Support penalty in overlap mode; return logprob with chunked prefill; improve benchmark scripts (#3988)

Co-authored-by: SangBin Cho <rkooo567@gmail.com>
Co-authored-by: dhou-xai <dhou@x.ai>
Co-authored-by: Hanming Lu <hanming_lu@berkeley.edu>

python/sglang/srt/layers/quantization/fp8.py

@@ -771,6 +771,8 @@ class Fp8MoEMethod:
         custom_routing_function: Optional[Callable] = None,
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
+        inplace: bool = True,
+        no_combine: bool = False,
     ) -> torch.Tensor:
         from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_experts
         from sglang.srt.layers.moe.topk import select_experts
@@ -793,6 +795,7 @@ class Fp8MoEMethod:
             from aiter.fused_moe import fused_experts_ck
 
             assert activation == "silu", f"{activation=} is not supported."
+            assert not no_combine, f"{no_combine=} is not supported."
 
             return fused_experts_ck(
                 x,
@@ -823,7 +826,7 @@ class Fp8MoEMethod:
                 layer.w2_weight,
                 topk_weights=topk_weights,
                 topk_ids=topk_ids,
-                inplace=True,
+                inplace=inplace and not no_combine,
                 activation=activation,
                 use_fp8_w8a8=True,
                 w1_scale=(
@@ -839,6 +842,7 @@ class Fp8MoEMethod:
                 a1_scale=layer.w13_input_scale,
                 a2_scale=layer.w2_input_scale,
                 block_shape=self.quant_config.weight_block_size,
+                no_combine=no_combine,
             )