recipes/docker-images/mistral-large4-vllm/Dockerfile

# syntax=docker/dockerfile:1
# Mistral Large 4 1T A52B NVFP4 — vLLM day-0 image (thin FROM wrapper).
#
# The official vllm/vllm-openai:mistral-large-4 carries vLLM's day-0 Mistral
# Large 4 support: DeepSeek-style sparse MLA + DSA top-k indexer, cross-layer
# KV sharing, NoPE layers, the FP8 latent KV cache (fp8_ds_mla) and the
# built-in DSpark speculative drafter. It also ships mistral-common 1.13
# (required to load the checkpoint's tekken.json tokenizer).
#
# The multi-arch manifest covers arm64 — the digest pin below IS the arm64
# variant, so the build on the aarch64 nodes resolves to the right platform
# without a manifest lookup. GB10 (sm_121a) is not in the vendor-verified
# matrix (B200/B300/GB200/GB300/Vera Rubin): first boot on the cluster is the
# real test — flashinfer JIT compiles any missing SM121 kernels at load time.
#
# No layers of our own: the wrapper exists so the launcher can stamp its
# sm121.checksum label and sync the image node-to-node like every other lane.
ARG BASE_IMAGE
FROM ${BASE_IMAGE}