wehub-resource-sync
50 lines
1.6 KiB
Python
50 lines
1.6 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
import torch
|
|
|
|
from vllm.model_executor.layers.quantization.utils.nvfp4_emulation_utils import (
|
|
kE2M1ToFloat_handle,
|
|
run_nvfp4_emulations,
|
|
)
|
|
|
|
from .base import NvFp4LinearKernel, NvFp4LinearLayerConfig
|
|
|
|
|
|
class EmulationNvFp4LinearKernel(NvFp4LinearKernel):
|
|
"""Software emulation fallback for NVFP4 (dequant → BF16 matmul)."""
|
|
|
|
@classmethod
|
|
def is_supported(
|
|
cls, compute_capability: int | None = None
|
|
) -> tuple[bool, str | None]:
|
|
# Always available as a last-resort fallback.
|
|
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:
|
|
# Move the E2M1 lookup table to the device now, because
|
|
# `.to(device)` is not allowed during CUDA graph capture.
|
|
kE2M1ToFloat_handle.val = kE2M1ToFloat_handle.val.to(layer.weight.device)
|
|
|
|
def apply_weights(
|
|
self,
|
|
layer: torch.nn.Module,
|
|
x: torch.Tensor,
|
|
bias: torch.Tensor | None = None,
|
|
) -> torch.Tensor:
|
|
out = run_nvfp4_emulations(
|
|
x=x,
|
|
input_global_scale=layer.input_global_scale_inv,
|
|
weight=layer.weight,
|
|
weight_scale_swizzled=layer.weight_scale,
|
|
weight_global_scale=layer.weight_global_scale,
|
|
swizzle=False,
|
|
)
|
|
if bias is not None:
|
|
out = out + bias
|
|
return out
|