Issues

• ai-waf: a fail-closed reverse proxy in front of ds4-server — integration proposal

  antirez/ds4#1318 · Filter by author Pierpaolo1223 opened on Nov 4, 2026

• server vision LRU cache thrash: the vision embedding cache holds ~12 images for qwen3.8-flash-next, so every turn re-encodes the whole image history if ≥12 images

  antirez/ds4#1317 · Filter by author wanderingmeow opened on Nov 4, 2026

• M5 Ultra: decode drift and repetition loops when M5_PARALLEL_FULL_FFN and M5_ROUTER_PROJECT_SELECT_FUSE are both enabled

  antirez/ds4#1316 · Filter by author Temikus opened on Nov 3, 2026

• Proposal: a model-free CI gate for pull requests

  antirez/ds4#1311 · Filter by author giuliowaitforitdavide opened on Nov 1, 2026
  2

• [SECURITY] Control-Token Injection (Template Injection / CWE-1336) in chat prompt renderer (ds4_tokenize_rendered_chat)

  antirez/ds4#1308 · Filter by author Pierpaolo1223 opened on Oct 31, 2026
  2

• CUDA: server process dies during long prefill at ~512K tokens (ctx=1048576, managed KV cache) instead of returning an error

  antirez/ds4#1305 · Filter by author potoior opened on Oct 30, 2026
  1

• V4.1 Flash support and ~10 t/s Flash Q2 decode on a cheap dual-socket Xeon. Interested in a PR?

  antirez/ds4#1302 · Filter by author 10sephus opened on Oct 29, 2026

• GLM 5.3 Flash: long agentic thinking turn degenerates into one repeated token at temperature 1.0; no server-side repetition control

  antirez/ds4#1301 · Filter by author nomishbhardwaj opened on Oct 29, 2026

• ROCm V4.1 SSD-streaming decode: 3-5 t/s with idle disk - routed-MoE vec path is the bottleneck (measurements + timing patch)

  antirez/ds4#1297 · Filter by author amittell opened on Oct 28, 2026
  4

• DeepSeek-V4.1-Flash-1017-abliterated-MXFP4-GGUF for ds4

  antirez/ds4#1294 · Filter by author varoudis opened on Oct 27, 2026

• Metal GLM MTP batch verification updates hotness without advancing its decay clock

  antirez/ds4#1271 · Filter by author Elvis-Luo opened on Oct 25, 2026

• Support for this magic dsv4.1 1017 quant

  antirez/ds4#1268 · Filter by author openSourcerer9000 opened on Oct 24, 2026

• Disk KV: the pre-lookup "evict" save deletes the cold prefix the same request is about to load

  antirez/ds4#1266 · Filter by author terribilissimo opened on Oct 23, 2026

• Support to FLINT

  antirez/ds4#1260 · Filter by author loretoparisi opened on Oct 23, 2026

• [Question] what is ds4 equivalent for -a/--alias in llama-server

  antirez/ds4#1259 · Filter by author dineshr93 opened on Oct 23, 2026

• Pre-tokenizer splits differ from the HF tokenizer for CJK punctuation + newline and for runs of spaces (DeepSeek V4 Flash)

  antirez/ds4#1255 · Filter by author IvanChangTW opened on Oct 23, 2026

• CUDA GLM-5.3 --mtp: MTP step fails on every token once the MTP window exceeds ~12k positions (staged decode weights kernel exceeds 48 KiB dynamic shared memory)

  antirez/ds4#1243 · Filter by author rfc1437 opened on Oct 21, 2026

• engram speedup patch achieved 28 tok sec deepseek 4.1 flash on m3 ultra 512 gb

  antirez/ds4#1240 · Filter by author badelon opened on Oct 20, 2026
  2

• DeepSeek V4.1 Flash: a ~3-bit (or mixed Q2/Q4) build for 256 GB Macs

  antirez/ds4#1239 · Filter by author maciekd opened on Oct 20, 2026
  2

• download_model.sh: stale SHA-256 for the 1017 Q2 checkpoint rejects a correct 341 GiB download

  antirez/ds4#1235 · Filter by author cropduster opened on Oct 19, 2026
  1

• Supporting Jinja Template (allow for model jailbreak)

  antirez/ds4#1231 · Filter by author SpentPoetry opened on Oct 19, 2026
  1

• Support 10/17 DeepSeek Flash V4.1 Update

  antirez/ds4#1224 · Filter by author tolidano opened on Oct 17, 2026
  2

• ds4_test: with DS4_TEST_DSPARK set, a full run exits at dspark-verify-depth on its own instance lock

  antirez/ds4#1223 · Filter by author francosax opened on Oct 16, 2026

• Windows + Volta (sm_70) port: runs DeepSeek-V4-Flash Q2 end-to-end — 27-line patch + small compat layer, first numbers

  antirez/ds4#1215 · Filter by author noahark opened on Oct 14, 2026
  2

• Qwen3.8-flash-next live kv cache miss due to "token-mismatch" in multi-turn vision conversations, causing unnecessary full prefill

  antirez/ds4#1214 · Filter by author wanderingmeow opened on Oct 14, 2026
  3