[TRTLLM] Remove the MoE GEMM weight name change (#30713)

Signed-off-by: Ming Yang <minos.future@gmail.com>
This commit is contained in:
Ming Yang
2025-12-16 11:01:38 -05:00
committed by GitHub
parent 59bd5f6a71
commit ce12b407f2
3 changed files with 16 additions and 40 deletions
@@ -1458,16 +1458,14 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
)
logger.debug_once("Finished shuffling weights for TRT-LLM MOE")
layer.gemm1_weights_fp4_shuffled = Parameter(
layer.w13_weight = Parameter(
gemm1_weights_fp4_shuffled, requires_grad=False
)
layer.gemm2_weights_fp4_shuffled = Parameter(
gemm2_weights_fp4_shuffled, requires_grad=False
)
layer.gemm1_scales_fp4_shuffled = Parameter(
layer.w2_weight = Parameter(gemm2_weights_fp4_shuffled, requires_grad=False)
layer.w13_weight_scale = Parameter(
gemm1_scales_fp4_shuffled, requires_grad=False
)
layer.gemm2_scales_fp4_shuffled = Parameter(
layer.w2_weight_scale = Parameter(
gemm2_scales_fp4_shuffled, requires_grad=False
)
@@ -1476,12 +1474,6 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
(layer.w2_input_scale_quant * layer.g1_alphas).to(torch.float32),
requires_grad=False,
)
# Clean up weights that won't be used by TRT-LLM
del layer.w2_weight
del layer.w2_weight_scale
del layer.w13_weight
del layer.w13_weight_scale
elif self.use_marlin:
# Marlin processing
prepare_moe_fp4_layer_for_marlin(layer)