[Bugfix] Disable shared expert overlap if Marlin MoE is used (#28410)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-11-11 23:16:12 +00:00
committed by GitHub
parent 28534b92b9
commit e5f599d4d1
6 changed files with 13 additions and 5 deletions
@@ -424,6 +424,7 @@ class AWQMoEMethod(FusedMoEMethodBase):
if self.quant_config.weight_bits != 4:
raise ValueError("AWQMoEMethod only supports 4bit now.")
self.quant_type = scalar_types.uint4
self.use_marlin = True
def create_weights(
self,
@@ -1342,6 +1342,7 @@ class CompressedTensorsWNA16MarlinMoEMethod(CompressedTensorsMoEMethod):
f"{WNA16_SUPPORTED_BITS}",
)
self.quant_type = WNA16_SUPPORTED_TYPES_MAP[self.num_bits]
self.use_marlin = True
def create_weights(
self,
@@ -482,6 +482,7 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
self.quant_type = scalar_types.uint8b128
else:
raise ValueError("GPTQMarlinMoEMethod only supports int4 and int8 now.")
self.use_marlin = True
def create_weights(
self,
@@ -216,6 +216,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
def __init__(self, moe: FusedMoEConfig):
super().__init__(moe)
self.mxfp4_backend = get_mxfp4_backend(moe.is_lora_enabled)
self.use_marlin = self.mxfp4_backend == Mxfp4Backend.MARLIN
self.max_capture_size = (
get_current_vllm_config().compilation_config.max_cudagraph_capture_size
)