Issues

• [Perf]: DeepSeek-V4.1 Engram n-gram hash falls back to ~150 eager torch kernels per step with multiple tokens per request (speculative decoding)

  PytorchSpeculative Decoding
  NVIDIA/TensorRT-LLM#20604 · Filter by author lucifer1004 opened on Oct 19, 2026
  2

• [Bug]: DeepSeek-V4.1 CSA2: intermittent illegal address in the dynamic-token-sparse MLA generation kernel, which loads -1 index padding past the sparse length

  bugCustomized kernelsPytorch
  NVIDIA/TensorRT-LLM#20603 · Filter by author lucifer1004 opened on Oct 19, 2026
  2

• [Bug]: DeepSeek-V4.1 vision tower fails on SM107: cuDNN SDPA "No valid execution plans built" during KV-cache profiling

  Customized kernelsPytorch
  NVIDIA/TensorRT-LLM#20586 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Performance]: CuTe DSL MLA decode split-KV reduction is load-latency bound at small batch

  General perf
  NVIDIA/TensorRT-LLM#20572 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Performance]: Kimi KDA verify runs the qkvg and [f_a | b] projections as two GEMMs on two streams

  General perf
  NVIDIA/TensorRT-LLM#20569 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Performance]: DFlash prefill context store blocks on a D2H read at every chunk, idling the GPU between chunks

  General perfSpeculative Decoding
  NVIDIA/TensorRT-LLM#20567 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Performance]: MLA chunked prefill builds context K with a cat per loop and a fixed buffer batch size

  General perf
  NVIDIA/TensorRT-LLM#20565 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Performance]: CuTe DSL MLA decode split-KV cap of 32 leaves the GPU idle at long context and small batch

  General perf
  NVIDIA/TensorRT-LLM#20561 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Performance]: copyBatchBlockOffsetsToDeviceKernel walks each long block-offset row with one CTA

  General perf
  NVIDIA/TensorRT-LLM#20553 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Performance]: MLA DSpark drafter runs each layer's q and kv paths serially on one stream

  General perf
  NVIDIA/TensorRT-LLM#20551 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Performance]: DFlash / MLA DSpark decode spends ~0.2 ms per step in small bookkeeping kernels around the drafter

  Speculative DecodingTriton backend
  NVIDIA/TensorRT-LLM#20548 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Performance]: Kimi K3 decode runs the MoE-tail additions as separate kernels before the next attention-residual

  General perfSpeculative Decoding
  NVIDIA/TensorRT-LLM#20545 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Performance]: Kimi K3 MoE decode runs the router GEMM and the routing kernel in front of FC1 on the main stream

  General perfSpeculative Decoding
  NVIDIA/TensorRT-LLM#20540 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Performance]: CuTe DSL MLA split-KV plan kernel takes ~75 us per decode step at small batch

  General perf
  NVIDIA/TensorRT-LLM#20538 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Performance]: Kimi K3 KDA prefill persistent kernels (fused K123, K4) hard-code a 148-block grid

  Customized kernels
  NVIDIA/TensorRT-LLM#20536 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Performance]: Kimi K2.5/K3 text-only prompts are tokenized through the slow-tokenizer pipeline (~0.5 s per 128K-token prompt)

  General perf
  NVIDIA/TensorRT-LLM#20533 · Filter by author lucifer1004 opened on Oct 19, 2026

• [Bug]: Expired etcd worker lease is reported as refreshed and blocks re-registration

  Disaggregated servingWindows
  NVIDIA/TensorRT-LLM#20529 · Filter by author codebreaker77 opened on Oct 18, 2026

• [Bug]: trtllm-gen FMHA JIT requires /usr/local/cuda/bin/nvcc (ignores CUDA_HOME / PATH) even with NVRTC, failing MLA generation and executor warmup when the toolkit lives elsewhere

  Customized kernels
  NVIDIA/TensorRT-LLM#20515 · Filter by author lucifer1004 opened on Oct 17, 2026

• [Bug]: DeepSeek-V4.1-Flash fails at model init on Hopper (SM90): "CSA2 native MXFP8 projections require SM100/SM103"

  NVIDIA/TensorRT-LLM#20511 · Filter by author Smallfu666 opened on Oct 17, 2026

• [Performance] CuTe DSL MLA FP8-KV decode kernel slower at long KV after #19338

  Customized kernels
  NVIDIA/TensorRT-LLM#20510 · Filter by author lucifer1004 opened on Oct 17, 2026
  1

• [Bug]: Kimi K3 / MLA unit tests on SM107: two KDA files skip SM107, a cute_dsl_mla routing case contradicts the perf table, and an indexed-pool decode test compares two code paths bit-exactly

  Testing
  NVIDIA/TensorRT-LLM#20509 · Filter by author lucifer1004 opened on Oct 17, 2026
  6

• [Bug]: FP8 block-scale Linear hits a device-side assert on SM107: DeepGEMM JIT compiles for sm_107f and its CUTLASS disables the TMEM paths

  Customized kernelsPytorch
  NVIDIA/TensorRT-LLM#20508 · Filter by author lucifer1004 opened on Oct 17, 2026

• [Bug]: trtllm-gen FMHA CudaArchDecl.h numbers Sm107a as 4, but the prebuilt libTrtLlmGen archives use 4 for 105a and 5 for 107a

  Customized kernels
  NVIDIA/TensorRT-LLM#20507 · Filter by author lucifer1004 opened on Oct 17, 2026
  1

• [Bug]: MLA generation-only thop.attention calls size the workspace by the context requests' total KV length (OOM on mixed steps with long cached contexts)

  Inference runtimePytorch
  NVIDIA/TensorRT-LLM#20506 · Filter by author lucifer1004 opened on Oct 17, 2026

• [Bug]: Breakable prefill CUDA graphs replay corrupted outputs for some buckets: the shared output buffer is only weakly referenced

  Customized kernels
  NVIDIA/TensorRT-LLM#20504 · Filter by author lucifer1004 opened on Oct 17, 2026
  1