license: other license_name: deepseek-license base_model:
- deepseek-ai/DeepSeek-V4.1-Flash
- dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8 tags:
- deepseek-v4.1
- moe
- affinity
- rocm
- rdna4
- quantized
- uncensored
- "1017" library_name: affinity pipeline_tag: text-generation
DeepSeek-V4.1-Flash-1017-UNCENSORED — affinity container (2.875 bpw)
A quantised affinity container of the uncensored DeepSeek-V4.1-Flash-1017 checkpoint (incorporating DeepSeek's official October 17 revision with native numeric reasoning-effort tokens and corrected DSML tool calling), made to run on two AMD Radeon AI PRO R9700 (gfx1201) GPUs.
It loads only in the affinity inference engine with our R9700 port and optimizations (available in the companion repo) — not in llama.cpp, vLLM or SGLang.
Full setup, engine patches, deployment unit, and v1.1.0 release live in the companion GitHub repo: 👉 https://github.com/drwolfen/deepseek-v4.1-flash-r9700-affinity
Hosted as 5 parts to comply with the Hub's per-file upload ceiling. Nothing here is an official DeepSeek or affinity release; it is a personal quantisation of an abliterated community checkpoint.
What is New in the 1017 Revision vs 0925
- Official 1017 Base: Calibrated and quantised from the October 17 uncensored checkpoint.
- Fixed DSML Tool Calling: Compatible with double-colon tool namespaces (
namespace::name) without corrupting XML tool envelopes. - Reasoning Effort Control: Supports numeric reasoning effort tokens (
low: 50,medium: 75,high: 75,max: 100). - Engine v1.1.0 Ready: Designed to pair with engine optimizations including
AFF_DRAFT_HEAD_INT4=1(3.88x smaller draft LM-head memory reads), fused DSpark taps (AFF_FUSE_HC_TAP=1), and length-faded capped rejection sampling (AFF_CACTUS_DELTA=0.35,AFF_CACTUS_CAP_R=0.5).
Measured Performance (Dual R9700, 512k Context)
Benchmarked on dual AMD Radeon AI PRO R9700 (32 GiB each, ROCm 7.14) running live agent sessions at 163.5k–170.7k context:
| Metric | Measured Value |
|---|---|
| Weighted Decode Throughput | 21.96 tok/s (+28.6% over v1.0.0 baseline of 17.08 tok/s) |
| Peak Burst Speed | 27.80 tok/s (e.g. 1,341 tokens in 48.15s at 166.3k context) |
| Speculative Block Latency | 166.7 ms/block median (-20.4%, down to 156.8 ms/block min) |
| Draft Acceptance | 79.9% (3.69 tokens/block average, up to 4.37 tokens/block) |
| Prefill Speed | ~1000 tok/s (persistent prefix cache restores 330+ blocks from NVMe) |
Files
| file | size | role |
|---|---|---|
model-uc-1017.aff.part-00 … part-04 | 191.32 GiB total | the container (routed experts 2.875 bpw codebook quant, dense FP8, embedding/head BF16). Split into 5 parts. |
model-uc-1017.dspark.aff | 5.22 GiB | DSpark speculative-decoding draft head |
model-uc-1017.engram | 506 KiB | engram n-gram sidecar |
uc-1017-cb.raw | 32 KiB | the 2×1024-entry codebook the experts were encoded against |
Reassembly
Reassemble the 5 parts into the full container using standard cat:
cat model-uc-1017.aff.part-* > model-uc-1017.aff
sha256sum -c SHA256SUMS.parts
Serving
Run with the optimized v1.1.0 affinity container:
docker run --name deepseek-affinity \
--device=/dev/kfd --device=/dev/dri --group-add video --ipc=host \
--cpuset-cpus=0-23,48-71 \
-e AFF_ENGRAM_WARM=1 \
-e AFF_SPEC_BACKOFF=2 \
-e AFF_ADAPTIVE_WIDTH=1 \
-e AFF_BOUNDED_REPLAY=1 \
-e AFF_FUSE_HC_TAP=1 \
-e AFF_DRAFT_HEAD_INT4=1 \
-e AFF_CACTUS_DELTA=0.35 \
-e AFF_CACTUS_CAP_R=0.5 \
-e AFF_RESIDENT_BIAS=0.008 \
-e AFF_NOBUDGET_MAX_TOKENS=16384 \
-p 8000:8000 \
-v /mnt/fast-storage:/mnt/fast-storage \
-v /path/to/affinity:/opt/affinity-src/build/affinity:ro \
affinity:v41-vision \
-m /mnt/fast-storage/model-uc-1017.aff \
--tokenizer /path/to/tokenizer.json \
--host 0.0.0.0 --port 8000 \
--gpu-headroom-mib 1024 \
--dspark on \
--dspark-model /mnt/fast-storage/model-uc-1017.dspark.aff \
--engram /mnt/fast-storage/model-uc-1017.engram \
--engram-tables /path/to/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8 \
--prefix-cache-dir /mnt/fast-storage/pcache \
--prefix-cache-gib 64 \
--persistent-prefix-cache \
--kv-size 524288 \
--kv-commit 262144 \
--host-pool-mib 200000 \
--host-pool-reserve-mib 1024 \
--placement-strategy demand \
--placement-save /mnt/fast-storage/placement.affprof \
--prefill-chunk-size 2816 \
--default-effort 75 \
--max-tokens-floor 0