[Scheduler] validate max_num_batched_tokens and max_model_len in AscendSchedulerConfig (#2434)

### What this PR does / why we need it?
Add configuration check logic for ascend scheduler: if chunked_prefill
is disabled, max_num_batched_tokens couldn't be less than max_model_len,
following vLLM;

### Does this PR introduce _any_ user-facing change?
users cannot set max_num_batched_tokens smaller than max_model_len with
ascend scheduler
### How was this patch tested?
CI and vllm serving passed

- vLLM version: v0.10.0
- vLLM main:
f77a0802b7

Signed-off-by: linfeng-yuan <1102311262@qq.com>

tests/e2e/singlecard/test_ascend_scheduler.py

@@ -16,7 +16,7 @@ def test_concurrent_partial_prefill():
                         },
                     },
                     max_num_seqs=3,
-                    max_num_batched_tokens=200,
+                    max_num_batched_tokens=2048,
                     enforce_eager=True,
                     max_model_len=2048,
                     gpu_memory_utilization=0.7) as vllm_model:
@@ -35,7 +35,7 @@ def test_prefix_cache_stats_is_recorded():
                         },
                     },
                     max_num_seqs=3,
-                    max_num_batched_tokens=200,
+                    max_num_batched_tokens=2048,
                     enforce_eager=True,
                     max_model_len=2048,
                     gpu_memory_utilization=0.7) as vllm_model:

tests/ut/core/test_schedule_config.py

@@ -24,6 +24,7 @@ class TestAscendSchedulerConfig(TestBase):
     def setUp(self):
         self.basic_scheduler_config = SchedulerConfig(
             max_num_batched_tokens=8192,
+            max_model_len=8192,
             is_multimodal_model=False,
             send_delta_data=False,
             scheduler_delay_factor=0,
@@ -51,6 +52,7 @@ class TestAscendSchedulerConfig(TestBase):
                 num_scheduler_steps=1,
                 scheduler_cls="vllm_ascend.core.scheduler.AscendScheduler",
                 max_num_batched_tokens=2048,
+                max_model_len=2048,
             ),
         )
         self.assertEqual(ascend_config.enable_chunked_prefill, False)
@@ -65,7 +67,11 @@ class TestAscendSchedulerConfig(TestBase):
         with self.assertRaises(NotImplementedError) as context:
             AscendSchedulerConfig.initialize_from_config(
                 self.basic_scheduler_config,
-                AscendSchedulerConfig(policy="custom_policy", ),
+                AscendSchedulerConfig(
+                    policy="custom_policy",
+                    max_num_batched_tokens=2048,
+                    max_model_len=2048,
+                ),
             )
         self.assertIn(
             "currently AscendScheduler only supports fcfs policy",
@@ -83,7 +89,11 @@ class TestAscendSchedulerConfig(TestBase):
         with self.assertRaises(NotImplementedError) as context:
             AscendSchedulerConfig.initialize_from_config(
                 self.basic_scheduler_config,
-                AscendSchedulerConfig(num_scheduler_steps=2),
+                AscendSchedulerConfig(
+                    num_scheduler_steps=2,
+                    max_num_batched_tokens=2048,
+                    max_model_len=2048,
+                ),
             )
         self.assertIn(
             "currently AscendScheduler doesn't support multi-step",
@@ -94,7 +104,12 @@ class TestAscendSchedulerConfig(TestBase):
         with self.assertRaises(NotImplementedError) as context:
             AscendSchedulerConfig.initialize_from_config(
                 self.basic_scheduler_config,
-                AscendSchedulerConfig(send_delta_data=True))
+                AscendSchedulerConfig(
+                    send_delta_data=True,
+                    max_num_batched_tokens=2048,
+                    max_model_len=2048,
+                ),
+            )
         self.assertIn(
             "currently AscendScheduler doesn't support send_delta_data",
             str(context.exception),
@@ -104,7 +119,12 @@ class TestAscendSchedulerConfig(TestBase):
         with self.assertRaises(NotImplementedError) as context:
             AscendSchedulerConfig.initialize_from_config(
                 self.basic_scheduler_config,
-                AscendSchedulerConfig(delay_factor=1))
+                AscendSchedulerConfig(
+                    delay_factor=1,
+                    max_num_batched_tokens=2048,
+                    max_model_len=2048,
+                ),
+            )
         self.assertIn(
             "currently AscendScheduler doesn't support scheduler_delay_factor",
             str(context.exception),
@@ -115,3 +135,33 @@ class TestAscendSchedulerConfig(TestBase):
             self.basic_scheduler_config, {})
         self.assertEqual(ascend_config.max_num_encoder_input_tokens, 8192)
         self.assertEqual(ascend_config.encoder_cache_size, 8192)
+
+    def test_valid_config_with_chunked_prefill(self):
+        ascend_config = AscendSchedulerConfig.initialize_from_config(
+            self.basic_scheduler_config,
+            AscendSchedulerConfig(
+                enable_chunked_prefill=True,
+                max_num_batched_tokens=2048,
+                max_model_len=4096,
+            ),
+        )
+        self.assertEqual(ascend_config.max_num_batched_tokens, 2048)
+        self.assertEqual(ascend_config.max_model_len, 4096)
+        self.assertTrue(ascend_config.enable_chunked_prefill)
+
+    def test_invalid_config_without_chunked_prefill(self):
+        with self.assertRaises(ValueError) as context:
+            AscendSchedulerConfig.initialize_from_config(
+                self.basic_scheduler_config,
+                AscendSchedulerConfig(
+                    enable_chunked_prefill=False,
+                    max_num_batched_tokens=2048,
+                    max_model_len=4096,
+                ),
+            )
+        self.assertIn(
+            "Ascend scheduler is enabled without chunked prefill feature",
+            str(context.exception),
+        )
+        self.assertIn("max_num_batched_tokens (2048)", str(context.exception))
+        self.assertIn("max_model_len (4096)", str(context.exception))

vllm_ascend/core/schedule_config.py

@@ -55,6 +55,16 @@ class AscendSchedulerConfig(SchedulerConfig):
         self.max_num_encoder_input_tokens = self.max_num_batched_tokens
         self.encoder_cache_size = self.max_num_batched_tokens
         self.chunked_prefill_enabled = self.enable_chunked_prefill
+        if (self.max_num_batched_tokens < self.max_model_len
+                and not self.chunked_prefill_enabled):
+            raise ValueError(
+                "Ascend scheduler is enabled without chunked prefill feature. "
+                f"Argument max_num_batched_tokens ({self.max_num_batched_tokens}) is "
+                f"smaller than max_model_len ({self.max_model_len}). "
+                "This effectively limits the maximum sequence length to "
+                "max_num_batched_tokens and makes vLLM reject longer "
+                "sequences. Please increase max_num_batched_tokens or "
+                "decrease max_model_len.")
         if self.policy != "fcfs":
             raise NotImplementedError(
                 f"currently AscendScheduler only supports fcfs policy, got {self.policy}"