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 15, 16:00 UTC
[V1] Revert the default `max_num_seqs` to V0 values for most hardware (#16158)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Apr 8, 2025 at 01:54 UTC
66d433b94ffd215820c6c90d97ca0d511dc12143
3 files changed
+6
-10
docs/source/getting_started/v1_user_guide.md
-7
@@ -156,10 +156,3 @@ vLLM V1 is currently optimized for decoder-only transformers. Models requiring
156
cross-attention between separate encoder and decoder are not yet supported (e.g., `BartForConditionalGeneration`, `MllamaForConditionalGeneration`).
157
158
For a complete list of supported models, see the [list of supported models](https://docs.vllm.ai/en/latest/models/supported_models.html).
159
-
160
-## Frequently Asked Questions
161
-
162
-**I'm using vLLM V1 and I'm getting CUDA OOM errors. What should I do?**
163
-The default `max_num_seqs` has been raised from `256` in V0 to `1024` in V1. If you encounter CUDA OOM only when using V1 engine, try setting a lower value of `max_num_seqs` or `gpu_memory_utilization`.
164
-
165
-On the other hand, if you get an error about insufficient memory for the cache blocks, you should increase `gpu_memory_utilization` as this indicates that your GPU has sufficient memory but you're not allocating enough to vLLM for KV cache blocks.
tests/v1/engine/test_engine_args.py
+4
-2
@@ -64,15 +64,17 @@ def test_defaults_with_usage_context():
64
# For H100 and H200, we use larger default values.
65
default_llm_tokens = 16384
66
default_server_tokens = 8192
67
+ default_max_num_seqs = 1024
68
else:
69
default_llm_tokens = 8192
70
default_server_tokens = 2048
71
+ default_max_num_seqs = 256
72
71
- assert vllm_config.scheduler_config.max_num_seqs == 1024
73
+ assert vllm_config.scheduler_config.max_num_seqs == default_max_num_seqs
74
assert vllm_config.scheduler_config.max_num_batched_tokens == default_llm_tokens # noqa: E501
75
76
engine_args = EngineArgs(model="facebook/opt-125m")
77
vllm_config = engine_args.create_engine_config(
78
UsageContext.OPENAI_API_SERVER)
77
- assert vllm_config.scheduler_config.max_num_seqs == 1024
79
+ assert vllm_config.scheduler_config.max_num_seqs == default_max_num_seqs
80
assert vllm_config.scheduler_config.max_num_batched_tokens == default_server_tokens # noqa: E501
vllm/engine/arg_utils.py
+2
-1
@@ -1666,12 +1666,14 @@ class EngineArgs:
1666
UsageContext.LLM_CLASS: 16384,
1667
UsageContext.OPENAI_API_SERVER: 8192,
1668
}
1669
+ default_max_num_seqs = 1024
1670
else:
1671
# TODO(woosuk): Tune the default values for other hardware.
1672
default_max_num_batched_tokens = {
1673
UsageContext.LLM_CLASS: 8192,
1674
UsageContext.OPENAI_API_SERVER: 2048,
1675
}
1676
+ default_max_num_seqs = 256
1677
1678
use_context_value = usage_context.value if usage_context else None
1679
if (self.max_num_batched_tokens is None
@@ -1682,7 +1684,6 @@ class EngineArgs:
1684
"Setting max_num_batched_tokens to %d for %s usage context.",
1685
self.max_num_batched_tokens, use_context_value)
1686
1685
- default_max_num_seqs = 1024
1687
if self.max_num_seqs is None:
1688
self.max_num_seqs = default_max_num_seqs
1689