fix(dflash): unblock Qwen3.5-MoE + DFlash + ContextPilot serving hang (multi-GPU device placement) - #147

Merged

drunkcoding merged 5 commits into

devfrom

fix/qwen35-dflash-multigpu-device-hang

Aug 13, 2026

Commits on Aug 11, 2026

fix(dflash): place native forward + speculator on the resident GPU
drunkcoding
committed
5ddd917
fix(serving): resolve Qwen3.5-MoE arch dims via get_text_config
drunkcoding
committed
54261d5
test(contextpilot): skip live tests when optional package is absent
drunkcoding
committed
8ae26c8
fix(ci): skip GLM routing tests when transformers drops route_tokens_to_experts
drunkcoding
committed
80adb3d
fix(test): accept speculative_draft in watchdog _FakeRuntimeEngine mock
drunkcoding
committed
3cb94cc