[Kernel] add Flashinfer cutedsl w4a16 linear (#53014)

Signed-off-by: Siyuan Fu <siyuanf@nvidia.com>
This commit is contained in:
Lain
2026-09-01 02:49:14 +00:00
committed by GitHub
parent 907b1a7f22
commit 882ca8d696
8 changed files with 230 additions and 10 deletions
+7 -3
View File
@@ -17,6 +17,8 @@ following `quantization.quant_algo` values:
- `FP8_PER_CHANNEL_PER_TOKEN`: per-channel weight scale and dynamic per-token activation quantization. - `FP8_PER_CHANNEL_PER_TOKEN`: per-channel weight scale and dynamic per-token activation quantization.
- `FP8_PB_WO` (ModelOpt may emit `fp8_pb_wo`): block-scaled FP8 weight-only (typically 128×128 blocks). - `FP8_PB_WO` (ModelOpt may emit `fp8_pb_wo`): block-scaled FP8 weight-only (typically 128×128 blocks).
- `NVFP4`: ModelOpt NVFP4 checkpoints (use `quantization="modelopt_fp4"`). - `NVFP4`: ModelOpt NVFP4 checkpoints (use `quantization="modelopt_fp4"`).
- `W4A16_NVFP4`: weight-only ModelOpt NVFP4 checkpoints with 16-bit
activations (use `quantization="modelopt_fp4"`).
- `MXFP8`: ModelOpt MXFP8 checkpoints (use `quantization="modelopt_mxfp8"`). - `MXFP8`: ModelOpt MXFP8 checkpoints (use `quantization="modelopt_mxfp8"`).
!!! note !!! note
@@ -28,10 +30,12 @@ following `quantization.quant_algo` values:
workloads. Use `--linear-backend` to override the automatic selection workloads. Use `--linear-backend` to override the automatic selection
(this replaces the deprecated `VLLM_NVFP4_GEMM_BACKEND` environment (this replaces the deprecated `VLLM_NVFP4_GEMM_BACKEND` environment
variable). Values relevant to NVFP4 include `cutlass`, variable). Values relevant to NVFP4 include `cutlass`,
`flashinfer_cutlass`, `flashinfer_trtllm`, `flashinfer_cudnn`, and `flashinfer_cutlass`, `flashinfer_cutedsl`, `flashinfer_trtllm`,
`marlin`; the full list is documented under `KernelConfig` on the `flashinfer_cudnn`, and `marlin`; see `KernelConfig` on the
[Engine Arguments](../../configuration/engine_args.md) page and shown by [Engine Arguments](../../configuration/engine_args.md) page and shown by
`vllm serve --help=KernelConfig`. `vllm serve --help=KernelConfig`. For `W4A16_NVFP4`, `auto` currently
selects Marlin. Models with BF16 activations can explicitly select the
FlashInfer CuTe-DSL backend with `--linear-backend flashinfer_cutedsl`.
!!! note !!! note
For models quantized to MXFP8 with BF16 activations on SM100-family GPUs, For models quantized to MXFP8 with BF16 activations on SM100-family GPUs,
@@ -2,6 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest import pytest
import torch import torch
import torch.nn.functional as F
from nvfp4_utils import ( from nvfp4_utils import (
FLOAT4_E2M1_MAX, FLOAT4_E2M1_MAX,
FLOAT8_E4M3_MAX, FLOAT8_E4M3_MAX,
@@ -10,6 +11,13 @@ from nvfp4_utils import (
) )
from vllm import _custom_ops as ops from vllm import _custom_ops as ops
from vllm.model_executor.kernels.linear.nvfp4 import NvFp4LinearLayerConfig
from vllm.model_executor.kernels.linear.nvfp4.flashinfer import (
FlashInferCuteDslNvFp4W4A16LinearKernel,
)
from vllm.model_executor.layers.quantization.utils.nvfp4_emulation_utils import (
dequantize_to_dtype,
)
from vllm.platforms import current_platform from vllm.platforms import current_platform
from vllm.utils.flashinfer import ( from vllm.utils.flashinfer import (
flashinfer_scaled_fp4_mm, flashinfer_scaled_fp4_mm,
@@ -166,3 +174,55 @@ def test_flashinfer_nvfp4_gemm(
) )
torch.testing.assert_close(out, expected_out.to(dtype=dtype), atol=1e-1, rtol=1e-1) torch.testing.assert_close(out, expected_out.to(dtype=dtype), atol=1e-1, rtol=1e-1)
@pytest.mark.parametrize("shape", SHAPES)
@pytest.mark.parametrize("seed", SEEDS)
@pytest.mark.parametrize("device", CUDA_DEVICES)
@torch.inference_mode()
def test_flashinfer_cutedsl_nvfp4_w4a16_linear(
shape: tuple[int, int, int],
seed: int,
device: str,
) -> None:
supported, reason = FlashInferCuteDslNvFp4W4A16LinearKernel.is_supported()
if not supported:
pytest.skip(reason)
set_random_seed(seed)
m, n, packed_k = shape
k = packed_k * 2
x = torch.randn((m, k), dtype=torch.bfloat16, device=device)
weight = torch.randn((n, k), dtype=torch.bfloat16, device=device)
weight_quant_scale = (FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / weight.abs().max()).to(
torch.float32
)
weight_global_scale = weight_quant_scale.reciprocal()
weight_fp4, weight_scale = ops.scaled_fp4_quant(
weight,
weight_quant_scale,
is_sf_swizzled_layout=False,
)
weight_ref = dequantize_to_dtype(
weight_fp4,
weight_scale,
weight_global_scale,
dtype=torch.bfloat16,
swizzle=False,
)
layer = torch.nn.Module()
layer.output_size_per_partition = n
layer.weight = torch.nn.Parameter(weight_fp4, requires_grad=False)
layer.weight_scale = torch.nn.Parameter(weight_scale, requires_grad=False)
layer.weight_global_scale = torch.nn.Parameter(
weight_global_scale, requires_grad=False
)
kernel = FlashInferCuteDslNvFp4W4A16LinearKernel(NvFp4LinearLayerConfig())
kernel.process_weights_after_loading(layer)
output = kernel.apply_weights(layer, x)
expected = F.linear(x, weight_ref)
assert output.shape == (m, n)
torch.testing.assert_close(output, expected, atol=1e-1, rtol=1e-1)
+11 -1
View File
@@ -16,6 +16,7 @@ from tests.quantization.utils import is_quant_method_supported
from vllm.config import VllmConfig, set_current_vllm_config from vllm.config import VllmConfig, set_current_vllm_config
from vllm.config.model import ModelConfig from vllm.config.model import ModelConfig
from vllm.model_executor.kernels.linear import ( from vllm.model_executor.kernels.linear import (
FlashInferCuteDslNvFp4W4A16LinearKernel,
HummingNvFp4LinearKernel, HummingNvFp4LinearKernel,
MarlinNvFp4LinearKernel, MarlinNvFp4LinearKernel,
) )
@@ -584,10 +585,19 @@ def test_modelopt_nvfp4_config_dispatches_w4a16_method():
@pytest.mark.parametrize( @pytest.mark.parametrize(
("linear_backend", "kernel_cls"), ("linear_backend", "kernel_cls"),
[("auto", MarlinNvFp4LinearKernel), ("humming", HummingNvFp4LinearKernel)], [
("auto", MarlinNvFp4LinearKernel),
("humming", HummingNvFp4LinearKernel),
("flashinfer_cutedsl", FlashInferCuteDslNvFp4W4A16LinearKernel),
],
) )
@pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only") @pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only")
def test_modelopt_w4a16_respects_linear_backend(linear_backend, kernel_cls): def test_modelopt_w4a16_respects_linear_backend(linear_backend, kernel_cls):
if linear_backend != "auto":
is_supported, reason = kernel_cls.is_supported()
if not is_supported:
pytest.skip(reason)
vllm_config = VllmConfig() vllm_config = VllmConfig()
vllm_config.kernel_config.linear_backend = linear_backend vllm_config.kernel_config.linear_backend = linear_backend
with set_current_vllm_config(vllm_config): with set_current_vllm_config(vllm_config):
+1 -1
View File
@@ -276,7 +276,7 @@ class KernelConfig:
- "cutlass": Use CUTLASS-based kernels - "cutlass": Use CUTLASS-based kernels
- "flashinfer_cutlass": Use FlashInfer with CUTLASS kernels - "flashinfer_cutlass": Use FlashInfer with CUTLASS kernels
- "flashinfer_cutedsl": Use FlashInfer with CuTe-DSL kernels - "flashinfer_cutedsl": Use FlashInfer with CuTe-DSL kernels
(BF16, NVFP4, MXFP8) (BF16, NVFP4, MXFP8, W4A16_NVFP4)
- "flashinfer_trtllm": Use FlashInfer with TensorRT-LLM kernels - "flashinfer_trtllm": Use FlashInfer with TensorRT-LLM kernels
- "flashinfer_cudnn": Use FlashInfer with cuDNN kernels - "flashinfer_cudnn": Use FlashInfer with cuDNN kernels
- "flashinfer_b12x": Use FlashInfer b12x CuteDSL NVFP4 GEMM (SM120+) - "flashinfer_b12x": Use FlashInfer b12x CuteDSL NVFP4 GEMM (SM120+)
+20 -3
View File
@@ -149,6 +149,7 @@ from vllm.model_executor.kernels.linear.nvfp4.flashinfer import (
FlashInferB12xNvFp4LinearKernel, FlashInferB12xNvFp4LinearKernel,
FlashInferCudnnNvFp4LinearKernel, FlashInferCudnnNvFp4LinearKernel,
FlashInferCuteDslNvFp4LinearKernel, FlashInferCuteDslNvFp4LinearKernel,
FlashInferCuteDslNvFp4W4A16LinearKernel,
FlashInferCutlassNvFp4LinearKernel, FlashInferCutlassNvFp4LinearKernel,
FlashInferTrtllmNvFp4LinearKernel, FlashInferTrtllmNvFp4LinearKernel,
) )
@@ -262,6 +263,7 @@ _LINEAR_BACKEND_KERNEL_MAP: dict[str, set[type]] = {
}, },
"flashinfer_cutedsl": { "flashinfer_cutedsl": {
FlashInferCuteDslNvFp4LinearKernel, FlashInferCuteDslNvFp4LinearKernel,
FlashInferCuteDslNvFp4W4A16LinearKernel,
FlashInferCutedslMxfp8LinearKernel, FlashInferCutedslMxfp8LinearKernel,
}, },
"flashinfer_trtllm": { "flashinfer_trtllm": {
@@ -531,6 +533,7 @@ _POSSIBLE_MXFP8_KERNELS: dict[PlatformEnum, list[type[Mxfp8LinearKernel]]] = {
_POSSIBLE_NVFP4_KERNELS: dict[PlatformEnum, list[type[NvFp4LinearKernel]]] = { _POSSIBLE_NVFP4_KERNELS: dict[PlatformEnum, list[type[NvFp4LinearKernel]]] = {
PlatformEnum.CUDA: [ PlatformEnum.CUDA: [
FlashInferCuteDslNvFp4LinearKernel, FlashInferCuteDslNvFp4LinearKernel,
FlashInferCuteDslNvFp4W4A16LinearKernel,
FlashInferCutlassNvFp4LinearKernel, FlashInferCutlassNvFp4LinearKernel,
FlashInferB12xNvFp4LinearKernel, FlashInferB12xNvFp4LinearKernel,
CutlassNvFp4LinearKernel, CutlassNvFp4LinearKernel,
@@ -1014,7 +1017,11 @@ def init_nvfp4_linear_kernel(use_a16: bool = False) -> NvFp4LinearKernel:
"""Select and instantiate the best NVFP4 linear kernel for the """Select and instantiate the best NVFP4 linear kernel for the
current platform.""" current platform."""
config = NvFp4LinearLayerConfig() config = NvFp4LinearLayerConfig()
a16_kernels = (MarlinNvFp4LinearKernel, HummingNvFp4LinearKernel) a16_kernels = (
FlashInferCuteDslNvFp4W4A16LinearKernel,
MarlinNvFp4LinearKernel,
HummingNvFp4LinearKernel,
)
# VLLM_BATCH_INVARIANT forces deterministic execution. Prefer the # VLLM_BATCH_INVARIANT forces deterministic execution. Prefer the
# batch-invariant CUTLASS implementation when available, otherwise fall # batch-invariant CUTLASS implementation when available, otherwise fall
@@ -1051,8 +1058,17 @@ def init_nvfp4_linear_kernel(use_a16: bool = False) -> NvFp4LinearKernel:
) )
force_kernel = EmulationNvFp4LinearKernel force_kernel = EmulationNvFp4LinearKernel
elif linear_backend == "auto" and use_a16: elif linear_backend == "auto" and use_a16:
# Force a16 (Marlin) when running weight-only quantization. _cc = current_platform.get_device_capability()
force_kernel = MarlinNvFp4LinearKernel compute_capability = _cc.to_int() if _cc is not None else None
# Weight-only: prefer FlashInfer CuTe-DSL W4A16 on SM100/103,
# otherwise Marlin.
cutedsl_ok, _ = FlashInferCuteDslNvFp4W4A16LinearKernel.is_supported(
compute_capability
)
if compute_capability in (100, 103) and cutedsl_ok:
force_kernel = FlashInferCuteDslNvFp4W4A16LinearKernel
else:
force_kernel = MarlinNvFp4LinearKernel
if force_kernel is not None: if force_kernel is not None:
if use_a16 and force_kernel not in a16_kernels: if use_a16 and force_kernel not in a16_kernels:
@@ -1232,6 +1248,7 @@ __all__ = [
"EmulationNvFp4LinearKernel", "EmulationNvFp4LinearKernel",
"FbgemmNvFp4LinearKernel", "FbgemmNvFp4LinearKernel",
"FlashInferCuteDslNvFp4LinearKernel", "FlashInferCuteDslNvFp4LinearKernel",
"FlashInferCuteDslNvFp4W4A16LinearKernel",
"FlashInferB12xNvFp4LinearKernel", "FlashInferB12xNvFp4LinearKernel",
"FlashInferCutlassNvFp4LinearKernel", "FlashInferCutlassNvFp4LinearKernel",
"FlashInferTrtllmNvFp4LinearKernel", "FlashInferTrtllmNvFp4LinearKernel",
@@ -20,14 +20,94 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
) )
from vllm.platforms import current_platform from vllm.platforms import current_platform
from vllm.utils.flashinfer import ( from vllm.utils.flashinfer import (
flashinfer_prepare_bf16_fp4_weights,
flashinfer_scaled_fp4_mm, flashinfer_scaled_fp4_mm,
has_flashinfer, has_flashinfer,
has_flashinfer_b12x_gemm, has_flashinfer_b12x_gemm,
has_flashinfer_bf16_fp4,
) )
from .base import NvFp4LinearKernel, NvFp4LinearLayerConfig from .base import NvFp4LinearKernel, NvFp4LinearLayerConfig
class FlashInferCuteDslNvFp4W4A16LinearKernel(NvFp4LinearKernel):
"""BF16 x NVFP4 GEMM via FlashInfer's CuTe-DSL backend."""
@classmethod
def is_supported(
cls, compute_capability: int | None = None
) -> tuple[bool, str | None]:
if compute_capability is None:
if not current_platform.is_cuda():
return False, "FlashInfer CuTe-DSL W4A16 requires CUDA"
capability = current_platform.get_device_capability()
if capability is None:
return False, "CUDA compute capability is unavailable"
compute_capability = capability.to_int()
if compute_capability not in (100, 103) and not (
compute_capability >= 120 and compute_capability < 130
):
return False, "FlashInfer CuTe-DSL W4A16 requires sm_100 or sm_12x"
if not has_flashinfer_bf16_fp4():
return False, "FlashInfer CuTe-DSL BF16 x FP4 GEMM is unavailable"
return True, None
@classmethod
def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
padded_weight, weights_padding_bytes = pad_nvfp4_weight_for_cutlass(
layer.weight.data, alignment=64
)
swizzled_scale = swizzle_blockscale(layer.weight_scale.data)
weight, weight_scale, global_scale = flashinfer_prepare_bf16_fp4_weights(
padded_weight,
swizzled_scale,
layer.weight_global_scale.data.reshape(1),
backend="cute-dsl",
)
assert global_scale is not None
layer.weight = torch.nn.Parameter(weight, requires_grad=False)
layer.weight_scale = torch.nn.Parameter(weight_scale, requires_grad=False)
layer.weight_global_scale = torch.nn.Parameter(
global_scale, requires_grad=False
)
layer.weights_padding_cols = weights_padding_bytes
def apply_weights(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
if x.dtype != torch.bfloat16:
raise ValueError(
f"FlashInfer CuTe-DSL W4A16 requires BF16 input, got {x.dtype}"
)
output_size = layer.output_size_per_partition
output_shape = [*x.shape[:-1], output_size]
x_2d = x.reshape(-1, x.shape[-1])
weights_padding_bytes = getattr(layer, "weights_padding_cols", 0)
if weights_padding_bytes:
x_2d = torch.nn.functional.pad(x_2d, (0, weights_padding_bytes * 2))
x_2d = x_2d.contiguous()
out = torch.ops.vllm.flashinfer_mm_bf16_fp4(
x_2d,
layer.weight,
layer.weight_scale,
layer.weight_global_scale,
)
out = slice_nvfp4_output(out, output_size)
if bias is not None:
out = out + bias
return out.view(*output_shape)
class FlashInferCuteDslNvFp4LinearKernel(NvFp4LinearKernel): class FlashInferCuteDslNvFp4LinearKernel(NvFp4LinearKernel):
"""NVFP4 GEMM via FlashInfer's cutedsl backend.""" """NVFP4 GEMM via FlashInfer's cutedsl backend."""
@@ -1038,7 +1038,7 @@ class ModelOptNvFp4Config(ModelOptQuantConfigBase):
# Select LinearMethod implementation based on quant_algo (FP8 pattern). # Select LinearMethod implementation based on quant_algo (FP8 pattern).
# NVFP4 -> W4A4: cutlass NVFP4 GEMM with input quantization # NVFP4 -> W4A4: cutlass NVFP4 GEMM with input quantization
# W4A16_NVFP4 -> W4A16: FP4 Marlin GEMM with bf16/fp16 activations # W4A16_NVFP4 -> W4A16: flashinfer cute-dsl or Marlin with 16-bit inputs
if quant_method == "NVFP4": if quant_method == "NVFP4":
self.LinearMethodCls = ModelOptNvFp4LinearMethod self.LinearMethodCls = ModelOptNvFp4LinearMethod
elif quant_method == "W4A16_NVFP4": elif quant_method == "W4A16_NVFP4":
@@ -1283,7 +1283,7 @@ class ModelOptNvFp4W4A16LinearMethod(LinearMethodBase):
self.quant_config = quant_config self.quant_config = quant_config
self.marlin_input_dtype = None self.marlin_input_dtype = None
# `init_nvfp4_linear_kernel(use_a16=True)` is best of both worlds: # `init_nvfp4_linear_kernel(use_a16=True)` is best of both worlds:
# 1. `use_a16=True` forces `Marlin`: https://github.com/vllm-project/vllm/commit/e68988a#diff-7135ab92aa94dfacb1ad3c77fc13f9c4ffe0b977f8eac5d86c2afe243e5f92a6R842-R889 # 1. `use_a16=True` forces flashinfer cutedsl for sm100 and `Marlin` for others.
# for `--linear-backend=auto`, avoiding a W4A4 kernel that requires input_scale. # for `--linear-backend=auto`, avoiding a W4A4 kernel that requires input_scale.
# 2. Specifying e.g. `--linear-backend=humming` will override. # 2. Specifying e.g. `--linear-backend=humming` will override.
self.kernel = init_nvfp4_linear_kernel(use_a16=True) self.kernel = init_nvfp4_linear_kernel(use_a16=True)
+49
View File
@@ -197,6 +197,9 @@ flashinfer_cutlass_fused_moe = _lazy_import_wrapper(
flashinfer_cutedsl_grouped_gemm_nt_masked = _lazy_import_wrapper( flashinfer_cutedsl_grouped_gemm_nt_masked = _lazy_import_wrapper(
"flashinfer.cute_dsl.blockscaled_gemm", "grouped_gemm_nt_masked" "flashinfer.cute_dsl.blockscaled_gemm", "grouped_gemm_nt_masked"
) )
flashinfer_prepare_bf16_fp4_weights = _lazy_import_wrapper(
"flashinfer.gemm", "prepare_bf16_fp4_weights"
)
flashinfer_fp4_quantize = _lazy_import_wrapper("flashinfer", "fp4_quantize") flashinfer_fp4_quantize = _lazy_import_wrapper("flashinfer", "fp4_quantize")
flashinfer_mxfp4_quantize = _lazy_import_wrapper("flashinfer", "mxfp4_quantize") flashinfer_mxfp4_quantize = _lazy_import_wrapper("flashinfer", "mxfp4_quantize")
nvfp4_batched_quantize = _lazy_import_wrapper("flashinfer", "nvfp4_batched_quantize") nvfp4_batched_quantize = _lazy_import_wrapper("flashinfer", "nvfp4_batched_quantize")
@@ -333,6 +336,17 @@ def has_flashinfer_cutedsl() -> bool:
) )
@functools.cache
def has_flashinfer_bf16_fp4() -> bool:
"""Return ``True`` if FlashInfer's CuTe-DSL W4A16 GEMM is available."""
if not has_flashinfer_cutedsl():
return False
mod = _get_submodule("flashinfer.gemm")
return mod is not None and all(
hasattr(mod, name) for name in ("mm_bf16_fp4", "prepare_bf16_fp4_weights")
)
@functools.cache @functools.cache
def has_flashinfer_trtllm_fused_moe() -> bool: def has_flashinfer_trtllm_fused_moe() -> bool:
"""Return `True` if FlashInfer TRTLLM fused MoE is available.""" """Return `True` if FlashInfer TRTLLM fused MoE is available."""
@@ -706,6 +720,39 @@ if has_flashinfer():
) -> torch.Tensor: ) -> torch.Tensor:
return torch.empty(A.shape[0], B.shape[1], dtype=dtype, device=A.device) return torch.empty(A.shape[0], B.shape[1], dtype=dtype, device=A.device)
@torch.library.custom_op(
"vllm::flashinfer_mm_bf16_fp4",
mutates_args=[],
device_types="cuda",
)
def flashinfer_mm_bf16_fp4(
A: torch.Tensor,
B: torch.Tensor,
B_scale: torch.Tensor,
global_scale: torch.Tensor,
) -> torch.Tensor:
from flashinfer.gemm import mm_bf16_fp4
return mm_bf16_fp4(
A,
B,
B_scale,
global_scale,
backend="cute-dsl",
)
@torch.library.register_fake(
"vllm::flashinfer_mm_bf16_fp4",
)
def flashinfer_mm_bf16_fp4_fake(
A: torch.Tensor,
B: torch.Tensor,
B_scale: torch.Tensor,
global_scale: torch.Tensor,
) -> torch.Tensor:
output_size = B.shape[0] if B.dtype == torch.uint8 else B.shape[1] // 2
return torch.empty(A.shape[0], output_size, dtype=A.dtype, device=A.device)
@torch.library.custom_op( @torch.library.custom_op(
"vllm::flashinfer_mxfp4_quantize", "vllm::flashinfer_mxfp4_quantize",
mutates_args=[], mutates_args=[],
@@ -1158,6 +1205,7 @@ __all__ = [
"flashinfer_trtllm_fp8_block_scale_moe", "flashinfer_trtllm_fp8_block_scale_moe",
"flashinfer_cutlass_fused_moe", "flashinfer_cutlass_fused_moe",
"flashinfer_cutedsl_grouped_gemm_nt_masked", "flashinfer_cutedsl_grouped_gemm_nt_masked",
"flashinfer_prepare_bf16_fp4_weights",
"flashinfer_fp4_quantize", "flashinfer_fp4_quantize",
"silu_and_mul_scaled_nvfp4_experts_quantize", "silu_and_mul_scaled_nvfp4_experts_quantize",
"scaled_fp4_grouped_quantize", "scaled_fp4_grouped_quantize",
@@ -1177,6 +1225,7 @@ __all__ = [
"has_flashinfer_cutlass_fused_moe", "has_flashinfer_cutlass_fused_moe",
"has_flashinfer_cutedsl_grouped_gemm_nt_masked", "has_flashinfer_cutedsl_grouped_gemm_nt_masked",
"has_flashinfer_cutedsl_moe_nvfp4", "has_flashinfer_cutedsl_moe_nvfp4",
"has_flashinfer_bf16_fp4",
"has_flashinfer_b12x_moe", "has_flashinfer_b12x_moe",
"has_flashinfer_b12x_gemm", "has_flashinfer_b12x_gemm",
"has_flashinfer_fp8_blockscale_gemm", "has_flashinfer_fp8_blockscale_gemm",