[Model][NVIDIA] Route DSA models to the CUDA non-compiled path (#52861)

This commit is contained in:
Woosuk Kwon
2026-08-19 00:13:52 -07:00
committed by GitHub
parent ee11730751
commit b09bd69b5b
16 changed files with 260 additions and 146 deletions
-13
View File
@@ -132,19 +132,6 @@ MODEL_SPECS = [
),
id="gpt_oss_120b",
),
# NOTE: DeepSeek-V3.2 requires sparse MLA (index_topk) which needs
# Hopper+ GPUs. This test must run on H100 (see pytorch.yaml).
pytest.param(
ModelStartupSpec(
model="deepseek-ai/DeepSeek-V3.2",
hf_overrides=_SMALL_MOE_OVERRIDES,
cold_artifacts_saved=9,
# https://github.com/vllm-project/vllm/issues/38051
warm_artifacts_saved=0 if is_torch_equal_or_newer("2.12.0") else 9,
warm_artifacts_loaded=9 if is_torch_equal_or_newer("2.12.0") else 0,
),
id="deepseek_v3.2",
),
pytest.param(
ModelStartupSpec(
model="moonshotai/Kimi-K2.5",