feat: extended EPLB support for Mistral Large 3 and additional MoE backends (#48355)
Signed-off-by: jdebache <jdebache@nvidia.com>
This commit is contained in:
@@ -899,7 +899,9 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
|
||||
|
||||
# Per tensor kernels require single activation scale. Use the max.
|
||||
w13_input_scale, w2_input_scale = process_fp8_input_tensor_strategy_moe(
|
||||
w13_input_scale, w2_input_scale
|
||||
w13_input_scale,
|
||||
w2_input_scale,
|
||||
layer.moe_config.moe_parallel_config.enable_eplb,
|
||||
)
|
||||
replace_parameter(layer, "w13_input_scale", w13_input_scale)
|
||||
replace_parameter(layer, "w2_input_scale", w2_input_scale)
|
||||
|
||||
Reference in New Issue
Block a user