license: mit
tags:
- engram
- uncensored
- abliterated
- deepseek-v4-1-flash
- dgx-spark
- sglang
DeepSeek-V4.1-Flash UNCENSORED — Engram tables, pre-packed for 4× DGX Spark (TP=4)
What is this? DeepSeek-V4.1-Flash carries two giant lookup tables the architecture calls
Engram (at layers 1 and 14): ~189 GiB of n-gram-conditioned embeddings the model reads with
a hash lookup on every token, in both prefill and decode. They are the reason the checkpoint is
510 GB but only ~305 GiB is "weights" — the tables are data, not matrix weights: never quantized,
never sharded by tensor shape, never computed on. On a DGX Spark that distinction is life-or-death:
host RAM is the GPU's memory (121.7 GiB unified per node). Park 189 GiB of tables in GPU memory
and nothing fits; put them on NVMe and read rows directly (the kit's Engram row-store, withOFFLOAD_MODE=nvme), and a 4-node Spark fleet serves the full frontier model with room for a
1M-token context on top.
These are those tables, already packed into each rank's NVMe-resident binary shards — the
output of Mia's pack_engram.py (the slow, all-read step you'd otherwise run once per node).
Download → drop in → boot.
Packed from the abliterated checkpointdealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8
with Mia AI Lab's 4×-Spark kit
(SGLang, TP=4, EP=4, RoCE rail). Engram content follows the weights: do not use these shards
with the standard deepseek-ai checkpoint — let the kit pack its own (./start.sh pack).
Who can use this
- 4× DGX Spark, TP=4, Mia's kit, the dealignai UNCENSORED-FP8 checkpoint: download, drop in,
boot. Skips the ~7–10 min/node pack step. Verify withsha256sum -c SHA256SUMS. - TP=3 or other engines: re-pack from the checkpoint with
pack_engram.py --tp Ninstead —
the shard layout partitions table rows by rank (each file = one rank's contiguous row range).
Provenance & gates
Packed 2026-09-15 from kit commit 79f656a. The world these shards came from passed our full
gate suite before serving: DSML tool-corruption repro 3/3, 30/30 structured outputs across
temp 0/0.7/1.0, tool round-trip 3/3, reasoning mode engages, context needle at 32k/400k.
Full protocol, numbers, and the operational fixes the recipe didn't mention are in the companion
repo: github.com/neko-legends/spark-bench —
section DeepSeek V4.1 Flash — SGLang.
Files
| file | layer | rank | size |
|---|---|---|---|
engram-l1-r{0..3}of4.bin |
1 | 0–3 | 23.6 GiB each |
engram-l14-r{0..3}of4.bin |
14 | 0–3 | 23.6 GiB each |
Total 192 GB. Each node needs only its own rank's two files (~48 GB).
Install
# on each node
mkdir -p ~/dsv41-engram-unc # ENGRAM_DIR (head) / $WORKER_ENGRAM_DIR (workers)
huggingface-cli download neko-legends/DeepSeek-V4.1-Flash-uncensored-engram-4x-spark \
--include "engram-*-r${RANK}of4.bin" --local-dir ~/dsv41-engram-unc
sha256sum -c SHA256SUMS --ignore-missing
# kit knobs: ENGRAM_DIR / WORKER_ENGRAM_DIR point there, DSV41_PACKED_DIR=/engram, SKIP_PREPARE=1
Exact hashes, sizes, and layout in manifest.json.