Repository Issue Activity (beta)

#12846 Rust preprocessor mis-tokenizes non-ASCII added/special tokens, silently dropping BOS (poolside/Laguna-S-2.1-FP8)

#14456 [BUG]: ThunderAgent at high concurrency, requests hang, states leak, Duplicate request ID

#14302 [BUG]: SGLang strips the stop token from nvext.completion_token_ids

#11320 [DEP] decouple Dynamo serving from discovery and request-plane activation

#14045 [CONTRIBUTION]: Add an agent-tree sibling placement benchmark

#11343 feat(router): parent-aware placement for subagent sessions (agentic workloads)

#11768 feat(request-trace): native S3 sink for request-trace records

#14418 [FEATURE]: Supports using abort as the finished_reason

#14354 DEP: Unified semantic error propagation across frontend and backends

#14520 KServe gRPC ModelMetadata leaks internal model_config.proto enum names instead of OIP wire datatype strings

#14519 tritonserver example POC: BYTES tensor bridge crashes on reshape (input) and can't read output

#14518 tritonserver example: --help text has stale KServe gRPC port (8787 vs actual 8000)

#13931 test(rl): validate the Dynamo rollout integration with verl

#13121 DEP: Dynamo-managed Elastic EP on Kubernetes

#14513 [BUG]: load-predictor pre-sweep scores a raising predictor as constant_last and lets two negative forecasts cancel

#13383 [FEATURE]: Document and test diffusion DGD deployments

#14476 DEP: Experimental incremental text input for /v1/realtime

#14310 DEP: One SelectionCore behind the frontend, the standalone selection service, and the EPP

#12905 [CONTRIBUTION]: fix unchecked integer multiplication in token sequence arithmetic

#12906 [CONTRIBUTION]: add RAII guard for CUDA memory pool allocation

#11661 DEP: EPP Embedded SelectionService Interface

#13390 feat(epp): add an SGLang tokenizer provider for standalone routing

#14213 [BUG]: Request-path overload marks never expire when rejection happens at the TCP ingress, causing permanent HTTP 529 on idle workers

#10664 [BUG] NATS reconnection fails with EAFNOSUPPORT after CRIU restore on IPv6-disabled Kubernetes pods

#13286 Unify graceful shutdown ordering and timeout budgets for Python workers and Rust sidecars