[BugFix][Fusion] Fix graph fusion failure problem (#5676)

Currently, the vllm pull request
(https://github.com/vllm-project/vllm/pull/24252) is causing operator
fusion to fail. This issue was previously fixed by patching the backend.
The root cause has been identified, and the problem can be resolved with
this pull request.
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef

---------

Signed-off-by: wxsIcey <1790571317@qq.com>

vllm_ascend/worker/worker.py

@@ -28,7 +28,7 @@ import torch_npu
 import vllm.envs as envs_vllm
 from torch_npu.op_plugin.atb._atb_ops import _register_atb_extensions
 from torch_npu.profiler import dynamic_profile as dp
-from vllm.config import VllmConfig, set_current_vllm_config
+from vllm.config import CUDAGraphMode, VllmConfig, set_current_vllm_config
 from vllm.distributed import (ensure_model_parallel_initialized,
                               init_distributed_environment)
 from vllm.distributed.ec_transfer import ensure_ec_transfer_initialized
@@ -381,10 +381,25 @@ class NPUWorker(WorkerBase):
         warmup_sizes = (self.vllm_config.compilation_config.compile_sizes
                         or []).copy()
         if not self.model_config.enforce_eager:
-            warmup_sizes = [
-                x for x in warmup_sizes if x not in
-                self.vllm_config.compilation_config.cudagraph_capture_sizes
-            ]
+            cg_capture_sizes: list[int] = []
+            if self.vllm_config.compilation_config.cudagraph_mode != CUDAGraphMode.NONE:
+                cg_sizes = self.vllm_config.compilation_config.cudagraph_capture_sizes
+                cg_capture_sizes = [] if cg_sizes is None else cg_sizes
+                warmup_sizes = [
+                    x for x in warmup_sizes if x not in cg_capture_sizes
+                ]
+
+            compile_ranges = self.vllm_config.compilation_config.get_compile_ranges(
+            )
+            # For each compile_range, if none of the batch sizes
+            # in warmup_sizes or cudagraph_capture_sizes are in the range,
+            # add the end of the range to ensure compilation/warmup.
+            all_sizes = set(cg_capture_sizes)
+            all_sizes.update([x for x in warmup_sizes if isinstance(x, int)])
+            for compile_range in compile_ranges:
+                if not any(x in compile_range for x in all_sizes):
+                    warmup_sizes.append(compile_range.end)
+
         for size in sorted(warmup_sizes, reverse=True):
             logger.info("Compile and warming up model for size %d", size)
             self.model_runner._dummy_run(size)