[Quantization] Add compressed-tensors NVFP4 MoE Support (#19990)

Signed-off-by: Dipika Sikka <dipikasikka1@gmail.com>
Signed-off-by: Dipika <dipikasikka1@gmail.com>
This commit is contained in:
Dipika Sikka
2025-06-29 22:05:40 +00:00
committed by GitHub
parent 7b1895e6ce
commit 6f2f53a82d
6 changed files with 295 additions and 22 deletions
@@ -2,9 +2,14 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
from vllm._custom_ops import cutlass_scaled_mm_supports_fp4
from vllm.platforms import current_platform
from vllm.scalar_type import scalar_types
__all__ = ["break_fp4_bytes", "dequantize_to_dtype", "ref_nvfp4_quant"]
__all__ = [
"break_fp4_bytes", "dequantize_to_dtype", "ref_nvfp4_quant",
"cutlass_fp4_supported"
]
FLOAT4_E2M1_MAX = scalar_types.float4_e2m1f.max()
@@ -12,6 +17,14 @@ kE2M1ToFloat = torch.tensor([0., 0.5, 1., 1.5, 2., 3., 4., 6.],
dtype=torch.float32)
def cutlass_fp4_supported() -> bool:
if not current_platform.is_cuda():
return False
capability_tuple = current_platform.get_device_capability()
capability = -1 if capability_tuple is None else capability_tuple.to_int()
return cutlass_scaled_mm_supports_fp4(capability)
def break_fp4_bytes(a, dtype):
assert a.dtype == torch.uint8
m, n = a.shape