Issues

• Remove qnn.lower, lower_cpu_reference and the .pte-writing parts, and QNN.md section 2 (D12, owner forbade any .pte pipeline)

  thisisthepy/torchnative#277 · Filter by author b-re-w opened on Nov 13, 2026

• torchnative.compile: backend="npu" and "inductor" run the graph eagerly; refuse by name

  thisisthepy/torchnative#274 · Filter by author b-re-w opened on Nov 13, 2026

• ci: dispatch-only workflow for the distributed suites

  thisisthepy/torchnative#273 · Filter by author b-re-w opened on Nov 13, 2026

• torch.compile: a view graph input cannot be fake-ified (TensorBase._base and four walls after it)

  thisisthepy/torchnative#272 · Filter by author b-re-w opened on Nov 13, 2026

• develop gate red: npufuse TypeDict guard fails on its own frame (since #208) and macOS rustc 1.100.0 abort (#266)

  thisisthepy/torchnative#270 · Filter by author b-re-w opened on Nov 13, 2026

• torchnative.compile: Dynamo tracing of a Qwen3 decode step with DynamicCache reaches Tensor._base, which the shim refuses

  thisisthepy/torchnative#269 · Filter by author b-re-w opened on Nov 12, 2026

• shim: torch.tensor(x, dtype=int) refuses Python's int as a dtype (blocks transformers StaticCache)

  thisisthepy/torchnative#268 · Filter by author b-re-w opened on Nov 12, 2026

• compile: StaticCache as the default KV cache on compiled devices

  thisisthepy/torchnative#263 · Filter by author b-re-w opened on Nov 12, 2026

• compile: dynamic shapes (SymInt) so a sequence-length range compiles once

  thisisthepy/torchnative#262 · Filter by author b-re-w opened on Nov 12, 2026

• QNN libraries route (a): never shipped in the wheel; found at run time on the device or in the app, refused by name when absent

  thisisthepy/torchnative#260 · Filter by author b-re-w opened on Nov 12, 2026

• eager tape after #251: a write-poisoned tape never recovers once its graph dies; save_for_backward(output) pins its graph

  thisisthepy/torchnative#259 · Filter by author b-re-w opened on Nov 12, 2026

• torchnative.compile: a failed capture on a compiled device falls back to eager silently (attach_to_module); refuse by name

  thisisthepy/torchnative#255 · Filter by author b-re-w opened on Nov 12, 2026

• CPU: in-place index_copy_ on a view costs O(base storage), making transformers continuous batching ~27 s/token

  thisisthepy/torchnative#254 · Filter by author b-re-w opened on Nov 12, 2026

• federated: secure aggregation by pairwise masking (Bonawitz-style) over the p2p mesh (part of #20)

  thisisthepy/torchnative#253 · Filter by author b-re-w opened on Nov 12, 2026

• eager tape: an abandoned forward plus an in-place write poisons every later backward, even on another model

  thisisthepy/torchnative#251 · Filter by author b-re-w opened on Nov 12, 2026
  1

• tape: nll_loss backward fails for reduction='none' and for class weights (reshape to [1])

  thisisthepy/torchnative#249 · Filter by author b-re-w opened on Nov 12, 2026

• distributed: DistributedDataParallel runs unchanged on backend torchnative through a Python Reducer (BACKEND.md D13, step 12 part)

  thisisthepy/torchnative#247 · Filter by author b-re-w opened on Nov 12, 2026

• publish-pypi: host and iOS wheel jobs still run on macos-14, which GitHub no longer offers

  thisisthepy/torchnative#244 · Filter by author b-re-w opened on Nov 12, 2026

• distributed: tensor-parallel inference of a real transformers model on backend torchnative (BACKEND.md step 9)

  thisisthepy/torchnative#242 · Filter by author b-re-w opened on Nov 12, 2026

• ProcessGroupLocal: _reduce_scatter_base / alltoall_base with async_op=True may publish an unfilled buffer (read, not run)

  thisisthepy/torchnative#241 · Filter by author b-re-w opened on Nov 12, 2026

• distributed: swappable server interface (rendezvous, address exchange) and opt-in pre-shared key (BACKEND.md step 5)

  thisisthepy/torchnative#233 · Filter by author b-re-w opened on Nov 12, 2026

• distributed: collectives of backend torchnative on the binary wire and the full mesh, parity with gloo (BACKEND.md step 3)

  thisisthepy/torchnative#232 · Filter by author b-re-w opened on Nov 12, 2026

• distributed: send_object_list / recv_object_list need UntypedStorage.from_buffer and ByteTensor(storage) in the shim

  thisisthepy/torchnative#229 · Filter by author b-re-w opened on Nov 12, 2026

• shim: Python torch.library kernels for custom namespaces dispatch through _aten_dispatch; truthful has_kernel_for_dispatch_key

  thisisthepy/torchnative#227 · Filter by author b-re-w opened on Nov 12, 2026

• distributed: new_group / split_group and an N-D DeviceMesh on backend torchnative, against gloo (BACKEND.md step 8)

  thisisthepy/torchnative#221 · Filter by author b-re-w opened on Nov 12, 2026