Compare commits

...
Author SHA1 Message Date
khluu 1b2570ec77 Add tasks/version/logging_utils/beam_search to basic_correctness deps
Following the deeper audit pattern from #42219:

- `vllm/tasks.py`: imported by `vllm/model_executor/models/registry.py`
  and other model-side code transitively reached by `LLM(...)`.
- `vllm/version.py`: imported by `vllm/v1/engine/core.py` and
  `vllm/engine/arg_utils.py`.
- `vllm/logging_utils/`: imported by `vllm/model_executor/models/registry.py`
  (`logtime`) and `vllm/v1/engine/core.py`.
- `vllm/beam_search.py`: imported by `vllm/entrypoints/llm.py` (the
  `LLM` class is the entry point under test).

These are first-order imports of the engine/LLM path that
`test_basic_correctness.py`, `test_cpu_offload.py`, and
`test_cumem.py` exercise.

Signed-off-by: Kevin H. Luu <kevin@anyscale.com>

Signed-off-by: khluu <khluu000@gmail.com>
2026-05-10 03:50:16 -07:00
khluu 8d82abca1d Add sequence.py and entrypoints/llm.py to basic_correctness deps
Address review on #42223:

- Add `vllm/sequence.py`: defines `IntermediateTensors` and other
  engine-API types used by the model forward path; an
  incompatible change can reach `LLM(...)` even without a direct
  test import.
- Add `vllm/entrypoints/llm.py`: tests do `from vllm import LLM`,
  which resolves to `vllm/entrypoints/llm.py`.

Skipped gemini's other suggestions:
- `vllm/worker/`, `vllm/executor/`, `vllm/attention/`: don't exist
  at the top level. Live under `vllm/v1/worker/`, `vllm/v1/executor/`,
  `vllm/v1/attention/`, all already covered by `vllm/v1/`.
- `vllm/forward_context.py` and `vllm/_custom_ops.py`: covered by
  compile-tests and kernel-tests respectively, which trigger on
  changes to those modules.

Signed-off-by: Kevin H. Luu <kevin@anyscale.com>

Signed-off-by: khluu <khluu000@gmail.com>
2026-05-10 03:38:02 -07:00
khluu 7f9edabf85 [CI] Narrow basic_correctness.yaml source dependencies
The Basic Correctness job in
`.buildkite/test_areas/basic_correctness.yaml` listed `vllm/` as a
source dependency, causing it to run on any change anywhere under
`vllm/`. The three tests it runs (`test_basic_correctness.py`,
`test_cpu_offload.py`, `test_cumem.py`) instantiate `LLM(...)` and
exercise the inference stack plus `vllm.device_allocator.cumem`.
Replace with the modules these tests actually exercise.

Excludes paths that have their own dedicated coverage and aren't
directly imported: `vllm/lora/`, `vllm/spec_decode/`, `vllm/tracing/`,
`vllm/profiler/`, `vllm/reasoning/`, `vllm/tool_parsers/`,
`vllm/renderers/`, `vllm/benchmarks/`, `vllm/entrypoints/openai/`,
`vllm/entrypoints/api_server.py`, `vllm/entrypoints/cli/`,
`vllm/compilation/`, `vllm/kernels/`, `vllm/ir/`, `vllm/plugins/`,
`vllm/triton_utils/`, `vllm/forward_context.py`, `vllm/sequence.py`,
`vllm/_aiter_ops.py`, `vllm/_custom_ops.py`.

Signed-off-by: Kevin H. Luu <kevin@anyscale.com>

Signed-off-by: khluu <khluu000@gmail.com>
2026-05-10 03:31:09 -07:00
+18 -1
View File
@@ -7,7 +7,24 @@ steps:
timeout_in_minutes: 30
device: h200_18gb
source_file_dependencies:
- vllm/
- vllm/beam_search.py
- vllm/config/
- vllm/device_allocator/
- vllm/distributed/
- vllm/engine/
- vllm/entrypoints/llm.py
- vllm/inputs/
- vllm/logging_utils/
- vllm/model_executor/
- vllm/multimodal/
- vllm/platforms/
- vllm/sampling_params.py
- vllm/sequence.py
- vllm/tasks.py
- vllm/transformers_utils/
- vllm/utils/
- vllm/v1/
- vllm/version.py
- tests/basic_correctness/test_basic_correctness
- tests/basic_correctness/test_cpu_offload
- tests/basic_correctness/test_cumem.py