Commit History

Автор SHA1 Съобщение Дата
  Alexander Korolev 5a9795c612 Rebase deepseek-v4-flash :fixed overlay onto production-3.76.1 преди 2 месеца
  Alexander Korolev 08871e3dfa Add --fit flag to serve: scale recipe to N nodes or whole cluster преди 2 месеца
  Alexander Korolev 551ea7d8d7 feat(nccl): auto-detect NCCL_IB_GID_INDEX per node, opt-out via config pin преди 2 месеца
  Alexander Korolev 42032fbe12 deepseek-v4-flash: Aiden base 3.75 + encoder fix baked in, 1M ctx, match 0rand compose tuning преди 2 месеца
  Alexander Korolev 7f5a038adc Tune deepseek-v4-flash-eugr: prefill threshold, prefix cache, reasoning_effort=max преди 2 месеца
  Alexander Korolev c6eda47a99 feat(deepseek-v4-flash-eugr): use high thinking by default преди 2 месеца
  Alexander Korolev 43c9bd346f recipe(deepseek-v4-flash-eugr): second-variant test config преди 2 месеца
  Alexander Korolev 5d90e025f0 revert: keep experimental eugr recipe changes out of main (user is still testing) преди 2 месеца
  Alexander Korolev 08d1e9fe03 feat(serve): add --mirror — serve from a per-node local HF mirror преди 2 месеца
  Alexander Korolev 4eff3e41b0 docs(skills): add nvidia-forum-reading skill преди 2 месеца
  Alexander Korolev a432ba42be docs(skills): add sm121-launcher and vllm-mod-patches skills преди 2 месеца
  Alexander Korolev 070bf6666a feat(deepseek-v4-eugr): optional reasoning_content key patch (default-off) преди 2 месеца
  Alexander Korolev 0c7b29fbd6 feat(deepseek-v4-eugr): minimal reasoning_effort low/high/max patch (not full rewrite) преди 2 месеца
  Alexander Korolev 25b18fcd56 fix(deepseek-v4): drafter b12x WO routing for real DSpark acceptance + GMU 0.78 преди 2 месеца
  Alexander Korolev e296685e8c fix(deepseek-v4): let DSpark survive mixed-context batches (tool-eval-bench) преди 2 месеца
  Alexander Korolev 5de97ee3aa fix(deepseek-v4): get DeepSeek-V4-Flash DSpark booting on production-3.8 (ds4) преди 2 месеца
  Alexander Korolev d2026c2ddf feat(deepseek-v4): raise max-model-len to 1M + pin KV cache bytes (eugr) преди 2 месеца
  Alexander Korolev 3aa9d380b2 feat(recipes): agent-serving profile + GMU/env fixes from benchmark findings преди 2 месеца
  Alexander Korolev 0362145ef5 docs(specs): investigation — production-3.8 image update failures преди 2 месеца
  Alexander Korolev a8e4932fc4 feat(deepseek-v4): add eugr b12x recipe variant with baked draft-loader mod преди 2 месеца
  Alexander Korolev 5bc841f211 tmp: rename to simplify преди 2 месеца
  Alexander Korolev 3a783ee58e Revert "feat(deepseek-v4): move :fixed overlay to production-3.8 base + enable DSPARK tiny decode" преди 2 месеца
  Alexander Korolev 72f40b3da2 feat(deepseek-v4): move :fixed overlay to production-3.8 base + enable DSPARK tiny decode преди 2 месеца
  Alexander Korolev e6f70ecef1 fix(deepseek-v4): forbid echoing injected content to close TC-34 преди 2 месеца
  Alexander Korolev 8269eeafca fix(deepseek-v4): honor thinking_mode kwarg; harden tool template vs prompt injection преди 2 месеца
  Alexander Korolev bd292d7508 fix(deepseek-v4): use canonical model-card dsv4 encoding over stale bundled copy преди 2 месеца
  Alexander Korolev a83c61cd7e fix(deepseek-v4): patch vendored dsv4 tokenizer reasoning_effort handling преди 2 месеца
  Alexander Korolev 1bae27bd44 fix(docker): complete DS4 src runtime lock with flashinfer transitive deps преди 2 месеца
  Alexander Korolev 5a6ec550d7 fix(ssh): pass through already single-quoted tokens in shq преди 2 месеца
  Alexander Korolev 0a21797aeb refactor(docker): extract model path-reference resolver to path_ref submodule преди 2 месеца