sleepyeldrazi/deepseek-v4-flash-reap-k128-Q2-GGUF

DeepSeek V4 Flash — REAP K128 (Uniform)

REAP-pruned DeepSeek V4 Flash at K128 (128 routed experts per MoE layer). Prunes 50% of routed experts via Cerebras REAP (Router-weighted Expert Activation Pruning), preserving all attention, embeddings, shared experts, router, and MTP components.

Drop-in compatible with the standard ds4-engine runtime — uniform IQ2XXS/Q2K expert quantization throughout all layers. No per-layer quant dispatch required.

At a Glance

This is the recommended variant for most users. Uniform expert quantization means it works out of the box with any ds4-engine build. The mixed-precision variant (52 GiB) has Q4_K in layers 37-42 but requires runtime per-layer quant dispatch.

Domain Split (Calibration)

8,000 prompts · 5.0M tokens · 1.3B routed expert observations

Calibration used the REAP activation_energy_sum2 score metric with 4,096 token context per prompt. Top-to-bottom expert score gap in layer 3: 2,200x (strong pruning signal).

How to Run

Requires eouya2/ds4-for-reaped (ds4 engine with compact GGUF support):

git clone https://github.com/eouya2/ds4-for-reaped
cd ds4-for-reaped
make cuda-spark -j$(nproc)  # DGX Spark / CUDA
# or: make                    # Metal / macOS

./ds4 --cuda -m DeepSeek-V4-Flash-REAP-K128-uniform.gguf --ctx 131072

API server mode:

./ds4-server --cuda -m DeepSeek-V4-Flash-REAP-K128-uniform.gguf \
  --host 0.0.0.0 --port 17777 --ctx 131072

How It Was Built

1. 1.Donor GGUF: Downloaded antirez IQ2XXS-w2Q2K-AProjQ8 variant (80.8 GiB) — uniform IQ2XXS/Q2K experts throughout
2. 2.Calibration: 8,000 prompts collected and run through ds4's imatrix collector on a DGX Spark (NVIDIA GB10) at 4,096 token context
3. 3.REAP scoring: Imatrix activation data converted to per-expert REAP scores using activation_energy_sum2 (same calibration used for the mixed-precision variant — REAP scores are quantization-independent)
4. 4.Pruning: 50% expert removal via ds4_prune_gguf.py from eouya2/reap-for-ds4. Layers 0-2 (hash-routed) preserved. Expert tensors copied byte-for-byte — no dequant/requant.
5. 5.Output: ds4-compact-v1 GGUF

No fine-tuning. Purely structural expert removal. Weights are unmodified — a subset of the original experts.

Comparison with Mixed-Precision Variant

Acknowledgments

• —DeepSeek — DeepSeek V4 Flash base model
• —antirez — ds4 engine and GGUF quants
• —eouya2 — ds4-for-reaped and reap-for-ds4
• —Cerebras Research — REAP (code)
• —NVIDIA — DGX Spark hardware