Repository Issue Activity (beta)

#201 Bayesian optimization: silently skipped when optuna is absent, and warm start enqueued outside its own search space

#193 Release qualified Qwen3.8-27B E03 GSQ 2-bit checkpoint

#194 Release qualified Qwen3.8-27B E03 GSQ ternary (1.58-bit) checkpoint

#31 NVIDIA Jetson AGX not supported natively

#132 research: design and validate capacity-hypothesis experiments

#110 test: track unavailable conditional runner waivers

#115 feat: support safe surgery on Accelerate-offloaded meta weights

#195 bug: Colab step 3 lacks gated-model authentication before loading

#185 research: Qwen3.8 architecture-aware causal search with held-out refusal and coherence gates

#190 Coherence verifier scores correct Qwen3.8 answers as failures after preview truncation

#15 Support native NVFP4 / ModelOpt checkpoints (e.g. Qwen3.5-9B-NVFP4)

#22 Chat not working

#198 Qualify claimed GGUF exports with pinned conversion, load, and generation gates

#197 Validate saved config against serialized tensor inventory before promotion

#191 release: qualify and deploy Qwen3.8-27B E03 BitsAndBytes 4-bit on Titan

#184 feature: stable headless experiment API with immutable manifests and durable run archives

#183 bug: spectral certification must be inconclusive for rank-deficient zero-noise estimates

#182 bug: replace KL proxy correction with measured sequence-token KL and exact rollback

#180 fix: bounded GPU admission diagnostics and actionable UI errors

#168 feat: validated Qwen3.8 hybrid surgery and chat-aware quality gates

#177 fix: require fast DeltaNet kernels and single-GPU placement for Qwen3.8

#175 Qwen3.8 probes override the checkpoint multi-EOS generation contract

#173 Allow bounded CUDA context residue when releasing benchmark lease

#171 fix: benchmark paths bypass GPU lifecycle admission

#165 feat: expose truthful post-load GPU readiness and lease resize lifecycle hooks