xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-03-25 14:32:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

[TPU] Avoid Triton Import (#15589)

Browse Source 
Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>
This commit is contained in:
Robert Shaw 2025-03-27 02:43:02 -04:00 committed by GitHub
parent df8d3d1287
commit e1e0fd7543
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 8 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/layers/fused_moe/layer.py
View File

@ -16,8 +16,6 @@ from vllm.distributed import (get_dp_group, get_tensor_model_parallel_rank,
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.logger import init_logger
from vllm.model_executor.custom_op import CustomOp
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
is_rocm_aiter_moe_enabled, shuffle_weights)
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig, QuantizeMethodBase)
from vllm.model_executor.utils import set_weight_attrs
@ -119,7 +117,9 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
layer.w2_weight = torch.nn.Parameter(self._maybe_pad_weight(
layer.w2_weight.data),
requires_grad=False)
# Lazy import to avoid importing triton.
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
is_rocm_aiter_moe_enabled, shuffle_weights)
if is_rocm_aiter_moe_enabled():
# reshaping weights is required for aiter moe kernel.
shuffled_w13, shuffled_w2 = shuffle_weights(

8
vllm/model_executor/layers/quantization/fp8.py
View File

@ -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"