sudoingX/dgx-spark-deepseek-v41 v1.2

v1.2: DeepSeek-V4.1-Flash-1017

on GitHub

DeepSeek's Oct 17 update, deepseek-ai/DeepSeek-V4.1-Flash-1017, keeps the launch checkpoint's config, tokenizer and tensor layout and changes the content of every shard. v1.2 re-runs the whole pipeline on it and makes the 1017 build the default. The engine and the serving image are the same as v1.1.

Weights: sudoingx/DeepSeek-V4.1-Flash-1017-EXL3-2.7bpw, 196 GB, plus 1017's own two Engram shards from DeepSeek's repo (the launch checkpoint's shards differ). The launch build stays at sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw, served by the same engine.

Quality against native 1017, on the same panel and reference implementation as the launch build:

• KLD 0.0243, top-1 agreement 95.47%. LibertAIDAI's 1017 EXL3 2.9 bpw pack scores 0.0304 / 94.76% there.
• HumanEval+ 96.3 / 90.2, MBPP+ 89.7 / 76.2, IFEval prompt strict 89.1, each within noise of native 1017 (paired, p >= 0.14).
• On 300 IFEval prompts with native replies, LibertAIDAI's pack is closer to native: KLD 0.1135 against 0.1199.

Speed, kit_bench.py set b, single stream, DSpark on, median of 9 reps a prompt:

• 100.7 / 59.2 / 141.3 tok/s on code / prose / structured; the launch build read 103.0 / 58.0 / 142.6 tok/s in the same session.
• Against TensorFold 8e0966e2 with the object-oriented-architects 1017 EXL3 2.7 bpw pack, on the same two Sparks with the same client, two boots each: code +2.0%, prose +0.9%, structured level.

Switching from v1.1: download the 1017 weights and 1017's Engram shards (quickstart step 3), then run up with MODEL_DIR on the new folder and DOCKER_ENV="VLLM_DSPARK_ROUND_MS=28.99,32.76,36.11,39.34,42.43,45.06", the depth table measured on 1017. No image rebuild.

Also in v1.2:

• data/: the 1017 plan, statistics, KLD, eval, set b and gate files, the same-box run, and two heavier dense / head compositions on the same experts (KLD 0.0227 and 0.0219, 2.4-7.8% slower than this build in one session; data/speed/variants_1017).
• FINDINGS 22 (memory fragmentation after days of uptime) and 23 (sensitivity across checkpoints).
• guards.sh is executable (#1, reported by @eduffd).

Not run on the 1017 build yet: the 256K needle rung (needles pass at 8K, 32K, 64K and 128K; the launch build passed 256K). The 2-hour agent soak has not run on either build.