[Quantization] add humming moe backend to all dense/moe oracles (#41652)

Signed-off-by: Jinzhen Lin <jinzhen.ljz@antgroup.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Jinzhen Lin
2026-07-06 13:36:07 -07:00
committed by GitHub
co-authored by mgoin
parent 04adc8843b
commit d891b9bd51
73 changed files with 1337 additions and 123 deletions
@@ -905,6 +905,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
fp8_backend=self.fp8_backend,
experts_cls=self.experts_cls,
routing_tables=layer._expert_routing_tables(),
layer=layer,
)
def process_weights_after_loading(self, layer: RoutedExperts) -> None:
@@ -951,6 +952,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
a1_scale=a1_scale,
a2_scale=a2_scale,
swiglu_limit=getattr(layer, "swiglu_limit", None),
layer=layer,
)
def apply_monolithic(
@@ -1602,7 +1604,9 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
moe_quant_config=self.moe_quant_config,
moe_config=self.moe,
experts_cls=self.experts_cls,
backend=self.nvfp4_backend,
routing_tables=layer._expert_routing_tables(),
layer=layer,
)
self.moe_kernel.fused_experts.process_weights_after_loading(layer)
@@ -1616,6 +1620,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
a13_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale,
swiglu_limit=getattr(layer, "swiglu_limit", None),
layer=layer,
)
@property
@@ -2162,6 +2167,7 @@ class ModelOptMxFp8FusedMoE(FusedMoEMethodBase):
fp8_backend=self.mxfp8_backend,
experts_cls=self.experts_cls,
routing_tables=layer._expert_routing_tables(),
layer=layer,
)
# No native MXFP8 MoE kernel on this device (e.g. gfx942): the emulation
@@ -2207,6 +2213,7 @@ class ModelOptMxFp8FusedMoE(FusedMoEMethodBase):
swiglu_limit=getattr(layer, "swiglu_limit", None),
gemm1_alpha=getattr(layer, "swiglu_alpha", None),
gemm1_beta=getattr(layer, "swiglu_beta", None),
layer=layer,
)
def apply_monolithic(