iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[CI] Build mamba-ssm with C++20 for torch 2.14 nightly compatibility (#49600)

Browse Source 
Co-authored-by: Shengqi Chen <harry-chen@outlook.com>
This commit is contained in:
Andrey Talman
2026-08-26 17:31:08 +08:00
committed by GitHub
co-authored by Shengqi Chen
parent 73bd7c83c4
commit b821d7b2a6
1 changed files with 9 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.buildkite/test_areas/models_language.yaml
+9 -3
View File
@@ -67,7 +67,9 @@ steps:
- tests/models/language/generation
commands:
# Install fast path packages for testing against transformers
- MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
# torch>=2.14 nightly requires C++20 for ATen headers (pytorch/pytorch#178150);
# mamba@v2.3.0 pins -std=c++17, so build from a patched checkout instead of the git URL.
- rm -rf /tmp/mamba-src && git clone --depth 1 --branch v2.3.0 https://github.com/state-spaces/mamba /tmp/mamba-src && sed -i 's/-std=c++17/-std=c++20/g' /tmp/mamba-src/setup.py && MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation /tmp/mamba-src
- CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
# Shard the hybrid language model tests that are numerically stable on Hopper.
- pytest -v -s models/language/generation -m hybrid_model -k 'not granite-4.0-tiny-preview' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
@@ -96,7 +98,9 @@ steps:
- "!vllm/distributed/kv_transfer/"
- tests/models/language/generation
commands:
- MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
# torch>=2.14 nightly requires C++20 for ATen headers (pytorch/pytorch#178150);
# mamba@v2.3.0 pins -std=c++17, so build from a patched checkout instead of the git URL.
- rm -rf /tmp/mamba-src && git clone --depth 1 --branch v2.3.0 https://github.com/state-spaces/mamba /tmp/mamba-src && sed -i 's/-std=c++17/-std=c++20/g' /tmp/mamba-src/setup.py && MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation /tmp/mamba-src
- CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
- pytest -v -s models/language/generation -m hybrid_model -k 'granite-4.0-tiny-preview'
@@ -111,7 +115,9 @@ steps:
- tests/models/language/generation
commands:
# Install fast path packages for testing against transformers
- MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
# torch>=2.14 nightly requires C++20 for ATen headers (pytorch/pytorch#178150);
# mamba@v2.3.0 pins -std=c++17, so build from a patched checkout instead of the git URL.
- rm -rf /tmp/mamba-src && git clone --depth 1 --branch v2.3.0 https://github.com/state-spaces/mamba /tmp/mamba-src && sed -i 's/-std=c++17/-std=c++20/g' /tmp/mamba-src/setup.py && MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation /tmp/mamba-src
- CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
- pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
mirror: