12月1月2月3月4月5月6月7月8月9月10月11月
周一周三周五
一年内 261608 次贡献
更少更多
- frostbyte_neo 推送了仓库 github-featured/unslothai--unsloth 的 grpo-fix-rlpy 分支比较 3 提交 »a3ce28ffbb Keep only GRPO state restore in rl.pyeb8ca4f4af [pre-commit.ci] auto fixes from pre-commit.com hooks9d2c1ff2b6 Restore GRPO eval mode in patched trainer
- frostbyte_neo 推送了仓库 github-featured/unslothai--unsloth 的 gemma4-fp16-force-float32 分支027b7f8551 Route Gemma-4 float16 through UNSLOTH_FORCE_FLOAT32
- frostbyte_neo 推送了仓库 github-featured/unslothai--unsloth 的 gemma4-eseries-generation-usecache 分支比较 4 提交 »b715eaad6d [pre-commit.ci] auto fixes from pre-commit.com hooks2ef4039911 is_gemma4_shared_kv_model: read config via _config_get for dict configs952fa1916e [pre-commit.ci] auto fixes from pre-commit.com hooks925d4b0675 Gemma 4 E-series: default generation to use_cache=False
- frostbyte_neo 推送了仓库 github-featured/unslothai--unsloth 的 followup-mtp-reload-guards 分支比较 4 提交 »068e129cb2 fix(mtp): gate legacy --draft-min with --draft-max on chained ngrama54396a675 test(mtp): prefer real httpx over stub so route-guard import chain worksfe61d6752b [pre-commit.ci] auto fixes from pre-commit.com hooks1b621cd9e7 studio: tighten MTP reload guards and asymmetric spec flags for #5582
- frostbyte_neo 推送了仓库 github-featured/unslothai--unsloth 的 followup-lockfile-audit-regressions 分支比较 2 提交 »a45536e218 [pre-commit.ci] auto fixes from pre-commit.com hooks70a090fd26 studio: harden lockfile supply-chain audit followups for #5604
- frostbyte_neo 推送了仓库 github-featured/unslothai--unsloth 的 followup-active-only-grouped 分支比较 4 提交 »dc42a0d82c [pre-commit.ci] auto fixes from pre-commit.com hooks2bbb15994b gemma4 moe 4bit: pure-PyTorch NF4 dequant variant (negative result, kept for docs)198c6232d1 [pre-commit.ci] auto fixes from pre-commit.com hooks6c76ae9f9d gemma4 moe 4bit: active-only + cache + static_bf16 forward variants (#5344 follow-up)
- frostbyte_neo 推送了仓库 github-featured/unslothai--unsloth 的 flexattn_refactor_fixed 分支比较 5 提交 »405bb58284 Fix config stamping when SDPA is not supported3b8860c039 Consistent config stamping in vision.py attention fallbacksf77a42e0ab Fix attention selection regressions in PR #42500bfe7e3de1 Merge remote-tracking branch 'origin/main' into flexattn_refactorc4f01056ae fix comment and match old behaviour
- frostbyte_neo 推送了仓库 github-featured/unslothai--unsloth 的 flash-attn-blackwell-6961 分支504e253674 Install Flash Attention on Blackwell when a CUDA >= 12.8 wheel exists
- frostbyte_neo 推送了仓库 github-featured/unslothai--unsloth 的 fla-gated-deltanet-advisory 分支比较 5 提交 »ceaa706efa [pre-commit.ci] auto fixes from pre-commit.com hooksbf974da5cd Cover olmo_hybrid in the gated-deltanet fla advisorya7d453175a Normalize model_types in fla install advisory for None and single stringf8b28b0024 Tighten comments11450bcf7b Note the bundled flash-linear-attention kernels for gated-deltanet models
