Repository navigation

[PyTorch] Add torch.compile support for fused MoE routing - #3691

Open

pggPL wants to merge 14 commits into

NVIDIA:mainfrom

pggPL:fused_router_compile

Open

[PyTorch] Add torch.compile support for fused MoE routing#3691
pggPL wants to merge 14 commits into
NVIDIA:mainfrom
pggPL:fused_router_compile

Commits on Oct 14, 2026
Commits on Oct 15, 2026
- Reuse router numerical tests for selected torch.compile configurations
  pggPL
  committed
  f5c02da
- Compile router buffer mutations through autograd.Function
  pggPL
  committed
  e0aa561
- Skip QB bounds value validation under torch.compile
  pggPL
  committed
  9f24dfc
- Simplify compiled router inference context handling
  pggPL
  committed
  a3a6a10
- Remove unrelated router test formatting change
  pggPL
  committed
  ea3275c
- Simplify fused router tests and helper names
  pggPL
  committed
  41219e9
- Remove dedicated compiled QB bounds validation tests
  pggPL
  committed
  80b7920
- Document router buffer constraints beside their parameters
  pggPL
  committed
  b8d8eec
- Use setup_context for compiled fused routing
  pggPL
  committed
  9282a4b
Commits on Oct 16, 2026
- Check that dynamic router inputs do not trigger recompilation
  pggPL
  committed
  eb9e922
- Check fused router backward custom operators with opcheck
  pggPL
  committed
  c2fb139