UPSTREAM.md 1.4 KB
Upstream provenance — DeepSeek-V4-Flash-Vision-Exp recipe
Day-0 port of the multimodal lane:
- repo: https://github.com/sfxnz/DeepSeek-V4-Flash-Vision-Exp-vLLM-2x-DGX-Spark
- commit:
7a958128b8(2026-08-31) - license: repo LICENSE (see tree); model deepseek-ai/DeepSeek-V4-Flash-Vision-Exp is MIT;
model revision served:
86f746b361
| local file | upstream path |
|---|---|
| Dockerfile | docker/Dockerfile.b12x-vision (FROM genericized to ARG BASE_IMAGE; original pins eugr/spark-vllm-b12x:latest, digest sha256:7dc02f162929943ba2e14514066ed2a04bb7e9ed3592d4eb460ebcbb1f8376bd) |
| patch_dspark_skip_bias_vl.py | docker/patch_dspark_skip_bias_vl.py (verbatim) |
| plugin/dsv4_vision/* | docker/plugin/dsv4_vision/* (verbatim; registered as VLLM_PLUGINS=dsv4_vision via the vllm.general_plugins entry point) |
| build.sh, versions.env | ours (generic driver; BASE_IMAGE/TAG overridable for the eugr-vs-aidendle94 fork comparison) |
Known upstream limitations (their README): image-token expert routing in the
MoE router is not wired for the ViT lane yet (ffn.gate.bias_vl gate
tensors are skipped by the loader patch); text decode unaffected.
VLLM_USE_B12X_MHC must stay OFF (this checkpoint's rms_norm_eps is 1e-20;
the fused Gram mHC kernel only accepts 1e-6).