[BugFix][Refactor] Fix bugs and refine codes for large scale simulato…

…r test (#93)

benchmark/benchmark_serving.py

@@ -372,7 +372,7 @@ def __init__(self):
         self._decode_sum_latencies = []
         self._all_decode_token_latencies = []
         self._inference_latencies = []
-        self._per_token_latencies_breakdown_dict = []
+        self._per_token_latency_breakdown_list = []
 
     def measure(self, f):
         async def measured(*args, **kwargs):
@@ -400,9 +400,9 @@ async def measured(*args, **kwargs):
                 self._all_token_latencies.append(lat_arr)
                 self._decode_sum_latencies.append(decode_sum_latency)
                 self._all_decode_token_latencies.extend(lat_arr[1:,1])
-            if 'per_token_latency_breakdown_dict' in output:
-                self._inference_latencies.append(np.mean(output['per_token_latency_breakdown_dict']['step_latency_engine']))
-                self._per_token_latencies_breakdown_dict.append(output['per_token_latency_breakdown_dict'])
+                self._inference_latencies.append(0.0)
+            if 'per_token_latency_breakdown_list' in output:
+                self._per_token_latency_breakdown_list.append(output['per_token_latency_breakdown_list'])
             return prompt, output
         return measured
 
@@ -494,7 +494,7 @@ async def benchmark(
            m._decode_sum_latencies, \
            m._request_lens, \
            m._all_decode_token_latencies, \
-           m._per_token_latencies_breakdown_dict
+           m._per_token_latency_breakdown_list
 
 def gen_random_response_lens(distribution: str, len_mean, len_range, num_prompts):
     if distribution == 'uniform':
@@ -785,7 +785,7 @@ def main():
     decode_sum_latencies, \
     request_lens, \
     all_decode_token_latencies, \
-    per_token_latencies_breakdown_dict = asyncio.run(benchmark(
+    per_token_latency_breakdown_list = asyncio.run(benchmark(
         backend,
         tokenizer,
         prompts,
@@ -823,8 +823,8 @@ def main():
                         "decode_sum_latencies": decode_sum_latencies,
                         "all_decode_token_latencies": all_decode_token_latencies,
                         "inference_latencies": inference_latencies,
-                        "per_token_latencies_breakdown_dict": per_token_latencies_breakdown_dict,
-                        "throughput": throughput,
+                        "per_token_latency_breakdown_list": per_token_latency_breakdown_list,
+                        "throughput": throughput, 
                         "instance_num": avg_instance_num})
         json.dump(results, f)
 

docs/Arguments.md

@@ -25,10 +25,11 @@ usage: -m llumnix.entrypoints.vllm.api_server [-h]
             [--scaling-load-metric {remaining_steps,usage_ratio}]
             [--polling-interval POLLING_INTERVAL]
             [--dispatch-policy {balanced,load,queue,rr}]
+            [--topk-random-dispatch TOPK_RANDOM_DISPATCH]
             [--enable-migration]
             [--enable-defrag]
             [--pair-migration-frequency PAIR_MIGRATION_FREQUENCY]
-            [--pair-migration-policy {balanced,defrag_constrained,defrag_relaxed}]
+            [--pair-migration-policy {balanced,defrag}]
             [--migrate-out-threshold MIGRATE_OUT_THRESHOLD]
             [--request-migration-policy {LCR,SR,LR,FCW,FCWSR}]
             [--enable-scaling]
@@ -139,6 +140,10 @@ usage: -m llumnix.entrypoints.vllm.api_server [-h]
 - Possible choices: balanced, load, queue, rr
 - Default: "load"
 
+`--topk-random-dispatch`
+- Number of candidate random dispatch instances for dispatch policy.
+- Default: 1
+
 `--enable-migration`
 - Enable migrate requests between instances.
 
@@ -151,8 +156,8 @@ usage: -m llumnix.entrypoints.vllm.api_server [-h]
 
 `--pair-migration-policy`
 - Pair migration policy.
-- Possible choices: balanced, defrag_constrained, defrag_relaxed
-- Default: "defrag_constrained"
+- Possible choices: balanced, defrag
+- Default: "defrag"
 
 `--migrate-out-threshold`
 - Migrate out instance load threshold.

docs/Simulator.md

@@ -3,7 +3,7 @@ Llumnix can generate latency data from logs. After run a real benchmark with `--
 
 After running profiling with `python llumnix.backends.profiling.py`. You can get a `$PROFILING_RESULT_FILE_PATH.pkl`
 
-Then, you can run simulator with `--profiling-result-file-path PROFILING_RESULT_FILE_PATH`.
+Then, you can run simulator with `--simulator-mode` and `--profiling-result-file-path PROFILING_RESULT_FILE_PATH`.
 
 
 ```

llumnix/arg_utils.py

@@ -44,6 +44,7 @@ def add_argument(self, *args, **kwargs):
                 kwargs['default'] = None
         super().add_argument(*args, **kwargs)
 
+
 @dataclass
 class EntrypointsArgs:
     host: str = None
@@ -112,13 +113,15 @@ def add_cli_args(parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
                             help="path to config file of arguments")
         return parser
 
+
 @dataclass
 class ManagerArgs:
     initial_instances: int = None
 
     polling_interval: float = None
     dispatch_policy: str = None
     scaling_load_metric: str = None
+    topk_random_dispatch: int = None
 
     enable_migration: bool = None
     pair_migration_frequency: int = None
@@ -174,6 +177,7 @@ def create_global_scheduler_config(self, is_group_kind_migration_backend) -> Tup
         # Create the GlobalScheduler Configuration.
         global_scheduler_config = GlobalSchedulerConfig(self.initial_instances,
                                                         self.dispatch_policy,
+                                                        self.topk_random_dispatch,
                                                         self.pair_migration_policy,
                                                         self.migrate_out_threshold,
                                                         self.scaling_policy,
@@ -205,6 +209,8 @@ def check_args(cls, args: 'ManagerArgs', parser: argparse.ArgumentParser):
         assert not args.enable_port_offset_store or args.enable_port_increment, \
             "Set enable_port_increment when enable_port_offset_store"
 
+        assert not args.enable_scaling, "Proactive auto-scaling is deprecated now, all auto-scaling related args will not take effects."
+
     @staticmethod
     def add_cli_args(parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
         parser.add_argument('--initial-instances',
@@ -226,6 +232,13 @@ def add_cli_args(parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
                             '* "queue" dispatch request to the instance with minimum waiting request queue length.\n'
                             '* "flood" dispatch request to the instance with maximum requests dispatched.\n'
                             '* "rr" dispatch requests with round-robin policy.\n')
+        parser.add_argument('--topk-random-dispatch',
+                            type=int,
+                            help='number of candidate random dispatch instances for dispatch policy.\n\n'
+                            'The candidate instances are first selected according to the load'
+                            '(including factors such as load, queue size, etc.) based on the dispatch policy,'
+                            'and then one of them is randomly chosen to receive the request for better load balancing.')
+
         parser.add_argument('--enable-migration',
                             action='store_true',
                             help='enable migrate requests between instances')
@@ -234,13 +247,11 @@ def add_cli_args(parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
                             help='pair migration frequency')
         parser.add_argument('--pair-migration-policy',
                             type=str,
-                            choices=['balanced', 'defrag_constrained', 'defrag_relaxed'],
+                            choices=['balanced', 'defrag'],
                             help='The pair migration policy.\n\n'
                             '* "balanced" pair migration to make the instance load of instance more balanced.\n'
-                            '* "defrag_constrained" pair migration without balanced constraint to '
-                            'achieve defragmentation thoroughly (with instance constraints).\n'
-                            '* "defrag_relaxed" pair migration to without balanced constraint '
-                            'to achieve defragmentation thoroughly (without instance constraints).\n')
+                            '* "defrag" pair migration without balanced constraint to '
+                            'achieve defragmentation thoroughly (with instance constraints).\n')
         parser.add_argument('--migrate-out-threshold',
                             type=float,
                             help='migrate out instance load threshold')
@@ -289,11 +300,13 @@ def add_cli_args(parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
                             help='the prefill decode ratio used in gloabl launch model e.g. "1:1"')
         return parser
 
+
 @dataclass
 class LaunchArgs:
     launch_mode: LaunchMode = None
     backend_type: BackendType = None
 
+
 @dataclass
 class InstanceArgs:
     instance_type: str = None

llumnix/backends/profiling.py

@@ -35,6 +35,7 @@
 def _pad_to_alignment(x, multiple_of):
     return x + ((-1*x) % multiple_of)
 
+
 @dataclasses.dataclass
 class LatencyMemData:
     # The latency of each stage
@@ -69,6 +70,7 @@ def get_prefill_dict_kv(self):
     def get_decode_dict_kv(self):
         return map(list, zip(*self.decode_latency.items()))
 
+
 @dataclasses.dataclass
 class ProfilingResult:
     """Store the profiling result of a model."""
@@ -127,6 +129,7 @@ def fit_from_database(self, parallel_config: SimParallelConfig):
             avg_loss += abs(sim_lat - latency_list[idx])
         print(f"decode sim avg_loss={avg_loss/len(latency_list)}")
 
+
 class ProfilingDatabase:
     """Store the profiling results of all the models"""
     def __init__(self, database_filename: str, new_database: bool = False):
@@ -198,6 +201,7 @@ def get_latency_mem(backend_type: BackendType, profiling_database: ProfilingData
         return latency_mem
     raise ValueError(f'Unsupported simulator backend: {backend_type}')
 
+
 if __name__ == "__main__":
     import argparse
     parser = argparse.ArgumentParser()

llumnix/backends/utils.py

@@ -27,6 +27,7 @@
 from llumnix.logging.logger import init_logger
 from llumnix.utils import get_instance_name
 from llumnix.internal_config import MigrationConfig
+from llumnix.metrics.timestamps import set_timestamp
 
 logger = init_logger(__name__)
 
@@ -55,16 +56,14 @@ async def put_nowait_to_servers(self,
         tasks = []
         for server_id, req_outputs in server_request_outputs.items():
             server_info = server_info_dict[server_id]
-            for req_output in req_outputs:
-                if hasattr(req_output, 'request_timestamps'):
-                    req_output.request_timestamps.engine_actor_put_queue_timestamp = time.time()
+            set_timestamp(req_outputs, 'engine_actor_put_queue_timestamp', time.time())
             tasks.append(asyncio.create_task(self.request_output_queue_client.put_nowait(req_outputs, server_info)))
         rets = await asyncio.gather(*tasks, return_exceptions=True)
         for idx, ret in enumerate(rets):
             if isinstance(ret, Exception):
                 server_id = list(server_request_outputs.keys())[idx]
                 server_info = server_info_dict[server_id]
-                logger.warning("Server {} is dead.".format(server_id))
+                logger.error("Server {} is dead, exception: {}".format(server_id, ret))
                 if self.request_output_queue_type == QueueType.ZMQ:
                     logger.warning("request output queue ip: {}, port: {}".format(server_info.request_output_queue_ip,
                                                                                   server_info.request_output_queue_port))

llumnix/backends/vllm/executor.py

@@ -12,7 +12,6 @@
 # limitations under the License.
 
 import time
-import asyncio
 
 from collections import defaultdict
 from typing import Callable, Dict, List, Optional, Tuple, Type
@@ -22,7 +21,6 @@
 from ray.util.placement_group import PlacementGroup
 
 from vllm.executor.executor_base import ExecutorBase
-from vllm.model_executor.layers.sampler import SamplerOutput, CompletionSequenceGroupOutput
 from vllm.executor.ray_gpu_executor import RayGPUExecutor, RayGPUExecutorAsync, RayWorkerWrapper, envs, \
                                            get_ip, get_vllm_instance_id, get_distributed_init_method, get_open_port
 from vllm.worker.worker_base import WorkerBase
@@ -263,68 +261,3 @@ async def execute_model_async(self, *args, **kwargs):
         t1 = time.time()
         self.last_inference_latency = (t1 - t0) * 1000
         return outputs
-
-class SimGPUExecutor(RayGPUExecutor):
-    latency_mem: LatencyMemData = None
-    def __init__(self, *args, **kwargs) -> None:
-        RayGPUExecutor.__init__(self, *args, **kwargs)
-        self.last_inference_latency = 0
-        self.migration_bandwidth = self.latency_mem.migration_bandwidth
-        # TODO(ZeldaHuang): add swap bandwidth
-
-        self.cache_block_size = get_cache_block_size(
-            self.cache_config.block_size, self.model_config, self.parallel_config)
-        self.cache_block_size /= GiB_bytes
-        self.sim_cache_config = SimCacheConfig(self.cache_config.gpu_memory_utilization,
-                                               self.cache_config.block_size,
-                                               self.scheduler_config.max_num_batched_tokens)
-
-    def _init_executor(self) -> None:
-        pass
-
-    def determine_num_available_blocks(self) -> Tuple[int, int]:
-        num_gpu_blocks = self.latency_mem.cache_dict.get(self.sim_cache_config, 880)
-        num_cpu_blocks = 2048
-        return (num_gpu_blocks, num_cpu_blocks)
-
-    def initialize_cache(self, num_gpu_blocks: int,
-                         num_cpu_blocks: int) -> None:
-        logger.info("# GPU blocks: {}, # CPU blocks: {}".format(num_gpu_blocks, num_cpu_blocks))
-
-    async def execute_model_async(
-            self,
-            execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]:
-        prefill_seq_len = 0
-        decode_seq_len = 0
-        decode_bs = 0
-        for meta_data in execute_model_req.seq_group_metadata_list:
-            if meta_data.is_prompt:
-                prefill_seq_len += meta_data.token_chunk_size
-            else:
-                decode_bs += meta_data.token_chunk_size
-                decode_seq_len += list(meta_data.seq_data.values())[0].get_len()
-        decode_bs = _pad_to_alignment(decode_bs, 8)
-        prefill_seq_len = _pad_to_alignment(prefill_seq_len, 8)
-        latency = 0
-        if prefill_seq_len:
-            latency += self.latency_mem.prefill_latency[prefill_seq_len][0] if prefill_seq_len in self.latency_mem.prefill_latency \
-                       else model_prefill(prefill_seq_len, *self.latency_mem.prefill_model_params)
-        if decode_bs:
-            decode_meta_data = (decode_bs, decode_seq_len)
-            latency += self.latency_mem.decode_latency[decode_meta_data][0] if decode_meta_data in self.latency_mem.decode_latency \
-                       else model_decode((decode_bs, decode_seq_len), *self.latency_mem.decode_model_params)
-        await asyncio.sleep(latency/1000)
-        sampler_outputs = []
-        for meta_data in execute_model_req.seq_group_metadata_list:
-            samples = []
-            for seq_id in meta_data.seq_data.keys():
-                dummy_sample_output = SequenceOutput(seq_id, 20, {20: Logprob(1.0)})
-                samples.append(dummy_sample_output)
-            if samples:
-                output = CompletionSequenceGroupOutput(samples, None)
-                sampler_outputs.append(output)
-        return [SamplerOutput(outputs=sampler_outputs)]
-
-    async def send_blocks(self, blocks_len) -> None:
-        migration_latency = (self.cache_block_size * blocks_len) / self.migration_bandwidth
-        await asyncio.sleep(migration_latency)