| 12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970 |
- # versions.env - authoritative build inputs for DeepSeek-V4-Flash DSpark (from source)
- # EXPERIMENTAL: builds vLLM + b12x from the local-inference-lab "eldritch" fork at the
- # newest DS4/DSpark branch, for DGX Spark (aarch64 / sm_121a). Every entry pinned.
- #
- # Reference stack inputs come from the fork's own tools/spark/versions.env (the
- # authoritative pairing for the pinned branch) plus the GLM-5.2 NF3-Hybrid harness
- # this directory is derived from (CUDA base, apt locks, checksums).
- # --- CUDA base ---
- 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) ---
- # NOTE: the newest DS4 fork line moved to torch 2.12.0. The python-build.txt lock
- # MUST be regenerated against these pins (see gen-locks.sh) before building.
- TORCH_VERSION=2.12.0
- TORCHVISION_VERSION=0.27.0
- TORCHAUDIO_VERSION=2.12.0
- TRITON_VERSION=3.7.0
- 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=0.1.11
- TILELANG_VERSION=0.1.9
- NUMBA_VERSION=0.65.0
- # --- NCCL (pinned from tools/spark/versions.env) ---
- NCCL_REPO=https://github.com/NVIDIA/nccl.git
- NCCL_REF=5067397c2676d5aed50042fc39e5c8ee96eb0027
- # --- vLLM (local-inference-lab b12x fork, newest DS4/DSpark line) ---
- VLLM_REPO=https://github.com/local-inference-lab/vllm.git
- VLLM_REF=feat/gg-ds4-0731-launcher-20260731
- VLLM_COMMIT=e63a190ee1b48d051ccebce485de343dbe2b3505
- # --- b12x (NVFP4 kernel library). Pin matches the fork's eldritch-fullstack line;
- # recheck against the branch's tools/spark/versions.env if it changes. ---
- B12X_REPO=https://github.com/voipmonitor/b12x.git
- B12X_REF=284a2eae83754ee1abd31c37b9ca66b68e20b8a8
- B12X_COMMIT=284a2eae83754ee1abd31c37b9ca66b68e20b8a8
- # --- FlashInfer (pinned from tools/spark/versions.env) ---
- FLASHINFER_REPO=https://github.com/flashinfer-ai/flashinfer.git
- FLASHINFER_REF=ce40d25a4cd05b2f28c935b99c9ea5091b4e64d2
- FLASHINFER_VERSION=0.6.15
- # --- DeepGEMM (informational; consumed as a b12x/flashinfer build dep) ---
- DEEPGEMM_REF=a6b593d2826719dcf4892609af7b84ee23aaf32a
- # --- cutlass-dsl (4.6.0 — matches the fork's requirements/cuda.txt pin) ---
- CUTLASS_DSL_VERSION=4.6.0
- # --- Runtime lock seeds ---
- RAY_VERSION=2.56.0
- FASTSAFETENSORS_VERSION=0.3.2
- INSTANTTENSOR_VERSION=0.1.9
- 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
- CUTE_DSL_ARCH=sm_121a
|
