Merge branch 'vllm-project:main' into main

.buildkite/nightly-benchmarks/kickoff-pipeline.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+# Install system packages
+apt update
+apt install -y curl jq
+
+# Install minijinja for templating
+curl -sSfL https://github.com/mitsuhiko/minijinja/releases/latest/download/minijinja-cli-installer.sh | sh
+source $HOME/.cargo/env
+
+# If BUILDKITE_PULL_REQUEST != "false", then we check the PR labels using curl and jq
+if [ "$BUILDKITE_PULL_REQUEST" != "false" ]; then
+  PR_LABELS=$(curl -s "https://api.github.com/repos/vllm-project/vllm/pulls/$BUILDKITE_PULL_REQUEST" | jq -r '.labels[].name')
+
+  if [[ $PR_LABELS == *"perf-benchmarks"* ]]; then
+    echo "This PR has the 'perf-benchmarks' label. Proceeding with the nightly benchmarks."
+  else
+    echo "This PR does not have the 'perf-benchmarks' label. Skipping the nightly benchmarks."
+    exit 0
+  fi
+fi
+
+# Upload sample.yaml
+buildkite-agent pipeline upload .buildkite/nightly-benchmarks/sample.yaml

.buildkite/nightly-benchmarks/sample.yaml

@@ -0,0 +1,39 @@
+steps:
+  # NOTE(simon): You can create separate blocks for different jobs
+  - label: "A100: NVIDIA SMI"
+    agents:
+      queue: A100
+    plugins:
+    - kubernetes:
+        podSpec:
+          containers:
+          # - image: us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:$BUILDKITE_COMMIT
+          # TODO(simon): check latest main branch or use the PR image.
+          - image: us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:45c35f0d58f4508bf43bd6af1d3d0d0ec0c915e6
+            command:
+            - bash -c 'nvidia-smi && nvidia-smi topo -m && pwd && ls'
+            resources:
+              limits:
+                nvidia.com/gpu: 8
+            volumeMounts:
+            - name: devshm
+              mountPath: /dev/shm
+          nodeSelector:
+            nvidia.com/gpu.product: NVIDIA-A100-SXM4-80GB
+          volumes:
+          - name: devshm
+            emptyDir:
+              medium: Memory
+  # TODO(simon): bring H100 online
+  # - label: "H100: NVIDIA SMI"
+  #   agents:
+  #     queue: H100
+  #   plugins:
+  #   - docker#v5.11.0:
+  #       image: us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:45c35f0d58f4508bf43bd6af1d3d0d0ec0c915e6
+  #       command:
+  #       - bash -c 'nvidia-smi && nvidia-smi topo -m'
+  #       propagate-environment: true
+  #       ipc: host
+  #       gpus: all
+

.buildkite/run-benchmarks.sh

@@ -50,7 +50,7 @@ echo "### Serving Benchmarks" >> benchmark_results.md
 sed -n '1p' benchmark_serving.txt >> benchmark_results.md # first line
 echo "" >> benchmark_results.md
 echo '```' >> benchmark_results.md
-tail -n 20 benchmark_serving.txt >> benchmark_results.md # last 20 lines
+tail -n 24 benchmark_serving.txt >> benchmark_results.md # last 24 lines
 echo '```' >> benchmark_results.md
 
 # if the agent binary is not found, skip uploading the results, exit 0

.buildkite/run-cpu-test.sh

@@ -11,7 +11,7 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Run the image
-docker run -itd -v ~/.cache/huggingface:/root/.cache/huggingface --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --name cpu-test cpu-test
+docker run -itd -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus=48-95 --cpuset-mems=1 --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --name cpu-test cpu-test
 
 # offline inference
 docker exec cpu-test bash -c "python3 examples/offline_inference.py"

.buildkite/test-pipeline.yaml

@@ -45,7 +45,8 @@ steps:
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
-  - pytest -v -s spec_decode/e2e/test_integration_dist.py 
+  - pytest -v -s spec_decode/e2e/test_integration_dist.py
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
 
 - label: Distributed Tests (Multiple Groups)
   #mirror_hardwares: [amd]
@@ -62,7 +63,6 @@ steps:
   mirror_hardwares: [amd]
 
   commands:
-  - pytest -v -s test_inputs.py
   - pytest -v -s entrypoints -m llm
   - pytest -v -s entrypoints -m openai
 
@@ -79,6 +79,13 @@ steps:
     - python3 llava_example.py
     - python3 tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
 
+- label: Inputs Test
+  #mirror_hardwares: [amd]
+  commands:
+    - bash ../.buildkite/download-images.sh
+    - pytest -v -s test_inputs.py
+    - pytest -v -s multimodal
+
 - label: Kernels Test %N
   #mirror_hardwares: [amd]
   command: pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
@@ -87,14 +94,13 @@ steps:
 - label: Models Test
   #mirror_hardwares: [amd]
   commands:
-    - bash ../.buildkite/download-images.sh
-    - pytest -v -s models --ignore=models/test_llava.py
+    - pytest -v -s models -m \"not llava\"
 
 - label: Llava Test
   mirror_hardwares: [amd]
   commands:
     - bash ../.buildkite/download-images.sh
-    - pytest -v -s models/test_llava.py
+    - pytest -v -s models -m llava
 
 - label: Prefix Caching Test
   mirror_hardwares: [amd]
@@ -118,7 +124,10 @@ steps:
 
 - label: Speculative decoding tests
   #mirror_hardwares: [amd]
-  command: pytest -v -s spec_decode
+  commands:
+    # See https://github.com/vllm-project/vllm/issues/5152
+    - export VLLM_ATTENTION_BACKEND=XFORMERS
+    - pytest -v -s spec_decode
 
 - label: LoRA Test %N
   #mirror_hardwares: [amd]
@@ -130,14 +139,7 @@ steps:
   num_gpus: 4
   # This test runs llama 13B, so it is required to run on 4 GPUs.
   commands:
-    # Temporarily run this way because we cannot clean up GPU mem usage
-    # for multi GPU tests.
-    # TODO(sang): Fix it.
-    - pytest -v -s lora/test_long_context.py::test_rotary_emb_replaced
-    - pytest -v -s lora/test_long_context.py::test_batched_rope_kernel
-    - pytest -v -s lora/test_long_context.py::test_self_consistency
-    - pytest -v -s lora/test_long_context.py::test_quality
-    - pytest -v -s lora/test_long_context.py::test_max_len
+    - pytest -v -s -x lora/test_long_context.py
 
 - label: Tensorizer Test
   #mirror_hardwares: [amd]

.buildkite/test-template-aws.j2

@@ -0,0 +1,59 @@
+{% set docker_image = "public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT" %}
+{% set default_working_dir = "/vllm-workspace/tests" %}
+
+steps:
+  - label: ":docker: build image"
+    agents:
+      queue: cpu_queue
+    commands:
+      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+      - "docker build --build-arg max_jobs=16 --tag {{ docker_image }} --target test --progress plain ."
+      - "docker push {{ docker_image }}"
+    env:
+      DOCKER_BUILDKIT: "1"
+    retry:
+      automatic:
+        - exit_status: -1  # Agent was lost
+          limit: 5
+        - exit_status: -10  # Agent was lost
+          limit: 5
+  - wait
+
+  {% for step in steps %}
+  - label: "{{ step.label }}"
+    agents:
+      {% if step.no_gpu %}
+      queue: cpu_queue
+      {% elif step.num_gpus == 2 or step.num_gpus == 4 %}
+      queue: gpu_4_queue
+      {% else %}
+      queue: gpu_1_queue
+      {% endif %}
+    soft_fail: true
+    {% if step.parallelism %}
+    parallelism: {{ step.parallelism }}
+    {% endif %}
+    retry:
+      automatic:
+        - exit_status: -1  # Agent was lost
+          limit: 5
+        - exit_status: -10  # Agent was lost
+          limit: 5
+    plugins:
+      - docker#v5.2.0:
+          image: {{ docker_image }}
+          always-pull: true
+          propagate-environment: true
+          {% if not step.no_gpu %}
+          gpus: all
+          {% endif %}
+          command: ["bash", "-c", "cd {{ (step.working_dir or default_working_dir) | safe  }} && {{ step.command  or (step.commands | join(' && ')) | safe }}"]
+          environment:
+            - VLLM_USAGE_SOURCE=ci-test
+            - HF_TOKEN
+            {% if step.label == "Speculative decoding tests" %}
+            - VLLM_ATTENTION_BACKEND=XFORMERS
+            {% endif %}
+          volumes:
+            - /dev/shm:/dev/shm
+  {% endfor %}

.buildkite/test-template.j2

@@ -4,7 +4,7 @@
 
 steps:
   - label: ":docker: build image"
-    commands: 
+    commands:
       - "docker build --build-arg max_jobs=16 --tag {{ docker_image }} --target test --progress plain ."
       - "docker push {{ docker_image }}"
     env:
@@ -28,6 +28,7 @@ steps:
         command: bash .buildkite/run-amd-test.sh "cd {{ (step.working_dir or default_working_dir) | safe  }} ; {{ step.command  or (step.commands | join(" ; ")) | safe }}"
         env:
           DOCKER_BUILDKIT: "1"
+        soft_fail: true
     {% endif %}
     {% endfor %}
 
@@ -36,7 +37,7 @@ steps:
     agents:
       queue: neuron
     command: bash .buildkite/run-neuron-test.sh
-    soft_fail: true
+    soft_fail: false
 
   - label: "Intel Test"
     depends_on: ~

Dockerfile.cpu

@@ -7,7 +7,7 @@ RUN apt-get update  -y \
     && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
 
 RUN pip install --upgrade pip \
-    && pip install wheel packaging ninja setuptools>=49.4.0 numpy
+    && pip install wheel packaging ninja "setuptools>=49.4.0" numpy
 
 FROM cpu-test-1 AS build
 

README.md

@@ -16,6 +16,13 @@ Easy, fast, and cheap LLM serving for everyone
 
 ---
 
+**Ray Summit CPF is Open (June 4th to June 20th)!**
+
+There will be a track for vLLM at the Ray Summit (09/30-10/02, SF) this year!
+If you have cool projects related to vLLM or LLM inference, we would love to see your proposals.
+This will be a great chance for everyone in the community to get together and learn.
+Please submit your proposal [here](https://raysummit.anyscale.com/flow/anyscale/raysummit2024/landing/page/eventsite)
+
 **The Fourth vLLM Bay Area Meetup (June 11th 5:30pm-8pm PT)**
 
 We are thrilled to announce our fourth vLLM Meetup!
@@ -104,6 +111,7 @@ vLLM is a community project. Our compute resources for development and testing a
 - Dropbox
 - Lambda Lab
 - NVIDIA
+- Sequoia Capital
 - Replicate
 - Roblox
 - RunPod

benchmarks/benchmark_serving.py

@@ -56,6 +56,9 @@ class BenchmarkMetrics:
     mean_tpot_ms: float
     median_tpot_ms: float
     p99_tpot_ms: float
+    mean_itl_ms: float
+    median_itl_ms: float
+    p99_itl_ms: float
 
 
 def sample_sharegpt_requests(
@@ -200,16 +203,24 @@ def calculate_metrics(
     actual_output_lens = []
     total_input = 0
     completed = 0
+    itls = []
     tpots = []
     ttfts = []
     for i in range(len(outputs)):
         if outputs[i].success:
-            output_len = len(tokenizer(outputs[i].generated_text).input_ids)
+            # We use the tokenizer to count the number of output tokens for all
+            # serving backends instead of looking at len(outputs[i].itl) since
+            # multiple output tokens may be bundled together
+            # Note: this may inflate the output token count slightly
+            output_len = len(
+                tokenizer(outputs[i].generated_text,
+                          add_special_tokens=False).input_ids)
             actual_output_lens.append(output_len)
             total_input += input_requests[i][1]
             if output_len > 1:
                 tpots.append(
                     (outputs[i].latency - outputs[i].ttft) / (output_len - 1))
+            itls += outputs[i].itl
             ttfts.append(outputs[i].ttft)
             completed += 1
         else:
@@ -234,6 +245,9 @@ def calculate_metrics(
         mean_tpot_ms=np.mean(tpots or 0) * 1000,
         median_tpot_ms=np.median(tpots or 0) * 1000,
         p99_tpot_ms=np.percentile(tpots or 0, 99) * 1000,
+        mean_itl_ms=np.mean(itls or 0) * 1000,
+        median_itl_ms=np.median(itls or 0) * 1000,
+        p99_itl_ms=np.percentile(itls or 0, 99) * 1000,
     )
 
     return metrics, actual_output_lens
@@ -333,6 +347,10 @@ async def benchmark(
     print("{:<40} {:<10.2f}".format("Median TPOT (ms):",
                                     metrics.median_tpot_ms))
     print("{:<40} {:<10.2f}".format("P99 TPOT (ms):", metrics.p99_tpot_ms))
+    print("{s:{c}^{n}}".format(s='Inter-token Latency', n=50, c='-'))
+    print("{:<40} {:<10.2f}".format("Mean ITL (ms):", metrics.mean_itl_ms))
+    print("{:<40} {:<10.2f}".format("Median ITL (ms):", metrics.median_itl_ms))
+    print("{:<40} {:<10.2f}".format("P99 ITL (ms):", metrics.p99_itl_ms))
     print("=" * 50)
 
     result = {
@@ -349,6 +367,9 @@ async def benchmark(
         "mean_tpot_ms": metrics.mean_tpot_ms,
         "median_tpot_ms": metrics.median_tpot_ms,
         "p99_tpot_ms": metrics.p99_tpot_ms,
+        "mean_itl_ms": metrics.mean_itl_ms,
+        "median_itl_ms": metrics.median_itl_ms,
+        "p99_itl_ms": metrics.p99_itl_ms,
         "input_lens": [output.prompt_len for output in outputs],
         "output_lens": actual_output_lens,
         "ttfts": [output.ttft for output in outputs],