[None][feat] add skip condition in AutoDeploy's triton fused moe kernel (NVIDIA#8632)

Signed-off-by: Suyog Gupta <41447211+suyoggupta@users.noreply.github.com>

Author: suyoggupta

tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/triton_moe.py

@@ -41,6 +41,7 @@ def fused_mlp_moe_kernel(
     topk_weights_ptr,
     sorted_token_ids_ptr,
     expert_ids_ptr,
+    num_tokens_post_padded_ptr,
     # Matrix dimensions
     N,
     K,
@@ -84,6 +85,10 @@ def fused_mlp_moe_kernel(
     pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
     pid_n = (pid % num_pid_in_group) // group_size_m
 
+    num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
+    if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
+        return
+
     offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
     # Bounds check: EM might not be a multiple of BLOCK_SIZE_M
     # so offs_token_id can exceed EM-1. Load with mask to avoid out-of-bounds.
@@ -270,6 +275,7 @@ def _grid(META):
         topk_weights if topk_weights is not None else C,
         sorted_token_ids,
         expert_ids,
+        num_tokens_post_padded,
         B.size(1),
         B.size(2),
         EM,