A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[V0 Deprecation] Remove args for multi-step scheduling (#22779)
Signed-off-by: Woosuk Kwon <woosuk@thinkingmachines.ai>
Woosuk Kwon committed
Aug 12, 2025 at 20:38 UTC
c5830381afbef44023ec1c97ae61ff02f22b1f9a
2 files changed
+1
-27
tests/utils_/test_utils.py
-1
@@ -161,7 +161,6 @@ def parser_with_config():
161
parser.add_argument('--port', type=int)
162
parser.add_argument('--tensor-parallel-size', type=int)
163
parser.add_argument('--trust-remote-code', action='store_true')
164
- parser.add_argument('--multi-step-stream-outputs', action=StoreBoolean)
164
return parser
165
166
vllm/config/scheduler.py
+1
-26
@@ -115,12 +115,6 @@ class SchedulerConfig:
115
(e.g., beam search), recomputation is not currently supported. In
116
such a case, we use swapping instead."""
117
118
- num_scheduler_steps: int = 1
119
- """Maximum number of forward steps per scheduler call."""
120
-
121
- multi_step_stream_outputs: bool = True
122
- """If False, then multi-step will stream outputs at the end of all steps"""
123
-
118
send_delta_data: bool = False
119
"""Private API. If used, scheduler sends delta data to
120
workers instead of an entire data. It should be enabled only
@@ -193,16 +187,7 @@ class SchedulerConfig:
187
188
if self.max_num_batched_tokens is None:
189
if self.enable_chunked_prefill:
196
- if self.num_scheduler_steps > 1:
197
- # Multi-step Chunked-Prefill doesn't allow prompt-chunking
198
- # for now. Have max_num_batched_tokens set to max_model_len
199
- # so we don't reject sequences on account of a short
200
- # max_num_batched_tokens.
201
- self.max_num_batched_tokens = max(
202
- self.max_model_len, DEFAULT_MAX_NUM_BATCHED_TOKENS)
203
- else:
204
- self.max_num_batched_tokens = (
205
- DEFAULT_MAX_NUM_BATCHED_TOKENS)
190
+ self.max_num_batched_tokens = DEFAULT_MAX_NUM_BATCHED_TOKENS
191
else:
192
# If max_model_len is too short, use
193
# DEFAULT_MAX_NUM_BATCHED_TOKENS as the default value
@@ -293,12 +278,6 @@ class SchedulerConfig:
278
f"({self.num_lookahead_slots}) must be greater than or "
279
"equal to 0.")
280
296
- if self.num_scheduler_steps < 1:
297
- raise ValueError(
298
- "num_scheduler_steps "
299
- f"({self.num_scheduler_steps}) must be greater than or "
300
- "equal to 1.")
301
-
281
if self.max_num_partial_prefills < 1:
282
raise ValueError(
283
f"max_num_partial_prefills ({self.max_num_partial_prefills}) "
@@ -323,7 +302,3 @@ class SchedulerConfig:
302
f"max_num_partial_prefills ({self.max_num_partial_prefills}).")
303
304
return self
326
-
327
- @property
328
- def is_multi_step(self) -> bool:
329
- return self.num_scheduler_steps > 1