[V1] [ROCm] Enable EP with AITER Fused MoE (#20270)

Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
TJian
2025-07-01 16:48:30 +00:00
committed by GitHub
parent 3d19d47d91
commit 02cabff207
4 changed files with 15 additions and 5 deletions
@@ -633,7 +633,8 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
a1_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale)
a2_scale=layer.w2_input_scale,
expert_map=expert_map)
if self.use_marlin:
assert activation == "silu", (
f"{activation} not supported for Marlin MoE.")
@@ -442,6 +442,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
"""
def __init__(self, quant_config: Fp8Config):
from vllm.model_executor.layers.fused_moe import fused_experts
self.quant_config = quant_config
self.block_quant = self.quant_config.weight_block_size is not None
@@ -879,7 +880,8 @@ class Fp8MoEMethod(FusedMoEMethodBase):
if self.block_quant else layer.w2_weight_scale),
a1_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale,
block_shape=self.quant_config.weight_block_size)
block_shape=self.quant_config.weight_block_size,
expert_map=expert_map)
elif self.use_marlin:
assert activation == "silu", (
f"{activation} not supported for Marlin MoE.")