Issues

• [Runtime] nrt_init intermittently aborts with "Assertion 'vcore != NULL' failed" in multi-process runs (2.35.30, 2.35.33)

  buginference
  aws-neuron/aws-neuron-sdk#1492 · Filter by author EricRobert opened on Nov 12, 2026
  1

• [Runtime] nrt_tensor_write never returns after a zero-copy page-pin failure

  bug
  aws-neuron/aws-neuron-sdk#1491 · Filter by author EricRobert opened on Nov 12, 2026
  1

• [Runtime/Docs] Collectives fail between trn3.48xlarge devices not linked in neuron-ls; error message and docs don't explain why

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1490 · Filter by author EricRobert opened on Nov 12, 2026
  1

• [Tooling] Access to partial execution traces for hung or incomplete kernels

  bugtrainingTrn1
  aws-neuron/aws-neuron-sdk#1489 · Filter by author EricRobert opened on Nov 12, 2026
  1

• [NKI] Loop-control tile may be reused before all consuming engines read it, causing inconsistent dynamic-loop trip counts (hang)

  bugtrainingTrn1
  aws-neuron/aws-neuron-sdk#1488 · Filter by author EricRobert opened on Nov 12, 2026
  1

• [NKI] Feature request: 2-D vector-offset scatter (DMA)

  bugtrainingTrn1
  aws-neuron/aws-neuron-sdk#1487 · Filter by author EricRobert opened on Nov 12, 2026
  1

• [Trn3 runtime] Feature request: one-sided device-to-device put (or working send/recv) in libnrt

  bugtrainingTrn1
  aws-neuron/aws-neuron-sdk#1486 · Filter by author EricRobert opened on Nov 12, 2026
  1

• [Trn3 runtime 2.35.30] nrt_cc_global_comm_init: undocumented g_id mod 4 rule; silent hang when g_count exceeds joined ids

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1485 · Filter by author EricRobert opened on Nov 12, 2026
  1

• [Trn3 runtime 2.35.30] nrt_async_sendrecv_* fails with "Unsupported intance type"; max communicators per lnc is 0

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1484 · Filter by author EricRobert opened on Nov 12, 2026
  1

• [neuronx-cc 2.28.7631 / Trn3] Second execution of a TF32 auto-cast matmul times out and terminates the process

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1482 · Filter by author john-horacsek-yasp opened on Nov 11, 2026
  1

• [NKI 0.7.0] nc_matmul rejects documented tfloat32 operands as !nisa.float32r

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1481 · Filter by author john-horacsek-yasp opened on Nov 11, 2026
  2

• [Trn3 runtime 2.35.30] Four-rank all_reduce aborts at Cayman architecture assertion in tdrv_get_bdf_workaround

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1480 · Filter by author john-horacsek-yasp opened on Nov 11, 2026
  1

• [NKI 0.7.0 / Trn3] DGE DMA silently misaddresses HBM at a 4 GiB partition span

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1479 · Filter by author john-horacsek-yasp opened on Nov 11, 2026
  1

• [neuronx-cc 2.28.7631] torch.compile fails with NCC_IBCG901 due to unaligned tensorcopy partitions

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1478 · Filter by author john-horacsek-yasp opened on Nov 11, 2026
  1

• [NKI 0.7.0] nl.while_loop with register_alloc(x=4) fails lowering with Could not find register

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1477 · Filter by author john-horacsek-yasp opened on Nov 11, 2026
  2

• [NKI 0.7.0] nc_transpose rejects tfloat32 input on Tensor and Vector engines

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1476 · Filter by author john-horacsek-yasp opened on Nov 11, 2026
  1

• [NKI 0.7.0] nl.power accepts scalar/broadcast forms at API level but lowering rejects op0 pow

  feature-requestinferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1475 · Filter by author john-horacsek-yasp opened on Nov 11, 2026
  1

• [NKI 0.7.0] local_gather with 32 partitions times out; 128-partition control passes

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1474 · Filter by author john-horacsek-yasp opened on Nov 11, 2026
  1

• [NKI 0.7.0] Explicit nc_matmul accumulate flags silently corrupt wide PSUM destinations after column 512

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1473 · Filter by author john-horacsek-yasp opened on Nov 11, 2026
  3

• [NKI 0.7.0] local_gather silently derives run length from dst shape instead of num_elem_per_idx

  buginferencetrainingTrn1
  aws-neuron/aws-neuron-sdk#1472 · Filter by author john-horacsek-yasp opened on Nov 11, 2026
  2

• [Trn3] 64-rank direct DMA times out with 16-row transfers; 8-row control passes

  bugtrainingTrn1
  aws-neuron/aws-neuron-sdk#1469 · Filter by author qdong-openai opened on Nov 10, 2026
  1

• [torch-neuronx 2.14.3] index_select / advanced indexing rejects valid indices with "index out of range" when float32(i) >= float32(N), for sources with more than 2^24 rows (bounds check compares in float32)

  bugtrainingTrn2
  aws-neuron/aws-neuron-sdk#1468 · Filter by author scttfrdmn opened on Nov 10, 2026
  1

• [torch-neuronx 2.14.3] Silent wrong results: device nonzero / where(mask) / argwhere / masked_select / A[mask] round flat indices through float32 on tensors with more than 2^24 elements

  bugtrainingTrn2
  aws-neuron/aws-neuron-sdk#1467 · Filter by author scttfrdmn opened on Nov 10, 2026
  1

• [torch-neuronx 2.14.3] Silent wrong results: device transpose/permute of int32 (and int64) tensors rounds values through float32 (|x| > 2^24 loses low bits); uint32/float32 of the same bits are exact

  bugtrainingTrn2
  aws-neuron/aws-neuron-sdk#1466 · Filter by author scttfrdmn opened on Nov 10, 2026
  2

• [Trn3 driver] Whole-chip reset fails after the buffer-handle table stops accepting allocations

  bug
  aws-neuron/aws-neuron-sdk#1465 · Filter by author qdong-openai opened on Nov 9, 2026
  1