Skip to content
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions docs/helion-hackathon.md
Original file line number Diff line number Diff line change
Expand Up @@ -291,11 +291,11 @@ def my_kernel(...):

The B200 instances also ship with **nvtriton**, NVIDIA's extended Triton compiler that includes a **TileIR** backend — an alternative compilation pipeline that bypasses LLVM and compiles directly to CUBIN via NVIDIA's `tileiras` compiler.

| | `ENABLE_TILE=0` (default) | `ENABLE_TILE=1` |
| | `ENABLE_TILE=0` (default) | `ENABLE_TILE=1` + `HELION_BACKEND=tileir` |
|---|---|---|
| **Helion backend** | `triton` | `tileir` |

**Step 1: Enable TileIR and autotune.** Set the env vars before importing Helion, then autotune as usual. Helion automatically adjusts the search space for the TileIR backend.
**Step 1: Enable TileIR and autotune.** Set both `ENABLE_TILE=1` and `HELION_BACKEND=tileir` env vars before importing Helion, then autotune as usual. Helion automatically adjusts the search space for the TileIR backend.

**Step 2: Hardcode the TileIR config in your submission.** Copy the best config from the autotuner output (it will include TileIR-specific fields like `num_ctas` and `occupancy`). The env vars must be set before imports:

Expand All @@ -320,7 +320,7 @@ def my_kernel(...):

### Which should I use?

Try both `ENABLE_TILE=0` and `ENABLE_TILE=1`, with and without ACFs, then submit whichever gives the best benchmark numbers.
Try both the default backend (`ENABLE_TILE=0`) and the TileIR backend (`ENABLE_TILE=1` + `HELION_BACKEND=tileir`), with and without ACFs, then submit whichever gives the best benchmark numbers.

## Tips

Expand Down
Loading