Merge pull request #5 from vllm-project/main

[pull] main from vllm-project:main

.buildkite/run-amd-test.sh

@@ -5,6 +5,19 @@ set -ex
 # Print ROCm version
 rocminfo
 
+
+echo "reset" > /opt/amdgpu/etc/gpu_state
+
+while true; do
+        sleep 3
+        if grep -q clean /opt/amdgpu/etc/gpu_state; then
+                echo "GPUs state is \"clean\""
+                break
+        fi
+done
+
+
+
 # Try building the docker image
 docker build -t rocm -f Dockerfile.rocm .
 
@@ -14,7 +27,8 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Run the image
-docker run --device /dev/kfd --device /dev/dri --network host --name rocm rocm python3 -m vllm.entrypoints.api_server &
+export HIP_VISIBLE_DEVICES=1
+docker run --device /dev/kfd --device /dev/dri --network host -e HIP_VISIBLE_DEVICES --name rocm rocm python3 -m vllm.entrypoints.api_server &
 
 # Wait for the server to start
 wait_for_server_to_start() {

.buildkite/test-pipeline.yaml

@@ -96,6 +96,9 @@ steps:
 - label: Metrics Test
   command: pytest -v -s metrics
 
+- label: Quantization Test
+  command: pytest -v -s quantization
+
 - label: Benchmarks
   working_dir: "/vllm-workspace/.buildkite"
   commands:

.github/workflows/mypy.yaml

@@ -32,19 +32,19 @@ jobs:
         pip install types-setuptools
     - name: Mypy
       run: |
-        mypy vllm/attention/*.py --follow-imports=skip --config-file pyproject.toml
-        mypy vllm/core/*.py --follow-imports=skip --config-file pyproject.toml
-        mypy vllm/distributed/*.py --follow-imports=skip --config-file pyproject.toml
-        mypy vllm/entrypoints/*.py --follow-imports=skip --config-file pyproject.toml
-        mypy vllm/executor/*.py --follow-imports=skip --config-file pyproject.toml
-        mypy vllm/usage/*.py --follow-imports=skip --config-file pyproject.toml
-        mypy vllm/*.py --follow-imports=skip --config-file pyproject.toml
-        mypy vllm/transformers_utils/*.py --follow-imports=skip --config-file pyproject.toml
+        mypy vllm/attention --config-file pyproject.toml
+        mypy vllm/distributed --config-file pyproject.toml
+        mypy vllm/entrypoints --config-file pyproject.toml
+        mypy vllm/executor --config-file pyproject.toml
+        mypy vllm/usage --config-file pyproject.toml
+        mypy vllm/*.py --config-file pyproject.toml
+        mypy vllm/transformers_utils --config-file pyproject.toml
+        mypy vllm/engine  --config-file pyproject.toml
+        mypy vllm/worker --config-file pyproject.toml
+        mypy vllm/spec_decode --config-file pyproject.toml
+        mypy vllm/lora --config-file pyproject.toml
 
-        mypy vllm/engine/*.py --follow-imports=skip --config-file pyproject.toml
-        mypy vllm/worker/*.py --follow-imports=skip --config-file pyproject.toml
-        mypy vllm/spec_decode/*.py --follow-imports=skip --config-file pyproject.toml
-        mypy vllm/model_executor/*.py --follow-imports=skip --config-file pyproject.toml
-        # TODO(sang): Follow up
-        # mypy vllm/lora/*.py --follow-imports=skip --config-file pyproject.toml
+        # TODO(sang): Fix nested dir
+        mypy vllm/model_executor/*.py  --config-file pyproject.toml
+        mypy vllm/core/*.py --follow-imports=skip --config-file pyproject.toml
 

.github/workflows/publish.yml

@@ -56,6 +56,9 @@ jobs:
       - name: Checkout
         uses: actions/checkout@v3
 
+      - name: Setup ccache
+        uses: hendrikmuhs/[email protected]
+
       - name: Set up Linux Env
         if: ${{ runner.os == 'Linux' }}
         run: |

.gitignore

@@ -70,6 +70,8 @@ instance/
 
 # Sphinx documentation
 docs/_build/
+docs/source/getting_started/examples/*.rst
+!**/*.template.rst
 
 # PyBuilder
 .pybuilder/

CMakeLists.txt

@@ -167,12 +167,14 @@ set(VLLM_EXT_SRC
   "csrc/layernorm_kernels.cu"
   "csrc/quantization/squeezellm/quant_cuda_kernel.cu"
   "csrc/quantization/gptq/q_gemm.cu"
+  "csrc/quantization/fp8/fp8_cuda_kernels.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/moe_align_block_size_kernels.cu"
   "csrc/pybind.cpp")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_EXT_SRC
+    "csrc/quantization/aqlm/gemm_kernels.cu"
     "csrc/quantization/awq/gemm_kernels.cu"
     "csrc/quantization/marlin/marlin_cuda_kernel.cu"
     "csrc/custom_all_reduce.cu")
@@ -210,23 +212,11 @@ define_gpu_extension_target(
 
 set(VLLM_PUNICA_EXT_SRC
   "csrc/punica/bgmv/bgmv_bf16_bf16_bf16.cu"
-  "csrc/punica/bgmv/bgmv_bf16_bf16_fp16.cu"
-  "csrc/punica/bgmv/bgmv_bf16_fp16_bf16.cu"
-  "csrc/punica/bgmv/bgmv_bf16_fp16_fp16.cu"
   "csrc/punica/bgmv/bgmv_bf16_fp32_bf16.cu"
-  "csrc/punica/bgmv/bgmv_bf16_fp32_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp16_bf16_bf16.cu"
-  "csrc/punica/bgmv/bgmv_fp16_bf16_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp16_fp16_bf16.cu"
   "csrc/punica/bgmv/bgmv_fp16_fp16_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp16_fp32_bf16.cu"
   "csrc/punica/bgmv/bgmv_fp16_fp32_fp16.cu"
   "csrc/punica/bgmv/bgmv_fp32_bf16_bf16.cu"
-  "csrc/punica/bgmv/bgmv_fp32_bf16_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp32_fp16_bf16.cu"
   "csrc/punica/bgmv/bgmv_fp32_fp16_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp32_fp32_bf16.cu"
-  "csrc/punica/bgmv/bgmv_fp32_fp32_fp16.cu"
   "csrc/punica/punica_ops.cc")
 
 #

README.md

@@ -74,10 +74,11 @@ vLLM seamlessly supports many Hugging Face models, including the following archi
 - Mistral (`mistralai/Mistral-7B-v0.1`, `mistralai/Mistral-7B-Instruct-v0.1`, etc.)
 - Mixtral (`mistralai/Mixtral-8x7B-v0.1`, `mistralai/Mixtral-8x7B-Instruct-v0.1`, `mistral-community/Mixtral-8x22B-v0.1`, etc.)
 - MPT (`mosaicml/mpt-7b`, `mosaicml/mpt-30b`, etc.)
-- OLMo (`allenai/OLMo-1B`, `allenai/OLMo-7B`, etc.)
+- OLMo (`allenai/OLMo-1B-hf`, `allenai/OLMo-7B-hf`, etc.)
 - OPT (`facebook/opt-66b`, `facebook/opt-iml-max-30b`, etc.)
 - Orion (`OrionStarAI/Orion-14B-Base`, `OrionStarAI/Orion-14B-Chat`, etc.)
 - Phi (`microsoft/phi-1_5`, `microsoft/phi-2`, etc.)
+- Phi-3 (`microsoft/Phi-3-mini-4k-instruct`, `microsoft/Phi-3-mini-128k-instruct`, etc.)
 - Qwen (`Qwen/Qwen-7B`, `Qwen/Qwen-7B-Chat`, etc.)
 - Qwen2 (`Qwen/Qwen1.5-7B`, `Qwen/Qwen1.5-7B-Chat`, etc.)
 - Qwen2MoE (`Qwen/Qwen1.5-MoE-A2.7B`, `Qwen/Qwen1.5-MoE-A2.7B-Chat`, etc.)

benchmarks/benchmark_serving.py

@@ -27,7 +27,7 @@
 import warnings
 from dataclasses import dataclass
 from datetime import datetime
-from typing import AsyncGenerator, List, Tuple
+from typing import AsyncGenerator, List, Optional, Tuple
 
 import numpy as np
 from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
@@ -58,7 +58,11 @@ def sample_sharegpt_requests(
     dataset_path: str,
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
+    fixed_output_len: Optional[int] = None,
 ) -> List[Tuple[str, int, int]]:
+    if fixed_output_len is not None and fixed_output_len < 4:
+        raise ValueError("output_len too small")
+
     # Load the dataset.
     with open(dataset_path) as f:
         dataset = json.load(f)
@@ -68,38 +72,32 @@ def sample_sharegpt_requests(
     dataset = [(data["conversations"][0]["value"],
                 data["conversations"][1]["value"]) for data in dataset]
 
-    # some of these will be filtered out, so sample more than we need
-    sampled_indices = random.sample(range(len(dataset)),
-                                    int(num_requests * 1.2))
-    dataset = [dataset[i] for i in sampled_indices]
-
-    # Tokenize the prompts and completions.
-    prompts = [prompt for prompt, _ in dataset]
-    prompt_token_ids = tokenizer(prompts).input_ids
-    completions = [completion for _, completion in dataset]
-    completion_token_ids = tokenizer(completions).input_ids
-    tokenized_dataset = []
-    for i in range(len(dataset)):
-        output_len = len(completion_token_ids[i])
-        tokenized_dataset.append((prompts[i], prompt_token_ids[i], output_len))
+    # Shuffle the dataset.
+    random.shuffle(dataset)
 
-    # Filter out too long sequences.
+    # Filter out sequences that are too long or too short
     filtered_dataset: List[Tuple[str, int, int]] = []
-    for prompt, prompt_token_ids, output_len in tokenized_dataset:
+    for i in range(len(dataset)):
+        if len(filtered_dataset) == num_requests:
+            break
+
+        # Tokenize the prompts and completions.
+        prompt = dataset[i][0]
+        prompt_token_ids = tokenizer(prompt).input_ids
+        completion = dataset[i][1]
+        completion_token_ids = tokenizer(completion).input_ids
         prompt_len = len(prompt_token_ids)
+        output_len = len(completion_token_ids
+                         ) if fixed_output_len is None else fixed_output_len
         if prompt_len < 4 or output_len < 4:
             # Prune too short sequences.
-            # This is because TGI causes errors when the input or output length
-            # is too short.
             continue
         if prompt_len > 1024 or prompt_len + output_len > 2048:
             # Prune too long sequences.
             continue
         filtered_dataset.append((prompt, prompt_len, output_len))
 
-    # Sample the requests.
-    sampled_requests = random.sample(filtered_dataset, num_requests)
-    return sampled_requests
+    return filtered_dataset
 
 
 def sample_sonnet_requests(
@@ -361,13 +359,15 @@ def main(args: argparse.Namespace):
             dataset_path=args.dataset,
             num_requests=args.num_prompts,
             tokenizer=tokenizer,
+            fixed_output_len=args.sharegpt_output_len,
         )
 
     elif args.dataset_name == "sharegpt":
         input_requests = sample_sharegpt_requests(
             dataset_path=args.dataset_path,
             num_requests=args.num_prompts,
             tokenizer=tokenizer,
+            fixed_output_len=args.sharegpt_output_len,
         )
 
     elif args.dataset_name == "sonnet":
@@ -524,6 +524,12 @@ def main(args: argparse.Namespace):
         default=1000,
         help="Number of prompts to process.",
     )
+    parser.add_argument(
+        "--sharegpt-output-len",
+        type=int,
+        default=None,
+        help="Output length for each request. Overrides the output length "
+        "from the ShareGPT dataset.")
     parser.add_argument(
         "--sonnet-input-len",
         type=int,

benchmarks/benchmark_throughput.py

@@ -103,25 +103,22 @@ def run_vllm(
     )
 
     # Add the requests to the engine.
+    prompts = []
+    sampling_params = []
     for prompt, _, output_len in requests:
-        sampling_params = SamplingParams(
-            n=n,
-            temperature=0.0 if use_beam_search else 1.0,
-            top_p=1.0,
-            use_beam_search=use_beam_search,
-            ignore_eos=True,
-            max_tokens=output_len,
-        )
-        # FIXME(woosuk): Do not use internal method.
-        llm._add_request(
-            prompt=prompt,
-            prompt_token_ids=None,
-            sampling_params=sampling_params,
-        )
+        prompts.append(prompt)
+        sampling_params.append(
+            SamplingParams(
+                n=n,
+                temperature=0.0 if use_beam_search else 1.0,
+                top_p=1.0,
+                use_beam_search=use_beam_search,
+                ignore_eos=True,
+                max_tokens=output_len,
+            ))
 
     start = time.perf_counter()
-    # FIXME(woosuk): Do not use internal method.
-    llm._run_engine(use_tqdm=True)
+    llm.generate(prompts, sampling_params, use_tqdm=True)
     end = time.perf_counter()
     return end - start