pkgs-set/pkgs · ark · Tangled

ark pkgs-set pkgs

2 folders · 82 files

at stolas-builder-fix

ptr.pet · improve deploy script · 3w ago

bluesky-tap.nix

pds-upload thingy

atlogin experiments

atspool, klbr stuff

buildGoApplication.nix

style(treewide): go my formatter

clickee-proxy.nix

feat(wolumonde): add clickee proxy

chernobog: add clickee service

cliproxyapi.nix

fix(cliproxy): retry kimi engine overloads

style: format everything tm

refactor(treewide): use nvfetcher to fetch all the dependencies (except nixpkgs cuz flakes sucks grrr) [skip ci]

improve deploy script

refactor: why did i put the dnsconfig in separate dirs lol [skip ci]

docker-credential-atcr.nix

replace faunu with dysnomia

generate-firefox-addons.nix

style(treewide): go my formatter

headroom-ai.nix

ik-llama-cpp.nix

knot-migrate.nix

feat(trimounts): migrate knot to knot2

feat(trimounts): migrate knot to knot2

llama-cpp-bonsai-bi-mma-f16.patch

valefar: bonsai on the 3060 with a 64K F16 cache, 2 slots, tensor-core attention in batch-invariant mode

llama-cpp-bonsai-cache-api.patch

valefar: bonsai cache page and retention policy (classes, TTLs, learned families, pins), FWHT writes MMQ q8 tiles (+3% prefill), Open WebUI

llama-cpp-bonsai-cache-family.patch

valefar: bonsai prompt families keyed on the system prompt, so web-extract's one-shot calls learn ephemeral

llama-cpp-bonsai-cache-policy.patch

valefar: bonsai cache page and retention policy (classes, TTLs, learned families, pins), FWHT writes MMQ q8 tiles (+3% prefill), Open WebUI

llama-cpp-bonsai-cache-tuning.patch

valefar: bonsai cache tuning (SCR reuse floor 512->128, keep warm system prompt prefix)

llama-cpp-bonsai-decode-verify.patch

valefar: bonsai MTP decode +8.7% (PTQ1_0 tensor-core verify mat-vec folds residue slots in registers)

llama-cpp-bonsai-fairness.patch

valefar: bonsai two-slot fairness, decode keeps flowing while the other slot prefills

llama-cpp-bonsai-fattn-q8-prefill.patch

valefar: bonsai prefill +8-13% (staged PTQ1_0 MMQ main loop, unpadded GQA q8_0 prefill attention), no core dumps on restart

llama-cpp-bonsai-fattn-verify-lead.patch

valefar: bonsai decode +3.6% (GDN state read in place, fused conv path, FA verify K/V lead), fix MTP draft desync under fairness

llama-cpp-bonsai-fwht-q8.patch

valefar: bonsai cache page and retention policy (classes, TTLs, learned families, pins), FWHT writes MMQ q8 tiles (+3% prefill), Open WebUI

llama-cpp-bonsai-gdn-conv.patch

valefar: bonsai decode +3.6% (GDN state read in place, fused conv path, FA verify K/V lead), fix MTP draft desync under fairness

llama-cpp-bonsai-gdn-prefill.patch

valefar: faster bonsai gated delta net prefill (bit-identical)

llama-cpp-bonsai-gdn-rows.patch

valefar: bonsai decode +3.6% (GDN state read in place, fused conv path, FA verify K/V lead), fix MTP draft desync under fairness

llama-cpp-bonsai-imgcache.patch

valefar: bonsai image conversations in the NVMe cache, image embedding cache, request telemetry on the status page, fix SCR consuming unrelated RAM entries

llama-cpp-bonsai-mmq-pipeline.patch

valefar: bonsai prefill +11-13% (PTQ1_0 MMQ pipeline and per-128 activation scales, fused SWIGLU into the Hadamard load, ADD+RMS_NORM+MUL fusion)

llama-cpp-bonsai-mmq-staged.patch

valefar: bonsai prefill +8-13% (staged PTQ1_0 MMQ main loop, unpadded GQA q8_0 prefill attention), no core dumps on restart

llama-cpp-bonsai-mmq-y128.patch

valefar: bonsai prefill +11-13% (PTQ1_0 MMQ pipeline and per-128 activation scales, fused SWIGLU into the Hadamard load, ADD+RMS_NORM+MUL fusion)

llama-cpp-bonsai-mtp-chain-h.patch

valefar: bonsai on an 88K q8_0 cache read in place by tensor-core attention, MTP drafting two tokens

llama-cpp-bonsai-mtp-decode.patch

valefar: bonsai decode, one MTP pass per round and the fast conv-state concat (output identical)

llama-cpp-bonsai-mtp-fair-deferred.patch

valefar: bonsai decode +3.6% (GDN state read in place, fused conv path, FA verify K/V lead), fix MTP draft desync under fairness

llama-cpp-bonsai-prefill-fusions.patch

valefar: bonsai prefill +11-13% (PTQ1_0 MMQ pipeline and per-128 activation scales, fused SWIGLU into the Hadamard load, ADD+RMS_NORM+MUL fusion)

llama-cpp-bonsai-prefix-entry.patch

valefar: Bonsai keeps learned prefixes learned through pin and keep

llama-cpp-bonsai-prefix-learn.patch

valefar: Bonsai learns and pins repeated system prefixes

llama-cpp-bonsai-prefix-persist.patch

valefar: Bonsai prefix learning outlasts restarts

llama-cpp-bonsai-prefix-replay.patch

valefar: Bonsai learns system prefixes that agents resend with shared turns

llama-cpp-bonsai-prefix-restore.patch

valefar: Bonsai splices restore from learned prefixes

llama-cpp-bonsai-prefix-share.patch

valefar: Bonsai stores conversations against their learned prefix

llama-cpp-bonsai-prefix-shared.patch

valefar: Bonsai pins the system prefix that albedo sessions share

llama-cpp-bonsai-prompt-cache-disk.patch

valefar: bonsai prompt cache with 24 GB of RAM and a 50 GB NVMe tier that survives restarts

llama-cpp-bonsai-ptq1-tc-verify.patch

valefar: bonsai on an 88K q8_0 cache read in place by tensor-core attention, MTP drafting two tokens

llama-cpp-bonsai-q8-kv.patch

valefar: bonsai on an 88K q8_0 cache read in place by tensor-core attention, MTP drafting two tokens

llama-cpp-bonsai-scr-mtp-flush.patch

valefar: flush MTP catch-up before SCR run and abort boundaries

llama-cpp-bonsai-scr-spans.patch

valefar: raise Bonsai SCR span cap from 6 to 64

llama-cpp-bonsai-scr.patch

valefar: bonsai suffix cache reuse on by default (--scr): edits in the middle of a cached prompt only prefill the new text

llama-cpp-bonsai-telemetry.patch

valefar: bonsai image conversations in the NVMe cache, image embedding cache, request telemetry on the status page, fix SCR consuming unrelated RAM entries

llama-cpp-bonsai-ttft.patch

valefar: cut bonsai time to first token on cached prompts (no per-request reserve, recycled checkpoints)

llama-cpp-bonsai.nix

valefar: Bonsai stores conversations against their learned prefix

llama-cpp-mtp.nix

llama-cpp-nanbeige.nix

llama-cpp-rocm.nix

mcp-auth-proxy-colab.nix

add chernobog host

nsid-tracker-server.nix

style(treewide): go my formatter

feat(wolumonde): perses dashboard as code with go

pi-coding-agent.nix

tangled-knot.nix

ohhh ahhh stuff idk

tangled-modules.nix

style(treewide): go my formatter

tangled-spindle.nix

ohhh ahhh stuff idk

tangled-sqlite-lib.nix

style(treewide): go my formatter

zen-browser.nix