[Refactor] refactor humming linear and moe backends to use explicit layer configs (#49610)

Signed-off-by: Jinzhen Lin <jinzhen.ljz@antgroup.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Jinzhen Lin
2026-08-07 10:03:19 -07:00
committed by GitHub
co-authored by OpenAI Codex
parent e229fdbc87
commit fcde8e1460
37 changed files with 448 additions and 341 deletions
@@ -28,7 +28,8 @@ class HummingLinearKernel(MPLinearKernel):
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
from vllm.model_executor.layers.quantization.utils.humming_utils import (
convert_linear_layer_to_humming_standard,
prepare_humming_layer,
get_humming_linear_compute_config,
prepare_humming_linear_layer_config,
)
name_map = {"weight": self.w_q_name, "weight_scale": self.w_s_name}
@@ -46,9 +47,11 @@ class HummingLinearKernel(MPLinearKernel):
convert_linear_layer_to_humming_standard(layer=layer, name_map=name_map)
input_quant_config = getattr(layer, "_humming_input_quant_config", None)
prepare_humming_layer(
self.layer_config = prepare_humming_linear_layer_config(
layer, quant_config, input_quant_config=input_quant_config
)
self.compute_config = get_humming_linear_compute_config()
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
def apply_weights(
self,
@@ -56,12 +59,14 @@ class HummingLinearKernel(MPLinearKernel):
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
from vllm.utils.humming import HummingMethod
from vllm.model_executor.layers.quantization.utils.humming_utils import (
apply_humming_linear,
)
flatten_inputs = x.view(-1, x.size(-1))
output = HummingMethod.forward_layer(
layer=layer,
inputs=flatten_inputs,
compute_config=layer.compute_config,
return apply_humming_linear(
layer,
x,
layer_config=self.layer_config,
compute_config=self.compute_config,
locks=self.locks,
)
return output.view(*x.shape[:-1], output.size(-1))