[ROCm] Expose kFp8DynamicTokenSym on AITER PTPC linears (#54248)

Signed-off-by: Rebecca Lee <rebecca.lee@amd.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Rebecca Lee
2026-09-16 11:48:30 -05:00
committed by GitHub
co-authored by Cursor
parent 8c1557a79c
commit 6ca2b23e22
5 changed files with 104 additions and 4 deletions
@@ -9,6 +9,10 @@ from torch.nn import Parameter
from vllm.config import get_current_vllm_config
from vllm.logger import init_logger
from vllm.model_executor.kernels.linear import init_fp8_linear_kernel
from vllm.model_executor.layers.fusion.quant_activation import (
QuantizedActivation,
expose_input_quant_key,
)
from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
create_fp8_scale_parameter,
@@ -195,10 +199,12 @@ class QuarkW8A8Fp8(QuarkScheme):
module_name=self.__class__.__name__,
)
expose_input_quant_key(layer, self.fp8_linear)
def apply_weights(
self,
layer: torch.nn.Module,
x: torch.Tensor,
x: torch.Tensor | QuantizedActivation,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
return self.fp8_linear.apply_weights(layer, x, bias)