Skip to content

Releases: Baekpica/ds4

Release list

v0.6.3-dfm

v0.6.3-dfm Latest
Latest

Choose a tag to compare

View all tags
@Baekpica Baekpica released this 23 Aug 23:16
v0.6.3-dfm
516456f

ds4-dfm release line: Entrpi v0.6.3 absorbed with the five model families (DeepSeek V4 Flash, Solar Open2 250B, K-EXAONE 236B A23B, Motif-3, dots3-note Preview) served from one binary on a DGX Spark GB10 (sm_121a).

Absorbed from Entrpi v0.6.3

  • Typed refusal for schema-constrained output (response_format and equivalents answer HTTP 400 in the endpoint's native envelope on all four API surfaces)
  • Chunked request bodies (Transfer-Encoding: chunked decoding; DS4_SERVER_CHUNKED=0 restores the old reader)
  • Think-dial observability: think=<none|low|high|max> on serving log lines, the continuous lane's per-request completion line, and the ds4_requests_think_total counter family in /v1/stats and /metrics
  • Full 1,048,576-token window decode dispatch: HG-before-cap, live-scalar fallback, exact-full persisted-bank restore
  • Whole-prompt depth fence: typed refusal for one-shot forwards past 8,192 rows (DS4_PREFILL_NOFENCE=1 lifts it for probe runs)
  • Best-fit trim victims under admission pressure (DS4_BATCH_TRIM_BESTFIT=0 restores the pure recency walk)

Family-side reconciliation

Upstream's exact-full bank restore fix (Inc 4 audit Finding 2) covered only the DeepSeek payload lane. The Solar, EXAONE, and Motif-3 cont bank restores carried the same >= seq_cap off-by-one and now accept the exactly-full payload a full family bank legitimately persists.

Gates (GB10, driver 610.43.02, CUDA 13.3, sm_121a cubins)

All passed on this binary: extractor self-test, ds4_test --server (including the new v0.6.3 refusal/fence/think units), split-GGUF, test-model-family-kernels, test-mmq-parity, cuda-regression (including the new substrate overflow leg), Motif loader/tokenizer/reference/CUDA, EXAONE kernels/reference, Solar loader/tokenizer/KDA/KDA-prefill/KDA-chunk/gates/KV/forward, dots3 loader/tokenizer.

A live VMM owner + worker gate on the Motif MQ87-88 artifact answered all four API surfaces on the continuous route (4 requests, 0 failures, 32 banks at -c 2048); the typed refusal, cont completion line, and think counters were observed live on the family lane.

Published Motif-3 and Solar performance numbers are unchanged; no remeasure was run for this cut.

Credit

Engine: antirez/ds4 via Entrpi/ds4 (also Entrpi/ds4-on-spark for the sm_121 target and aligned-artifact tier); GGUF and quant formats from ggml-org/llama.cpp. Details in docs/ds4-dfm-model-families.md and CHANGELOG.md.

Assets 2
Loading