DeepSeek-V4-Flash-moe-w2-planes

DeepSeek-V4-Flash — prepacked 2-bit MoE expert planes (vLLM-Moet / DGX Spark)

The routed experts of deepseek-ai/DeepSeek-V4-Flash (159B MoE), converted once, offline to the fragment-major 2-bit format consumed by vLLM-Moet's hand-written SM120/SM121 SASS kernels (sign-symmetric {-4,-1,1,4} codebook + UE8M0 block-32 scales — the author's sweep-validated scheme, produced with his exact conversion functions).

Why this exists: on unified-memory machines (NVIDIA DGX Spark / GB10), converting at load time OOMs the box. With these planes, the Sapid-Labs/vLLM-Moet Spark port serves this 159B model on one DGX Spark (~21 tok/s single-stream) — see spark/RUNBOOK.md.

These planes are not a standalone model. They replace only the routed experts; the dense stack/attention/embeddings load from the original checkpoint, which you also need.

Use

hf download deepseek-ai/DeepSeek-V4-Flash --local-dir ~/models/hf/DeepSeek-V4-Flash
hf download sapidlabs/DeepSeek-V4-Flash-moe-w2-planes \
  --local-dir ~/models/hf/DeepSeek-V4-Flash/moe_w2_planes
# then follow the runbook; the serve scripts set VLLM_MOE_W2_PREPACKED_DIR

Per layer: layer_NNN.{planes13,sc13,planes2,sc2}.npy + layer_NNN.meta.json (43 layers, 73 GiB total). Reproducible from the base checkpoint with spark/prepack_planes.py.

Credits

• Base model: DeepSeek (MIT)
• Quantization scheme, kernels, vLLM patch: kacper-daftcode/vLLM-Moet
• DGX Spark port + prepack: Sapid-Labs

Quality note: this is the bare 2-bit tier (no FP4 delta/gate) — upstream measured 89% next-token agreement with the FP4 reference and MTP acceptance at/above the FP4 baseline. Independent evals to follow at howtospark.com.