Card: first-start note, same-machine vision baseline, Typed Decisions note; drop unmeasured throughput line
#1
by TheoHsiung - opened
README only; no hashed file changes.
- Remove the "Throughput under load: not measured yet" bullet.
- Run with vLLM Semantic Router: note that the first start downloads the package and, if it stops at the runtime's 10-minute ready timeout, the same command run again only loads.
- Explain the LocalLLaMA/typed-decisions result in the page's evaluation widget (separate run, score_model.py, d3 0.7485 on the same setup).
- Highlights: give d3's same-machine vision score (73.93) next to 74.15.
- Highlights: add the Typed Decisions line the sister cards have (99.08%, 83.24% against 83.25%).
Second commit: d3q/load.py docstring no longer names d3-FP8 or the personal repo TheoHsiung/d3-W4A16-G128 (now vllm-sr/d3-W4A16-GPTQ); the README links bench/d3-quant-eval at a fixed commit (7908956a) instead of a branch; MODEL_MANIFEST.json carries the new hashes and sizes of README.md and d3q/load.py, so the manifest matches this PR. Identity (model_sha256) unchanged; checked with vllm_srun decision3 package.verify (upstream main 4e37669a) and d3_runtime.verify_package(release) on the full package.
Xunzhuo changed pull request status to merged