[TPU] Avoid Triton Import (#15589)

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>
This commit is contained in:
Robert Shaw
2025-03-27 06:43:02 +00:00
committed by GitHub
parent df8d3d1287
commit e1e0fd7543
2 changed files with 8 additions and 6 deletions
@@ -13,9 +13,6 @@ from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase,
FusedMoeWeightScaleSupported)
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
expand_weights, is_rocm_aiter_block_scaled_moe_enabled,
is_rocm_aiter_moe_enabled, shuffle_weights)
from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase,
UnquantizedLinearMethod)
from vllm.model_executor.layers.quantization.base_config import (
@@ -532,6 +529,11 @@ class Fp8MoEMethod(FusedMoEMethodBase):
layer.w2_input_scale = None
def process_weights_after_loading(self, layer: Module) -> None:
# Lazy import to avoid importing triton too early.
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
expand_weights, is_rocm_aiter_block_scaled_moe_enabled,
is_rocm_aiter_moe_enabled, shuffle_weights)
# TODO (rob): refactor block quant into separate class.
if self.block_quant:
assert self.quant_config.activation_scheme == "dynamic"