Merge branch 'lab/vllm_version' into 'export/v1-2-0'

Do not install vllm 0.22.0 due to unsupporting exllama kernel

See merge request onecomp/onecomp-lab!84

Author: FKKimura

CHANGELOG.md

@@ -86,6 +86,10 @@
 - Fixed `tests/onecomp/quantizer/test_module.py` to feed `y_replaced` consistently into `q_proj` / `k_proj` / `v_proj` after quantized weights are applied, aligning the replacement-path forward test with the intended residual update flow
 - Extracted the duplicated attention+MLP forward loop in `test_quantize_error` into `TestModel.forward()` (`tests/onecomp/quantizer/test_module.py`); both the pre-quantization and post-quantization inference paths now call `model(inp)` directly, eliminating 34 duplicate lines
 
+### Dependencies
+
+- Pinned the `vllm` optional dependency to `vllm>=0.10,<0.22` in `pyproject.toml` (and regenerated `uv.lock`). vLLM 0.22.0 removed the legacy Exllama GPTQ kernel that OneComp's GPTQ serving uses for low bit-widths (2-/3-bit, and 4-/8-bit models that are asymmetric or use `desc_act`), so **vLLM 0.22 and later are not supported** — serving affected models on 0.22+ fails at runtime. Documented in `docs/user-guide/vllm-inference.md` and `docs/getting-started/installation.md`.
+
 ### Documentation
 
 - Documented save/load and vLLM compatibility for the newly-supported **JointQ**, **RTN**, and **OneBit** quantizers across the docs:

docs/getting-started/installation.md

@@ -119,6 +119,9 @@ uv sync --extra cu130 --extra dev --extra visualize --extra vllm
 !!! note "vLLM requires the `cu130` extra"
     Recent vLLM releases depend on `torch>=2.10`, whose wheels are only published for the `cu130` index. The `--extra vllm` declaration in `pyproject.toml` therefore conflicts with `cpu`, `cu118`, `cu121`, `cu124`, `cu126`, and `cu128`; combining any of these with `--extra vllm` is rejected by `uv` at lock time.
 
+!!! warning "vLLM 0.22+ is not supported"
+    vLLM 0.22.0 removed the legacy Exllama GPTQ kernel that OneComp's GPTQ serving relies on for low bit-widths (2-/3-bit, and Marlin-ineligible 4-/8-bit), so `pyproject.toml` pins `vllm>=0.10,<0.22`. See [vLLM Inference](../user-guide/vllm-inference.md#installation) for details.
+
 !!! warning
     Do **not** install vLLM with `uv pip install vllm` after `uv sync`. Packages installed via `uv pip` are not tracked by the lockfile and will be removed or overwritten by subsequent `uv sync` or `uv run` commands. Always use `--extra vllm` instead.
 

docs/user-guide/vllm-inference.md

@@ -43,6 +43,9 @@ vLLM is available as an optional dependency:
     pip install vllm
     ```
 
+!!! warning "vLLM 0.22+ is not supported"
+    OneComp's GPTQ serving relies on the Exllama GPTQ kernel for low bit-widths — 2-/3-bit, and 4-/8-bit when the model is asymmetric or uses activation reordering (`desc_act`). vLLM 0.22.0 removed the legacy Exllama GPTQ kernel, so these configurations fail at runtime. `pyproject.toml` therefore pins `vllm>=0.10,<0.22`; use a vLLM version below 0.22.
+
 !!! note
     vLLM requires CUDA and a compatible GPU. See the [vLLM documentation](https://docs.vllm.ai/) for detailed installation instructions and system requirements.
 

pyproject.toml

@@ -71,6 +71,8 @@ visualize = [
 hydra = [
     "hydra-core",
 ]
+# vllm v0.22.0 removes Exllama legacy gptq
+vllm = ["vllm>=0.10,<0.22"]
 eval = [
     "hydra-core>=1.3",
     "omegaconf>=2.3",
@@ -79,7 +81,6 @@ eval = [
     "japanize-matplotlib>=1.1",
     "shortuuid>=1.0",
 ]
-vllm = ["vllm>=0.10"]
 docs = [
     "mkdocs-material",
     "mkdocstrings[python]",