DeepSeek-V4.1-Flash-1017, uncensored β€” radiance container

DeepSeek-V4.1-Flash-1017 as a single .rad container for the radiance inference engine (AMD RDNA4, ROCm), with the refusal direction edited out of 18 attention-output projections.

Files deepseek-v4.1-flash-1017-uncensored.rad.part-00 … .part-05 β€” 251.07 GiB in six parts (45 GiB each, last 26.1 GiB; the Hub's per-file limit is 50 GB)
Reassemble cat deepseek-v4.1-flash-1017-uncensored.rad.part-0[0-5] > deepseek-v4.1-flash-1017-uncensored.rad (sha256 7444650de88519317b8d626559f2675e47b93969aa78502c281669534916315d)
Base DeepSeek-V4.1-Flash-1017 (763B parameters; fp8 weights, fp4 routed experts; MIT)
Abliteration attn_output_b of layers 11–28 re-quantised from the official fp8 at alpha 0.75 of huihui's abliteration strength; the other 18 of 36 layers stay stock. It removes the refusal direction without touching attention q/k/v, the MLP or the Engram tables
Quantisation per-tensor, from Unsloth's UD-IQ1_S mix: routed experts iq1_s/iq2_xxs (gate_up 34 layers iq1_s + 6 iq2_xxs; down 20 + 20), attention and shared experts q6_K/q5_K/q8_0, the Engram table iq4_nl, norms f32, the DSpark drafter's experts MXFP4 as released
Drafter the release's own DSpark mtp.* head (MXFP4 experts, 6.72 GiB)
Engram the 2 Engram layers (103 GiB table) rebuilt from the release tokenizer by data/engram_tables.py; the release ships none
Needs radiance with the DeepSeek-V4.1 architecture and TP4 kernels; four 32 GB cards (R9700 gfx1201) and about 4 GiB of pinned host pool per rank; the Engram table on a disk tier
Measured on the 4Γ—R9700 lab at --max-model-len 409600, greedy essay 50.1 / code 53.4 tok/s (n-gram 5), 16k-token time to first token 4.7 s; 0 GPU resets

Quantisation

Per-tensor mix, not one global type. The .rad directory has 32,701 tensors; sizes below are from rad-info on the 251.13 GiB container.

Component Quantiser Bits/weight Size Tensors
Token embedding (token_embd) q6_K β‰ˆ6.56 0.51 GiB 1
Engram table (per_layer_token_embd) iq4_nl β‰ˆ4.50 103.00 GiB 1
Routed experts β€” gate_up iq1_s (34 layers), iq2_xxs (6 layers) β‰ˆ2.07 / 2.22 88.25 GiB 15,360
Routed experts β€” down iq1_s (20 layers), iq2_xxs (20 layers) β‰ˆ2.08 / 2.22 45.30 GiB 15,360
Shared experts q6_K (54), q8_0 (26) β‰ˆ6.56 / 8.50 1.19 GiB 80
Attention projections β€” layers 0–26 q6_K β‰ˆ6.56 2.61 GiB 324
Attention projections β€” layers 27–39 q5_K β‰ˆ5.50 1.05 GiB 156
Attention norms / indexer / compressor (main) f32 / bf16 32 / 16 0.19 GiB 312
Norms / router / HC f32 32 0.37 GiB 240
Output (output) q6_K β‰ˆ6.56 0.51 GiB 1
Engram KV / gate q6_K / f32 β‰ˆ6.57 / 32 0.24 GiB 10
DSpark drafter β€” routed experts fp4_e2m1 + e8m0 scale (MXFP4) β‰ˆ4.25 6.72 GiB 768
DSpark drafter β€” linears / norms / attention bf16 / f32 16 / 32 1.19 GiB 86

iq1_s and iq2_xxs are imatrix-weighted GGML searches; q6_K/q5_K/q8_0/iq4_nl are the canonical GGML encodings. The DSpark experts keep the release's MXFP4 (fp4_e2m1 with e8m0 scales, 1Γ—32 block), not a recipe rule.

Serve

Reassemble first (the six parts are split at the Hub's 50 GB per-file limit):

cat deepseek-v4.1-flash-1017-uncensored.rad.part-0[0-5] > deepseek-v4.1-flash-1017-uncensored.rad
radiance --model deepseek-v4.1-flash-1017-uncensored.rad \
  --tp 4 --tp-wire exact --max-model-len 409600 --max-num-seqs 4 \
  --max-num-batched-tokens 4096 --kv-cache-dtype bf16 \
  --placement expert_tiered --host-pool-mib 4096 --ngram-placement disk \
  --num-speculative-tokens 5 --spec-ngram 5 \
  --expert-profile expert-stats.json --expert-pin-top 0.20 \
  --host 0.0.0.0 --port 8010

expert-stats.json is radiance's --expert-stats output for this model (per-layer routed-expert usage); --expert-pin-top 0.20 keeps the hottest fifth resident while the rest streams. Without it, the same recipe runs with the expert plane on the host pool alone. The Engram table is read from the disk tier by --ngram-placement disk; it is 103 GiB, so keep it on NVMe.

The server speaks OpenAI's chat API (/v1/chat/completions, /v1/models); the served model name is DeepSeek-V4.1-Flash. The chat template is embedded in the container.

How this file was made

data/engram_tables.py DeepSeek-V4.1-Flash-1017 engram-tables.safetensors
rad-convert DeepSeek-V4.1-Flash-1017 \
  --add engram-tables.safetensors \
  --recipe data/recipes/deepseek41-ud-iq1s.recipe \
  --imatrix imatrix_unsloth.gguf_file \
  -o deepseek-v4.1-flash-1017-uncensored.rad

The abliteration is applied after conversion: for layers 11–28, attn_output_b is rebuilt as W + alpha * (huihui's W - W), then re-quantised to the container's q6_K; alpha 0.75 gives the abl075 file. The recipe (deepseek41-ud-iq1s.recipe in radiance) names every weight's quantiser; first match wins, and anything it does not name keeps the checkpoint's own dtype. iq1_s and iq2_xxs need the imatrix; Unsloth's UD-IQ1_S imatrix is used, with ffn_gate_exps/ffn_up_exps fused by data/imatrix_merge.py because radiance fuses gate_up.

License

DeepSeek-V4.1-Flash-1017 is MIT licensed (Copyright Β© DeepSeek). This container is a modified copy: the routed experts and linears are re-encoded as the quantisers above, the Engram table is generated from the tokenizer, and 18 attention-output projections carry the abliteration edit. MIT applies to the original and to these modifications.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Sunbert/DeepSeek-V4.1-Flash-1017-uncensored-rad

Finetuned
(93)
this model