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:
Julien Debache
2026-08-07 05:32:52 -07:00
committed by GitHub
parent 8d9b52f7c2
commit ae934ba8a5
16 changed files with 500 additions and 96 deletions
@@ -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)