fix: ascend_scheduler adapt v0.9.0

Signed-off-by: zzzzwwjj <1183291235@qq.com>

Author: zzzzwwjj

.github/workflows/vllm_ascend_test.yaml

@@ -112,7 +112,7 @@ jobs:
             # pytest -sv tests/singlecard/tets_schedule.py
             # guided decoding doesn't work, fix it later
             # pytest -sv tests/singlecard/test_guided_decoding.py.py
-            pytest -sv tests/singlecard/ --ignore=tests/singlecard/test_offline_inference.py --ignore=tests/singlecard/test_scheduler.py --ignore=tests/singlecard/test_guided_decoding.py
+            pytest -sv tests/singlecard/ --ignore=tests/singlecard/test_offline_inference.py --ignore=tests/singlecard/test_guided_decoding.py
           else
             pytest -sv tests/multicard/test_ilama_lora_tp2.py
             VLLM_USE_MODELSCOPE=True pytest -sv tests/multicard/ --ignore=tests/multicard/test_ilama_lora_tp2.py

tests/singlecard/test_scheduler.py

@@ -30,6 +30,7 @@
 from vllm.v1.request import Request, RequestStatus
 from vllm.v1.structured_output import StructuredOutputManager
 
+from tests.conftest import VllmRunner
 from vllm_ascend.core.scheduler import AscendScheduler
 
 EOS_TOKEN_ID = 50256
@@ -394,3 +395,27 @@ def test_stop_via_update_from_output():
     assert len(scheduler.running) == 1
     assert not requests[0].is_finished()
     assert list(requests[0].output_token_ids) == [EOS_TOKEN_ID, 10, 11]
+
+MODELS = [
+    "Qwen/Qwen3-0.6B-Base",
+]
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float16"])
+@pytest.mark.parametrize("max_tokens", [5])
+def test_models(model: str, dtype: str, max_tokens: int) -> None:
+    # 5042 tokens for gemma2
+    # gemma2 has alternating sliding window size of 4096
+    # we need a prompt with more than 4096 tokens to test the sliding window
+    prompt = "The following numbers of the sequence " + ", ".join(
+        str(i) for i in range(1024)) + " are:"
+    example_prompts = [prompt]
+
+    with VllmRunner(model,
+                    max_model_len=8192,
+                    dtype=dtype,
+                    enforce_eager=True,
+                    gpu_memory_utilization=0.7,
+                    enable_prefix_caching=False,
+                    additional_config={'ascend_schduler_config': {}}) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)

vllm_ascend/core/scheduler.py

@@ -130,14 +130,15 @@ def skip_cur_request():
 
             assert num_new_tokens > 0
             watermark = getattr(self.scheduler_config, "watermark", 0.01)
-            if not self._check_watermark_for_prefill(
-                    request, num_new_tokens, computed_blocks, watermark):
+            if not self._check_watermark_for_prefill(request, num_new_tokens,
+                                                     computed_blocks.blocks,
+                                                     watermark):
                 # Scheduling would exceed watermark, skip.
                 skip_cur_request()
                 continue
 
             new_blocks = self.kv_cache_manager.allocate_slots(
-                request, num_new_tokens, computed_blocks)
+                request, num_new_tokens, new_computed_blocks=computed_blocks)
             if new_blocks is None:
                 # The request cannot be scheduled.
                 break
@@ -155,9 +156,8 @@ def skip_cur_request():
 
             if self.lora_config and request.lora_request:
                 scheduled_loras.add(request.lora_request.lora_int_id)
-            req_to_new_block_ids[request.request_id] = [
-                b.block_id for b in computed_blocks + new_blocks
-            ]
+            req_to_new_block_ids[request.request_id] = (
+                self.kv_cache_manager.get_block_ids(request.request_id))
             # Update request info.
             num_scheduled_tokens[request.request_id] = num_new_tokens
             token_budget -= num_new_tokens
@@ -215,9 +215,8 @@ def skip_cur_request():
                 # Schedule the request.
                 scheduled_running_reqs.append(request)
                 self.scheduled_req_ids.add(request.request_id)
-                req_to_new_block_ids[request.request_id] = [
-                    b.block_id for b in new_blocks
-                ]
+                req_to_new_block_ids[request.request_id] = (
+                    new_blocks.get_block_ids())
                 num_scheduled_tokens[request.request_id] = num_new_tokens
                 token_budget -= num_new_tokens
                 req_index += 1
@@ -326,7 +325,8 @@ def _check_watermark_for_prefill(self,
                                len(computed_blocks) * self.block_size)
         num_required_blocks = cdiv(num_new_tokens + num_computed_tokens,
                                    self.block_size)
-        req_blocks = self.kv_cache_manager.req_to_blocks[request.request_id]
+        req_blocks = self.kv_cache_manager.single_type_manager.req_to_blocks[
+            request.request_id]
         num_new_blocks = (num_required_blocks - len(req_blocks) -
                           len(computed_blocks))
         num_evictable_computed_blocks = sum(1 for blk in computed_blocks