[BUGFIX] mrope 二维表示情况下，在dcu上kernel无法提前下发

456e8c10 · guanyu1 · zhangzbb · b233584a · 456e8c10
Commit 456e8c10 authored Apr 03, 2026 by guanyu1 Committed by zhangzbb Apr 03, 2026
Hide whitespace changes
Inline Side-by-side

Showing with 5 additions and 1 deletion

vllm/v1/worker/gpu_model_runner.py vllm/v1/worker/gpu_model_runner.py +5 -1

No files found.
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -816,8 +816,12 @@ class GPUModelRunner(
                non_blocking=True,
            )
            return
+        # self.mrope_positions.gpu[:, :num_tokens].copy_(
+        #     self.mrope_positions.cpu[:, :num_tokens],
+        #     non_blocking=True,
+        # )
        self.mrope_positions.gpu[:, :num_tokens].copy_(
-            self.mrope_positions.cpu[:, :num_tokens],
+            self.mrope_positions.cpu[:, :num_tokens].contiguous().pin_memory(),
            non_blocking=True,
        )