Skip to content

Commit bdabf6c

Browse files
authored
Add Hopper mega moe test to runner (#46)
1 parent b5238ad commit bdabf6c

1 file changed

Lines changed: 19 additions & 5 deletions

File tree

sgl_deep_gemm/run_tests.sh

Lines changed: 19 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -37,7 +37,7 @@ if [ "${NUM_GPUS}" -eq 0 ]; then
3737
echo "No GPUs visible to nvidia-smi — DeepGEMM tests require a GPU." >&2
3838
exit 1
3939
fi
40-
# arch major: 9 == Hopper (SM90), 10 == Blackwell (SM100).
40+
# arch major: 9 == Hopper (SM90), 10 == Blackwell (SM100/SM103).
4141
COMPUTE_CAP=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null | head -1)
4242
ARCH_MAJOR=${COMPUTE_CAP%%.*}
4343

@@ -123,15 +123,23 @@ if [ -f "${TESTS_DIR}/test_lazy_init.py" ]; then
123123
skip_test test_lazy_init.py "tvm_ffi eager CUDA init (upstream)"
124124
fi
125125

126-
# mega_moe family uses SM100 fp4 + symmetric-memory kernels (SM100-only).
126+
# mega_moe family covers Blackwell fp4 + symmetric-memory kernels and the SM90
127+
# Hopper fp8 path.
127128
# test_mega_moe.py additionally needs deep_ep (with ElasticBuffer); the l1 and
128129
# pre_dispatch tests use deep_gemm's own symmetric buffer.
129-
MEGA_MOE_ALL=(
130+
MEGA_MOE_BLACKWELL=(
130131
test_mega_moe.py
131132
test_mega_moe_l1_fp4_accuracy.py
132133
test_mega_moe_l1_sentinel.py
133134
test_mega_moe_pre_dispatch.py
134135
)
136+
MEGA_MOE_HOPPER=(
137+
test_mega_moe_hopper.py
138+
)
139+
MEGA_MOE_ALL=(
140+
"${MEGA_MOE_BLACKWELL[@]}"
141+
"${MEGA_MOE_HOPPER[@]}"
142+
)
135143
MEGA_MOE_L1=(
136144
test_mega_moe_l1_fp4_accuracy.py
137145
test_mega_moe_l1_sentinel.py
@@ -156,9 +164,15 @@ elif [ "${ARCH_MAJOR}" -ge 10 ]; then
156164
[ -f "${TESTS_DIR}/${t}" ] && skip_test "${t}" "fp4 kernel failures, see comment (confirm on B200)"
157165
done
158166
[ -f "${TESTS_DIR}/test_mega_moe_pre_dispatch.py" ] && run_test test_mega_moe_pre_dispatch.py
167+
for t in "${MEGA_MOE_HOPPER[@]}"; do
168+
[ -f "${TESTS_DIR}/${t}" ] && skip_test "${t}" "SM90-only, arch major ${ARCH_MAJOR}"
169+
done
159170
else
160-
for t in "${MEGA_MOE_ALL[@]}"; do
161-
[ -f "${TESTS_DIR}/${t}" ] && skip_test "${t}" "SM100-only, arch major ${ARCH_MAJOR}"
171+
for t in "${MEGA_MOE_BLACKWELL[@]}"; do
172+
[ -f "${TESTS_DIR}/${t}" ] && skip_test "${t}" "Blackwell-only, arch major ${ARCH_MAJOR}"
173+
done
174+
for t in "${MEGA_MOE_HOPPER[@]}"; do
175+
[ -f "${TESTS_DIR}/${t}" ] && run_test "${t}" --num-processes "${NPROC}"
162176
done
163177
fi
164178

0 commit comments

Comments
 (0)