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 `_set_default_args_v0` function (#25409)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Sep 23, 2025 at 09:52 UTC
c625f9043c5beb5921e94c6a5b3ac18372bab4db
1 file changed
+11
-72
vllm/engine/arg_utils.py
+11
-72
@@ -1147,20 +1147,15 @@ class EngineArgs:
1147
else:
1148
envs.set_vllm_use_v1(use_v1)
1149
1150
- # Set default arguments for V0 or V1 Engine.
1151
- if use_v1:
1152
- self._set_default_args_v1(usage_context, model_config)
1153
- # Disable chunked prefill for POWER (ppc64le)/ARM/s390x CPUs in V1
1154
- if current_platform.is_cpu(
1155
- ) and current_platform.get_cpu_architecture() in (
1156
- CpuArchEnum.POWERPC, CpuArchEnum.S390X, CpuArchEnum.ARM):
1157
- logger.info(
1158
- "Chunked prefill is not supported for ARM and POWER "
1159
- "and S390X CPUs; "
1160
- "disabling it for V1 backend.")
1161
- self.enable_chunked_prefill = False
1162
- else:
1163
- self._set_default_args_v0(model_config)
1150
+ # Set default arguments for V1 Engine.
1151
+ self._set_default_args(usage_context, model_config)
1152
+ # Disable chunked prefill for POWER (ppc64le)/ARM/s390x CPUs in V1
1153
+ if current_platform.is_cpu() and current_platform.get_cpu_architecture(
1154
+ ) in (CpuArchEnum.POWERPC, CpuArchEnum.S390X, CpuArchEnum.ARM):
1155
+ logger.info("Chunked prefill is not supported for ARM and POWER "
1156
+ "and S390X CPUs; "
1157
+ "disabling it for V1 backend.")
1158
+ self.enable_chunked_prefill = False
1159
assert self.enable_chunked_prefill is not None
1160
1161
sliding_window: Optional[int] = None
@@ -1528,64 +1523,8 @@ class EngineArgs:
1523
1524
return True
1525
1531
- def _set_default_args_v0(self, model_config: ModelConfig) -> None:
1532
- """Set Default Arguments for V0 Engine."""
1533
-
1534
- max_model_len = model_config.max_model_len
1535
- use_long_context = max_model_len > 32768
1536
- if self.enable_chunked_prefill is None:
1537
- # Chunked prefill not supported for Multimodal or MLA in V0.
1538
- if model_config.is_multimodal_model or model_config.use_mla:
1539
- self.enable_chunked_prefill = False
1540
-
1541
- # Enable chunked prefill by default for long context (> 32K)
1542
- # models to avoid OOM errors in initial memory profiling phase.
1543
- elif use_long_context:
1544
- is_gpu = current_platform.is_cuda()
1545
- use_sliding_window = (model_config.get_sliding_window()
1546
- is not None)
1547
- use_spec_decode = self.speculative_config is not None
1548
-
1549
- if (is_gpu and not use_sliding_window and not use_spec_decode
1550
- and not self.enable_lora):
1551
- self.enable_chunked_prefill = True
1552
- logger.warning(
1553
- "Chunked prefill is enabled by default for models "
1554
- "with max_model_len > 32K. Chunked prefill might "
1555
- "not work with some features or models. If you "
1556
- "encounter any issues, please disable by launching "
1557
- "with --enable-chunked-prefill=False.")
1558
-
1559
- if self.enable_chunked_prefill is None:
1560
- self.enable_chunked_prefill = False
1561
-
1562
- if not self.enable_chunked_prefill and use_long_context:
1563
- logger.warning(
1564
- "The model has a long context length (%s). This may cause"
1565
- "OOM during the initial memory profiling phase, or result "
1566
- "in low performance due to small KV cache size. Consider "
1567
- "setting --max-model-len to a smaller value.", max_model_len)
1568
-
1569
- # Disable prefix caching for multimodal models for VLLM_V0.
1570
- if self.enable_prefix_caching and model_config.is_multimodal_model:
1571
- logger.warning(
1572
- "--enable-prefix-caching is not supported for multimodal "
1573
- "models in V0 and has been disabled.")
1574
- self.enable_prefix_caching = False
1575
-
1576
- if self.enable_prompt_embeds:
1577
- logger.warning(
1578
- "--enable-prompt-embeds and --enable-prefix-caching "
1579
- "are not supported together in V0. Prefix caching has "
1580
- "been disabled.")
1581
- self.enable_prefix_caching = False
1582
-
1583
- # Set max_num_seqs to 256 for VLLM_V0.
1584
- if self.max_num_seqs is None:
1585
- self.max_num_seqs = 256
1586
-
1587
- def _set_default_args_v1(self, usage_context: UsageContext,
1588
- model_config: ModelConfig) -> None:
1526
+ def _set_default_args(self, usage_context: UsageContext,
1527
+ model_config: ModelConfig) -> None:
1528
"""Set Default Arguments for V1 Engine."""
1529
1530
# V1 always uses chunked prefills and prefix caching