[CI] Fix failing CUDA graph capture in Triton MOE (#46735)
Signed-off-by: Felix Marty <Felix.Marty@amd.com>
This commit is contained in:
@@ -372,6 +372,11 @@ class Nvfp4QuantizationEmulationTritonExperts(TritonExperts):
|
||||
def expects_unquantized_inputs(self) -> bool:
|
||||
return True
|
||||
|
||||
@property
|
||||
def a1_scale(self) -> torch.Tensor:
|
||||
# Used in experts/triton_moe.py and passed to moe_kernel_quantize_input.
|
||||
return self.a1_gscale
|
||||
|
||||
@staticmethod
|
||||
def supports_lora() -> bool:
|
||||
return False
|
||||
|
||||
@@ -245,7 +245,7 @@ class TritonExperts(LoRAExpertsMixin, mk.FusedMoEExpertsModular):
|
||||
lora_unquantized_hidden_states = hidden_states
|
||||
hidden_states, a1q_scale = moe_kernel_quantize_input(
|
||||
hidden_states,
|
||||
self.a1_scale or self.a1_gscale,
|
||||
self.a1_scale,
|
||||
self.quant_dtype,
|
||||
self.per_act_token_quant,
|
||||
self.block_shape,
|
||||
|
||||
Reference in New Issue
Block a user