Skip to content

Hopper: build joyomni_ops for sm_90a, not sm_90 (FP8 GEMM aborts at launch on H100) #27

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
stayinalive181 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Hopper: build joyomni_ops for sm_90a, not sm_90 (FP8 GEMM aborts at launch on H100) #27

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
2 changes: 1 addition & 1 deletion DEPLOYMENT.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -126,7 +126,7 @@ Then install the attention and kernel dependencies:
git -C deploy/tmp/cutlass checkout dcf215af
# build only this machine's arch (the default is a 5-arch fat binary —
# sm_80..120a — which multiplies compile time ~5x); auto-detected:
export JOYOMNI_OPS_CUDA_ARCHS=$(python -c "import torch; cc = torch.cuda.get_device_capability(0); print(f'{cc[0]}{cc[1]}a' if cc[0] >= 10 else f'{cc[0]}{cc[1]}')")
export JOYOMNI_OPS_CUDA_ARCHS=$(python -c "import torch; cc = torch.cuda.get_device_capability(0); print(f'{cc[0]}{cc[1]}a' if cc[0] >= 9 else f'{cc[0]}{cc[1]}')")
echo "building joyomni_ops for sm_$JOYOMNI_OPS_CUDA_ARCHS"
JOYOMNI_OPS_CUTLASS_DIR=$(pwd)/deploy/tmp/cutlass \
python -m pip install --no-build-isolation ./deploy/joyomni_ops
Expand Down
9 changes: 7 additions & 2 deletions deploy/joyomni_ops/setup.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@
- fp8_scaled_mm : FP8 per-token x per-channel scaled GEMM (cutlass)

GPU arch coverage is chosen from the local nvcc version:
- always: sm_80, sm_89, sm_90
- always: sm_80, sm_89, sm_90a
- CUDA >= 12.8: also sm_100a (B200) and sm_120a (RTX PRO 6000 / RTX 5090)
So building on a CUDA 12.8+ toolchain automatically yields Blackwell support.
"""
Expand Down Expand Up @@ -57,7 +57,12 @@ def _gencodes():
flags = [
"-gencode=arch=compute_80,code=sm_80",
"-gencode=arch=compute_89,code=sm_89",
"-gencode=arch=compute_90,code=sm_90",
# Hopper must be sm_90a: fp8_scaled_mm selects cutlass'
# KernelTmaWarpSpecializedPingpongFP8FastAccum, whose WGMMA path is
# arch-conditional. Built as plain sm_90 it compiles and loads, then
# aborts every launch ("Arch conditional MMA instruction used without
# targeting appropriate compute capability").
"-gencode=arch=compute_90a,code=sm_90a",
]
if (mj, mn) >= (12, 8):
flags += [
Expand Down