Skip to content

[https://nvbugs/6644644][fix] Replay both verified orphan payloads onto current origin/main as a single… #18142

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

Merged
2ez4bz merged 3 commits into from
Aug 28, 2026
Merged

[https://nvbugs/6644644][fix] Replay both verified orphan payloads onto current origin/main as a single… #18142

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
b60e874
[nvbugs/6644644][fix] Import the CuTe DSL MoE runners where they are …
trtllm-agent Aug 24, 2026
3715cc6
Merge branch 'main' into repair-bot-bug6644644
2ez4bz Aug 28, 2026
983462a
Fix autoflake issue
2ez4bz Aug 28, 2026
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
33 changes: 22 additions & 11 deletions tensorrt_llm/_torch/modules/fused_moe/fused_moe_cute_dsl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,7 @@
from ...autotuner import (AutoTuner, ConstraintSpec, DynamicTensorSpec,
OptimizationProfile, TunableRunner, TuningConfig)
from ...custom_ops.cute_dsl_custom_ops import GroupedGemmInputsHelper
from ...cute_dsl_utils import IS_CUTLASS_DSL_AVAILABLE
from ...model_config import ModelConfig
from ...utils import (ActivationType, AuxStreamType, EventType,
Fp4QuantizedTensor,
Expand All @@ -39,6 +40,26 @@
from .quantization import MoEWeightLoadingMode, NVFP4CuteDslFusedMoEMethod
from .routing import BaseMoeRoutingMethod

# These runners are defined inside cute_dsl_custom_ops' ``if
# IS_CUTLASS_DSL_AVAILABLE:`` block, which has no else-branch, so importing them
# unconditionally would break every importer of this file -- and create_moe
# imports it eagerly under _torch.models, so that reaches all model startup
# rather than just this backend. Guard the same way the sibling custom_ops
# modules do, and leave the tuple empty when the DSL is absent: no CuteDSL
# runner can be tuned in that case, so nothing can match it.
_TILE_SIZE_CHECKED_RUNNERS: Tuple[type, ...] = ()
if IS_CUTLASS_DSL_AVAILABLE:
from ...custom_ops.cute_dsl_custom_ops import (
Sm100BlockScaledContiguousGatherGroupedGemmActFusionRunner,
Sm100BlockScaledContiguousGroupedGemmFinalizeFusionRunner,
Sm100BlockScaledContiguousGroupedGemmRunner,
Sm100BlockScaledContiguousGroupedGemmSwigluFusionRunner)
_TILE_SIZE_CHECKED_RUNNERS = (
Sm100BlockScaledContiguousGroupedGemmRunner,
Sm100BlockScaledContiguousGroupedGemmFinalizeFusionRunner,
Sm100BlockScaledContiguousGroupedGemmSwigluFusionRunner,
Sm100BlockScaledContiguousGatherGroupedGemmActFusionRunner)


@dataclass
class NvFp4WeightView:
Expand Down Expand Up @@ -328,19 +349,9 @@ def runner_tactic_comb_checker(
# eagerly under _torch.models, so that reaches all model startup rather
# than just this backend. Reaching this line means a CuteDSL runner is
# already being tuned, so the DSL is installed.
from ...custom_ops.cute_dsl_custom_ops import (
Sm100BlockScaledContiguousGatherGroupedGemmActFusionRunner,
Sm100BlockScaledContiguousGroupedGemmFinalizeFusionRunner,
Sm100BlockScaledContiguousGroupedGemmRunner,
Sm100BlockScaledContiguousGroupedGemmSwigluFusionRunner)

for runner, tactic in comb:
if isinstance(
runner,
(Sm100BlockScaledContiguousGroupedGemmRunner,
Sm100BlockScaledContiguousGroupedGemmFinalizeFusionRunner,
Sm100BlockScaledContiguousGroupedGemmSwigluFusionRunner,
Sm100BlockScaledContiguousGatherGroupedGemmActFusionRunner)):
if isinstance(runner, _TILE_SIZE_CHECKED_RUNNERS):
mma_tiler_mn, *_ = tactic
if mma_tiler_mn[0] != tile_size:
return False
Expand Down
5 changes: 5 additions & 0 deletions tests/unittest/_torch/thop/parallel/test_cute_dsl_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,7 @@
from utils.util import check_accuracy

from tensorrt_llm._torch.custom_ops.cute_dsl_custom_ops import GroupedGemmInputsHelper
from tensorrt_llm._torch.cute_dsl_utils import IS_CUTLASS_DSL_AVAILABLE
from tensorrt_llm._torch.modules.fused_moe.fused_moe_cute_dsl import cute_dsl_nvfp4_grouped_gemm_ref
from tensorrt_llm._torch.modules.fused_moe.quantization import interleave_linear_and_gate
from tensorrt_llm._torch.utils import (
Expand Down Expand Up @@ -413,6 +414,7 @@ def test_moe_gelu(dtype: str, num_tokens: int, top_k: int, tile_size: int):
get_sm_version() not in (100, 103),
reason="This test is only supported on SM 100 and SM 103 GPUs",
)
@pytest.mark.skipif(not IS_CUTLASS_DSL_AVAILABLE, reason="cutlass-dsl is not available")
@pytest.mark.parametrize("tile_size", [128, 256])
@pytest.mark.parametrize("ep_size", [1, 8, 32])
@pytest.mark.parametrize("top_k", [1, 2, 8])
Expand Down Expand Up @@ -509,6 +511,7 @@ def test_nvfp4_grouped_gemm_blackwell(num_tokens: int, top_k: int, ep_size: int,
get_sm_version() not in (100, 103),
reason="This test is only supported on SM 100 and SM 103 GPUs",
)
@pytest.mark.skipif(not IS_CUTLASS_DSL_AVAILABLE, reason="cutlass-dsl is not available")
@pytest.mark.parametrize("tile_size", [128, 256])
@pytest.mark.parametrize("ep_size", [1, 8, 32])
@pytest.mark.parametrize("top_k", [1, 2, 8])
Expand Down Expand Up @@ -610,6 +613,7 @@ def test_nvfp4_grouped_gemm_finalize_blackwell(
get_sm_version() not in (100, 103),
reason="This test is only supported on SM 100 and SM 103 GPUs",
)
@pytest.mark.skipif(not IS_CUTLASS_DSL_AVAILABLE, reason="cutlass-dsl is not available")
@pytest.mark.parametrize("tile_size", [128, 256])
@pytest.mark.parametrize("ep_size", [1, 8, 32])
@pytest.mark.parametrize("top_k", [1, 2, 8])
Expand Down Expand Up @@ -735,6 +739,7 @@ def test_nvfp4_grouped_gemm_swiglu_blackwell(
get_sm_version() not in (100, 103),
reason="This test is only supported on SM 100 and SM 103 GPUs",
)
@pytest.mark.skipif(not IS_CUTLASS_DSL_AVAILABLE, reason="cutlass-dsl is not available")
@pytest.mark.parametrize(
"activation_type",
[ActivationType.Swiglu, ActivationType.Relu2],
Expand Down
Loading