Change the name to vLLM (vllm-project#150)

CONTRIBUTING.md

@@ -1,6 +1,6 @@
-# Contributing to CacheFlow
+# Contributing to vLLM
 
-Thank you for your interest in contributing to CacheFlow!
+Thank you for your interest in contributing to vLLM!
 Our community is open to everyone and welcomes all kinds of contributions, no matter how small or large.
 There are several ways you can contribute to the project:
 
@@ -11,9 +11,9 @@ There are several ways you can contribute to the project:
 However, remember that contributions aren't just about code.
 We believe in the power of community support; thus, answering queries, assisting others, and enhancing the documentation are highly regarded and beneficial contributions.
 
-Finally, one of the most impactful ways to support us is by raising awareness about CacheFlow.
+Finally, one of the most impactful ways to support us is by raising awareness about vLLM.
 Talk about it in your blog posts, highlighting how it's driving your incredible projects.
-Express your support on Twitter if CacheFlow aids you, or simply offer your appreciation by starring our repository.
+Express your support on Twitter if vLLM aids you, or simply offer your appreciation by starring our repository.
 
 
 ## Setup for development
@@ -70,5 +70,5 @@ If a comment isn't clear or you disagree with a suggestion, feel free to ask for
 
 ### Thank You
 
-Finally, thank you for taking the time to read these guidelines and for your interest in contributing to CacheFlow.
-Your contributions make CacheFlow a great tool for everyone!
+Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM.
+Your contributions make vLLM a great tool for everyone!

README.md

@@ -1,4 +1,4 @@
-# CacheFlow
+# vLLM
 
 ## Build from source
 
@@ -28,7 +28,7 @@ python examples/simple_server.py --help
 To start the server:
 ```bash
 ray start --head
-python -m cacheflow.entrypoints.fastapi_server # --model <your_model>
+python -m vllm.entrypoints.fastapi_server # --model <your_model>
 ```
 
 To test the server:
@@ -45,9 +45,9 @@ pip install gradio
 
 Start the server:
 ```bash
-python -m cacheflow.http_frontend.fastapi_frontend
+python -m vllm.http_frontend.fastapi_frontend
 # At another terminal
-python -m cacheflow.http_frontend.gradio_webserver
+python -m vllm.http_frontend.gradio_webserver
 ```
 
 ## Load LLaMA weights
@@ -62,5 +62,5 @@ Since LLaMA weight is not fully public, we cannot directly download the LLaMA we
 2. For all the commands above, specify the model with `--model /output/path/llama-7b` to load the model. For example:
     ```bash
     python simple_server.py --model /output/path/llama-7b
-    python -m cacheflow.http_frontend.fastapi_frontend --model /output/path/llama-7b
+    python -m vllm.http_frontend.fastapi_frontend --model /output/path/llama-7b
     ```

benchmarks/README.md

@@ -1,4 +1,4 @@
-# Benchmarking CacheFlow
+# Benchmarking vLLM
 
 ## Downloading the ShareGPT dataset
 

benchmarks/benchmark_async_llm_server.py

@@ -11,7 +11,7 @@ def main(args: argparse.Namespace):
               for i in range(args.n_threads)]
 
     api_url = f"http://{args.host}:{args.port}/generate"
-    headers = {"User-Agent": "CacheFlow Benchmark Client"}
+    headers = {"User-Agent": "vLLM Benchmark Client"}
     ploads = [{
         "prompt": p,
         "max_tokens": args.max_tokens,

benchmarks/benchmark_latency.py

@@ -6,7 +6,7 @@
 import torch
 from tqdm import tqdm
 
-from cacheflow import LLM, SamplingParams
+from vllm import LLM, SamplingParams
 
 
 def main(args: argparse.Namespace):

benchmarks/benchmark_serving.py

@@ -1,8 +1,8 @@
 """Benchmark online serving throughput.
 
 On the server side, run one of the following commands:
-    (CacheFlow backend)
-    python -m cacheflow.entrypoints.api_server \
+    (vLLM backend)
+    python -m vllm.entrypoints.api_server \
         --disable-log-requests --model <your_model>
 
     (TGI backend)
@@ -114,7 +114,7 @@ async def send_request(
     request_start_time = time.time()
 
     headers = {"User-Agent": "Benchmark Client"}
-    if backend == "cacheflow":
+    if backend == "vllm":
         pload = {
             "prompt": prompt,
             "n": 1,
@@ -213,8 +213,8 @@ def main(args: argparse.Namespace):
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(
         description="Benchmark the online serving throughput.")
-    parser.add_argument("--backend", type=str, default="cacheflow",
-                        choices=["cacheflow", "tgi"])
+    parser.add_argument("--backend", type=str, default="vllm",
+                        choices=["vllm", "tgi"])
     parser.add_argument("--host", type=str, default="localhost")
     parser.add_argument("--port", type=int, default=8001)
     parser.add_argument("--dataset", type=str, required=True,

benchmarks/benchmark_throughput.py

@@ -5,12 +5,13 @@
 import time
 from typing import List, Tuple
 
-from cacheflow import LLM, SamplingParams
 import torch
 from transformers import (AutoConfig, AutoTokenizer, AutoModelForCausalLM,
                           PreTrainedTokenizerBase)
 from tqdm import tqdm
 
+from vllm import LLM, SamplingParams
+
 
 def get_tokenizer(model_name: str) -> PreTrainedTokenizerBase:
     config = AutoConfig.from_pretrained(model_name)
@@ -70,7 +71,7 @@ def sample_requests(
     return sampled_requests
 
 
-def run_cacheflow(
+def run_vllm(
     requests: List[Tuple[str, int, int]],
     model: str,
     tensor_parallel_size: int,
@@ -172,8 +173,8 @@ def main(args: argparse.Namespace):
     tokenizer = get_tokenizer(args.model)
     requests = sample_requests(args.dataset, args.num_prompts, tokenizer)
 
-    if args.backend == "cacheflow":
-        elapsed_time = run_cacheflow(
+    if args.backend == "vllm":
+        elapsed_time = run_vllm(
             requests, args.model, args.tensor_parallel_size, args.seed, args.n,
             args.use_beam_search)
     elif args.backend == "hf":
@@ -192,8 +193,8 @@ def main(args: argparse.Namespace):
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description="Benchmark the throughput.")
-    parser.add_argument("--backend", type=str, choices=["cacheflow", "hf"],
-                        default="cacheflow")
+    parser.add_argument("--backend", type=str, choices=["vllm", "hf"],
+                        default="vllm")
     parser.add_argument("--dataset", type=str, required=True,
                         help="Path to the dataset.")
     parser.add_argument("--model", type=str, default="facebook/opt-125m")
@@ -207,7 +208,7 @@ def main(args: argparse.Namespace):
     parser.add_argument("--hf-max-batch-size", type=int, default=None,
                         help="Maximum batch size for HF backend.")
     args = parser.parse_args()
-    if args.backend == "cacheflow":
+    if args.backend == "vllm":
         if args.hf_max_batch_size is not None:
             raise ValueError("HF max batch size is only for HF backend.")
     elif args.backend == "hf":

csrc/activation_kernels.cu

@@ -1,7 +1,7 @@
 #include <torch/extension.h>
 #include <ATen/cuda/CUDAContext.h>
 
-namespace cacheflow {
+namespace vllm {
 
 template<typename T>
 __device__ __forceinline__ T silu(const T& x) {
@@ -22,7 +22,7 @@ __global__ void silu_and_mul_kernel(
   }
 }
 
-} // namespace cacheflow
+} // namespace vllm
 
 void silu_and_mul(
   torch::Tensor& out,      // [num_tokens, d]
@@ -40,7 +40,7 @@ void silu_and_mul(
     input.scalar_type(),
     "silu_and_mul_kernel",
     [&] {
-      cacheflow::silu_and_mul_kernel<scalar_t><<<grid, block, 0, stream>>>(
+      vllm::silu_and_mul_kernel<scalar_t><<<grid, block, 0, stream>>>(
         out.data_ptr<scalar_t>(),
         input.data_ptr<scalar_t>(),
         d);

csrc/attention/attention_generic.cuh

@@ -1,6 +1,6 @@
 /*
  * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
- * Copyright (c) 2023, The CacheFlow team.
+ * Copyright (c) 2023, The vLLM team.
  * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -19,7 +19,7 @@
 
 #include <stdint.h>
 
-namespace cacheflow {
+namespace vllm {
 
 // A vector type to store Q, K, V elements.
 template<typename T, int VEC_SIZE>
@@ -61,4 +61,4 @@ inline __device__ void zero(T& dst) {
   dst = tmp.raw;
 }
 
-} // namespace cacheflow
+} // namespace vllm

csrc/attention/attention_kernels.cu

@@ -1,6 +1,6 @@
 /*
  * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
- * Copyright (c) 2023, The CacheFlow team.
+ * Copyright (c) 2023, The vLLM team.
  * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -27,7 +27,7 @@
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 
-namespace cacheflow {
+namespace vllm {
 
 // Utility function for attention softmax.
 template<int NUM_WARPS>
@@ -315,10 +315,10 @@ __global__ void single_query_cached_kv_attention_kernel(
   }
 }
 
-} // namespace cacheflow
+} // namespace vllm
 
 #define LAUNCH_ATTENTION_KERNEL(T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS)                        \
-  cacheflow::single_query_cached_kv_attention_kernel<T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS>   \
+  vllm::single_query_cached_kv_attention_kernel<T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS>        \
   <<<grid, block, shared_mem_size, stream>>>(                                                 \
     out_ptr,                                                                                  \
     query_ptr,                                                                                \

csrc/attention/attention_utils.cuh

@@ -1,6 +1,6 @@
 /*
  * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
- * Copyright (c) 2023, The CacheFlow team.
+ * Copyright (c) 2023, The vLLM team.
  * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -22,7 +22,7 @@
 #include <float.h>
 #include <type_traits>
 
-namespace cacheflow {
+namespace vllm {
 
 // Q*K^T operation.
 template<int THREAD_GROUP_SIZE, typename Vec, int N>
@@ -52,4 +52,4 @@ struct Qk_dot {
   }
 };
 
-} // namespace cacheflow
+} // namespace vllm

csrc/attention/dtype_bfloat16.cuh

@@ -1,7 +1,7 @@
 /*
  * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
  * and https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
- * Copyright (c) 2023, The CacheFlow team.
+ * Copyright (c) 2023, The vLLM team.
  * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -25,7 +25,7 @@
 #include <cuda_fp16.h>
 #include <stdint.h>
 
-namespace cacheflow {
+namespace vllm {
 
 // Define custom BF16 vector data types.
 struct bf16_4_t {
@@ -420,4 +420,4 @@ inline __device__ void from_float(bf16_8_t& dst, Float8_ src) {
 #endif
 }
 
-} // namespace cacheflow
+} // namespace vllm

csrc/attention/dtype_float16.cuh

@@ -1,7 +1,7 @@
 /*
  * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
  * and https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
- * Copyright (c) 2023, The CacheFlow team.
+ * Copyright (c) 2023, The vLLM team.
  * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -23,7 +23,7 @@
 
 #include <stdint.h>
 
-namespace cacheflow {
+namespace vllm {
 
 // FP16 vector types for Q, K, V.
 template<>
@@ -441,4 +441,4 @@ inline __device__ Float8_ to_float(uint4 u) {
   return tmp;
 }
 
-} // namespace cacheflow
+} // namespace vllm

csrc/attention/dtype_float32.cuh

@@ -1,7 +1,7 @@
 /*
  * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
  * and https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
- * Copyright (c) 2023, The CacheFlow team.
+ * Copyright (c) 2023, The vLLM team.
  * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -22,7 +22,7 @@
 
 #include <stdint.h>
 
-namespace cacheflow {
+namespace vllm {
 
 // Define custom FP32 vector data types.
 struct Float4_ {
@@ -265,4 +265,4 @@ inline __device__ Float8_ to_float(Float8_ u) {
   return u;
 }
 
-} // namespace cacheflow
+} // namespace vllm