[PyTorch][torch.compile] Add TensorProto mechanism - #3153

Merged

ptrendx merged 35 commits into

NVIDIA:mainfrom

pggPL:tensor_proto_mechanism

Aug 5, 2026

Commits on Jul 7, 2026

Add TensorProto mechanism for data-free quantized tensor allocation

pggPL
and
claude
committed
6d3eece
[PyTorch] torch.compile: dedup cached FP8 weight from saved-for-backward

pggPL
and
claude
committed
4bdaa60
[PyTorch] nvfp4: emit _describe_buffers in canonical flatten order

pggPL
and
claude
committed
3be6fc7
Address review: error on undescribed buffers, gate nvfp4 test on HW support

pggPL
and
claude
committed
8a9d90c
[PyTorch] Workaround torch.compile staticmethod guard bug in NVFP4 _describe_buffers

pggPL
and
claude
committed
e36cf6d

Commits on Jul 13, 2026

[PyTorch] Document TensorOrQuantized union (review feedback)
pggPL
committed
90fc494
[PyTorch] Add shape property to QuantizedTensorStorage
pggPL
committed
a3bf5c1
[PyTorch] Test parity of Python alloc vs C++ make_empty
pggPL
committed
37b0d17
[PyTorch] Simplify comment on captured requires_grad flags
pggPL
committed
4feacdc
[PyTorch] Drop redundant None guards in wt_save alias dedup
pggPL
committed
5adb6e4
[PyTorch] Use torch._prims_common.make_contiguous_strides_for
pggPL
committed
40d2d7b
[PyTorch] Raise on update_usage of a non-quantized TensorProto
pggPL
committed
78e95a6
[PyTorch] Collapse to_tensor_proto branches
pggPL
committed
b3229cc
[PyTorch] Drop dead weight_fp8 fallback in fake backward
pggPL
committed
2b19368
[PyTorch] Drop redundant comment on saved_weight
pggPL
committed
c3dec83
Merge branch 'main' of https://github.com/NVIDIA/TransformerEngine into tensor_proto_mechanism
pggPL
committed
cb53cc3

Commits on Jul 15, 2026

[PyTorch] Add TensorProto.assemble for rebuilding from ready-made buffers
pggPL
committed
dd99206

Commits on Jul 27, 2026

Merge remote-tracking branch 'upstream/main' into tensor_proto_mechanism
pggPL
committed
2c28aa1

Commits on Jul 28, 2026

Commits on Jul 29, 2026

Commits on Jul 31, 2026

Merge remote-tracking branch 'upstream/main' into tensor_proto_mechanism
pggPL
committed
0228c9c
Raise when a parameter vanishes during _apply instead of skipping it
pggPL
committed
a24e4e0
Drop the vanished-buffer check from _apply
pggPL
committed
a8b8873
Declare flat buffers on the field instead of in a parallel tuple
pggPL
committed
6bb030f
Name the flat buffers after PyTorch's own term for them
pggPL
committed
e0bbb18
Carry the storage class itself through the flatten context
pggPL
committed
9a4fa8a

Commits on Aug 4, 2026

Rename TensorProto to TensorSpec and tighten the inner-tensor contract

pggPL
and
claude
committed
0fbb4af
Restore three unrelated comments mangled by the buffers/inner-tensors rename

pggPL
and
claude
committed
9de57b3
Merge branch 'main' into tensor_proto_mechanism

pggPL
and
claude
committed
5612f13