Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
88 lines
2.9 KiB
Python
88 lines
2.9 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.humming import (
|
|
apply_humming_linear,
|
|
convert_linear_layer_to_humming_standard,
|
|
get_humming_linear_compute_config,
|
|
prepare_humming_linear_layer_config,
|
|
quant_key_to_input_schema,
|
|
)
|
|
from vllm.platforms import current_platform
|
|
from vllm.utils.humming import create_humming_locks
|
|
from vllm.utils.import_utils import has_humming
|
|
|
|
from .base import MxFp4LinearKernel, MxFp4LinearLayerConfig
|
|
|
|
|
|
class HummingMxFp4LinearKernel(MxFp4LinearKernel):
|
|
"""Humming GEMM Kernel for MXFP4."""
|
|
|
|
@classmethod
|
|
def is_supported(
|
|
cls, compute_capability: int | None = None
|
|
) -> tuple[bool, str | None]:
|
|
if not current_platform.is_cuda():
|
|
return False, "Humming only supported on CUDA"
|
|
|
|
if not has_humming():
|
|
return False, "Humming is not installed"
|
|
|
|
if not current_platform.has_device_capability(75):
|
|
return False, "Humming only supported on SM75+"
|
|
|
|
return True, None
|
|
|
|
@classmethod
|
|
def can_implement(cls, config: MxFp4LinearLayerConfig) -> tuple[bool, str | None]:
|
|
try:
|
|
quant_key_to_input_schema(config.activation_quant_key)
|
|
except ValueError as error:
|
|
return False, str(error)
|
|
return True, None
|
|
|
|
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
|
layer.weight_scale.data = layer.weight_scale.data.view(torch.float8_e8m0fnu)
|
|
name_map = {"weight": "weight", "weight_scale": "weight_scale"}
|
|
|
|
quant_config = {
|
|
"quant_method": "humming",
|
|
"dtype": "float4e2m1",
|
|
"scale_dtype": "float8e8m0",
|
|
"group_size": 32,
|
|
"weight_scale_type": "group",
|
|
}
|
|
|
|
convert_linear_layer_to_humming_standard(layer=layer, name_map=name_map)
|
|
act_key = self.config.activation_quant_key
|
|
input_schema = quant_key_to_input_schema(act_key)
|
|
# Keep checkpoint-declared FP8 activations instead of falling back to
|
|
# A16 for group-scaled weights.
|
|
self.layer_config = prepare_humming_linear_layer_config(
|
|
layer,
|
|
quant_config,
|
|
input_schema=input_schema,
|
|
allow_input_schema_fallback=(
|
|
act_key is None or act_key.dtype != current_platform.fp8_dtype()
|
|
),
|
|
)
|
|
self.compute_config = get_humming_linear_compute_config()
|
|
self.locks = create_humming_locks(layer.weight.device)
|
|
|
|
def apply_weights(
|
|
self,
|
|
layer: torch.nn.Module,
|
|
x: torch.Tensor,
|
|
bias: torch.Tensor | None = None,
|
|
) -> torch.Tensor:
|
|
return apply_humming_linear(
|
|
layer,
|
|
x,
|
|
skip_bias_add=bias is None,
|
|
layer_config=self.layer_config,
|
|
compute_config=self.compute_config,
|
|
locks=self.locks,
|
|
)
|