Repository navigation

[PyTorch] Avoid concatenation copies for compiled split Linear parameters - #3623

Draft

pggPL wants to merge 13 commits into

NVIDIA:mainfrom

pggPL:split_parameters_linear

Draft

[PyTorch] Avoid concatenation copies for compiled split Linear parameters#3623
pggPL wants to merge 13 commits into
NVIDIA:mainfrom
pggPL:split_parameters_linear

Commits on Oct 5, 2026
- [PyTorch] Avoid concatenation copies for compiled split Linear parameters
  pggPL
  committed
  fdaea9e
- [PyTorch] Make deferred parameter concatenation explicit
  pggPL
  committed
  289cff1
- [PyTorch] Simplify split Linear parameter handling
  pggPL
  committed
  43a33bb
- [PyTorch] Limit the split-parameter API refactor
  pggPL
  committed
  5d5c180
Commits on Oct 12, 2026
- Merge main into split_parameters_linear
  pggPL
  committed
  bdb5bc1
- [PyTorch] Keep the frozen-weight bias fix out of split parameter support
  pggPL
  committed
  73d58af
Commits on Oct 14, 2026
- [PyTorch] Read split parameter shape and gradient flags directly
  pggPL
  committed
  122a2c9
- [PyTorch] Pass split parameters as tensor lists
  pggPL
  committed
  156576e
- [PyTorch] Use representation-neutral names for Linear operands
  pggPL
  committed
  dd01b3a
Commits on Oct 16, 2026
- [PyTorch] Pass split Linear weights in separate tensor-list fields
  pggPL
  committed
  88b47c3
- [PyTorch] Skip empty split-parameter bookkeeping in eager Linear
  pggPL
  committed
  bcf13f6
- [PyTorch] Make deferred Linear parameter lists optional
  pggPL
  committed
  623baf8
- [PyTorch] Trim split parameter tests
  pggPL
  committed
  7b28093