[CORE] concurrent partial prefills

docs/source/user_guide/configuration/additional_config.md

@@ -53,6 +53,8 @@ The details of each config option are as follows:
 | Name | Type | Default | Description |
 | ---- | ---- | ------- | ----------- |
 | `enabled` | bool | `False` | Whether to enable ascend scheduler for V1 engine|
+| `max_long_partial_prefills` | Union[int, float] | `float('inf')` | the maximum number of prompts longer than long_prefill_token_threshold that will be prefilled concurrently. |
+| `long_prefill_token_threshold` | Union[int, float] | `False` | a request is considered long if the prompt is longer than this number of tokens. |
 
 ascend_scheduler_config also support the options from [vllm scheduler config](https://docs.vllm.ai/en/stable/api/vllm/config.html#vllm.config.SchedulerConfig). For example, you can add `enable_chunked_prefill: True` to ascend_scheduler_config as well.
 
@@ -73,6 +75,8 @@ An example of additional configuration is as follows:
     "ascend_scheduler_config": {
         "enabled": True,
         "enable_chunked_prefill": True,
+        "max_long_partial_prefills": 1,
+        "long_prefill_token_threshold": 4096,
     },
     "refresh": False,
 }

vllm_ascend/core/schedule_config.py

@@ -16,14 +16,16 @@
 #
 
 from dataclasses import dataclass, fields
-from typing import Type, Union
+from typing import Optional, Type, Union
 
 from vllm.config import SchedulerConfig
 
 
 @dataclass
 class AscendSchedulerConfig(SchedulerConfig):
     enable_chunked_prefill: bool = False
+    max_long_partial_prefills: Optional[Union[int, float]] = None
+    long_prefill_token_threshold: Optional[Union[int, float]] = None
     policy: str = "fcfs"
     num_scheduler_steps: int = 1
     scheduler_cls: Union[str, Type[object]] = (
@@ -41,6 +43,8 @@ def initialize_from_config(
         }
         # Override default values into original SchedulerConfig
         scheduler_config["enable_chunked_prefill"] = False
+        scheduler_config["max_long_partial_prefills"] = None
+        scheduler_config["long_prefill_token_threshold"] = None
         scheduler_config["policy"] = "fcfs"
         scheduler_config["num_scheduler_steps"] = 1
         scheduler_config["scheduler_cls"] = (
@@ -55,6 +59,17 @@ def __post_init__(self) -> None:
         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
+        # concurrent partial prefills. Default is inf
+        if self.max_long_partial_prefills is None:
+            self.max_long_partial_prefills = float('inf')
+            self.long_prefill_token_threshold = float('inf')
+        else:
+            if self.long_prefill_token_threshold is None:
+                self.long_prefill_token_threshold = \
+                    int(self.max_model_len * 0.04)
+
+        assert (self.max_long_partial_prefills > 0)
+        assert (self.long_prefill_token_threshold > 0)
         if self.policy != "fcfs":
             raise NotImplementedError(
                 f"currently AscendScheduler only supports fcfs policy, got {self.policy}"

vllm_ascend/core/scheduler.py

@@ -75,6 +75,10 @@ def schedule(self) -> SchedulerOutput:
         # and put back at the head of the waiting queue later
         skipped_waiting_requests: deque[Request] = deque()
 
+        # Skip long prompt requests in prefill stage.
+        # long_prefill_budget is float('inf') if not use.
+        long_prefill_budget = self.vllm_config.scheduler_config.max_long_partial_prefills
+
         # Schedule prefill requests first.
         while self.waiting and token_budget > 0:
             if len(self.running) == self.max_num_running_reqs:
@@ -173,6 +177,11 @@ def skip_cur_request():
                 skip_cur_request()
                 continue
 
+            if  num_new_tokens > self.vllm_config.scheduler_config.long_prefill_token_threshold \
+                and long_prefill_budget <= 0:
+                skip_cur_request()
+                continue
+
             new_blocks = self.kv_cache_manager.allocate_slots(
                 request,
                 num_new_tokens + num_external_computed_tokens,
@@ -222,6 +231,7 @@ def skip_cur_request():
             # Update request info.
             num_scheduled_tokens[request.request_id] = num_new_tokens
             token_budget -= num_new_tokens
+            long_prefill_budget -= 1
             request.status = RequestStatus.RUNNING
             request.num_computed_tokens = num_computed_tokens
             # Count the number of prefix cached tokens.