Issues

• [Feature]: Report prefix tokens lost to evicted recurrent-state snapshots (hybrid Mamba, KV cache manager V2)

  KV-Cache Management
  NVIDIA/TensorRT-LLM#21884 · Filter by author hokwangchoi opened on Nov 8, 2026

• [Performance]: SM120 FP8 block-scale MoE uses 16-row tiles at every batch size, which halves prefill MoE throughput

  Model optimization
  NVIDIA/TensorRT-LLM#21882 · Filter by author hokwangchoi opened on Nov 8, 2026

• LongBench V2 blog command breaks on inline comments after continuations

  Doc
  NVIDIA/TensorRT-LLM#21879 · Filter by author ddlw opened on Nov 8, 2026

• DSpark rejects requested token and prompt log-probability outputs

  Speculative Decoding
  NVIDIA/TensorRT-LLM#21859 · Filter by author yifjiang opened on Nov 7, 2026

• Precompiled package extraction omits direct NIXL and UCX libraries

  Infra
  NVIDIA/TensorRT-LLM#21858 · Filter by author yifjiang opened on Nov 7, 2026

• [Bug]: V2 scheduler deadlocks when a recompute-paused request's replay exceeds max_num_tokens (chunked prefill off)

  KV-Cache Management
  NVIDIA/TensorRT-LLM#21826 · Filter by author maxwbuckley opened on Nov 6, 2026
  1

• [Bug]: Suffix automaton returns other requests' draft tokens when the draft length is below the workspace width (draft_len_schedule)

  Speculative Decoding
  NVIDIA/TensorRT-LLM#21820 · Filter by author maxwbuckley opened on Nov 6, 2026
  1

• [Bug]: Ray orchestrator: the DeepEP buffers bootstrap from an MPI_COMM_WORLD split, so DeepEPLowLatency gets a 1-rank group (same pattern as #20083)

  Disaggregated serving
  NVIDIA/TensorRT-LLM#21809 · Filter by author dmvevents opened on Nov 6, 2026

• [Feature]: Port the FLA Triton kernels off tl.make_block_ptr before the move to Triton 3.9

  Triton backend
  NVIDIA/TensorRT-LLM#21801 · Filter by author waynehacking8 opened on Nov 6, 2026

• [Bug]: Eagle3 loads a d2t that does not match t2d without any error

  Speculative Decoding
  NVIDIA/TensorRT-LLM#21800 · Filter by author waynehacking8 opened on Nov 6, 2026

• Export speculative verification rounds for aggregate acceptance length

  Speculative Decoding
  NVIDIA/TensorRT-LLM#21731 · Filter by author yifjiang opened on Nov 5, 2026

• [Bug]: MiniMax-M3 streaming tool parser leaks the next call as text when one chunk ends a call and starts another

  NVIDIA/TensorRT-LLM#21728 · Filter by author paolo-perrone opened on Nov 5, 2026

• [Bug] Kimi K2 parser drops tool calls with multiline JSON arguments

  NVIDIA/TensorRT-LLM#21648 · Filter by author 0z5a opened on Nov 4, 2026

• [Bug] Chat benchmark counts metadata-only stream events as output tokens

  NVIDIA/TensorRT-LLM#21647 · Filter by author 0z5a opened on Nov 4, 2026

• [Bug]: Poolside streaming parser drops buffered text at completion

  NVIDIA/TensorRT-LLM#21643 · Filter by author 0z5a opened on Nov 4, 2026

• [Bug]: Non-string enum tool arguments are returned as strings

  OpenAI API
  NVIDIA/TensorRT-LLM#21642 · Filter by author 0z5a opened on Nov 4, 2026

• [Bug] MiniMax M3 streaming reuses tool-call indices across blocks

  NVIDIA/TensorRT-LLM#21636 · Filter by author 0z5a opened on Nov 4, 2026

• [Bug] MiniMax M2 streaming parser loses buffered text at finish

  NVIDIA/TensorRT-LLM#21634 · Filter by author 0z5a opened on Nov 4, 2026

• [Bug]: SamplingParams with an explicit end_id fails when stop or bad strings are set

  bugLLM API
  NVIDIA/TensorRT-LLM#21613 · Filter by author Anshtyagi1729 opened on Nov 4, 2026

• [Bug]: /v1/completions logprobs text_offset points at the wrong character for every token after the first

  NVIDIA/TensorRT-LLM#21577 · Filter by author paolo-perrone opened on Nov 3, 2026

• [Bug]: MiniMax-M3 tool parser drops the end of a streamed response

  NVIDIA/TensorRT-LLM#21573 · Filter by author paolo-perrone opened on Nov 3, 2026

• [Performance] Nemotron-H prefill on B300: five measured speedups, which would you like as PRs?

  General perf
  NVIDIA/TensorRT-LLM#21571 · Filter by author jvpflum opened on Nov 3, 2026

• [Bug]: trtllm-bench multi-turn acceptance statistics use only the last turn's decoding iterations

  NVIDIA/TensorRT-LLM#21519 · Filter by author codebreaker77 opened on Nov 3, 2026

• [Bug]: Executor crashes with "User-provided base page indices is too short" when the KV pool is smaller than one max_num_tokens context chunk

  Customized kernelsKV-Cache Management
  NVIDIA/TensorRT-LLM#21470 · Filter by author benchislett opened on Nov 2, 2026
  3

• [Bug] DFlash drafter with attention_backend VANILLA ignores a reused prefix: acceptance roughly halves on cached turns

  Speculative Decoding
  NVIDIA/TensorRT-LLM#21462 · Filter by author jvpflum opened on Nov 2, 2026