[NPU] perf update with kvcache nz & w4a8 quant (#14423)

2025-12-13 17:39:55 +08:00
parent 0e7d7969d5
commit d36299ad77
6 changed files with 219 additions and 108 deletions
--- a/python/sglang/srt/model_executor/forward_batch_info.py
+++ b/python/sglang/srt/model_executor/forward_batch_info.py
@@ -541,11 +541,6 @@ class ForwardBatch:
            num_tokens_per_dp=num_tokens_per_dp,
        )

-        self.num_token_non_padded_cpu = compute_local_num_token_non_padded(
-            global_num_token_non_padded=self.num_token_non_padded_cpu,
-            num_tokens_per_dp=num_tokens_per_dp,
-        )
-
    def merge_mm_inputs(self) -> Optional[MultimodalInputs]:
        """
        Merge all multimodal inputs in the batch into a single MultiModalInputs object.