Fix DeepSeek chunked prefill memory issue (#11149)

2025-10-02 14:56:59 +08:00
parent c0dbbdd12b
commit 44b1fbe258
1 changed files with 1 additions and 0 deletions
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -1965,6 +1965,7 @@ class DeepseekV2AttentionMLA(nn.Module):
            tmp_lse = torch.empty_like(accum_lse)
            merge_state_v2(output, lse, accum_output, accum_lse, tmp_output, tmp_lse)
            accum_output, accum_lse = tmp_output, tmp_lse
+            del kv, k, v, output, lse, tmp_output, tmp_lse

        return accum_output