# syntax=docker/dockerfile:1.7
# DeepSeek-V4-Flash-DSpark — NCCL ABI fix + instanttensor on top of
# aidendle94/sparkrun-vllm-ds4-gb10:production-3.75
#
# The base image ships nvidia-nccl-cu13 2.29.7 but the rest of the
# CUDA 12.1 stack expects the 2.30.4 ABI.  This mismatch breaks NCCL
# collectives (AllGather etc.).  Re-align and add instanttensor for
# faster model loading (load-format instanttensor).
#
# Build: cd docker-images/deepseek-v4-flash-dspark && docker build -t sparkrun-vllm-ds4-gb10:fixed .
# Tag:   docker tag sparkrun-vllm-ds4-gb10:fixed aidendle94/sparkrun-vllm-ds4-gb10:production-3.75

FROM aidendle94/sparkrun-vllm-ds4-gb10:production-3.75

RUN pip install --no-cache-dir --no-deps --force-reinstall nvidia-nccl-cu13==2.30.4 \
 && pip install --no-cache-dir --no-deps --force-reinstall instanttensor==0.1.9