[Kernel][Helion] Disable unsafe B200 RMS reduction warp specialization (#50345)

Signed-off-by: Shangdi Yu <shangdiy@meta.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Shangdi Yu
2026-07-30 20:37:52 +00:00
committed by GitHub
co-authored by OpenAI Codex
parent 0eec856cc3
commit 45b60e3914
2 changed files with 19 additions and 6 deletions
@@ -136,6 +136,19 @@ class TestRmsNormPerBlockQuantConfigPicker:
{"hidden_size": 4096, "group_size": 128, "num_tokens": 32}
)
def test_b200_configs_disable_reduction_warp_specialization(self):
config_set = ConfigManager.get_instance().load_config_set(
"rms_norm_per_block_quant"
)
configs = config_set.to_dict()["nvidia_b200"].values()
# TODO: Remove once the Triton pin includes
# https://github.com/triton-lang/triton/pull/9716. Tracked by
# https://github.com/triton-lang/triton/issues/10901.
assert all(
config["range_warp_specializes"][1] is not True for config in configs
)
DTYPES = [torch.bfloat16, torch.float]
QUANT_DTYPES = [torch.int8, FP8_DTYPE]
@@ -234,7 +234,7 @@
],
"range_warp_specializes": [
null,
true,
false,
null,
null
],
@@ -1354,7 +1354,7 @@
],
"range_warp_specializes": [
null,
true,
false,
null,
null
],
@@ -1564,7 +1564,7 @@
],
"range_warp_specializes": [
null,
true,
false,
null,
null
],
@@ -1634,7 +1634,7 @@
],
"range_warp_specializes": [
null,
true,
false,
null,
null
],
@@ -2546,7 +2546,7 @@
],
"range_warp_specializes": [
null,
true,
false,
null,
null
],
@@ -2686,7 +2686,7 @@
],
"range_warp_specializes": [
null,
true,
false,
null,
null
],