FazBrowse GitHub Viewer | Trending | | Home

Tools: [Download Repo ZIP] [Original HTTPS Page]

History for python - gpu-cli/zerostart · GitHub

Commits

History for

Commits on Mar 13, 2026

Perf: async shared cache, parallel_downloads=32, inference benchmarks

lightsofapollo
and
claude
committed
8cff975

Commits on Mar 12, 2026

Fix system package isolation and accelerate() ordering in zerostart run

lightsofapollo
and
claude
committed
742e44f
Improve accelerate() for large MoE models (Qwen3.5-35B-A3B)

lightsofapollo
and
claude
committed
515aa2e

Commits on Mar 10, 2026

vLLM loader: use safetensors_load_strategy="eager" on network volumes

lightsofapollo
and
claude
committed
1f3d10e
Fix vLLM --load-format zerostart: rewrite load_format before parent init

lightsofapollo
and
claude
committed
fe408f2
Fix vLLM integration: use register_model_loader API, subclass BaseModelLoader

lightsofapollo
and
claude
committed
0450f12
Add diffusers support, ComfyUI and vLLM integrations

lightsofapollo
and
claude
committed
37affb5
Add zerostart.accelerate() — transparent 9x faster model loading

lightsofapollo
and
claude
committed
715ab3c
Rewrite snapshot v2: JSON-only, no cloudpickle, Rust tokenizer

lightsofapollo
and
claude
committed
b8fa30f
Fix snapshot: HF cache discovery, e2e benchmark on volume

lightsofapollo
and
claude
committed
5a56bf2
Materialize meta tensors after hydrate (rotary embeddings, etc.)

lightsofapollo
and
claude
committed
a952517
Use meta device for model creation — zero memory allocation

lightsofapollo
and
claude
committed
17db591
Skip random weight init during hydrate via _no_init_weights

lightsofapollo
and
claude
committed
de440cd
Add tie_weights() after load_state_dict in hydrate

lightsofapollo
and
claude
committed
410b8ba
Fix model reconstruction: use normal init instead of empty weights

lightsofapollo
and
claude
committed
d633889
Fix tensor name matching: strip all prefix levels for safetensors lookup

lightsofapollo
and
claude
committed
88abfce
Add snapshot/hydrate for fast model restore via safetensors mmap

lightsofapollo
and
claude
committed
f2b75b8

Commits on Mar 9, 2026

Fix GPU test suite: daemon binary fallback, abi3 wheel support, cache isolation

lightsofapollo
and
claude
committed
80fd905
Clean up repo for open source readiness

lightsofapollo
and
claude
committed
77945c8
Move cold path to Rust, use pylock.toml for single-call resolution

lightsofapollo
and
claude
committed
72dda6d

Commits on Mar 8, 2026

Add package mode: zerostart as a faster uvx alternative

lightsofapollo
and
claude
committed
d4abafc
Fix: wait for uv when no daemon (all small packages)

lightsofapollo
and
claude
committed
43800b4
Fix GPU test runner: remove unpublished zs-fast-wheel dep, fix install order

lightsofapollo
and
claude
committed
7084886
Add PEP 723 inline script metadata support + test plan

lightsofapollo
and
claude
committed
f8a1bff
Prepare for open source: MIT license, zerostart CLI, README rewrite

lightsofapollo
and
claude
committed
3a13623
Remove CRIU snapshot code, refocus on fast install + progressive loading

lightsofapollo
and
claude
committed
d92b559
Add Tier 1 CRIU process snapshot infrastructure

lightsofapollo
and
claude
committed
1cb0a2e
Add warm command: resolve requirements and download without Python

lightsofapollo
and
claude
committed
dd6fba6
Rewrite orchestrator and import hook to use in-memory DaemonHandle

lightsofapollo
and
claude
committed
fdd6203
Remove dead file-based IPC (StatusDir, DemandWatcher, status_dir)

lightsofapollo
and
claude
committed
1f6d5bc
Add Python resolver, cache, and tests for fast-wheel progressive loading

lightsofapollo
and
claude
committed
8b9132e

Commits on Mar 7, 2026

Add lazy import hook with progressive loading and demand signaling

lightsofapollo
and
claude
committed
c002500

Loading

Back | FazBrowse Home | New Git URL