Commit b2dc738
authored
[Refactor][Patch] Remove obsolete parser patches (vllm-project#11453)
### What this PR does / why we need it?
vLLM `v0.24.0` and the verified main revision no longer need these
compatibility patches:
- `patch_glm47_tool_call_parser.py`
- `patch_glm_tool_call_streaming.py`
- `patch_tool_choice_none_content.py`
This PR removes the patches, their registrations, patch documentation,
and dedicated unit tests.
The supported vLLM baselines now include the GLM streaming parser engine
from vllm-project/vllm#45915 and the empty
`tool_calls` response fix from
vllm-project/vllm#44105.
### Does this PR introduce _any_ user-facing change?
No. It removes obsolete compatibility code now covered by upstream vLLM.
### How was this patch tested?
- Built vLLM Ascend from source with `MAX_JOBS=32 python setup.py
build_ext --inplace`.
- `TORCH_DEVICE_BACKEND_AUTOLOAD=0 python -m pytest -q
tests/ut/patch/platform`: 69 passed.
- `python -m py_compile vllm_ascend/patch/platform/__init__.py
vllm_ascend/patch/__init__.py`: passed.
- `git diff upstream/main...HEAD --check`: passed.
- `bash format.sh ci`: passed.
- Ran a real GLM-5 TP16 server on `/models/GLM-5-w4a8` with
`--tool-call-parser glm47` and `--reasoning-parser glm45`. Streaming and
non-streaming cases passed for content-only responses,
`tool_choice=none`, zero-argument tools, nested JSON arguments, required
and named tools, automatic tool choice, multiple tool calls,
thinking-only output, and thinking-plus-tool output. Streamed tool
metadata appeared once per call and reconstructed arguments were valid
JSON.
- vLLM version: v0.24.0
- vLLM main:
vllm-project/vllm@85c09e9
Signed-off-by: QwertyJack <7554089+QwertyJack@users.noreply.github.com>
Co-authored-by: QwertyJack <7554089+QwertyJack@users.noreply.github.com>1 parent acdaf76 commit b2dc738
8 files changed
Lines changed: 0 additions & 817 deletions
Lines changed: 0 additions & 139 deletions
This file was deleted.
Lines changed: 0 additions & 144 deletions
This file was deleted.
0 commit comments