[Examples] Resettle features examples. (#40995)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
wang.yuqi
2026-04-28 00:33:41 -07:00
committed by GitHub
parent 76c9cccc36
commit a8208e6a81
61 changed files with 234 additions and 234 deletions
@@ -15,7 +15,7 @@ compares the generation time for two queries that share the same prefix
but ask different questions.
Run:
python examples/offline_inference/automatic_prefix_caching.py
python examples/features/automatic_prefix_caching/automatic_prefix_caching_offline.py
"""
import time
@@ -6,7 +6,7 @@ of a Qwen model using the YARN method (rope_parameters)
and run a simple chat example.
Usage:
python examples/offline_inference/context_extension.py
python examples/features/context_extension/context_extension_offline.py
"""
from vllm import LLM, RequestOutput, SamplingParams
@@ -3,14 +3,14 @@
"""
Usage:
Single node:
python examples/offline_inference/data_parallel.py \
python examples/features/data_parallel/data_parallel_offline.py \
--model="ibm-research/PowerMoE-3b" \
-dp=2 \
-tp=2
Multi-node:
Node 0 (assume the node has ip of 10.99.48.128):
python examples/offline_inference/data_parallel.py \
python examples/features/data_parallel/data_parallel_offline.py \
--model="ibm-research/PowerMoE-3b" \
-dp=2 \
-tp=2 \
@@ -19,7 +19,7 @@ Multi-node:
--dp-master-addr=10.99.48.128 \
--dp-master-port=13345
Node 1:
python examples/offline_inference/data_parallel.py \
python examples/features/data_parallel/data_parallel_offline.py \
--model="ibm-research/PowerMoE-3b" \
-dp=2 \
-tp=2 \
@@ -12,7 +12,7 @@ from vllm.v1.metrics.loggers import AggregatedLoggingStatLogger
"""
To run this example, run the following commands simultaneously with
different CUDA_VISIBLE_DEVICES:
python examples/online_serving/multi_instance_data_parallel.py
python examples/features/data_parallel/multi_instance_data_parallel.py
vllm serve ibm-research/PowerMoE-3b -dp 2 -dpr 1 \
--data-parallel-address 127.0.0.1 --data-parallel-rpc-port 62300 \
@@ -9,7 +9,7 @@ This directory contains examples demonstrating how to use custom logits processo
Demonstrates how to instantiate vLLM with a custom logits processor class that operates at the batch level. The example uses a `DummyLogitsProcessor` that masks out all tokens except a specified `target_token` when passed via `SamplingParams.extra_args`.
```bash
python examples/offline_inference/logits_processor/custom.py
python examples/features/logits_processor/custom.py
```
### `custom_req.py` — Request-level logits processor wrapper
@@ -17,7 +17,7 @@ python examples/offline_inference/logits_processor/custom.py
Shows how to wrap a request-level logits processor (which operates on individual requests) to be compatible with vLLM's batch-level logits processing interface.
```bash
python examples/offline_inference/logits_processor/custom_req.py
python examples/features/logits_processor/custom_req.py
```
### `custom_req_init.py` — Request-level processor with engine config
@@ -25,7 +25,7 @@ python examples/offline_inference/logits_processor/custom_req.py
A special case of wrapping a request-level logits processor where the processor needs access to engine configuration or model metadata during initialization (e.g., vocabulary size, tokenizer info).
```bash
python examples/offline_inference/logits_processor/custom_req_init.py
python examples/features/logits_processor/custom_req_init.py
```
## Key Concepts
@@ -8,7 +8,7 @@ This is a guide to performing batch inference using the OpenAI batch file format
The OpenAI batch file format consists of a series of json objects on new lines.
[See here for an example file.](https://github.com/vllm-project/vllm/blob/main/examples/offline_inference/openai_batch/openai_example_batch.jsonl)
[See here for an example file.](https://github.com/vllm-project/vllm/blob/main/examples/features/openai_batch/openai_example_batch.jsonl)
Each line represents a separate request. See the [OpenAI package reference](https://platform.openai.com/docs/api-reference/batch/requestInput) for more details.
@@ -30,13 +30,13 @@ We currently support `/v1/chat/completions`, `/v1/embeddings`, and `/v1/score` e
To follow along with this example, you can download the example batch, or create your own batch file in your working directory.
```bash
wget https://raw.githubusercontent.com/vllm-project/vllm/main/examples/offline_inference/openai_batch/openai_example_batch.jsonl
wget https://raw.githubusercontent.com/vllm-project/vllm/main/examples/features/openai_batch/openai_example_batch.jsonl
```
Once you've created your batch file it should look like this
```bash
cat offline_inference/openai_batch/openai_example_batch.jsonl
cat features/openai_batch/openai_example_batch.jsonl
{"custom_id": "request-1", "method": "POST", "url": "/v1/chat/completions", "body": {"model": "meta-llama/Meta-Llama-3-8B-Instruct", "messages": [{"role": "system", "content": "You are a helpful assistant."},{"role": "user", "content": "Hello world!"}],"max_completion_tokens": 1000}}
{"custom_id": "request-2", "method": "POST", "url": "/v1/chat/completions", "body": {"model": "meta-llama/Meta-Llama-3-8B-Instruct", "messages": [{"role": "system", "content": "You are an unhelpful assistant."},{"role": "user", "content": "Hello world!"}],"max_completion_tokens": 1000}}
```
@@ -49,7 +49,7 @@ You can run the batch with the following command, which will write its results t
```bash
python -m vllm.entrypoints.openai.run_batch \
-i offline_inference/openai_batch/openai_example_batch.jsonl \
-i features/openai_batch/openai_example_batch.jsonl \
-o results.jsonl \
--model meta-llama/Meta-Llama-3-8B-Instruct
```
@@ -58,7 +58,7 @@ or use command-line:
```bash
vllm run-batch \
-i offline_inference/openai_batch/openai_example_batch.jsonl \
-i features/openai_batch/openai_example_batch.jsonl \
-o results.jsonl \
--model meta-llama/Meta-Llama-3-8B-Instruct
```
@@ -77,11 +77,11 @@ cat results.jsonl
The batch runner supports remote input and output urls that are accessible via http/https.
For example, to run against our example input file located at `https://raw.githubusercontent.com/vllm-project/vllm/main/examples/offline_inference/openai_batch/openai_example_batch.jsonl`, you can run
For example, to run against our example input file located at `https://raw.githubusercontent.com/vllm-project/vllm/main/examples/features/openai_batch/openai_example_batch.jsonl`, you can run
```bash
python -m vllm.entrypoints.openai.run_batch \
-i https://raw.githubusercontent.com/vllm-project/vllm/main/examples/offline_inference/openai_batch/openai_example_batch.jsonl \
-i https://raw.githubusercontent.com/vllm-project/vllm/main/examples/features/openai_batch/openai_example_batch.jsonl \
-o results.jsonl \
--model meta-llama/Meta-Llama-3-8B-Instruct
```
@@ -90,7 +90,7 @@ or use command-line:
```bash
vllm run-batch \
-i https://raw.githubusercontent.com/vllm-project/vllm/main/examples/offline_inference/openai_batch/openai_example_batch.jsonl \
-i https://raw.githubusercontent.com/vllm-project/vllm/main/examples/features/openai_batch/openai_example_batch.jsonl \
-o results.jsonl \
--model meta-llama/Meta-Llama-3-8B-Instruct
```
@@ -113,13 +113,13 @@ To integrate with cloud blob storage, we recommend using presigned urls.
To follow along with this example, you can download the example batch, or create your own batch file in your working directory.
```bash
wget https://raw.githubusercontent.com/vllm-project/vllm/main/examples/offline_inference/openai_batch/openai_example_batch.jsonl
wget https://raw.githubusercontent.com/vllm-project/vllm/main/examples/features/openai_batch/openai_example_batch.jsonl
```
Once you've created your batch file it should look like this
```bash
cat offline_inference/openai_batch/openai_example_batch.jsonl
cat features/openai_batch/openai_example_batch.jsonl
{"custom_id": "request-1", "method": "POST", "url": "/v1/chat/completions", "body": {"model": "meta-llama/Meta-Llama-3-8B-Instruct", "messages": [{"role": "system", "content": "You are a helpful assistant."},{"role": "user", "content": "Hello world!"}],"max_completion_tokens": 1000}}
{"custom_id": "request-2", "method": "POST", "url": "/v1/chat/completions", "body": {"model": "meta-llama/Meta-Llama-3-8B-Instruct", "messages": [{"role": "system", "content": "You are an unhelpful assistant."},{"role": "user", "content": "Hello world!"}],"max_completion_tokens": 1000}}
```
@@ -127,7 +127,7 @@ cat offline_inference/openai_batch/openai_example_batch.jsonl
Now upload your batch file to your S3 bucket.
```bash
aws s3 cp offline_inference/openai_batch/openai_example_batch.jsonl s3://MY_BUCKET/MY_INPUT_FILE.jsonl
aws s3 cp features/openai_batch/openai_example_batch.jsonl s3://MY_BUCKET/MY_INPUT_FILE.jsonl
```
### Step 2: Generate your presigned urls
@@ -1,135 +1,135 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Test pause/resume with Data Parallel (DP) via HTTP API.
This example demonstrates coordinated pause/resume across multiple DP ranks.
The pause synchronizes across all DP engines via all-reduce.
Prerequisites:
Start a vLLM server with data parallelism:
$ VLLM_SERVER_DEV_MODE=1 vllm serve facebook/opt-125m \
--enforce-eager \
--data-parallel-size 4 \
--tensor-parallel-size 1
Then run this script:
$ python data_parallel_pause_resume.py
The test verifies pause works by:
1. Starting a streaming generation request
2. Pausing the server mid-generation
3. Sleeping for PAUSE_DURATION seconds
4. Resuming the server
5. Verifying there was a gap in token generation matching the pause duration
"""
import argparse
import threading
import time
import requests
from openai import OpenAI
BASE_URL = "http://localhost:8000"
MODEL_NAME = "facebook/opt-125m"
PAUSE_DURATION = 3.0
def pause_generation(base_url: str, mode: str = "keep") -> None:
"""Pause generation via HTTP endpoint."""
url = f"{base_url}/pause"
response = requests.post(url, params={"mode": mode}, timeout=60)
response.raise_for_status()
print("Server paused")
def resume_generation(base_url: str) -> None:
"""Resume generation via HTTP endpoint."""
url = f"{base_url}/resume"
response = requests.post(url, timeout=60)
response.raise_for_status()
print("Server resumed")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--base-url", default=BASE_URL)
parser.add_argument("--model", default=MODEL_NAME)
args = parser.parse_args()
client = OpenAI(
base_url=f"{args.base_url}/v1",
api_key="EMPTY",
)
prompt = "Write a long story about a dragon. Once upon a time"
token_times: list[float] = []
pause_token_idx = 0
pause_triggered = threading.Event()
def generator_thread():
"""Stream tokens and record timestamps."""
stream = client.completions.create(
model=args.model,
prompt=prompt,
max_tokens=50,
stream=True,
)
for chunk in stream:
if chunk.choices[0].text:
token_times.append(time.monotonic())
token_count = len(token_times)
print(f"Token {token_count}: {chunk.choices[0].text!r}")
# Signal controller after some tokens
if token_count >= 5 and not pause_triggered.is_set():
pause_triggered.set()
def controller_thread():
"""Pause and resume the server."""
nonlocal pause_token_idx
# Wait for some tokens
pause_triggered.wait()
print(f"\nPausing server (keep mode) at token {len(token_times)}...")
pause_generation(args.base_url, mode="keep")
pause_token_idx = len(token_times)
print(f"Sleeping for {PAUSE_DURATION}s...")
time.sleep(PAUSE_DURATION)
print("Resuming server...")
resume_generation(args.base_url)
print("Resumed!\n")
# Run both threads
gen_thread = threading.Thread(target=generator_thread)
ctrl_thread = threading.Thread(target=controller_thread)
gen_thread.start()
ctrl_thread.start()
gen_thread.join()
ctrl_thread.join()
# Check gap at the pause point
if pause_token_idx < len(token_times):
pause_gap = token_times[pause_token_idx] - token_times[pause_token_idx - 1]
print(
f"\nGap after pause (token {pause_token_idx} -> "
f"{pause_token_idx + 1}): {pause_gap:.3f}s"
)
if pause_gap >= PAUSE_DURATION * 0.9:
print("Test passed! Pause synchronized across DP ranks.")
else:
print(f"Test failed! Expected ~{PAUSE_DURATION}s gap, got {pause_gap:.3f}s")
else:
print("Test failed! No tokens were generated after resuming.")
if __name__ == "__main__":
main()
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Test pause/resume with Data Parallel (DP) via HTTP API.
This example demonstrates coordinated pause/resume across multiple DP ranks.
The pause synchronizes across all DP engines via all-reduce.
Prerequisites:
Start a vLLM server with data parallelism:
$ VLLM_SERVER_DEV_MODE=1 vllm serve facebook/opt-125m \
--enforce-eager \
--data-parallel-size 4 \
--tensor-parallel-size 1
Then run this script:
$ python data_parallel_pause_resume.py
The test verifies pause works by:
1. Starting a streaming generation request
2. Pausing the server mid-generation
3. Sleeping for PAUSE_DURATION seconds
4. Resuming the server
5. Verifying there was a gap in token generation matching the pause duration
"""
import argparse
import threading
import time
import requests
from openai import OpenAI
BASE_URL = "http://localhost:8000"
MODEL_NAME = "facebook/opt-125m"
PAUSE_DURATION = 3.0
def pause_generation(base_url: str, mode: str = "keep") -> None:
"""Pause generation via HTTP endpoint."""
url = f"{base_url}/pause"
response = requests.post(url, params={"mode": mode}, timeout=60)
response.raise_for_status()
print("Server paused")
def resume_generation(base_url: str) -> None:
"""Resume generation via HTTP endpoint."""
url = f"{base_url}/resume"
response = requests.post(url, timeout=60)
response.raise_for_status()
print("Server resumed")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--base-url", default=BASE_URL)
parser.add_argument("--model", default=MODEL_NAME)
args = parser.parse_args()
client = OpenAI(
base_url=f"{args.base_url}/v1",
api_key="EMPTY",
)
prompt = "Write a long story about a dragon. Once upon a time"
token_times: list[float] = []
pause_token_idx = 0
pause_triggered = threading.Event()
def generator_thread():
"""Stream tokens and record timestamps."""
stream = client.completions.create(
model=args.model,
prompt=prompt,
max_tokens=50,
stream=True,
)
for chunk in stream:
if chunk.choices[0].text:
token_times.append(time.monotonic())
token_count = len(token_times)
print(f"Token {token_count}: {chunk.choices[0].text!r}")
# Signal controller after some tokens
if token_count >= 5 and not pause_triggered.is_set():
pause_triggered.set()
def controller_thread():
"""Pause and resume the server."""
nonlocal pause_token_idx
# Wait for some tokens
pause_triggered.wait()
print(f"\nPausing server (keep mode) at token {len(token_times)}...")
pause_generation(args.base_url, mode="keep")
pause_token_idx = len(token_times)
print(f"Sleeping for {PAUSE_DURATION}s...")
time.sleep(PAUSE_DURATION)
print("Resuming server...")
resume_generation(args.base_url)
print("Resumed!\n")
# Run both threads
gen_thread = threading.Thread(target=generator_thread)
ctrl_thread = threading.Thread(target=controller_thread)
gen_thread.start()
ctrl_thread.start()
gen_thread.join()
ctrl_thread.join()
# Check gap at the pause point
if pause_token_idx < len(token_times):
pause_gap = token_times[pause_token_idx] - token_times[pause_token_idx - 1]
print(
f"\nGap after pause (token {pause_token_idx} -> "
f"{pause_token_idx + 1}): {pause_gap:.3f}s"
)
if pause_gap >= PAUSE_DURATION * 0.9:
print("Test passed! Pause synchronized across DP ranks.")
else:
print(f"Test failed! Expected ~{PAUSE_DURATION}s gap, got {pause_gap:.3f}s")
else:
print("Test failed! No tokens were generated after resuming.")
if __name__ == "__main__":
main()
@@ -15,7 +15,7 @@ vllm serve meta-llama/Llama-3.2-1B-Instruct \
--enable-prompt-embeds
Run the client:
python examples/online_serving/prompt_embed_inference_with_openai_client.py
python examples/features/prompt_embed/prompt_embed_inference_with_openai_client.py
Model: meta-llama/Llama-3.2-1B-Instruct
Note: This model is gated on Hugging Face Hub.
@@ -15,7 +15,7 @@ Requirements:
- transformers
Run:
python examples/offline_inference/prompt_embed_inference.py
python examples/features/prompt_embed/prompt_embed_offline.py
"""
import torch
@@ -3,16 +3,16 @@
"""
Validates the loading of a model saved with the sharded_state format.
This script demonstrates how to load a model that was previously saved
using save_sharded_state.py and validates it by running inference.
using save_sharded_state_offline.py and validates it by running inference.
Example usage:
(First need to save a sharded_state mode)
python save_sharded_state.py \
python save_sharded_state_offline.py \
--model /path/to/load \
--tensor-parallel-size 8 \
--output /path/to/save/sharded/model
python load_sharded_state.py \
python load_sharded_state_offline.py \
--model /path/to/saved/sharded/model \
--load-format sharded_state \
--tensor-parallel-size 8 \
@@ -7,7 +7,7 @@ read its own shard rather than the entire checkpoint.
Example usage:
python save_sharded_state.py \
python save_sharded_state_offline.py \
--model /path/to/load \
--tensor-parallel-size 8 \
--output /path/to/save
@@ -20,7 +20,7 @@ vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-7B \
If you want to run this script standalone with `uv`, you can use the following:
```bash
uvx --from git+https://github.com/vllm-project/vllm#subdirectory=examples/online_serving/structured_outputs \
uvx --from git+https://github.com/vllm-project/vllm#subdirectory=examples/features/structured_outputs \
structured-outputs
```
@@ -34,19 +34,19 @@ See [feature docs](https://docs.vllm.ai/en/latest/features/structured_outputs.ht
Run all constraints, non-streaming:
```bash
uv run structured_outputs.py
uv run structured_outputs_offline.py
```
Run all constraints, streaming:
```bash
uv run structured_outputs.py --stream
uv run structured_outputs_offline.py --stream
```
Run certain constraints, for example `structural_tag` and `regex`, streaming:
```bash
uv run structured_outputs.py \
uv run structured_outputs_offline.py \
--constraint structural_tag regex \
--stream
```
@@ -54,5 +54,5 @@ uv run structured_outputs.py \
Run all constraints, with reasoning models and streaming:
```bash
uv run structured_outputs.py --reasoning --stream
uv run structured_outputs_offline.py --reasoning --stream
```
@@ -7,15 +7,15 @@ no internal lb supported in external_launcher mode.
To run this example:
```bash
$ torchrun --nproc-per-node=2 examples/offline_inference/torchrun_dp_example.py
$ torchrun --nproc-per-node=2 examples/features/torchrun/torchrun_dp_example_offline.py
```
With custom parallelism settings:
```bash
$ torchrun --nproc-per-node=8 examples/offline_inference/torchrun_dp_example.py \
$ torchrun --nproc-per-node=8 examples/features/torchrun/torchrun_dp_example_offline.py \
--tp-size=2 --pp-size=1 --dp-size=4 --enable-ep
```
"""
""" # noqa: E501
import argparse
@@ -4,7 +4,7 @@
experimental support for tensor-parallel inference with torchrun,
see https://github.com/vllm-project/vllm/issues/11400 for
the motivation and use case for this example.
run the script with `torchrun --nproc-per-node=4 torchrun_example.py`,
run the script with `torchrun --nproc-per-node=4 torchrun_example_offline.py`,
the argument `4` should match the product of `tensor_parallel_size` and
`pipeline_parallel_size` below. see `tests/distributed/test_torchrun_example.py`
for the unit test.
@@ -9,7 +9,7 @@ Validates that:
3. Results are deterministic across runs (baseline vs reference).
Usage:
python examples/offline_inference/routed_experts_e2e.py \
python examples/rl/routed_experts_e2e.py \
--model Qwen/Qwen3-30B-A3B \
--tp 4 \
--max-model-len 4096 \