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 `VLLM_USE_V1` from docs and scripts (#26336)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Oct 7, 2025 at 16:46 UTC
7e4cd070b01dc1f8666fa7ded3fa7ec3ade270c2
11 files changed
+17
-26
.buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
-5
@@ -454,11 +454,6 @@ main() {
454
fi
455
check_hf_token
456
457
- # Set to v1 to run v1 benchmark
458
- if [[ "${ENGINE_VERSION:-v0}" == "v1" ]]; then
459
- export VLLM_USE_V1=1
460
- fi
461
-
457
# dependencies
458
(which wget && which curl) || (apt-get update && apt-get install -y wget curl)
459
(which jq) || (apt-get update && apt-get -y install jq)
.buildkite/scripts/hardware_ci/run-tpu-v1-test-part2.sh
+1
-2
@@ -64,10 +64,9 @@ python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git
64
&& python3 -m pip install --progress-bar off "lm-eval @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
65
&& python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
66
echo "--- Python dependencies installed ---"
67
-export VLLM_USE_V1=1
67
+
68
export VLLM_XLA_CHECK_RECOMPILATION=1
69
export VLLM_XLA_CACHE_PATH=
70
-echo "Using VLLM V1"
70
71
echo "--- Hardware Information ---"
72
# tpu-info
.buildkite/scripts/hardware_ci/run-tpu-v1-test.sh
+1
-2
@@ -64,10 +64,9 @@ python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git
64
&& python3 -m pip install --progress-bar off "lm-eval @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
65
&& python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
66
echo "--- Python dependencies installed ---"
67
-export VLLM_USE_V1=1
67
+
68
export VLLM_XLA_CHECK_RECOMPILATION=1
69
export VLLM_XLA_CACHE_PATH=
70
-echo "Using VLLM V1"
70
71
echo "--- Hardware Information ---"
72
# tpu-info
.buildkite/scripts/tpu/run_bm.sh
+1
-1
@@ -42,7 +42,7 @@ echo "lanching vllm..."
42
echo "logging to $VLLM_LOG"
43
echo
44
45
-VLLM_USE_V1=1 vllm serve $MODEL \
45
+vllm serve $MODEL \
46
--seed 42 \
47
--max-num-seqs $MAX_NUM_SEQS \
48
--max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \
benchmarks/auto_tune/auto_tune.sh
+2
-2
@@ -96,11 +96,11 @@ start_server() {
96
# This correctly passes each element as a separate argument.
97
if [[ -n "$profile_dir" ]]; then
98
# Start server with profiling enabled
99
- VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir \
99
+ VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir \
100
vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
101
else
102
# Start server without profiling
103
- VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 \
103
+ VLLM_SERVER_DEV_MODE=1 \
104
vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
105
fi
106
local server_pid=$!
docs/design/p2p_nccl_connector.md
+8
-8
@@ -97,7 +97,7 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
97
??? console "Command"
98
99
```shell
100
- VLLM_USE_V1=1 CUDA_VISIBLE_DEVICES=0 vllm serve {your model directory} \
100
+ CUDA_VISIBLE_DEVICES=0 vllm serve {your model directory} \
101
--host 0.0.0.0 \
102
--port 20001 \
103
--tensor-parallel-size 1 \
@@ -118,7 +118,7 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
118
??? console "Command"
119
120
```shell
121
- VLLM_USE_V1=1 CUDA_VISIBLE_DEVICES=1 vllm serve {your model directory} \
121
+ CUDA_VISIBLE_DEVICES=1 vllm serve {your model directory} \
122
--host 0.0.0.0 \
123
--port 20002 \
124
--tensor-parallel-size 1 \
@@ -139,7 +139,7 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
139
??? console "Command"
140
141
```shell
142
- VLLM_USE_V1=1 CUDA_VISIBLE_DEVICES=2 vllm serve {your model directory} \
142
+ CUDA_VISIBLE_DEVICES=2 vllm serve {your model directory} \
143
--host 0.0.0.0 \
144
--port 20003 \
145
--tensor-parallel-size 1 \
@@ -160,7 +160,7 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
160
??? console "Command"
161
162
```shell
163
- VLLM_USE_V1=1 CUDA_VISIBLE_DEVICES=3 vllm serve {your model directory} \
163
+ CUDA_VISIBLE_DEVICES=3 vllm serve {your model directory} \
164
--host 0.0.0.0 \
165
--port 20004 \
166
--tensor-parallel-size 1 \
@@ -190,7 +190,7 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
190
??? console "Command"
191
192
```shell
193
- VLLM_USE_V1=1 CUDA_VISIBLE_DEVICES=0 vllm serve {your model directory} \
193
+ CUDA_VISIBLE_DEVICES=0 vllm serve {your model directory} \
194
--host 0.0.0.0 \
195
--port 20001 \
196
--tensor-parallel-size 1 \
@@ -211,7 +211,7 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
211
??? console "Command"
212
213
```shell
214
- VLLM_USE_V1=1 CUDA_VISIBLE_DEVICES=1 vllm serve {your model directory} \
214
+ CUDA_VISIBLE_DEVICES=1 vllm serve {your model directory} \
215
--host 0.0.0.0 \
216
--port 20002 \
217
--tensor-parallel-size 1 \
@@ -232,7 +232,7 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
232
??? console "Command"
233
234
```shell
235
- VLLM_USE_V1=1 CUDA_VISIBLE_DEVICES=2 vllm serve {your model directory} \
235
+ CUDA_VISIBLE_DEVICES=2 vllm serve {your model directory} \
236
--host 0.0.0.0 \
237
--port 20003 \
238
--tensor-parallel-size 1 \
@@ -253,7 +253,7 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
253
??? console "Command"
254
255
```shell
256
- VLLM_USE_V1=1 CUDA_VISIBLE_DEVICES=3 vllm serve {your model directory} \
256
+ CUDA_VISIBLE_DEVICES=3 vllm serve {your model directory} \
257
--host 0.0.0.0 \
258
--port 20004 \
259
--tensor-parallel-size 1 \
docs/design/torch_compile.md
+1
-1
@@ -2,7 +2,7 @@
2
3
In vLLM's V1 architecture, `torch.compile` is enabled by default and is a critical part of the framework. This document gives a simple walk-through example to show how to understand the `torch.compile` usage.
4
5
-Throughout the example, we will run a common Llama model using v1, and turn on debug level logging to show all the details. The command to be used is `VLLM_USE_V1=1 VLLM_LOGGING_LEVEL=DEBUG vllm serve meta-llama/Llama-3.2-1B`.
5
+Throughout the example, we will run a common Llama model, and turn on debug level logging to show all the details. The command to be used is `VLLM_LOGGING_LEVEL=DEBUG vllm serve meta-llama/Llama-3.2-1B`.
6
7
## Compilation Cache
8
examples/online_serving/disaggregated_serving_p2p_nccl_xpyd/disagg_example_p2p_nccl_xpyd.sh
+2
-2
@@ -166,7 +166,7 @@ main() {
166
local kv_port=$((21001 + i))
167
168
echo " Prefill server $((i+1)): GPU $gpu_id, Port $port, KV Port $kv_port"
169
- CUDA_VISIBLE_DEVICES=$gpu_id VLLM_USE_V1=1 vllm serve $MODEL \
169
+ CUDA_VISIBLE_DEVICES=$gpu_id vllm serve $MODEL \
170
--enforce-eager \
171
--host 0.0.0.0 \
172
--port $port \
@@ -194,7 +194,7 @@ main() {
194
local kv_port=$((22001 + i))
195
196
echo " Decode server $((i+1)): GPU $gpu_id, Port $port, KV Port $kv_port"
197
- VLLM_USE_V1=1 CUDA_VISIBLE_DEVICES=$gpu_id vllm serve $MODEL \
197
+ CUDA_VISIBLE_DEVICES=$gpu_id vllm serve $MODEL \
198
--enforce-eager \
199
--host 0.0.0.0 \
200
--port $port \
examples/online_serving/elastic_ep/serve_deepseek_v2.sh
-1
@@ -55,7 +55,6 @@ done
55
echo "Starting vLLM server for $MODEL_NAME with data parallel size: $DATA_PARALLEL_SIZE and redundant experts: $REDUNDANT_EXPERTS"
56
57
export RAY_DEDUP_LOGS=0
58
-export VLLM_USE_V1=1
58
export VLLM_ALL2ALL_BACKEND="pplx"
59
export VLLM_USE_DEEP_GEMM=1
60
examples/online_serving/openai_chat_completion_client_with_tools_required.py
+1
-1
@@ -5,7 +5,7 @@ To run this example, you can start the vLLM server
5
without any specific flags:
6
7
```bash
8
-VLLM_USE_V1=0 vllm serve unsloth/Llama-3.2-1B-Instruct \
8
+vllm serve unsloth/Llama-3.2-1B-Instruct \
9
--structured-outputs-config.backend outlines
10
```
11
examples/online_serving/ray_serve_deepseek.py
-1
@@ -36,7 +36,6 @@ llm_config = LLMConfig(
36
},
37
# Set to the node's accelerator type.
38
accelerator_type="H100",
39
- runtime_env={"env_vars": {"VLLM_USE_V1": "1"}},
39
# Customize engine arguments as required (for example, vLLM engine kwargs).
40
engine_kwargs={
41
"tensor_parallel_size": 8,