Issues

• [NKI] nki_matmul_fully_optimized_ (matmul tutorial) rounds its K-block partial sums to BF16; error grows with K

  buginferencetrainingTrn1Trn2
  aws-neuron/aws-neuron-sdk#1422 · Filter by author jithendra1798 opened on Oct 10, 2026
  2

• neuron-monitor: a period such as "1.0s" is silently replaced by the 5 s default

  bugInf2training
  aws-neuron/aws-neuron-sdk#1419 · Filter by author saqibkh opened on Oct 9, 2026
  1

• nl.fori_loop raise an error about defining tensor outside the loop, even when all tensors are already defined outside the loop

  buginferenceTrn2
  aws-neuron/aws-neuron-sdk#1418 · Filter by author OswaldHe opened on Oct 9, 2026
  2

• nisa.tensor_scalar does not support different data width between data and dst for bitvec operation

  buginferenceTrn2
  aws-neuron/aws-neuron-sdk#1417 · Filter by author OswaldHe opened on Oct 7, 2026
  2

• torch.cumsum on bool/integer XLA tensors keeps the input dtype instead of promoting to int64 (wrong position ids in SmolVLA and pi0.5 with torch-neuronx 2.9)

  bugInf2inference
  aws-neuron/aws-neuron-sdk#1415 · Filter by author quinncheong opened on Oct 6, 2026
  2

• nisa.activation silently compile reduce_op to nl.add no matter what operation is used

  buginferenceTrn2
  aws-neuron/aws-neuron-sdk#1414 · Filter by author OswaldHe opened on Oct 5, 2026
  2

• Backward Compatibility of nl.fori_loop: Missing support for LNC=1 and carried buffer in PSUM across loop iterations.

  buginferencetrainingTrn2
  aws-neuron/aws-neuron-sdk#1408 · Filter by author OswaldHe opened on Oct 3, 2026
  2

• Simplifier error NCC_ISMP902 compiling Qwen3-0.6B on trn1 and inf2 (not sequence-length related, unlike #1391)

  bugInf1Inf2inferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1407 · Filter by author istiak133 opened on Oct 3, 2026
  1

• nki.isa.nc_matmul with accumulation=False reset the whole PSUM buffer instead of overwrite.

  buginferenceTrn2
  aws-neuron/aws-neuron-sdk#1405 · Filter by author OswaldHe opened on Oct 1, 2026
  2

• torch_neuronx.trace silently miscomputes x.permute(0, 2, 1) + weight: the [T, C] weight is read as [C, T] - wrong result (buffer read as transposed)

  bugInf1Inf2inferencetraining
  aws-neuron/aws-neuron-sdk#1403 · Filter by author varuntej07 opened on Sep 28, 2026
  1

• nisa.scalar_tensor_tensor supports both full arithmetic operators and full bitvec operators.

  bugdocumentation
  aws-neuron/aws-neuron-sdk#1402 · Filter by author OswaldHe opened on Sep 28, 2026
  2

• Faster MXFP4 dequantization on Trn 2

  bugfeature-requestTrn2
  aws-neuron/aws-neuron-sdk#1401 · Filter by author OswaldHe opened on Sep 28, 2026
  3

• AwsNeuronTopK returns wrong indices for some reduction sizes when batch > 1 (silent; e.g. 151936)

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1400 · Filter by author liuhanzuo opened on Sep 28, 2026
  1

• Matrix transpose miss full support for e4m3fn data type on TRN2

  bugTrn2
  aws-neuron/aws-neuron-sdk#1399 · Filter by author OswaldHe opened on Sep 25, 2026
  2

• Trn3 Interconnect Bandwidth Documentation Clarification

  bugdocumentationinference
  aws-neuron/aws-neuron-sdk#1396 · Filter by author sikkaaman opened on Sep 18, 2026
  1

• neuronx-cc compile --help crashes with "unsupported format character"

  bugTrn2
  aws-neuron/aws-neuron-sdk#1394 · Filter by author dadmobile opened on Sep 9, 2026
  2

• Is there a supported FP8 storage dtype on TRN2? F8E4M3FN is rejected even when never used as an operand

  buginferencetrainingTrn1Trn2
  aws-neuron/aws-neuron-sdk#1393 · Filter by author dadmobile opened on Sep 9, 2026
  3

• NCC_ILNK020: "tensor_map verification failed after linking" above ~12 layers

  buginferencetrainingTrn2
  aws-neuron/aws-neuron-sdk#1392 · Filter by author dadmobile opened on Sep 9, 2026
  1

• neuronx-cc 2.27.5334: Simplifier is_subset() crash (NCC_ISMP902, exit 70) on token_generation graph for sliding-window model at any seq_len != 4096

  bugInf1Inf2inference
  aws-neuron/aws-neuron-sdk#1391 · Filter by author barrhawk opened on Sep 3, 2026
  2

• [neuron runtime] indirect dma_transpose with oob_mode=skip causes DMA abort and persistent device failure

  buginferenceTrn2
  aws-neuron/aws-neuron-sdk#1390 · Filter by author hongsuh-oai opened on Sep 1, 2026
  3

• Silent miscompilation of masked assignment, plus 3 defects and a bf16 regression, on a dilated CNN (trn2, neuronxcc-2.26.6360.0)

  bugtrainingTrn1Trn2
  aws-neuron/aws-neuron-sdk#1381 · Filter by author cralbors opened on Aug 15, 2026
  5

• [neuron compiler backend] reg_dce incorrectly removes dependencies

  buginferenceTrn2
  aws-neuron/aws-neuron-sdk#1380 · Filter by author hongsuh-oai opened on Aug 14, 2026
  3

• Extremely long compile times and high compiler memory usage for some large NKI kernels

  bugcompilerTrn1Trn2
  aws-neuron/aws-neuron-sdk#1374 · Filter by author natetyoung opened on Aug 5, 2026
  2

• [NCC_ICFG001] shard_on_i_hybrid MoE kernel fails to compile at seq_len > 2048 (multiple exit nodes in sg0000)

  buginferenceTrn2
  aws-neuron/aws-neuron-sdk#1372 · Filter by author UbaidHunts opened on Jul 30, 2026
  1

• neuronx-cc INTERNAL_ERROR NCC_IBIR297 (bir::samePartitionsAll in ModuleForkPass) compiling an MLA + 256-expert MoE decode graph on trn2

  buginferenceTrn2
  aws-neuron/aws-neuron-sdk#1362 · Filter by author dadmobile opened on Jul 14, 2026
  2