| 12345678910111213141516171819202122232425262728293031323334353637383940414243 |
- # versions.env - authoritative build inputs for Solar Open2 250B NVFP4 (Nota)
- # Every entry is a tag or exact version - no :latest, no main, no ranges.
- #
- # Stack mirrors glm52-nf3-hybrid (cu130 / aarch64 / sm_121a) minus the b12x
- # suite: Solar is a stock vLLM serve using the UpstageAI solar_open2 fork
- # (solar_open2 reasoning/tool parsers + SolarOpen2TemplateLogitsProcessor).
- # NVFP4 (compressed-tensors) is auto-detected from the model config.
- #
- # locks/ and checksums/ are shared with glm52-nf3-hybrid (same base stack).
- CUDA_BASE_IMAGE=nvidia/cuda:13.2.0-devel-ubuntu24.04
- CUDA_BASE_DIGEST=sha256:a5b6256e470196fc1d5f8f62139d57d3662867746dfe1cb352d7652024047020
- # Python / uv
- UV_VERSION=0.11.28
- # PyTorch stack (cu130 index works on CUDA 13.2 for aarch64)
- TORCH_VERSION=2.11.0
- TORCHVISION_VERSION=0.26.0
- TORCHAUDIO_VERSION=2.11.0
- TRITON_VERSION=3.6.0
- PYTORCH_INDEX_URL=https://download.pytorch.org/whl/cu130
- PYPI_INDEX_URL=https://pypi.org/simple
- # NCCL
- NCCL_REPO=https://github.com/NVIDIA/nccl.git
- NCCL_REF=v2.30.7-1
- # vLLM (UpstageAI fork — solar_open2 parsers + logits processor)
- VLLM_REPO=https://github.com/UpstageAI/vllm.git
- VLLM_REF=v0.22.0-solar-open2
- VLLM_COMMIT=00907fc9b982f973808a0193ad0f8c61fbb62854
- # FlashInfer
- FLASHINFER_REPO=https://github.com/flashinfer-ai/flashinfer.git
- FLASHINFER_REF=v0.6.12
- # cutlass-dsl (4.5.3 for firmware 580.159.03 compat)
- CUTLASS_DSL_VERSION=4.5.3
- # Target GPU architecture (DGX Spark = sm_121a)
- TORCH_CUDA_ARCH_LIST=12.1a
- FLASHINFER_CUDA_ARCH_LIST=12.1a
|
