': 1, 'base_model': 'decoder',
        'checkpoint_repository': {'repo': 'TinyLlama/TinyLlama-1.1B-Chat-v1.0', 'source': 'HF'}}},
})
print('accepted python_version =', cfg.python_version)
print('but forced to run on:', TRTLLM_PYTHON_EXECUTABLE)
"
accepted python_version = py39
but forced to run on: /usr/local/briton/venv/bin/python

No warning was logged for this mismatch on main.

This override is intentional (traced to #1548 — TRT-LLM base images are backend-supplied with their own bundled Python/TensorRT-LLM runtime), so this PR doesn't change that behavior. It only surfaces it, since nothing currently does.

💻 How

Added a logger.warning in trt_llm_common_validation() (truss/base/trt_llm_config.py), which already runs for every TRT-LLM config (both v1 and v2, via trt_llm_validation) and already contains a similar warning for another misconfiguration (WEIGHTS_ONLY_INT8 quantization on A100). The new check fires when config.python_version differs from the field's default, using the same idiom already established in this function.

Considered raising a ValueError instead, since a few other checks in this function do raise. Went with a warning because setting python_version alongside trt_llm isn't actually invalid — just inert — so hard-failing would break any existing config that harmlessly sets both. Happy to switch to a hard error if you'd prefer that here.

🔬 Testing

Added 3 tests to truss/tests/trt_llm/test_trt_llm_config.py, reusing the existing trtllm_config/trtllm_config_v2 fixtures and the caplog.at_level("WARNING") pattern already used in truss/tests/test_config.py:

• test_trt_llm_non_default_python_version_warns — fails on main (no warning exists), passes with this change.
• test_trt_llm_default_python_version_no_warning — the default (py313) config produces no warning; guards against noise on the common case.
• test_trt_llm_v2_non_default_python_version_warns — same check against a v2 config, since trt_llm_common_validation is shared between both inference stacks.

truss/tests/trt_llm/ -q → 24 passed, 1 skipped

Also ran the full non-integration suite: 1651 passed, 19 skipped, 0 failed (no regressions vs. the main baseline of 1648 passed, 19 skipped), plus ruff check ., ruff format . --check, and pre-commit run --all-files (mypy-local, uv-lock, config-schema) — all clean.

Out of scope: the original issue also reported a torch.compile/Triton InductorError crash on newer PyTorch/Triton versions — per the issue thread itself, that's a deployment-environment issue, not something fixable in truss, so this PR only addresses the python_version part.

[CLAassistant]

All committers have signed the CLA.