[Build] Target torch 2.10 + add the Triton codegen route (draft) - #305

Draft

YWHyuk wants to merge 35 commits into

feature/triton-codegen

Commits on Jul 27, 2026

[Frontend] Add the Triton codegen route (Inductor triton + tnpu passes)
YWHyuk
committed
4c077b4
[Build] Target torch 2.10 so the frontend and triton-npu agree on triton 3.6
YWHyuk
committed
5686735
[CI] Add the triton-npu toolchain image and its workflow
YWHyuk
committed
ef76c24
[CI] Source triton-npu from PSAL-POSTECH and drop the local workarounds
YWHyuk
committed
2173834
[CI] Note only the token as a prerequisite
YWHyuk
committed
e4af91b
[CI] Bump the triton-npu pin to de2c767
YWHyuk
committed
5f632d6
[CI] Bump the triton-npu pin to e113819
YWHyuk
committed
8faa1e5
[CI] Bump the triton-npu pin to fd423b0
YWHyuk
committed
b442aa0
[Frontend] Rename print_operation to visit_operation
YWHyuk
committed
600a6a9
[Frontend] Give the trace pipeline a kernel whose body is one work-item
YWHyuk
committed
c0df618
[Frontend] Run the Triton route's kernel through TOGSim
YWHyuk
committed
15ce2a1
[Frontend] Measure the Triton route's tile cost with gem5
YWHyuk
committed
3c82969
[Frontend] Let the route ask for a multi-axis grid
YWHyuk
committed
85c4e66
[Frontend] Take the grid at run time, so one trace serves every shape
YWHyuk
committed
9657d38
[Frontend] Pin only the blocks the kernel takes as parameters
YWHyuk
committed
8da4bbc
[Frontend] Pass the kernel's scalar arguments to the tnpu wrapper
YWHyuk
committed
cdd53c0
[Frontend] Run the Triton route's kernel on the launch's own tensors
YWHyuk
committed
5e77a5f
[CI] Bump the triton-npu pin to 22df065
YWHyuk
committed
68a6aae

Commits on Jul 28, 2026

[Docs] Report the Triton route against the MLIR one
YWHyuk
committed
5f52306
[Docs] State where the port ends and the lowering pass begins
YWHyuk
committed
05e4edc
[Docs] Make the next step coverage parity, measured rather than guessed
YWHyuk
committed
6670e5f
[Docs] Rewrite the report as one piece
YWHyuk
committed
c099add
[Docs] Say lowering pass in prose, keep tnpu only where it is a real name
YWHyuk
committed
e3672a8
[Docs] Call the lowering layer the NPU lowering pass
YWHyuk
committed
83bce51
[Docs] Show the grid being rebuilt as the outer loop
YWHyuk
committed
6ddab17
[Docs] Make model coverage the goal, op coverage the step to it
YWHyuk
committed
f8f3d0a
[Docs] Keep the remaining work high level
YWHyuk
committed
ebbd113

Commits on Aug 3, 2026

[CI] Bump the triton-npu pin to 4835b38, on the released spike
YWHyuk
committed
31baf29
[CI] Bump the triton-npu pin to 5d84caf
YWHyuk
committed
d37d471
[CI] Set the spike override in the image, not in each job
YWHyuk
committed
65ada99
[CI] Run the whole test suite on the Triton route, and report what fails
YWHyuk
committed
f8e744d
[Frontend] Say why a kernel was rejected, and run the sweep in parallel
YWHyuk
committed
d356e35
[Docs] Record the first Triton route coverage measurement
YWHyuk
committed
8e17519
[Docs] Back the coverage report with the artifacts it came from
YWHyuk
committed
6e3bd7e
[Docs] Add the Korean coverage report
YWHyuk
committed
33a222a