 kilork
|
267b9daae7
build works outside the launcher repo; bundles gain a build.sh
|
hai 1 mes |
|
kilork
|
caae1b0ead
4p67: eager vs CUDA-graphs A/B measured (tool-eval-bench, seed 42)
|
hai 1 mes |
|
kilork
|
5ab8a05080
restore recipes/ in-tree — recipe + docker-images must share one revision
|
hai 1 mes |
|
kilork
|
ee2855da38
config errors name every searched path
|
hai 1 mes |
|
kilork
|
d0d210e466
move recipe data to git.kilork.org/sm121/recipes (registry data repo)
|
hai 1 mes |
 kilork
|
848772a3e7
glm53-flash-dflash2: target-side EAGLE3 port + rigorous KV-page blocker writeup
|
hai 1 mes |
|
kilork
|
2599e39186
glm53-flash-nvfp4-4p67: MXFP8-attention support patchset + boot verified on ds4
|
hai 1 mes |
 Alexander Korolev
|
30268d382b
glm53-flash-nvfp4-4p67: dedicated image with packed_modules_mapping fix
|
hai 1 mes |
|
Alexander Korolev
|
bb98052579
dflash2 build.sh: strip shell quotes from versions.env values before --build-arg
|
hai 1 mes |
|
Alexander Korolev
|
49549333c0
dflash2 build.sh: repair a redaction placeholder that corrupted the env-parsing loop
|
hai 1 mes |
|
Alexander Korolev
|
1c1d3455f8
glm53-flash: NVFP4-4p67 + DFlash2-MXFP8 recipes (model cards followed, GB10-adapted)
|
hai 1 mes |
|
Alexander Korolev
|
796e8b862c
repo moved to the sm121 org: remote + doc links now git.kilork.org/sm121/sm121-launcher
|
hai 1 mes |
|
Alexander Korolev
|
51912523c5
feat(registry): publish --registry upload + sm121 registry list/show/install
|
hai 1 mes |
|
Alexander Korolev
|
79d1a76c74
feat(cli): sm121 publish — recipe -> standalone reproducible bundle (and gist)
|
hai 1 mes |
|
Alexander Korolev
|
ed4e0f99ea
build(images): digest-pin glm53 base image — tag glm53-flash-arm64-cu130 is mutable; pin sha256:905c0293 (the base our verified images were built on). Checksum changes → one no-op rebuild on next serve.
|
hai 1 mes |
|
Alexander Korolev
|
96c6bfe777
feat(recipes): lab defaults to --enforce-eager — MTP+CUDA-graphs quality bug confirmed by A/B bench
|
hai 1 mes |
|
Alexander Korolev
|
28e08afa73
feat(recipes): redhat @ 524288 ctx — pinned 10 GiB KV, autotune off, 4096 chunks, lab image for CC-12.0 topk guards
|
hai 1 mes |
|
Alexander Korolev
|
3d30d85655
docs(specs): lab /174 explained — TC-53 excluded on 120s infra timeout (score unchanged at 89 either way)
|
hai 1 mes |
|
Alexander Korolev
|
a0383302e3
docs(specs): redhat MTP reality check — RedHatAI ships mtp as separate -mtp-FP8 variant repos (GLM-5.2 pattern); none for 5.3-Flash yet
|
hai 1 mes |
|
Alexander Korolev
|
5730abbfee
docs(specs): GLM-5.3-Flash three-way quant benchmark on ds4 (measured)
|
hai 1 mes |
|
Alexander Korolev
|
9d72c8b804
feat(recipes): lab GLM-5.3 recipe = FujitsuPolycom proven stack, verified on ds4
|
hai 1 mes |
|
Alexander Korolev
|
b710a09bcb
feat(recipes): GLM-5.3-Flash NVFP4 variants — local-inference-lab + RedHatAI
|
hai 1 mes |
|
Alexander Korolev
|
01fd6bd817
chore(deps): drop unused russh-keys dependency
|
hai 1 mes |
|
Alexander Korolev
|
25c53b5f50
refactor(recipes): drop redundant entrypoint override from glm53-flash-nvfp4
|
hai 1 mes |
|
Alexander Korolev
|
095dec1d73
refactor(recipes): quant-agnostic served-model-name + rename glm53 to -nvfp4
|
hai 1 mes |
|
Alexander Korolev
|
995e45c6c6
fix(recipes): GLM-5.3-Flash -> vLLM NVFP4 on 2x Spark with NoPE-MLA mod
|
hai 1 mes |
|
Alexander Korolev
|
cdedc90e5b
feat(recipes): Qwen3.8-Flash-Next FP8 on vLLM via expert parallelism
|
hai 1 mes |
|
Alexander Korolev
|
5ee2326c40
fix(recipes): switch Qwen3.8-Flash-Next to NVFP4 for 2x Spark
|
hai 1 mes |
|
Alexander Korolev
|
08d005ffb5
feat(config): global [before_serve."*"] hooks + optional master_ib_ip
|
hai 1 mes |
|
Alexander Korolev
|
e9212feb72
feat(recipes): add Qwen3.8-Flash-Next-FP8 and GLM-5.3-Flash SGLang day-0 recipes
|
hai 1 mes |
