[annotations]
"aipcc.component" = "Accelerator Enablement"
"aipcc.accelerator-enablement.libtorch" = "true"
"aipcc.accelerator-enablement.cuda" = "true"
"aipcc.accelerator-enablement.gaudi" = "true"
"aipcc.accelerator-enablement.rocm" = "true"
"aipcc.accelerator-enablement.spyre" = "true"
"aipcc.accelerator-enablement.tpu" = "true"

[changelog]
"0.5.2.2" = ["2024-09-12: replace patches with project_override", "2024-09-14: verify PyTorch build before building vLLM"]
"0.5.2.3" = ["2024-09-12: replace patches with project_override", "2024-09-13: Fix vllm-0.5.2.3 patches.", "2024-09-14: verify PyTorch build before building vLLM", "2024-09-27: Add ROCm triton dependency patch"]
"0.5.2.4" = ["2024-10-12: Add new version for CUDA and ROCm variants", "2024-10-12: fix cutlass include and force cuDNN and cuSPARSELt"]
"0.5.3.post1" = ["2024-09-13: remove dependency on Ray from HPU", "2024-09-14: verify PyTorch build before building vLLM", "2024-10-28: build 1.18.0 fork", "2024-11-06: Fix the missing plugin args", "2025-01-14: Remove this version"]
"0.6.2" = ["2024-10-18: forward port relevant patches from 0.5.2.4", "2024-11-07: Update patches for ROCm build on 0.6.2 from the ROCm fork", "2024-11-08: Move ROCm patches for variant specific fork changes", "2024-11-08: Removed 0.5.2.4 support", "2024-11-21: Made rpdTracer support optional for ROCm", "2024-11-26: Add amdsmi, Triton and flash_attn as dependencies for ROCm", "2024-11-27: Add GPU_ARCHS for ROCm vLLM build", "2024-12-02: rebuild for ROCm gfx908;gfx942", "2024-12-04: Move ROCm build to the same code from the fork but hosed on the ODH repo", "2025-01-06: Remove temporary workaround for ROCm ODH repo", "2025-01-20: Remove CUDA variant patches"]
"0.6.4.post1" = ["2024-12-09: forward port relevant patches from 0.6.2"]
"0.6.4.post2" = ["2025-01-13: forward port relevant patches from 0.6.4.post1", "2025-01-14: Gaudi - add patch to enable OMP for multiprocessing on single node", "2025-02-07: Add patches for RHELAI-3303 / CVE-2025-24357", "2025-02-11: Fix 0002-Remove-ray-dependency.patch after Intel emergency patch and update upstream pointer"]
"0.6.6.post1" = ["2024-01-08: forward port relevant patches from 0.6.4.post1", "2025-01-30: Updated to fromager 0.37.0, update resolve_specifier()", "2025-02-05: Use DOWNLOAD_RELEASE instead of REPO", "2025-02-06: Cleanup resolve_specifier()", "2025-02-06: Move 0.6.2, 0.6.4 and 0.6.6 variants to apply fix for CVE-2025-24357. Change relevant for vLLM < 0.7", "2025-03-02: Move Gaudi to 0.6.6.post1 from fork", "2025-03-13: Use OpenDataHub Gaudi fork rather than upstream", "2025-03-24: Add in patch to use updated patch version of vllm-hpu-extension"]
"0.7.2" = ["2025-02-07: forward port relevant patches from 0.6.6.post1", "2025-02-20: build with correct version, AIPCC-432", "2025-03-04: Update CUDA execution provider to use Neural Magic repository", "2025-03-06: remove cmake dependency removal patch", "2025-03-06: use the dynamically generated PKG-INFO for vLLM", "2025-03-06: remove CUDA compute capabilities for Blackwell (10.0)", "2025-03-06: downgrade torch to 2.5.1", "2025-03-26: Pull in API changes from 0.7.3 for CUDA provider"]
"0.8.3" = ["2025-04-15: forward port patches to 0.8.3 and add support for ROCm", "2025-04-16: Remove support for 0.6.2"]
"0.8.4" = ["2025-04-22: forward port patches to 0.8.4 and add support for ROCm builds", "2025-04-24: use single NM_VERSION_MAP for CPU, CUDA, and ROCm builds", "2025-04-25: fix missing vllm.benchmark subpackage", "2025-04-25: build with numpy 2.x", "2025-05-06: Update to vLLM 0.8.4.2_downstream for CUDA and ROCm", "2025-05-07: build for cpu-ubi9", "2025-05-21: Update to use downstream mirror for vllm-gaudi"]
"0.8.5" = ["2025-05-22: forward port patches to 0.8.5 and add support for ROCm and CUDA builds", "2025-05-22: Move cutlass clone into plugin", "2025-06-01: limit architectures for vLLM FA with VLLM_FA_CMAKE_GPU_ARCHES and define it for vLLM builds", "2025-06-02: Move clones into plugin"]
"0.9.0" = ["2025-06-09: forward port patches to 0.9.0"]
"0.9.0.1" = ["2025-06-19: update vLLM to midstream's v0.9.0.1.0", "2025-06-23: updated map to include v0.9.0.1 for cpu-ubi9", "2025-06-25: remove quotes from VLLM_FA_CMAKE_GPU_ARCHES", "2025-06-25: Only build CUTLASS MoE kernels on Hopper", "2025-06-26: Remove INCLUDES and PUNICA kernel vars", "2025-06-26: Add external sources to sdist", "2025-06-26: Add 100-real and 120-real to VLLM_FA_CMAKE_GPU_ARCHES", "2025-06-26: Update to midstream-cuda-v0.9.0.1.1", "2025-06-26: Update to midstream-cuda-v0.9.0.1.3", "2025-07-02: Update to midstream-cuda-v0.9.0.1.4, including hopper llama4 scout fix", "2025-07-03: mark package as exclusive build", "2025-07-04: remove VLLM_FA_CMAKE_GPU_ARCHES that is not needed anymore"]
"0.9.1" = ["2025-07-10: update vLLM to midstream's v0.9.1.1"]
"0.9.2" = ["2025-07-11: update vLLM to midstream's v0.9.2.0", "2025-07-14: update vLLM to midstream's v0.9.2.1", "2025-07-16: update vLLM to midstream's v0.9.2.2", "2025-07-24: updated package_plugins/vllm.py for better midstream version handling, AIPCC-3173"]
"0.10.0" = ["2025-07-28: add support for spyre", "2025-08-04: add support for TPU", "2025-08-07: update CUDA vLLM to midstream's v0.10.0.0", "2025-08-07: Add filelock as a build dependency for vLLM", "2025-08-11: TORCH_CUDA_ARCH_LIST change, compress fatbin", "2025-08-11: Remove the patch that installs Triton", "2025-08-11: Updated vLLM package plugin to use 'rhai' tagging instead of 'midstream' for GitLab tags", "2025-08-13: Resolver provider fix, AIPCC-4340", "2025-08-13: update ROCm vLLM to midstream's v0.10.0.1", "2025-08-13: Remove version munging and non-cpu custom patches", "2025-08-13: Fix torch_xla dependency for TPU", "2025-08-14: removed ray patches, added ray[cgraph] and building cupy as cupy-cuda12x", "2025-08-27: Set VLLM_TARGET_DEVICE to empty for spyre", "2025-09-12: Remove ray dependencies from ROCm build requirements", "2025-10-02: Add support for pre-release versions in the vLLM package plugin"]
"0.10.2" = ["2025-11-20: update vLLM to use upstream for Spyre only, because it's not mirrored internally"]
"0.11.0rc4" = ["2025-10-03: update vLLM to midstream v0.11.0rc4 on CUDA", "2025-10-03: Remove TPU dependency fixes for versions >= 0.11.0", "2025-10-03: Remove torch_xla dependency fix for versions >= 0.11.0", "2025-10-03: update vLLM to midstream v0.11.0rc4 on TPU", "2025-10-06: update vLLM to midstream v0.11.0rc4 on ROCm"]
"0.11.0" = ["2025-10-07: update vLLM to include tensorizer on TPU and update TPU dependency fixes", "2025-10-28: update cutlass repo URL to use GitLab mirror", "2025-11-20: update vLLM to use midstream for Spyre", "2025-12-10: add neuron-ubi9 variant with CPU build settings (VLLM_CPU_DISABLE_AVX512)", "2026-02-10: force rebuild for vllm plugin changes to support midstream tags and switch to the midstream tag provider for the neuron variant", "2026-02-13: force rebuild of vllm plugin to remove unnecessary github resolver for old version of vllm"]
"0.11.1rc6" = ["2025-11-10:  update vLLM to midstream v0.11.1rc6 on CUDA", "2025-11-12: modify _fix_cuda_xformers_dependency to unconstrain xformers", "2025-11-25: PCC-ADR114 – apply tagging pattern for midstream mirrors"]
"0.13.0rc1" = ["2026-01-06: Replace fastsafetensors git URL with PyPI package for ROCm builds", "2026-01-06: Clone triton_kernels during source preparation for CUDA and ROCm builds"]
"0.13.0" = ["2026-01-13: update vLLM to midstream v0.13.0 on tpu and update TPU dependency fixes", "2026-01-22: support for building vLLM cpu variants"]
"0.14.0" = ["2026-02-03: add support for vLLM 0.14.0 on ROCm"]
"0.14.1" = ["2026-02-27: fix Spyre midstream tag provider", "2026-03-04: fix CPU midstream tag provider"]
"0.16.0" = ["2026-02-27: Add Gaudi support with VLLM_TARGET_DEVICE=empty (vllm-gaudi plugin handles HPU)", "2026-02-27: Update vllm plugin to use default resolver for Gaudi (vllm-gaudi is now a separate package)", "2026-03-11: fix midstream matcher for TPU and Spyre", "2026-03-18: use new avx build flags"]
"0.17.1" = ["2026-04-07: Add gaudi support to midstream tagging mechanism"]
"0.18.0" = ["2026-03-19: Add vLLM 0.18.0 to torch-2.10.0/rocm7.1-ubi9 collection for RHAI 3.4", "2026-03-19: Removed amdsmi from ROCm collections as it is unused", "2026-04-03: Add vLLM 0.18.0 to accelerated/spyre-ubi9 collection", "2026-04-08: enable gcc 14 for x86 cpu", "2026-04-30: Remove per-package gcc-toolset-14 opt-in, handle missing oneDNN/ACL gracefully (AIPCC-10154, AIPCC-15146)", "2026-05-08: remove stale and incorrect env vars across variants", "2026-05-28: Add DGX Spark (sm_121) CUDA arch support"]
"0.21.0" = ["2026-05-27: Add vLLM 0.21.0 to torch-2.10.0/tpu-ubi9 collection", "2026-05-27: Update Gaudi to vLLM 0.21.0 with PyTorch 2.11.0", "2026-05-29: Pre-clone DeepGEMM during source preparation for CUDA builds (AIPCC-16435)", "2026-06-02: Fix atomicAdd ambiguity with ROCm 7.14 by removing USE_ROCM guards from GPTQ compat.cuh AIPCC-15160", "2026-06-22: Unconstrain tpu-inference pin in TPU requirements to use version from constraints"]
"0.24.0" = ["2026-07-09: Update vLLM constraint for ROCm 7.14 (torch-2.11.0)", "2026-07-13: Update Gaudi to vLLM 0.24.0 with Gaudi 1.24.1", "2026-07-14: Update plugin for vLLM 0.24 cmake patterns and fmha_sm100 support (AIPCC-19337)", "2026-07-28: Fix FA4 cute import rewrite via FETCHCONTENT_SOURCE_DIR for vllm-flash-attn (INFERENG-9370)"]

[env]
CMAKE_BUILD_TYPE = "Release"

[download_source]
destination_filename = "vllm-${version}.tar.gz"

[resolver_dist]
min_release_age = 0

[build_options]
memory_per_job_gb = 8.0
exclusive_build = true

[project_override]
update_build_requires = ["torch", "filelock"]
remove_build_requires = ["cmake"]

[variants.cpu-ubi9.env]
VLLM_TARGET_DEVICE = "cpu"

[variants.neuron-ubi9.env]
VLLM_TARGET_DEVICE = "empty"
VLLM_CPU_DISABLE_AVX512 = "true"

[variants.cuda-ubi9.env]
NVCC_THREADS = "2"
VLLM_TARGET_DEVICE = "cuda"

[variants.gaudi-ubi9.env]
VLLM_TARGET_DEVICE = "empty"

[variants.rocm-ubi9.env]
VLLM_TARGET_DEVICE = "rocm"

[variants.spyre-ubi9.env]
VLLM_TARGET_DEVICE = "empty"

[variants.tpu-ubi9.env]
VLLM_TARGET_DEVICE = "tpu"
