dsv4-iq2xxs-expert-major

DeepSeek V4 Flash IQ2_XXS SSD Sidecar for ds4-ssd

This repository contains a prebuilt SSD-streaming sidecar package for Anemll/ds4-ssd, an alpha fork of antirez's DwarfStar 4 (ds4) runtime for DeepSeek V4 Flash.

This is not a standalone Transformers model. It is meant to be used with ds4-ssd sidecar mode: dense tensors are stored in dense/model-dense.gguf, while routed-MoE expert tensors are stored in layer-major sidecar files and streamed from SSD through DS4's slot-bank cache.

Contents

dsv4-iq2xxs-expert-major/
  manifest.json
  dense/
    model-dense.gguf
    flashmoe-package.json
  layer_000.bin
  ...
  layer_042.bin

• Sidecar layout: layer_major_expert
• Architecture: deepseek4
• Expert count: 256
• Active routed experts per token: 6
• Expert quantization: IQ2_XXS / Q2_K as recorded in manifest.json
• Dense model: GGUF under dense/model-dense.gguf

Download

hf download anemll/dsv4-iq2xxs-expert-major \
  --local-dir /path/to/dsv4-iq2xxs-expert-major

The package is large, so place it on a fast local SSD.

Run with ds4-ssd

git clone https://github.com/Anemll/ds4-ssd
cd ds4-ssd
make

export DS4_SIDECAR_DIR=/path/to/dsv4-iq2xxs-expert-major

./ds4 \
  -m "$DS4_SIDECAR_DIR/dense/model-dense.gguf" \
  --moe-sidecar "$DS4_SIDECAR_DIR" \
  --moe-mode slot-bank \
  --moe-slot-bank 8 \
  --ctx 9000 \
  -p "Hello"

--ctx is the KV window. DS4_METAL_PREFILL_CHUNK= is the prefill chunk cap used by the alpha sidecar path. Leave DS4_METAL_GRAPH_RAW_CAP unset so DS4 can auto-size the raw KV graph cap.

To verify that SSD streaming is active, startup logs should include:

applied sidecar tuning profile
Flash-MoE sidecar loaded
Flash-MoE slot banks allocated

If you pass only a full resident GGUF with -m, DS4 is not in SSD-streaming mode. Sidecar mode requires both -m "$DS4_SIDECAR_DIR/dense/model-dense.gguf" and --moe-sidecar "$DS4_SIDECAR_DIR".

Smoke Test

From the ds4-ssd checkout:

DS4_SIDECAR_DIR=/path/to/dsv4-iq2xxs-expert-major make sidecar-smoke

The alpha smoke uses a 16K prefill prompt and one deterministic generated token.

Provenance

This sidecar was created from the DeepSeek V4 Flash GGUF model distributed by antirez/deepseek-v4-gguf for use with the ds4-ssd runtime. Use this artifact according to the upstream model and GGUF distribution terms.

For runtime documentation, build instructions, and release notes, see Anemll/ds4-ssd.