v1.0

inference/vllm/tests/async_engine/test_request_tracker.py

+import pytest
+
+from vllm.engine.async_llm_engine import RequestTracker
+from vllm.outputs import RequestOutput
+
+
+class DummyEvent:
+
+    def __init__(self):
+        self.flag = False
+
+    def set(self):
+        self.flag = True
+
+    def clear(self):
+        self.flag = False
+
+
+def test_request_tracker():
+    tracker = RequestTracker()
+    tracker.new_requests_event = DummyEvent()
+    stream_1 = tracker.add_request("1")
+    assert tracker.new_requests_event.flag
+    new, finished = tracker.get_new_and_finished_requests()
+    assert not tracker.new_requests_event.flag
+    assert len(new) == 1
+    assert new[0]["request_id"] == "1"
+    assert not finished
+    assert not stream_1.finished
+
+    stream_2 = tracker.add_request("2")
+    stream_3 = tracker.add_request("3")
+    assert tracker.new_requests_event.flag
+    new, finished = tracker.get_new_and_finished_requests()
+    assert not tracker.new_requests_event.flag
+    assert len(new) == 2
+    assert new[0]["request_id"] == "2"
+    assert new[1]["request_id"] == "3"
+    assert not finished
+    assert not stream_2.finished
+    assert not stream_3.finished
+
+    # request_ids must be unique
+    with pytest.raises(KeyError):
+        tracker.add_request("1")
+    assert not tracker.new_requests_event.flag
+
+    tracker.abort_request("1")
+    new, finished = tracker.get_new_and_finished_requests()
+    assert len(finished) == 1
+    assert "1" in finished
+    assert not new
+    assert stream_1.finished
+
+    stream_4 = tracker.add_request("4")
+    tracker.abort_request("4")
+    assert tracker.new_requests_event.flag
+    new, finished = tracker.get_new_and_finished_requests()
+    assert len(finished) == 1
+    assert "4" in finished
+    assert not new
+    assert stream_4.finished
+
+    stream_5 = tracker.add_request("5")
+    assert tracker.new_requests_event.flag
+    tracker.process_request_output(
+        RequestOutput("2", "output", [], [], [], finished=True))
+    new, finished = tracker.get_new_and_finished_requests()
+    assert not tracker.new_requests_event.flag
+    assert len(finished) == 1
+    assert "2" in finished
+    assert len(new) == 1
+    assert new[0]["request_id"] == "5"
+    assert stream_2.finished
+    assert not stream_5.finished

inference/vllm/tests/conftest.py

+from typing import List, Optional, Tuple
+
+import pytest
+import torch
+from transformers import AutoModelForCausalLM
+
+from vllm import LLM, SamplingParams
+from vllm.transformers_utils.tokenizer import get_tokenizer
+
+_TEST_PROMPTS = [
+    # pylint: disable=line-too-long
+    "vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs.",
+    "Briefly describe the major milestones in the development of artificial intelligence from 1950 to 2020.",
+    "Compare and contrast artificial intelligence with human intelligence in terms of processing information.",
+    "Describe the basic components of a neural network and how it can be trained.",
+    "Write a short story about a robot that dreams for the first time.",
+    "Analyze the impact of the COVID-19 pandemic on global economic structures and future business models.",
+    "Explain the cultural significance of the Mona Lisa painting, and how its perception might vary in Western versus Eastern societies.",
+    "Translate the following English sentence into Japanese, French, and Swahili: 'The early bird catches the worm.'",
+]
+
+
+@pytest.fixture
+def example_prompts() -> List[str]:
+    return _TEST_PROMPTS
+
+
+_STR_DTYPE_TO_TORCH_DTYPE = {
+    "half": torch.half,
+    "bfloat16": torch.bfloat16,
+    "float": torch.float,
+}
+
+
+class HfRunner:
+
+    def __init__(
+        self,
+        model_name: str,
+        tokenizer_name: Optional[str] = None,
+        dtype: str = "half",
+    ) -> None:
+        assert dtype in _STR_DTYPE_TO_TORCH_DTYPE
+        torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
+        self.model = AutoModelForCausalLM.from_pretrained(
+            model_name,
+            torch_dtype=torch_dtype,
+            trust_remote_code=True,
+        ).cuda()
+        if tokenizer_name is None:
+            tokenizer_name = model_name
+        self.tokenizer = get_tokenizer(tokenizer_name, trust_remote_code=True)
+
+    def generate(
+        self,
+        prompts: List[str],
+        **kwargs,
+    ) -> List[Tuple[List[int], str]]:
+        outputs: List[Tuple[List[int], str]] = []
+        for prompt in prompts:
+            input_ids = self.tokenizer(prompt, return_tensors="pt").input_ids
+            output_ids = self.model.generate(
+                input_ids.cuda(),
+                use_cache=True,
+                **kwargs,
+            )
+            output_str = self.tokenizer.batch_decode(
+                output_ids,
+                skip_special_tokens=True,
+                clean_up_tokenization_spaces=False,
+            )
+            output_ids = output_ids.cpu().tolist()
+            outputs.append((output_ids, output_str))
+        return outputs
+
+    def generate_greedy(
+        self,
+        prompts: List[str],
+        max_tokens: int,
+    ) -> List[Tuple[List[int], str]]:
+        outputs = self.generate(prompts,
+                                do_sample=False,
+                                max_new_tokens=max_tokens)
+        for i in range(len(outputs)):
+            output_ids, output_str = outputs[i]
+            outputs[i] = (output_ids[0], output_str[0])
+        return outputs
+
+    def generate_beam_search(
+        self,
+        prompts: List[str],
+        beam_width: int,
+        max_tokens: int,
+    ) -> List[Tuple[List[int], str]]:
+        outputs = self.generate(prompts,
+                                do_sample=False,
+                                max_new_tokens=max_tokens,
+                                num_beams=beam_width,
+                                num_return_sequences=beam_width)
+        for i in range(len(outputs)):
+            output_ids, output_str = outputs[i]
+            for j in range(len(output_ids)):
+                output_ids[j] = [
+                    x for x in output_ids[j]
+                    if x != self.tokenizer.pad_token_id
+                ]
+            outputs[i] = (output_ids, output_str)
+        return outputs
+
+    def generate_greedy_logprobs(
+        self,
+        prompts: List[str],
+        max_tokens: int,
+    ) -> List[List[torch.Tensor]]:
+        all_logprobs = []
+        for prompt in prompts:
+            input_ids = self.tokenizer(prompt, return_tensors="pt").input_ids
+            output = self.model.generate(
+                input_ids.cuda(),
+                use_cache=True,
+                do_sample=False,
+                max_new_tokens=max_tokens,
+                output_hidden_states=True,
+                return_dict_in_generate=True,
+            )
+            seq_logprobs = []
+            for hidden_states in output.hidden_states:
+                last_hidden_states = hidden_states[-1][0]
+                logits = torch.matmul(
+                    last_hidden_states,
+                    self.model.get_output_embeddings().weight.t(),
+                )
+                if self.model.get_output_embeddings().bias is not None:
+                    logits += self.model.get_output_embeddings(
+                    ).bias.unsqueeze(0)
+                logprobs = torch.nn.functional.log_softmax(logits,
+                                                           dim=-1,
+                                                           dtype=torch.float32)
+                seq_logprobs.append(logprobs)
+            all_logprobs.append(seq_logprobs)
+        return all_logprobs
+
+
+@pytest.fixture
+def hf_runner():
+    return HfRunner
+
+
+class VllmRunner:
+
+    def __init__(
+        self,
+        model_name: str,
+        tokenizer_name: Optional[str] = None,
+        dtype: str = "half",
+    ) -> None:
+        self.model = LLM(
+            model=model_name,
+            tokenizer=tokenizer_name,
+            trust_remote_code=True,
+            dtype=dtype,
+            swap_space=0,
+        )
+
+    def generate(
+        self,
+        prompts: List[str],
+        sampling_params: SamplingParams,
+    ) -> List[Tuple[List[int], str]]:
+        req_outputs = self.model.generate(prompts,
+                                          sampling_params=sampling_params)
+        outputs = []
+        for req_output in req_outputs:
+            prompt_str = req_output.prompt
+            prompt_ids = req_output.prompt_token_ids
+            req_sample_output_ids = []
+            req_sample_output_strs = []
+            for sample in req_output.outputs:
+                output_str = sample.text
+                output_ids = sample.token_ids
+                req_sample_output_ids.append(prompt_ids + output_ids)
+                req_sample_output_strs.append(prompt_str + output_str)
+            outputs.append((req_sample_output_ids, req_sample_output_strs))
+        return outputs
+
+    def generate_greedy(
+        self,
+        prompts: List[str],
+        max_tokens: int,
+    ) -> List[Tuple[List[int], str]]:
+        greedy_params = SamplingParams(temperature=0.0, max_tokens=max_tokens)
+        outputs = self.generate(prompts, greedy_params)
+        return [(output_ids[0], output_str[0])
+                for output_ids, output_str in outputs]
+
+    def generate_beam_search(
+        self,
+        prompts: List[str],
+        beam_width: int,
+        max_tokens: int,
+    ) -> List[Tuple[List[int], str]]:
+        beam_search_params = SamplingParams(n=beam_width,
+                                            use_beam_search=True,
+                                            temperature=0.0,
+                                            max_tokens=max_tokens)
+        outputs = self.generate(prompts, beam_search_params)
+        return outputs
+
+
+@pytest.fixture
+def vllm_runner():
+    return VllmRunner

inference/vllm/tests/distributed/test_comm_ops.py

+"""Test the communication operators.
+
+Run `pytest tests/distributed/test_comm_ops.py --forked`.
+"""
+from multiprocessing import Process, set_start_method
+
+import pytest
+import torch
+
+from vllm.config import ParallelConfig
+from vllm.engine.ray_utils import get_open_port
+from vllm.model_executor.parallel_utils.communication_op import (
+    tensor_model_parallel_all_reduce,
+    tensor_model_parallel_all_gather,
+)
+from vllm.worker.worker import _init_distributed_environment
+
+
+def init_test_distributed_environment(pipeline_parallel_size: int,
+                                      tensor_parallel_size: int, rank: int,
+                                      distributed_init_port: str):
+    parallel_config = ParallelConfig(pipeline_parallel_size,
+                                     tensor_parallel_size,
+                                     worker_use_ray=True)
+    distributed_init_method = f"tcp://localhost:{distributed_init_port}"
+    torch.cuda.set_device(rank)
+    _init_distributed_environment(parallel_config, rank,
+                                  distributed_init_method)
+
+
+def all_reduce_test_worker(tensor_parallel_size: int, rank: int,
+                           distributed_init_port: str):
+    init_test_distributed_environment(1, tensor_parallel_size, rank,
+                                      distributed_init_port)
+    num_elements = 8
+    all_tensors = [
+        torch.arange(num_elements, dtype=torch.float32, device="cuda") *
+        (r + 1) for r in range(tensor_parallel_size)
+    ]
+    expected = torch.sum(torch.stack(all_tensors, dim=0), dim=0)
+    t = all_tensors[rank]
+    t = tensor_model_parallel_all_reduce(t)
+    assert torch.allclose(t, expected)
+
+
+def all_gather_test_worker(tensor_parallel_size: int, rank: int,
+                           distributed_init_port: str):
+    init_test_distributed_environment(1, tensor_parallel_size, rank,
+                                      distributed_init_port)
+    num_dimensions = 3
+    tensor_size = list(range(2, num_dimensions + 2))
+    total_size = 1
+    for s in tensor_size:
+        total_size *= s
+    for all_gather_dimension in range(num_dimensions):
+        all_tensors = [
+            torch.arange(total_size, dtype=torch.float32,
+                         device="cuda").reshape(tensor_size) * (r + 1)
+            for r in range(tensor_parallel_size)
+        ]
+        expected = torch.cat(all_tensors, dim=all_gather_dimension)
+        t = all_tensors[rank]
+        t = tensor_model_parallel_all_gather(t, all_gather_dimension)
+        assert torch.allclose(t, expected)
+
+
+@pytest.mark.skipif(torch.cuda.device_count() < 2,
+                    reason="Need at least 2 GPUs to run the test.")
+@pytest.mark.parametrize("tensor_parallel_size", [2])
+@pytest.mark.parametrize("test_target",
+                         [all_reduce_test_worker, all_gather_test_worker])
+def test_multi_process_tensor_parallel(tensor_parallel_size, test_target):
+    set_start_method("spawn", force=True)
+    distributed_init_port = get_open_port()
+    processes = []
+    for rank in range(tensor_parallel_size):
+        p = Process(target=test_target,
+                    args=(tensor_parallel_size, rank, distributed_init_port))
+        p.start()
+        processes.append(p)
+    for p in processes:
+        p.join()
+    assert all(p.exitcode == 0 for p in processes)

inference/vllm/tests/engine/test_detokenize.py

+import pytest
+
+from transformers import AutoTokenizer
+
+from vllm.transformers_utils.tokenizer import detokenize_incrementally
+
+TRUTH = [
+    # pylint: disable=line-too-long
+    "Hello here, this is a simple test",
+    "vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. It is designed to be used in production environments, where inference and serving",
+    "我很感谢你的热情"
+]
+TOKENIZERS = [
+    "facebook/opt-125m",
+    "gpt2",
+    "bigcode/tiny_starcoder_py",
+    "EleutherAI/gpt-j-6b",
+    "EleutherAI/pythia-70m",
+    "bigscience/bloom-560m",
+    "mosaicml/mpt-7b",
+    "tiiuae/falcon-7b",
+    "meta-llama/Llama-2-7b-hf",
+    "codellama/CodeLlama-7b-hf",
+]
+
+
+def _run_incremental_decode(tokenizer, all_input_ids,
+                            skip_special_tokens: bool):
+    decoded_text = ""
+    offset = 0
+    token_offset = 0
+    prev_tokens = None
+    for i in range(len(all_input_ids)):
+        new_tokens, text, offset, token_offset = detokenize_incrementally(
+            tokenizer,
+            all_input_ids[:i + 1],
+            prev_tokens,
+            offset,
+            token_offset,
+            skip_special_tokens=skip_special_tokens)
+        decoded_text += text
+        if prev_tokens is None:
+            prev_tokens = new_tokens
+        else:
+            prev_tokens += new_tokens
+    return decoded_text
+
+
+@pytest.mark.parametrize("truth", TRUTH)
+@pytest.mark.parametrize("tokenizer_id", TOKENIZERS)
+@pytest.mark.parametrize("skip_special_tokens", (True, False))
+def test_decode_streaming(tokenizer_id, truth, skip_special_tokens):
+    tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)
+    all_input_ids = tokenizer(truth, add_special_tokens=False)["input_ids"]
+    if skip_special_tokens:
+        all_input_ids = ([tokenizer.bos_token_id]
+                         if tokenizer.bos_token_id is not None else
+                         []) + all_input_ids + [tokenizer.eos_token_id]
+
+    decoded_text = _run_incremental_decode(
+        tokenizer, all_input_ids, skip_special_tokens=skip_special_tokens)
+
+    assert decoded_text == truth

inference/vllm/tests/kernels/conftest.py

+from typing import List, Tuple
+
+import pytest
+import torch
+
+
+def create_kv_caches(
+    num_blocks: int,
+    block_size: int,
+    num_layers: int,
+    num_heads: int,
+    head_size: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> Tuple[List[torch.Tensor], List[torch.Tensor]]:
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    scale = head_size**-0.5
+    x = 16 // torch.tensor([], dtype=dtype).element_size()
+    key_cache_shape = (num_blocks, num_heads, head_size // x, block_size, x)
+    key_caches = []
+    for _ in range(num_layers):
+        key_cache = torch.empty(size=key_cache_shape,
+                                dtype=dtype,
+                                device='cuda')
+        key_cache.uniform_(-scale, scale)
+        key_caches.append(key_cache)
+
+    value_cache_shape = (num_blocks, num_heads, head_size, block_size)
+    value_caches = []
+    for _ in range(num_layers):
+        value_cache = torch.empty(size=value_cache_shape,
+                                  dtype=dtype,
+                                  device='cuda')
+        value_cache.uniform_(-scale, scale)
+        value_caches.append(value_cache)
+    return key_caches, value_caches
+
+
+@pytest.fixture()
+def kv_cache_factory():
+    return create_kv_caches

inference/vllm/tests/kernels/test_activation.py

+import pytest
+import torch
+import torch.nn.functional as F
+from transformers.activations import get_activation
+
+from vllm import activation_ops
+
+DTYPES = [torch.half, torch.bfloat16, torch.float]
+NUM_TOKENS = [7, 83, 2048]  # Arbitrary values for testing
+D = [512, 4096, 5120, 13824]  # Arbitrary values for testing
+SEEDS = [0]
+
+
+def ref_silu_and_mul(x: torch.Tensor) -> torch.Tensor:
+    x1, x2 = x.chunk(chunks=2, dim=1)
+    return F.silu(x1) * x2
+
+
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("d", D)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_silu_and_mul(
+    num_tokens: int,
+    d: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    x = torch.randn(num_tokens, 2 * d, dtype=dtype, device="cuda")
+    out = torch.empty(num_tokens, d, dtype=dtype, device="cuda")
+    activation_ops.silu_and_mul(out, x)
+    ref_out = ref_silu_and_mul(x)
+    assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)
+
+
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("d", D)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_gelu_new(
+    num_tokens: int,
+    d: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    x = torch.randn(num_tokens, d, dtype=dtype, device="cuda")
+    out = torch.empty(num_tokens, d, dtype=dtype, device="cuda")
+    activation_ops.gelu_new(out, x)
+    ref_out = get_activation("gelu_new")(x)
+    assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)
+
+
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("d", D)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+def test_gelu_fast(
+    num_tokens: int,
+    d: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    x = torch.randn(num_tokens, d, dtype=dtype, device="cuda")
+    out = torch.empty(num_tokens, d, dtype=dtype, device="cuda")
+    activation_ops.gelu_fast(out, x)
+    ref_out = get_activation("gelu_fast")(x)
+    assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)

inference/vllm/tests/kernels/test_attention.py

+import random
+from typing import List, Optional, Tuple
+
+import pytest
+import torch
+from xformers import ops as xops
+from xformers.ops.fmha.attn_bias import BlockDiagonalCausalMask
+
+from vllm import attention_ops
+from vllm.utils import get_max_shared_memory_bytes
+
+FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
+# This will change depending on the compute capability.
+# - 512 as a buffer
+MAX_SEQ_LEN = get_max_shared_memory_bytes() // FLOAT32_BYTES - 512
+NUM_BLOCKS = 40000  # Arbitrary values for testing
+PARTITION_SIZE = 512
+
+DTYPES = [torch.half, torch.bfloat16, torch.float]
+NUM_GEN_SEQS = [7]  # Arbitrary values for testing
+NUM_PREFILL_SEQS = [3]  # Arbitrary values for testing
+NUM_HEADS = [(40, 40), (64, 8)]  # Arbitrary values for testing
+HEAD_SIZES = [64, 80, 96, 112, 128, 256]
+BLOCK_SIZES = [16, 32]
+USE_ALIBI = [False, True]
+SEEDS = [0]
+
+
+def ref_masked_attention(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    scale: float,
+    attn_mask: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    attn_weights = scale * torch.einsum("qhd,khd->hqk", query, key).float()
+    if attn_mask is not None:
+        attn_weights = attn_weights + attn_mask.float()
+    attn_weights = torch.softmax(attn_weights, dim=-1).to(value.dtype)
+    out = torch.einsum("hqk,khd->qhd", attn_weights, value)
+    return out
+
+
+def ref_single_query_cached_kv_attention(
+    output: torch.Tensor,
+    query: torch.Tensor,
+    num_queries_per_kv: int,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    block_tables: torch.Tensor,
+    context_lens: torch.Tensor,
+    scale: float,
+    alibi_slopes: Optional[torch.Tensor],
+) -> None:
+    num_query_heads = query.shape[1]
+    num_kv_heads = value_cache.shape[1]
+    head_size = value_cache.shape[2]
+    block_size = value_cache.shape[3]
+    num_seqs = query.shape[0]
+
+    block_tables = block_tables.cpu().tolist()
+    context_lens = context_lens.cpu().tolist()
+    for i in range(num_seqs):
+        q = query[i].unsqueeze(0)
+        block_table = block_tables[i]
+        context_len = int(context_lens[i])
+
+        keys = []
+        values = []
+        for j in range(context_len):
+            block_number = int(block_table[j // block_size])
+            block_offset = j % block_size
+
+            k = key_cache[block_number, :, :, block_offset, :]
+            k = k.reshape(num_kv_heads, head_size)
+            keys.append(k)
+
+            v = value_cache[block_number, :, :, block_offset]
+            values.append(v)
+        keys = torch.stack(keys, dim=0)
+        values = torch.stack(values, dim=0)
+        if num_queries_per_kv > 1:
+            # Handle MQA and GQA
+            keys = torch.repeat_interleave(keys, num_queries_per_kv, dim=1)
+            values = torch.repeat_interleave(values, num_queries_per_kv, dim=1)
+
+        alibi_bias = None
+        if alibi_slopes is not None:
+            # Create the ALiBi bias used in the paged attention kernel.
+            position_ids = torch.arange(context_len, device="cuda").int()
+            alibi_bias = (position_ids - context_len + 1).float()
+            alibi_bias = alibi_slopes.view(-1, 1, 1) * alibi_bias.view(
+                1, 1, -1)
+
+        out = ref_masked_attention(q, keys, values, scale, alibi_bias)
+        out = out.view(num_query_heads, head_size)
+        output[i].copy_(out, non_blocking=True)
+
+
+@pytest.mark.parametrize("version", ["v1", "v2"])
+@pytest.mark.parametrize("num_seqs", NUM_GEN_SEQS)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("use_alibi", USE_ALIBI)
+@pytest.mark.parametrize("block_size", BLOCK_SIZES)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+def test_paged_attention(
+    kv_cache_factory,
+    version: str,
+    num_seqs: int,
+    num_heads: Tuple[int, int],
+    head_size: int,
+    use_alibi: bool,
+    block_size: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    random.seed(seed)
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    scale = float(1.0 / (head_size**0.5))
+    num_query_heads, num_kv_heads = num_heads
+    query = torch.empty(num_seqs,
+                        num_query_heads,
+                        head_size,
+                        dtype=dtype,
+                        device="cuda")
+    query.uniform_(-scale, scale)
+
+    assert num_query_heads % num_kv_heads == 0
+    num_queries_per_kv = num_query_heads // num_kv_heads
+    head_mapping = torch.repeat_interleave(
+        torch.arange(num_kv_heads, dtype=torch.int32, device="cuda"),
+        num_queries_per_kv)
+    alibi_slopes = None
+    if use_alibi:
+        alibi_slopes = torch.randn(num_query_heads,
+                                   dtype=torch.float,
+                                   device="cuda")
+
+    context_lens = [random.randint(1, MAX_SEQ_LEN) for _ in range(num_seqs)]
+    context_lens[-1] = MAX_SEQ_LEN
+    max_context_len = max(context_lens)
+    context_lens = torch.tensor(context_lens, dtype=torch.int, device="cuda")
+
+    # Create the block tables.
+    max_num_blocks_per_seq = (max_context_len + block_size - 1) // block_size
+    block_tables = []
+    for _ in range(num_seqs):
+        block_table = [
+            random.randint(0, NUM_BLOCKS - 1)
+            for _ in range(max_num_blocks_per_seq)
+        ]
+        block_tables.append(block_table)
+    block_tables = torch.tensor(block_tables, dtype=torch.int, device="cuda")
+
+    # Create the KV caches.
+    key_caches, value_caches = kv_cache_factory(NUM_BLOCKS, block_size, 1,
+                                                num_kv_heads, head_size, dtype,
+                                                seed)
+    key_cache, value_cache = key_caches[0], value_caches[0]
+
+    # Call the paged attention kernel.
+    output = torch.empty_like(query)
+    if version == "v1":
+        attention_ops.paged_attention_v1(
+            output,
+            query,
+            key_cache,
+            value_cache,
+            head_mapping,
+            scale,
+            block_tables,
+            context_lens,
+            block_size,
+            max_context_len,
+            alibi_slopes,
+        )
+    elif version == "v2":
+        num_partitions = ((max_context_len + PARTITION_SIZE - 1) //
+                          PARTITION_SIZE)
+        assert PARTITION_SIZE % block_size == 0
+        num_seqs, num_heads, head_size = output.shape
+        tmp_output = torch.empty(
+            size=(num_seqs, num_heads, num_partitions, head_size),
+            dtype=output.dtype,
+            device=output.device,
+        )
+        exp_sums = torch.empty(
+            size=(num_seqs, num_heads, num_partitions),
+            dtype=torch.float32,
+            device=output.device,
+        )
+        max_logits = torch.empty_like(exp_sums)
+        attention_ops.paged_attention_v2(
+            output,
+            exp_sums,
+            max_logits,
+            tmp_output,
+            query,
+            key_cache,
+            value_cache,
+            head_mapping,
+            scale,
+            block_tables,
+            context_lens,
+            block_size,
+            max_context_len,
+            alibi_slopes,
+        )
+    else:
+        assert False, f"Unknown version: {version}"
+
+    # Run the reference implementation.
+    ref_output = torch.empty_like(query)
+    ref_single_query_cached_kv_attention(
+        ref_output,
+        query,
+        num_queries_per_kv,
+        key_cache,
+        value_cache,
+        block_tables,
+        context_lens,
+        scale,
+        alibi_slopes,
+    )
+
+    # NOTE(woosuk): Due to the kernel-level differences in the two
+    # implementations, there is a small numerical difference in the two
+    # outputs. Thus, we use a relaxed tolerance for the test.
+    assert torch.allclose(output, ref_output, atol=1e-3, rtol=1e-5)
+
+
+def ref_multi_query_kv_attention(
+    cu_seq_lens: List[int],
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    scale: float,
+    dtype: torch.dtype,
+) -> torch.Tensor:
+    num_seqs = len(cu_seq_lens) - 1
+    ref_outputs = []
+    for i in range(num_seqs):
+        start_idx = cu_seq_lens[i]
+        end_idx = cu_seq_lens[i + 1]
+        seq_len = end_idx - start_idx
+
+        # Create attention mask.
+        attn_mask = torch.triu(torch.ones(seq_len, seq_len, dtype=dtype),
+                               diagonal=1)
+        attn_mask = attn_mask * torch.finfo(dtype).min
+        attn_mask = attn_mask.to(dtype=dtype, device="cuda")
+
+        ref_output = ref_masked_attention(
+            query[start_idx:end_idx],
+            key[start_idx:end_idx],
+            value[start_idx:end_idx],
+            scale,
+            attn_mask=attn_mask,
+        )
+        ref_outputs.append(ref_output)
+    ref_output = torch.cat(ref_outputs, dim=0)
+    return ref_output
+
+
+# TODO(woosuk): Add tests for USE_ALIBI=True.
+@pytest.mark.parametrize("num_seqs", NUM_PREFILL_SEQS)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_multi_query_kv_attention(
+    num_seqs: int,
+    num_heads: Tuple[int, int],
+    head_size: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    random.seed(seed)
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    # MAX_SEQ_LEN sometimes causes OOM in the reference implementation.
+    # As the xformers library is already tested with its own tests, we can use
+    # a smaller MAX_SEQ_LEN here.
+    max_len = min(MAX_SEQ_LEN, 4096)
+    seq_lens = random.sample(range(1, max_len), num_seqs)
+    num_tokens = sum(seq_lens)
+
+    scale = float(1.0 / (head_size**0.5))
+    num_query_heads, num_kv_heads = num_heads
+    qkv = torch.empty(num_tokens,
+                      num_query_heads + 2 * num_kv_heads,
+                      head_size,
+                      dtype=dtype,
+                      device="cuda")
+    qkv.uniform_(-scale, scale)
+    query, key, value = qkv.split(
+        [num_query_heads, num_kv_heads, num_kv_heads], dim=1)
+
+    num_queries_per_kv = num_query_heads // num_kv_heads
+    if num_queries_per_kv > 1:
+        # Handle MQA and GQA
+        key = torch.repeat_interleave(key, num_queries_per_kv, dim=1)
+        value = torch.repeat_interleave(value, num_queries_per_kv, dim=1)
+    attn_bias = BlockDiagonalCausalMask.from_seqlens(seq_lens)
+    output = xops.memory_efficient_attention_forward(
+        query.unsqueeze(0),
+        key.unsqueeze(0),
+        value.unsqueeze(0),
+        attn_bias=attn_bias,
+        p=0.0,
+        scale=scale,
+    )
+    output = output.squeeze(0)
+
+    cu_seq_lens = [0]
+    for seq_len in seq_lens:
+        cu_seq_lens.append(cu_seq_lens[-1] + seq_len)
+    ref_output = ref_multi_query_kv_attention(
+        cu_seq_lens,
+        query,
+        key,
+        value,
+        scale,
+        dtype,
+    )
+    assert torch.allclose(output, ref_output, atol=1e-3, rtol=1e-5)

inference/vllm/tests/kernels/test_cache.py

+import random
+
+import pytest
+import torch
+
+from vllm import cache_ops
+
+DTYPES = [torch.half, torch.bfloat16, torch.float]
+NUM_TOKENS = [83]  # Arbitrary values for testing
+NUM_LAYERS = [1]  # Arbitrary values for testing
+NUM_HEADS = [8]  # Arbitrary values for testing
+HEAD_SIZES = [64, 80, 96, 112, 128, 256]
+BLOCK_SIZES = [8, 16, 32]
+NUM_BLOCKS = [1024, 36000]  # Arbitrary values for testing
+NUM_MAPPINGS = [256]  # Arbitrary values for testing
+SEEDS = [0]
+
+
+@pytest.mark.parametrize("num_mappings", NUM_MAPPINGS)
+@pytest.mark.parametrize("num_layers", NUM_LAYERS)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("block_size", BLOCK_SIZES)
+@pytest.mark.parametrize("num_blocks", NUM_BLOCKS)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_copy_blocks(
+    kv_cache_factory,
+    num_mappings: int,
+    num_layers: int,
+    num_heads: int,
+    head_size: int,
+    block_size: int,
+    num_blocks: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    random.seed(seed)
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    # Generate random block mappings where each source block is mapped to two
+    # destination blocks.
+    assert 2 * num_mappings <= num_blocks
+    src_blocks = random.sample(range(num_blocks), num_mappings)
+    remainig_blocks = list(set(range(num_blocks)) - set(src_blocks))
+    dst_blocks = random.sample(remainig_blocks, 2 * num_mappings)
+    block_mapping = {}
+    for i in range(num_mappings):
+        src = src_blocks[i]
+        dst1 = dst_blocks[2 * i]
+        dst2 = dst_blocks[2 * i + 1]
+        block_mapping[src] = [dst1, dst2]
+
+    # Create the KV caches.
+    key_caches, value_caches = kv_cache_factory(num_blocks, block_size,
+                                                num_layers, num_heads,
+                                                head_size, dtype, seed)
+
+    # Clone the KV caches.
+    cloned_key_caches = [key_cache.clone() for key_cache in key_caches]
+    cloned_value_caches = [value_cache.clone() for value_cache in value_caches]
+
+    # Call the copy blocks kernel.
+    cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
+
+    # Run the reference implementation.
+    for src, dsts in block_mapping.items():
+        for dst in dsts:
+            for cloned_key_cache in cloned_key_caches:
+                cloned_key_cache[dst].copy_(cloned_key_cache[src])
+            for cloned_value_cache in cloned_value_caches:
+                cloned_value_cache[dst].copy_(cloned_value_cache[src])
+
+    # Compare the results.
+    for key_cache, cloned_key_cache in zip(key_caches, cloned_key_caches):
+        assert torch.allclose(key_cache, cloned_key_cache)
+    for value_cache, cloned_value_cache in zip(value_caches,
+                                               cloned_value_caches):
+        assert torch.allclose(value_cache, cloned_value_cache)
+
+
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("block_size", BLOCK_SIZES)
+@pytest.mark.parametrize("num_blocks", NUM_BLOCKS)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_reshape_and_cache(
+    kv_cache_factory,
+    num_tokens: int,
+    num_heads: int,
+    head_size: int,
+    block_size: int,
+    num_blocks: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    random.seed(seed)
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    # Create a random slot mapping.
+    num_slots = block_size * num_blocks
+    slot_mapping = random.sample(range(num_slots), num_tokens)
+    slot_mapping = torch.tensor(slot_mapping, dtype=torch.long, device="cuda")
+
+    qkv = torch.randn(num_tokens,
+                      3,
+                      num_heads,
+                      head_size,
+                      dtype=dtype,
+                      device="cuda")
+    _, key, value = qkv.unbind(dim=1)
+
+    # Create the KV caches.
+    key_caches, value_caches = kv_cache_factory(num_blocks, block_size, 1,
+                                                num_heads, head_size, dtype,
+                                                seed)
+    key_cache, value_cache = key_caches[0], value_caches[0]
+
+    # Clone the KV caches.
+    cloned_key_cache = key_cache.clone()
+    cloned_value_cache = value_cache.clone()
+
+    # Call the reshape_and_cache kernel.
+    cache_ops.reshape_and_cache(key, value, key_cache, value_cache,
+                                slot_mapping)
+
+    # Run the reference implementation.
+    reshaped_key = key.reshape(num_tokens, *key_cache[0, :, :, 0, :].shape)
+    block_indicies = torch.div(slot_mapping, block_size, rounding_mode="floor")
+    block_indicies = block_indicies.cpu().tolist()
+    block_offsets = slot_mapping % block_size
+    block_offsets = block_offsets.cpu().tolist()
+    for i in range(num_tokens):
+        block_idx = block_indicies[i]
+        block_offset = block_offsets[i]
+        cloned_key_cache[block_idx, :, :, block_offset, :] = reshaped_key[i]
+        cloned_value_cache[block_idx, :, :, block_offset] = value[i]
+
+    assert torch.allclose(key_cache, cloned_key_cache)
+    assert torch.allclose(value_cache, cloned_value_cache)

inference/vllm/tests/kernels/test_layernorm.py

+import pytest
+import torch
+import torch.nn as nn
+
+from vllm import layernorm_ops
+
+DTYPES = [torch.half, torch.bfloat16, torch.float]
+HIDDEN_SIZES = [67, 768, 2048, 5120, 8192]  # Arbitrary values for testing
+NUM_TOKENS = [7, 83, 4096]  # Arbitrary values for testing
+SEEDS = [0]
+
+
+class RefRMSNorm(nn.Module):
+
+    def __init__(self, hidden_size, eps=1e-6):
+        super().__init__()
+        weight = torch.empty(hidden_size)
+        weight.normal_(mean=1.0, std=0.1)
+        self.weight = nn.Parameter(weight)
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance +
+                                                    self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_rms_norm(
+    num_tokens: int,
+    hidden_size: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    scale = float(hidden_size**-0.5)
+    x = torch.empty(num_tokens, hidden_size, dtype=dtype, device="cuda")
+    x.uniform_(-scale, scale)
+    ref = RefRMSNorm(hidden_size).to(dtype).cuda()
+
+    out = torch.empty_like(x)
+    layernorm_ops.rms_norm(
+        out,
+        x,
+        ref.weight.data,
+        ref.variance_epsilon,
+    )
+    ref_out = ref(x)
+    assert torch.allclose(out, ref_out, atol=1e-2, rtol=1e-5)

inference/vllm/tests/kernels/test_pos_encoding.py

+from typing import Optional, Tuple
+
+import pytest
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from vllm import pos_encoding_ops
+
+IS_NEOX_STYLE = [True, False]
+DTYPES = [torch.half, torch.bfloat16, torch.float]
+HEAD_SIZES = [64, 80, 96, 112, 128, 256]
+ROTARY_DIMS = [None, 32]  # None means rotary dim == head size
+NUM_HEADS = [7, 12, 40, 52]  # Arbitrary values for testing
+NUM_TOKENS = [11, 83, 2048]  # Arbitrary values for testing
+SEEDS = [0]
+
+
+def rotate_neox(x: torch.Tensor) -> torch.Tensor:
+    x1 = x[..., :x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2:]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def rotate_gptj(x: torch.Tensor) -> torch.Tensor:
+    x1 = x[..., ::2]
+    x2 = x[..., 1::2]
+    x = torch.stack((-x2, x1), dim=-1)
+    return x.flatten(-2)
+
+
+def apply_rope(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    is_neox_style: bool,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    rotate_fn = rotate_neox if is_neox_style else rotate_gptj
+    q_embed = (q * cos) + (rotate_fn(q) * sin)
+    k_embed = (k * cos) + (rotate_fn(k) * sin)
+    return q_embed, k_embed
+
+
+class RefRotaryEmbedding(nn.Module):
+    """Reference implementation of rotary embedding."""
+
+    def __init__(
+        self,
+        dim: int,
+        is_neox_style: bool,
+        max_position_embeddings: int = 8192,
+        base: int = 10000,
+    ) -> None:
+        super().__init__()
+        self.rotary_dim = dim
+        self.is_neox_style = is_neox_style
+        self.max_position_embeddings = max_position_embeddings
+
+        # Create cos and sin embeddings.
+        inv_freq = 1.0 / (base**(torch.arange(0, dim, 2) / dim))
+        t = torch.arange(max_position_embeddings).float()
+        freqs = torch.einsum("i,j->ij", t, inv_freq.float())
+        if is_neox_style:
+            emb = torch.cat((freqs, freqs), dim=-1)
+        else:
+            emb = torch.repeat_interleave(freqs, 2, -1)
+        cos = emb.cos().to(dtype=inv_freq.dtype)
+        sin = emb.sin().to(dtype=inv_freq.dtype)
+        self.register_buffer("cos_cached", cos, persistent=False)
+        self.register_buffer("sin_cached", sin, persistent=False)
+
+    def forward(
+        self,
+        positions: torch.Tensor,  # [num_tokens]
+        query: torch.Tensor,  # [num_tokens, num_heads, head_size]
+        key: torch.Tensor,  # [num_tokens, num_heads, head_size]
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        query_rot = query[..., :self.rotary_dim]
+        query_pass = query[..., self.rotary_dim:]
+        key_rot = key[..., :self.rotary_dim]
+        key_pass = key[..., self.rotary_dim:]
+
+        query_rot = query_rot.transpose(0, 1)
+        key_rot = key_rot.transpose(0, 1)
+        cos = F.embedding(positions, self.cos_cached)
+        sin = F.embedding(positions, self.sin_cached)
+
+        query_rot, key_rot = apply_rope(query_rot, key_rot, cos, sin,
+                                        self.is_neox_style)
+        query_rot = query_rot.transpose(0, 1).contiguous()
+        key_rot = key_rot.transpose(0, 1).contiguous()
+
+        query = torch.cat((query_rot, query_pass), dim=-1)
+        key = torch.cat((key_rot, key_pass), dim=-1)
+
+        # Output query/key shape: [num_tokens, num_tokens, head_size]
+        return query, key
+
+
+@pytest.mark.parametrize("is_neox_style", IS_NEOX_STYLE)
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("rotary_dim", ROTARY_DIMS)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_rotary_embedding(
+    is_neox_style: bool,
+    num_tokens: int,
+    num_heads: int,
+    head_size: int,
+    rotary_dim: Optional[int],
+    dtype: torch.dtype,
+    seed: int,
+    max_position: int = 8192,
+    base: int = 10000,
+) -> None:
+    if rotary_dim is None:
+        rotary_dim = head_size
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    positions = torch.randint(0, max_position, (num_tokens, ), device="cuda")
+    query = torch.randn(num_tokens,
+                        num_heads * head_size,
+                        dtype=dtype,
+                        device="cuda")
+    key = torch.randn(num_tokens,
+                      num_heads * head_size,
+                      dtype=dtype,
+                      device="cuda")
+
+    # Create the rotary embedding.
+    inv_freq = 1.0 / (base**(
+        torch.arange(0, rotary_dim, 2, dtype=torch.float) / rotary_dim))
+    t = torch.arange(max_position).float()
+    freqs = torch.einsum("i,j -> ij", t, inv_freq)
+    cos = freqs.cos()
+    sin = freqs.sin()
+    cos_sin_cache = torch.cat((cos, sin), dim=-1)
+    cos_sin_cache = cos_sin_cache.to(dtype=dtype, device="cuda")
+
+    # Run the kernel. The kernel is in-place, so we need to clone the inputs.
+    out_query = query.clone()
+    out_key = key.clone()
+    pos_encoding_ops.rotary_embedding(
+        positions,
+        out_query,
+        out_key,
+        head_size,
+        cos_sin_cache,
+        is_neox_style,
+    )
+
+    # Run the reference implementation.
+    ref_rotary_embedding = RefRotaryEmbedding(
+        dim=rotary_dim,
+        is_neox_style=is_neox_style,
+        max_position_embeddings=max_position,
+        base=base,
+    ).to(dtype=dtype, device="cuda")
+    ref_query, ref_key = ref_rotary_embedding(
+        positions,
+        query.view(num_tokens, num_heads, head_size),
+        key.view(num_tokens, num_heads, head_size),
+    )
+    ref_query = ref_query.view(num_tokens, num_heads * head_size)
+    ref_key = ref_key.view(num_tokens, num_heads * head_size)
+
+    # Compare the results.
+    assert torch.allclose(out_query, ref_query, atol=1e-5, rtol=1e-5)
+    assert torch.allclose(out_key, ref_key, atol=1e-5, rtol=1e-5)

inference/vllm/tests/models/test_models.py

+"""Compare the outputs of HF and vLLM when using greedy sampling.
+
+Run `pytest tests/models/test_models.py --forked`.
+"""
+import pytest
+
+MODELS = [
+    "facebook/opt-125m",
+    "meta-llama/Llama-2-7b-hf",
+    "mistralai/Mistral-7B-v0.1",
+    "tiiuae/falcon-7b",
+    "gpt2",
+    "bigcode/tiny_starcoder_py",
+    "EleutherAI/gpt-j-6b",
+    "EleutherAI/pythia-70m",
+    "bigscience/bloom-560m",
+    "mosaicml/mpt-7b",
+    "microsoft/phi-1_5",
+]
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["half"])
+@pytest.mark.parametrize("max_tokens", [128])
+def test_models(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+    hf_model = hf_runner(model, dtype=dtype)
+    hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
+    del hf_model
+
+    vllm_model = vllm_runner(model, dtype=dtype)
+    vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+    del vllm_model
+
+    for i in range(len(example_prompts)):
+        hf_output_ids, hf_output_str = hf_outputs[i]
+        vllm_output_ids, vllm_output_str = vllm_outputs[i]
+        assert hf_output_str == vllm_output_str, (
+            f"Test{i}:\nHF: {hf_output_str!r}\nvLLM: {vllm_output_str!r}")
+        assert hf_output_ids == vllm_output_ids, (
+            f"Test{i}:\nHF: {hf_output_ids}\nvLLM: {vllm_output_ids}")

inference/vllm/tests/samplers/test_beam_search.py

+"""Compare the outputs of HF and vLLM when using beam search.
+
+Run `pytest tests/samplers/test_beam_search.py --forked`.
+"""
+import pytest
+
+# FIXME(zhuohan): The test can not pass if we:
+#   1. Increase max_tokens to 256.
+#   2. Increase beam_width to 8.
+#   3. Use the model "huggyllama/llama-7b".
+MAX_TOKENS = [128]
+BEAM_WIDTHS = [4]
+MODELS = ["facebook/opt-125m"]
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["half"])
+@pytest.mark.parametrize("max_tokens", MAX_TOKENS)
+@pytest.mark.parametrize("beam_width", BEAM_WIDTHS)
+def test_beam_search_single_input(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+    beam_width: int,
+) -> None:
+    hf_model = hf_runner(model, dtype=dtype)
+    hf_outputs = hf_model.generate_beam_search(example_prompts, beam_width,
+                                               max_tokens)
+    del hf_model
+
+    vllm_model = vllm_runner(model, dtype=dtype)
+    vllm_outputs = vllm_model.generate_beam_search(example_prompts, beam_width,
+                                                   max_tokens)
+    del vllm_model
+
+    for i in range(len(example_prompts)):
+        hf_output_ids, _ = hf_outputs[i]
+        vllm_output_ids, _ = vllm_outputs[i]
+        assert len(hf_output_ids) == len(vllm_output_ids)
+        for j in range(len(hf_output_ids)):
+            assert hf_output_ids[j] == vllm_output_ids[j], (
+                f"Test{i} output{j}:\nHF: {hf_output_ids}\n"
+                f"vLLM: {vllm_output_ids}")

inference/vllm/tests/samplers/test_logprobs.py

+import pytest
+import torch
+
+from vllm import SamplingParams
+
+MODELS = ["facebook/opt-125m"]
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["half"])
+def test_get_prompt_logprobs(
+    hf_runner,
+    vllm_runner,
+    model,
+    dtype,
+    example_prompts,
+):
+    max_tokens = 5
+    hf_model = hf_runner(model, dtype=dtype)
+    hf_logprobs = hf_model.generate_greedy_logprobs(
+        example_prompts,
+        max_tokens=max_tokens,
+    )
+    del hf_model
+
+    vllm_model = vllm_runner(model, dtype=dtype)
+    vllm_sampling_params = SamplingParams(max_tokens=max_tokens,
+                                          logprobs=5,
+                                          prompt_logprobs=5,
+                                          temperature=0.0)
+    vllm_results = vllm_model.model.generate(
+        example_prompts, sampling_params=vllm_sampling_params)
+
+    # Test whether logprobs are included in the results.
+    for result in vllm_results:
+        assert result.prompt_logprobs is not None
+        assert result.outputs[0].logprobs is not None
+
+    # Test whether prompt logprobs are consistent with HF
+    for vllm_result, hf_logprob in zip(vllm_results, hf_logprobs):
+        # Check prompt logprobs
+        vllm_prompt_logprobs = vllm_result.prompt_logprobs[1:]
+        for i, vllm_prompt_logprob_dict in enumerate(vllm_prompt_logprobs):
+            for token_id, logprob in vllm_prompt_logprob_dict.items():
+                torch.testing.assert_close(logprob,
+                                           hf_logprob[0][i][token_id].item(),
+                                           atol=1e-2,
+                                           rtol=1e-2)
+        vllm_sample_logprobs = vllm_result.outputs[0].logprobs
+        for i, vllm_sample_logprob_dict in enumerate(vllm_sample_logprobs):
+            for token_id, logprob in vllm_sample_logprob_dict.items():
+                torch.testing.assert_close(logprob,
+                                           hf_logprob[i][-1][token_id].item(),
+                                           atol=1e-2,
+                                           rtol=1e-2)

inference/vllm/tests/samplers/test_sampler.py

+# pylint: disable=protected-access
+import random
+from typing import Tuple
+from unittest.mock import patch
+
+import pytest
+import torch
+
+from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.utils import set_random_seed
+from vllm.sequence import SamplingParams, SequenceData, SequenceGroupMetadata
+from vllm.worker.worker import Worker
+
+
+class MockLogitsSampler(Sampler):
+
+    def __init__(self, vocab_size: int, fake_logits: torch.Tensor):
+        super().__init__(vocab_size=vocab_size)
+        self.fake_logits = fake_logits
+
+    def forward(self, *args, **kwargs):
+        with patch("vllm.model_executor.layers.sampler._prune_hidden_states",
+                   lambda x, y: x):
+            with patch("vllm.model_executor.layers.sampler._get_logits",
+                       lambda *args, **kwargs: self.fake_logits):
+                return super().forward(*args, **kwargs)
+
+
+def _prepare_test(
+    batch_size: int
+) -> Tuple[torch.Tensor, torch.Tensor, MockLogitsSampler, Worker]:
+    vocab_size = 32000
+    input_tensor = torch.rand((batch_size, 1024),
+                              device="cuda",
+                              dtype=torch.float16)
+    fake_logits = torch.full((batch_size, vocab_size),
+                             1e-2,
+                             device=input_tensor.device,
+                             dtype=input_tensor.dtype)
+    sampler = MockLogitsSampler(32000, fake_logits)
+    worker = Worker(None, None, None)
+    worker.block_size = 16
+    return input_tensor, fake_logits, sampler, worker
+
+
+RANDOM_SEEDS = list(range(128))
+
+
+@pytest.mark.parametrize("seed", RANDOM_SEEDS)
+def test_sampler_all_greedy(seed: int):
+    set_random_seed(seed)
+    batch_size = random.randint(1, 256)
+    input_tensor, fake_logits, sampler, worker = _prepare_test(batch_size)
+
+    seq_group_metadata_list = []
+    for i in range(batch_size):
+        seq_group_metadata_list.append(
+            SequenceGroupMetadata(
+                request_id=f"test_{i}",
+                is_prompt=True,
+                seq_data={0: SequenceData([1, 2, 3])},
+                sampling_params=SamplingParams(temperature=0, ),
+                block_tables={0: [1]},
+            ))
+
+    _, _, input_metadata = worker._prepare_inputs(seq_group_metadata_list)
+    sampler_output = sampler(embedding=None,
+                             hidden_states=input_tensor,
+                             input_metadata=input_metadata)
+    expected = torch.argmax(fake_logits, dim=-1)
+    for i, sequence_output in enumerate(sampler_output):
+        for nth_output in sequence_output.samples:
+            assert nth_output.output_token == expected[i].item()
+
+
+@pytest.mark.parametrize("seed", RANDOM_SEEDS)
+def test_sampler_all_random(seed: int):
+    set_random_seed(seed)
+    batch_size = random.randint(1, 256)
+    input_tensor, fake_logits, sampler, worker = _prepare_test(batch_size)
+
+    for i in range(batch_size):
+        fake_logits[i, i] = 1e2
+
+    seq_group_metadata_list = []
+    for i in range(batch_size):
+        seq_group_metadata_list.append(
+            SequenceGroupMetadata(
+                request_id=f"test_{i}",
+                is_prompt=True,
+                seq_data={0: SequenceData([1, 2, 3])},
+                sampling_params=SamplingParams(
+                    temperature=1.0,
+                    n=random.randint(1, 10),
+                ),
+                block_tables={0: [1]},
+            ))
+
+    _, _, input_metadata = worker._prepare_inputs(seq_group_metadata_list)
+    sampler_output = sampler(embedding=None,
+                             hidden_states=input_tensor,
+                             input_metadata=input_metadata)
+    for i, sequence_output in enumerate(sampler_output):
+        for nth_output in sequence_output.samples:
+            assert nth_output.output_token == i
+
+
+@pytest.mark.parametrize("seed", RANDOM_SEEDS)
+def test_sampler_all_beam(seed: int):
+    set_random_seed(seed)
+    batch_size = random.randint(1, 256)
+    input_tensor, _, sampler, worker = _prepare_test(batch_size)
+
+    seq_group_metadata_list = []
+    for i in range(batch_size):
+        seq_group_metadata_list.append(
+            SequenceGroupMetadata(
+                request_id=f"test_{i}",
+                is_prompt=True,
+                seq_data={0: SequenceData([1, 2, 3])},
+                sampling_params=SamplingParams(
+                    temperature=0,
+                    best_of=2,
+                    use_beam_search=True,
+                ),
+                block_tables={0: [1]},
+            ))
+
+    _, _, input_metadata = worker._prepare_inputs(seq_group_metadata_list)
+    sampler(embedding=None,
+            hidden_states=input_tensor,
+            input_metadata=input_metadata)
+    # no assertion here as I am not sure how to determine whether
+    # the outputs are expected - in other words, this just tests
+    # whether there are no exceptions in the sampler
+    # when handling an all-beam search case.
+
+
+@pytest.mark.parametrize("seed", RANDOM_SEEDS)
+def test_sampler_mixed(seed: int):
+    set_random_seed(seed)
+    batch_size = random.randint(1, 256)
+    input_tensor, fake_logits, sampler, worker = _prepare_test(batch_size)
+
+    seq_group_metadata_list = []
+    expected_tokens = []
+    for i in range(batch_size):
+        n = 1
+        sampling_type = random.randint(0, 2)
+        if sampling_type == 0:
+            sampling_params = SamplingParams(temperature=0)
+        elif sampling_type == 1:
+            n = random.randint(1, 10)
+            sampling_params = SamplingParams(
+                temperature=random.random() + 0.1,
+                top_p=min(random.random() + 0.1, 1),
+                top_k=random.randint(0, 10) or -1,
+                n=n,
+                presence_penalty=random.randint(0, 1),
+            )
+        else:
+            sampling_params = SamplingParams(temperature=0,
+                                             use_beam_search=True,
+                                             best_of=2)
+        for idx in range(n):
+            fake_logits[i, i + idx] = 1e2
+            expected_tokens.append(i + idx)
+        seq_group_metadata_list.append(
+            SequenceGroupMetadata(
+                request_id=f"test_{i}",
+                is_prompt=True,
+                seq_data={0: SequenceData([1, 2, 3])},
+                sampling_params=sampling_params,
+                block_tables={0: [1]},
+            ))
+
+    _, _, input_metadata = worker._prepare_inputs(seq_group_metadata_list)
+    sampler_output = sampler(embedding=None,
+                             hidden_states=input_tensor,
+                             input_metadata=input_metadata)
+    for i, sequence_output in enumerate(sampler_output):
+        if seq_group_metadata_list[i].sampling_params.use_beam_search:
+            continue
+        for nth_output in sequence_output.samples:
+            assert nth_output.output_token in expected_tokens
+
+
+@pytest.mark.parametrize("seed", RANDOM_SEEDS)
+def test_sampler_logits_processors(seed: int):
+    set_random_seed(seed)
+    batch_size = random.randint(1, 256)
+    input_tensor, _, sampler, worker = _prepare_test(batch_size)
+
+    # This sample logits processor gives infinite score to the i-th token,
+    # where i is the length of the input sequence.
+    # We therefore expect the output token sequence to be [0, 1, 2, ...]
+    def pick_ith(token_ids, logits):
+        logits[len(token_ids)] = float("inf")
+        return logits
+
+    seq_group_metadata_list = []
+    for i in range(batch_size):
+        seq_group_metadata_list.append(
+            SequenceGroupMetadata(
+                request_id=f"test_{i}",
+                is_prompt=True,
+                seq_data={0: SequenceData([1, 2, 3])},
+                sampling_params=SamplingParams(temperature=0,
+                                               logits_processors=[pick_ith]),
+                block_tables={0: [1]},
+            ))
+
+    _, _, input_metadata = worker._prepare_inputs(seq_group_metadata_list)
+    sampler_output = sampler(embedding=None,
+                             hidden_states=input_tensor,
+                             input_metadata=input_metadata)
+    for i, sequence_output in enumerate(sampler_output):
+        for idx, nth_output in enumerate(sequence_output.samples):
+            assert nth_output.output_token == idx

inference/vllm/tests/test_regression.py

+"""Containing tests that check for regressions in vLLM's behavior.
+
+It should include tests that are reported by users and making sure they
+will never happen again.
+
+"""
+from vllm import LLM, SamplingParams
+
+
+def test_duplicated_ignored_sequence_group():
+    """https://github.com/vllm-project/vllm/issues/1655"""
+
+    sampling_params = SamplingParams(temperature=0.01,
+                                     top_p=0.1,
+                                     max_tokens=256)
+    llm = LLM(model="facebook/opt-125m",
+              max_num_batched_tokens=4096,
+              tensor_parallel_size=1)
+    prompts = ["This is a short prompt", "This is a very long prompt " * 1000]
+    outputs = llm.generate(prompts, sampling_params=sampling_params)
+
+    assert len(prompts) == len(outputs)
+
+
+if __name__ == "__main__":
+    import pytest
+    pytest.main([__file__])

inference/vllm/tests/worker/test_worker.py

+# pylint: disable=protected-access
+import random
+import torch
+
+from vllm.sequence import SamplingParams, SequenceData, SequenceGroupMetadata
+from vllm.worker.worker import Worker
+
+
+def test_worker_prepare_inputs_for_prompt():
+    worker = Worker(None, None, None)
+    worker.block_size = 16
+    batch_size = random.randint(1, 256)
+    prompt_lens = []
+    seq_group_metadata_list = []
+    for i in range(batch_size):
+        # make sure all tokens fit into one block
+        prompt_len = i % (worker.block_size - 1) + 1
+        prompt_lens.append(prompt_len)
+        seq_data = list(range(prompt_len))
+        seq_group_metadata_list.append(
+            SequenceGroupMetadata(
+                request_id=f"test_{i}",
+                is_prompt=True,
+                seq_data={0: SequenceData(seq_data)},
+                sampling_params=SamplingParams(temperature=0),
+                block_tables={0: [1]},
+            ))
+    expected_selected_token_indices = []
+    selected_token_start_idx = 0
+    max_seq_len = max(prompt_lens)
+    for prompt_len in prompt_lens:
+        expected_selected_token_indices.append(selected_token_start_idx +
+                                               prompt_len - 1)
+        selected_token_start_idx += max_seq_len
+    input_tokens, input_positions, input_metadata = worker._prepare_inputs(
+        seq_group_metadata_list)
+    assert input_tokens.shape == input_positions.shape == (batch_size,
+                                                           max_seq_len)
+    torch.testing.assert_close(input_tokens, input_positions)
+    actual = input_metadata.selected_token_indices
+    expected = torch.tensor(expected_selected_token_indices,
+                            device=actual.device,
+                            dtype=actual.dtype)
+    torch.testing.assert_close(actual, expected)

inference/vllm/vllm.egg-info/PKG-INFO

+Metadata-Version: 2.1
+Name: vllm
+Version: 0.2.2
+Summary: A high-throughput and memory-efficient inference and serving engine for LLMs
+Home-page: https://github.com/vllm-project/vllm
+Author: vLLM Team
+License: Apache 2.0
+Project-URL: Homepage, https://github.com/vllm-project/vllm
+Project-URL: Documentation, https://vllm.readthedocs.io/en/latest/
+Classifier: Programming Language :: Python :: 3.8
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.8
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: ninja
+Requires-Dist: psutil
+Requires-Dist: ray>=2.5.1
+Requires-Dist: pandas
+Requires-Dist: pyarrow
+Requires-Dist: sentencepiece
+Requires-Dist: numpy
+Requires-Dist: einops
+Requires-Dist: torch>=2.1.0
+Requires-Dist: transformers>=4.34.0
+Requires-Dist: xformers>=0.0.22.post7
+Requires-Dist: fastapi
+Requires-Dist: uvicorn[standard]
+Requires-Dist: pydantic==1.10.13
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/logos/vllm-logo-text-dark.png">
+    <img alt="vLLM" src="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/logos/vllm-logo-text-light.png" width=55%>
+  </picture>
+</p>
+
+<h3 align="center">
+Easy, fast, and cheap LLM serving for everyone
+</h3>
+
+<p align="center">
+| <a href="https://vllm.readthedocs.io/en/latest/"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://discord.gg/jz7wjKhh6g"><b>Discord</b></a> |
+
+</p>
+
+---
+
+*Latest News* 🔥
+- [2023/10] We hosted [the first vLLM meetup](https://lu.ma/first-vllm-meetup) in SF! Please find the meetup slides [here](https://docs.google.com/presentation/d/1QL-XPFXiFpDBh86DbEegFXBXFXjix4v032GhShbKf3s/edit?usp=sharing).
+- [2023/09] We created our [Discord server](https://discord.gg/jz7wjKhh6g)! Join us to discuss vLLM and LLM serving! We will also post the latest announcements and updates there.
+- [2023/09] We released our [PagedAttention paper](https://arxiv.org/abs/2309.06180) on arXiv!
+- [2023/08] We would like to express our sincere gratitude to [Andreessen Horowitz](https://a16z.com/2023/08/30/supporting-the-open-source-ai-community/) (a16z) for providing a generous grant to support the open-source development and research of vLLM.
+- [2023/07] Added support for LLaMA-2! You can run and serve 7B/13B/70B LLaMA-2s on vLLM with a single command!
+- [2023/06] Serving vLLM On any Cloud with SkyPilot. Check out a 1-click [example](https://github.com/skypilot-org/skypilot/blob/master/llm/vllm) to start the vLLM demo, and the [blog post](https://blog.skypilot.co/serving-llm-24x-faster-on-the-cloud-with-vllm-and-skypilot/) for the story behind vLLM development on the clouds.
+- [2023/06] We officially released vLLM! FastChat-vLLM integration has powered [LMSYS Vicuna and Chatbot Arena](https://chat.lmsys.org) since mid-April. Check out our [blog post](https://vllm.ai).
+
+---
+
+vLLM is a fast and easy-to-use library for LLM inference and serving.
+
+vLLM is fast with:
+
+- State-of-the-art serving throughput
+- Efficient management of attention key and value memory with **PagedAttention**
+- Continuous batching of incoming requests
+- Optimized CUDA kernels
+
+vLLM is flexible and easy to use with:
+
+- Seamless integration with popular Hugging Face models
+- High-throughput serving with various decoding algorithms, including *parallel sampling*, *beam search*, and more
+- Tensor parallelism support for distributed inference
+- Streaming outputs
+- OpenAI-compatible API server
+
+vLLM seamlessly supports many Hugging Face models, including the following architectures:
+
+- Aquila & Aquila2 (`BAAI/AquilaChat2-7B`, `BAAI/AquilaChat2-34B`, `BAAI/Aquila-7B`, `BAAI/AquilaChat-7B`, etc.)
+- Baichuan (`baichuan-inc/Baichuan-7B`, `baichuan-inc/Baichuan-13B-Chat`, etc.)
+- BLOOM (`bigscience/bloom`, `bigscience/bloomz`, etc.)
+- ChatGLM (`THUDM/chatglm2-6b`, `THUDM/chatglm3-6b`, etc.)
+- Falcon (`tiiuae/falcon-7b`, `tiiuae/falcon-40b`, `tiiuae/falcon-rw-7b`, etc.)
+- GPT-2 (`gpt2`, `gpt2-xl`, etc.)
+- GPT BigCode (`bigcode/starcoder`, `bigcode/gpt_bigcode-santacoder`, etc.)
+- GPT-J (`EleutherAI/gpt-j-6b`, `nomic-ai/gpt4all-j`, etc.)
+- GPT-NeoX (`EleutherAI/gpt-neox-20b`, `databricks/dolly-v2-12b`, `stabilityai/stablelm-tuned-alpha-7b`, etc.)
+- InternLM (`internlm/internlm-7b`, `internlm/internlm-chat-7b`, etc.)
+- LLaMA & LLaMA-2 (`meta-llama/Llama-2-70b-hf`, `lmsys/vicuna-13b-v1.3`, `young-geng/koala`, `openlm-research/open_llama_13b`, etc.)
+- Mistral (`mistralai/Mistral-7B-v0.1`, `mistralai/Mistral-7B-Instruct-v0.1`, etc.)
+- MPT (`mosaicml/mpt-7b`, `mosaicml/mpt-30b`, etc.)
+- OPT (`facebook/opt-66b`, `facebook/opt-iml-max-30b`, etc.)
+- Phi-1.5 (`microsoft/phi-1_5`, etc.)
+- Qwen (`Qwen/Qwen-7B`, `Qwen/Qwen-7B-Chat`, etc.)
+- Yi (`01-ai/Yi-6B`, `01-ai/Yi-34B`, etc.)
+
+Install vLLM with pip or [from source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source):
+
+```bash
+pip install vllm
+```
+
+## Getting Started
+
+Visit our [documentation](https://vllm.readthedocs.io/en/latest/) to get started.
+- [Installation](https://vllm.readthedocs.io/en/latest/getting_started/installation.html)
+- [Quickstart](https://vllm.readthedocs.io/en/latest/getting_started/quickstart.html)
+- [Supported Models](https://vllm.readthedocs.io/en/latest/models/supported_models.html)
+
+## Contributing
+
+We welcome and value any contributions and collaborations.
+Please check out [CONTRIBUTING.md](./CONTRIBUTING.md) for how to get involved.
+
+## Citation
+
+If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs/2309.06180):
+```bibtex
+@inproceedings{kwon2023efficient,
+  title={Efficient Memory Management for Large Language Model Serving with PagedAttention},
+  author={Woosuk Kwon and Zhuohan Li and Siyuan Zhuang and Ying Sheng and Lianmin Zheng and Cody Hao Yu and Joseph E. Gonzalez and Hao Zhang and Ion Stoica},
+  booktitle={Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles},
+  year={2023}
+}
+```

inference/vllm/vllm.egg-info/SOURCES.txt

+LICENSE
+MANIFEST.in
+README.md
+pyproject.toml
+requirements.txt
+setup.py
+csrc/activation.cpp
+csrc/activation_kernels.cu
+csrc/attention.cpp
+csrc/cache.cpp
+csrc/cache_kernels.cu
+csrc/cuda_utils.cpp
+csrc/cuda_utils_kernels.cu
+csrc/dispatch_utils.h
+csrc/layernorm.cpp
+csrc/layernorm_kernels.cu
+csrc/pos_encoding.cpp
+csrc/pos_encoding_kernels.cu
+csrc/quantization.cpp
+csrc/reduction_utils.cuh
+csrc/attention/attention_dtypes.h
+csrc/attention/attention_generic.cuh
+csrc/attention/attention_kernels.cu
+csrc/attention/attention_utils.cuh
+csrc/attention/dtype_bfloat16.cuh
+csrc/attention/dtype_float16.cuh
+csrc/attention/dtype_float32.cuh
+csrc/quantization/awq/dequantize.cuh
+csrc/quantization/awq/gemm_kernels.cu
+csrc/quantization/squeezellm/quant_cuda_kernel.cu
+tests/test_regression.py
+vllm/__init__.py
+vllm/block.py
+vllm/config.py
+vllm/logger.py
+vllm/outputs.py
+vllm/py.typed
+vllm/sampling_params.py
+vllm/sequence.py
+vllm/utils.py
+vllm.egg-info/PKG-INFO
+vllm.egg-info/SOURCES.txt
+vllm.egg-info/dependency_links.txt
+vllm.egg-info/requires.txt
+vllm.egg-info/top_level.txt
+vllm/core/__init__.py
+vllm/core/block_manager.py
+vllm/core/policy.py
+vllm/core/scheduler.py
+vllm/engine/__init__.py
+vllm/engine/arg_utils.py
+vllm/engine/async_llm_engine.py
+vllm/engine/llm_engine.py
+vllm/engine/ray_utils.py
+vllm/entrypoints/__init__.py
+vllm/entrypoints/api_server.py
+vllm/entrypoints/llm.py
+vllm/entrypoints/openai/__init__.py
+vllm/entrypoints/openai/api_server.py
+vllm/entrypoints/openai/protocol.py
+vllm/model_executor/__init__.py
+vllm/model_executor/input_metadata.py
+vllm/model_executor/model_loader.py
+vllm/model_executor/utils.py
+vllm/model_executor/weight_utils.py
+vllm/model_executor/layers/__init__.py
+vllm/model_executor/layers/activation.py
+vllm/model_executor/layers/attention.py
+vllm/model_executor/layers/layernorm.py
+vllm/model_executor/layers/linear.py
+vllm/model_executor/layers/rotary_embedding.py
+vllm/model_executor/layers/sampler.py
+vllm/model_executor/layers/vocab_parallel_embedding.py
+vllm/model_executor/layers/quantization/__init__.py
+vllm/model_executor/layers/quantization/awq.py
+vllm/model_executor/layers/quantization/base_config.py
+vllm/model_executor/layers/quantization/squeezellm.py
+vllm/model_executor/models/__init__.py
+vllm/model_executor/models/aquila.py
+vllm/model_executor/models/baichuan.py
+vllm/model_executor/models/bloom.py
+vllm/model_executor/models/chatglm.py
+vllm/model_executor/models/cpm.py
+vllm/model_executor/models/cpm_mistral.py
+vllm/model_executor/models/cpm_old.py
+vllm/model_executor/models/cpmmistral.py
+vllm/model_executor/models/falcon.py
+vllm/model_executor/models/gpt2.py
+vllm/model_executor/models/gpt_bigcode.py
+vllm/model_executor/models/gpt_j.py
+vllm/model_executor/models/gpt_neox.py
+vllm/model_executor/models/internlm.py
+vllm/model_executor/models/llama.py
+vllm/model_executor/models/mistral.py
+vllm/model_executor/models/mpt.py
+vllm/model_executor/models/opt.py
+vllm/model_executor/models/phi_1_5.py
+vllm/model_executor/models/qwen.py
+vllm/model_executor/models/yi.py
+vllm/model_executor/parallel_utils/__init__.py
+vllm/model_executor/parallel_utils/communication_op.py
+vllm/model_executor/parallel_utils/parallel_state.py
+vllm/model_executor/parallel_utils/utils.py
+vllm/transformers_utils/__init__.py
+vllm/transformers_utils/config.py
+vllm/transformers_utils/tokenizer.py
+vllm/transformers_utils/configs/__init__.py
+vllm/transformers_utils/configs/aquila.py
+vllm/transformers_utils/configs/baichuan.py
+vllm/transformers_utils/configs/chatglm.py
+vllm/transformers_utils/configs/cpm.py
+vllm/transformers_utils/configs/cpm_mistral.py
+vllm/transformers_utils/configs/cpmmistral.py
+vllm/transformers_utils/configs/falcon.py
+vllm/transformers_utils/configs/mpt.py
+vllm/transformers_utils/configs/qwen.py
+vllm/transformers_utils/configs/yi.py
+vllm/worker/__init__.py
+vllm/worker/cache_engine.py
+vllm/worker/worker.py
\ No newline at end of file

inference/vllm/vllm.egg-info/dependency_links.txt

+

inference/vllm/vllm.egg-info/requires.txt

+ninja
+psutil
+ray>=2.5.1
+pandas
+pyarrow
+sentencepiece
+numpy
+einops
+torch>=2.1.0
+transformers>=4.34.0
+xformers>=0.0.22.post7
+fastapi
+uvicorn[standard]
+pydantic==1.10.13