| 1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556 |
- # versions.env - authoritative build inputs for GLM-5.2 NF3-Hybrid
- # Every entry is a tag or exact version - no :latest, no main, no ranges.
- CUDA_BASE_IMAGE=nvidia/cuda:13.2.0-devel-ubuntu24.04
- CUDA_BASE_DIGEST=sha256:a5b6256e470196fc1d5f8f62139d57d3662867746dfe1cb352d7652024047020
- GB10_BUILD=1
- # 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
- # CUDA-side companion packages
- NVSHMEM_VERSION=3.4.5
- TVM_FFI_VERSION=0.1.9
- TILELANG_VERSION=0.1.9
- NUMBA_VERSION=0.65.0
- # NCCL
- NCCL_REPO=https://github.com/NVIDIA/nccl.git
- NCCL_REF=v2.30.7-1
- # vLLM (local-inference-lab fork with b12x, DCP, MTP support)
- VLLM_REPO=https://github.com/local-inference-lab/vllm.git
- VLLM_REF=codex/dcp-globaltopk-sharddraft-defaults-20260622
- VLLM_COMMIT=e232d262369b8c918cf478a7a96a0fcf8127cf65
- # b12x (NVFP4 kernel library)
- B12X_REPO=https://github.com/voipmonitor/b12x.git
- B12X_REF=f3686b55
- B12X_COMMIT=f3686b555d639823b276c2080f173145eed7f007
- # 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
- # Runtime lock seeds
- RAY_VERSION=2.56.0
- FASTSAFETENSORS_VERSION=0.2.2
- INSTANTTENSOR_VERSION=0.1.8
- BITSANDBYTES_VERSION=0.49.2
- ACCELERATE_VERSION=1.14.0
- # Target GPU architecture (DGX Spark = sm_121a)
- TORCH_CUDA_ARCH_LIST=12.1a
- FLASHINFER_CUDA_ARCH_LIST=12.1a
|
