-
Format standardization: ladder/cross-link blocks
-
Add config.json so Hub download stats are counted
-
Copy edit: state facts plainly, drop self-referential honesty language
-
Fix tilde strikethrough rendering (~ -> ≈)
-
Two-model repro: --model flag, qwen3 8/32/128K support
-
Rebuild with qwen3 qk-norm adapter fix + mixed-step routing fix (batch>=2 needle-drop, 2026-07-03 root cause)
-
Pin torch==2.11.0 (validated stack); --local branch GPU-tested
-
README: reflect source integrations layer
-
integrations/ as source for torch.compile parity (kernel binary + calibration exclusion unchanged); script drops profiler override
-
Final repro script: median-of-3 timing, receipt pool accounting, retries
-
Ship vllm_model_adapters as source (torch-dynamo must trace patched forwards); rest stays binary
-
Rebuild wheel with annotation_typing=False (Cython enum-default fix); sync repro script (--contexts flag, 32-token timing)
-
Option-B repro kit: prebuilt SM80 binary wheel (no CUDA/calibration source, audited), build receipt, repro script, license
-
initial commit
